View in EDS HTML Full Text PDF Full Text

Bayesian Estimation of Hierarchical Linear Models from Incomplete Data: Cluster-Level Interaction Effects and Small Sample Sizes

Saved in:
Bibliographic Details
Title: Bayesian Estimation of Hierarchical Linear Models from Incomplete Data: Cluster-Level Interaction Effects and Small Sample Sizes
Language: English
Authors: Dongho Shin (ORCID 0009-0006-5754-6815), Yongyun Shin (ORCID 0000-0003-0654-2620), Nao Hagiwara (ORCID 0000-0003-3933-8917)
Source: Grantee Submission. 2025 44.
Peer Reviewed: Y
Page Count: 11
Publication Date: 2025
Sponsoring Agency: National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) (DHHS/NIH)
National Cancer Institute (NCI) (DHHS/NIH)
National Institute of Nursing Research (NINR) (DHHS/NIH)
Institute of Education Sciences (ED)
Contract Number: R01DK112009
R01CA263501
Document Type: Journal Articles
Reports - Research
Descriptors: Bayesian Statistics, Hierarchical Linear Modeling, Multivariate Analysis, Data Analysis, Sample Size, Mathematical Formulas, Statistical Distributions
DOI: 10.1002/sim.70051
Abstract: We consider Bayesian estimation of a hierarchical linear model (HLM) from partially observed data, assumed to be missing at random, and small sample sizes. A vector of continuous covariates C includes cluster-level partially observed covariates with interaction effects. Due to small sample sizes from 37 patient-physician encounters repeatedly measured at four time points, maximum-likelihood estimation is suboptimal. Existing Gibbs samplers impute missing values of C by a Metropolis algorithm using proposal densities that have constant variances while the target posterior distributions have nonconstant variances. Therefore, these samplers may not ensure compatibility with the HLM and, as a result, may not guarantee unbiased estimation of the HLM. We introduce a compatible Gibbs sampler that imputes parameters and missing values directly from the exact posterior distributions. We apply our Gibbs sampler to the longitudinal patient-physician encounter data and compare our estimators with those from existing methods by simulation.
Abstractor: As Provided
IES Funded: Yes
Entry Date: 2025
Accession Number: ED673559
Database: ERIC
Full text is not displayed to guests.
FullText Links:
  – Type: pdflink
    Url: https://content.ebscohost.com/cds/retrieve?content=AQICAHj0k_4E0hTGH8RJwT4gCJyBsGNe_WN95AvKlDbXJGqwxwH9LPzSnmLCa0oLwVsQoZQCAAAA4zCB4AYJKoZIhvcNAQcGoIHSMIHPAgEAMIHJBgkqhkiG9w0BBwEwHgYJYIZIAWUDBAEuMBEEDCgoYokzkyxtJuu7XwIBEICBm55tbiCx7msOg_6gTFhKK2YG26NAkSot3afJNmxBcQtl5Tv4AZwDHbdz_JFSk2c69_mZLWYWerQfJ5RSMdd211mmwOnROu2bcjWgztpPKHQJDK1saAIcBH_DDReWgdZ2OcmuAVl5xiyGKLeMDcGmFjsEXh07WnaEZpSqWikXAPluELM8OQr-sfwRhHIsxR6P7cR-du9qgpv3m1Of
Text:
  Availability: 1
  Value: <anid>AN0185452270;2dz01may.25;2025May29.06:15;v2.2.500</anid> <title id="AN0185452270-1">Bayesian Estimation of Hierarchical Linear Models From Incomplete Data: Cluster‐Level Interaction Effects and Small Sample Sizes </title> <p>We consider Bayesian estimation of a hierarchical linear model (HLM) from partially observed data, assumed to be missing at random, and small sample sizes. A vector of continuous covariates C$$ C $$ includes cluster‐level partially observed covariates with interaction effects. Due to small sample sizes from 37 patient–physician encounters repeatedly measured at four time points, maximum‐likelihood estimation is suboptimal. Existing Gibbs samplers impute missing values of C$$ C $$ by a Metropolis algorithm using proposal densities that have constant variances while the target posterior distributions have nonconstant variances. Therefore, these samplers may not ensure compatibility with the HLM and, as a result, may not guarantee unbiased estimation of the HLM. We introduce a compatible Gibbs sampler that imputes parameters and missing values directly from the exact posterior distributions. We apply our Gibbs sampler to the longitudinal patient–physician encounter data and compare our estimators with those from existing methods by simulation.</p> <p>Keywords: Compatible Gibbs sampler; Exact posterior distributions; Interaction effects of continuous covariates; Provisionally known random effects; Robust estimation</p> <hd id="AN0185452270-2">Introduction</hd> <p>Medical researchers frequently confront four challenges of interest in this paper. First, patients are repeatedly measured or nested within clusters, such as physicians, clinics and neighborhoods. A hierarchical linear model (HLM) is appropriate to analyze such data [[<reflink idref="bib1" id="ref1">1</reflink>]]. Second, missing data commonly occur at the cluster level, posing a challenge to efficient and unbiased estimation of the HLM. Third, partially observed covariates may have nonlinear effects at the cluster level to make the estimation even more challenging. In particular, when two interactive cluster‐level covariates are partially observed, complete‐case analysis drops the clusters and all their nested units having either of the covariates missing. The resulting inference is inefficient and may be substantially biased [[<reflink idref="bib3" id="ref2">3</reflink>]]. Finally, small sample sizes constrain estimability of the HLM.</p> <p>Hierarchical missing data are routinely handled by multiple imputation (MI) [[<reflink idref="bib4" id="ref3">4</reflink>]] under the assumption of data missing at random (MAR) [[<reflink idref="bib5" id="ref4">5</reflink>]]. A popular joint modeling approach conceives of the response and partially observed covariates as the outcomes in a multivariate HLM, estimates their joint distribution, and imputes missing values from the estimated distribution [[<reflink idref="bib6" id="ref5">6</reflink>]]. Missing data may also be imputed from their conditional distributions by chained regressions, which is also known as fully conditional specification (FCS) [[<reflink idref="bib8" id="ref6">8</reflink>], [<reflink idref="bib10" id="ref7">10</reflink>]]. FCS fits each partially observed variable conditional on all other variables and imputes missing values from the fitted model. FCS is a Gibbs sampler if the chained regressions are compatible with an underlying joint distribution of all variables MAR, including the outcome [[<reflink idref="bib12" id="ref8">12</reflink>], [<reflink idref="bib14" id="ref9">14</reflink>]]. These approaches have been shown to work well when the joint distribution of partially observed variables are reasonably multivariate normal (MN) [[<reflink idref="bib14" id="ref10">14</reflink>], [<reflink idref="bib16" id="ref11">16</reflink>]].</p> <p>Given the nonlinear effects of covariates <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0003" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>C</mi></mrow><annotation encoding="application/x-tex">$$ C $$</annotation></semantics></math> </ephtml> on a response <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0004" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>Y</mi></mrow><annotation encoding="application/x-tex">$$ Y $$</annotation></semantics></math> </ephtml> in a HLM, however, multivariate normality of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0005" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>Y</mi><mo>,</mo><mi>C</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(Y,C\right) $$</annotation></semantics></math> </ephtml> is not possible even if the factorized conditional distributions are normal [[<reflink idref="bib17" id="ref12">17</reflink>]]. A concern then involves the bias that can arise from the incompatibility or conflicting assumptions between the fully conditional distributions that generate the imputations and the assumed joint distribution of the observed data [[<reflink idref="bib14" id="ref13">14</reflink>], [<reflink idref="bib17" id="ref14">17</reflink>], [<reflink idref="bib19" id="ref15">19</reflink>], [<reflink idref="bib21" id="ref16">21</reflink>]]. Kim et al. [[<reflink idref="bib17" id="ref17">17</reflink>], [<reflink idref="bib22" id="ref18">22</reflink>]] developed a Gibbs sampler to impute missing values of a partially observed continuous covariate interactive with a known covariate in a single‐level regression model from the exact conditional distribution. Carpenter and Kenward [[<reflink idref="bib19" id="ref19">19</reflink>]] took a MN approach to handling continuous and categorical data MAR by joint normality of continuous variables and latent continuous variables underlying categorical ones. Goldstein et al. [[<reflink idref="bib23" id="ref20">23</reflink>]] imputed missing values of covariates having nonlinear effects on a continuous or binary outcome by the MN approach via the Gibbs sampler and Metropolis algorithm [[<reflink idref="bib24" id="ref21">24</reflink>]]. Enders et al. [[<reflink idref="bib21" id="ref22">21</reflink>]] extended the MN‐Metropolis Gibbs sampler to handling categorical and continuous predictors MAR having nonlinear effects in two‐ and three‐level hierarchical models, and illustrated how FCS using its software MICE [[<reflink idref="bib10" id="ref23">10</reflink>]] may lead to an imputation model incompatible with the HLM to produce biased estimates.</p> <p>Shin and Raudenbush [[<reflink idref="bib18" id="ref24">18</reflink>]] extended the joint modeling approach to estimation of a HLM with the nonlinearities from continuous data <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0006" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>Y</mi><mo>,</mo><mi>C</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(Y,C\right) $$</annotation></semantics></math> </ephtml> MAR. The HLM implied a compatible nonstandard joint distribution <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0007" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>h</mi><mo>(</mo><mi>Y</mi><mo>,</mo><mi>C</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ h\left(Y,C\right) $$</annotation></semantics></math> </ephtml> such that the likelihood function <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0008" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>L</mi><mo>(</mo><mi>θ</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ L\left(\theta \right) $$</annotation></semantics></math> </ephtml> of the joint parameters <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0009" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>θ</mi></mrow><annotation encoding="application/x-tex">$$ \theta $$</annotation></semantics></math> </ephtml> is an integral with respect to high‐dimensional random effects <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0010" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>u</mi><mo>,</mo><mi>ν</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(u,\nu \right) $$</annotation></semantics></math> </ephtml> given the observed data <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0011" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left({Y}_{obs},{C}_{obs}\right) $$</annotation></semantics></math> </ephtml> . They selected "provisionally known random effects (PKREs)" <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0012" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>u</mi></mrow><annotation encoding="application/x-tex">$$ u $$</annotation></semantics></math> </ephtml> such that <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0013" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>h</mi><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>|</mo><mi>u</mi><mo>)</mo><mo>=</mo><mo>∫</mo><mi>h</mi><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><mi>ν</mi><mo>|</mo><mi>u</mi><mo>)</mo><mi>d</mi><mi>ν</mi></mrow><annotation encoding="application/x-tex">$$ h\left({Y}_{obs},{C}_{obs}|u\right)=\int h\left({Y}_{obs},{C}_{obs},\nu |u\right) d\nu $$</annotation></semantics></math> </ephtml> is analytically tractable with respect to high dimensional <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0014" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>ν</mi></mrow><annotation encoding="application/x-tex">$$ \nu $$</annotation></semantics></math> </ephtml> and, then, <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0015" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>L</mi><mo>(</mo><mi>θ</mi><mo>)</mo><mo>=</mo><mo>∫</mo><mi>h</mi><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>|</mo><mi>u</mi><mo>)</mo><mi>g</mi><mo>(</mo><mi>u</mi><mo>)</mo><mi>d</mi><mi>u</mi></mrow><annotation encoding="application/x-tex">$$ L\left(\theta \right)=\int h\left({Y}_{obs},{C}_{obs}|u\right)g(u) du $$</annotation></semantics></math> </ephtml> is approximated numerically with respect to low dimensional <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0016" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>u</mi></mrow><annotation encoding="application/x-tex">$$ u $$</annotation></semantics></math> </ephtml> by adaptive Gauss‐Hermite quadrature (AGHQ). They maximized the likelihood by the EM algorithm to produce the estimates <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0017" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mover accent="true"><mrow><mi>θ</mi></mrow><mo>^</mo></mover></mrow><annotation encoding="application/x-tex">$$ \hat{\theta} $$</annotation></semantics></math> </ephtml> at ML, and translated the <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0018" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mover accent="true"><mrow><mi>θ</mi></mrow><mo>^</mo></mover></mrow><annotation encoding="application/x-tex">$$ \hat{\theta} $$</annotation></semantics></math> </ephtml> to the ML estimates of the HLM. They illustrated efficient ML estimation of random and fixed interaction effects involving lower‐level covariates and their latent cluster means. Given the estimates <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0019" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mover accent="true"><mrow><mi>θ</mi></mrow><mo>^</mo></mover></mrow><annotation encoding="application/x-tex">$$ \hat{\theta} $$</annotation></semantics></math> </ephtml> and associated <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0020" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>v</mi><mi>a</mi><mi>r</mi><mo>(</mo><mover accent="true"><mrow><mi>θ</mi></mrow><mo>^</mo></mover><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathit{\operatorname{var}}\left(\hat{\theta}\right) $$</annotation></semantics></math> </ephtml> at ML, Shin and Raudenbush [[<reflink idref="bib26" id="ref25">26</reflink>]] generated MI of missing data, including random effects, from a nonstandard predictive model <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0021" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>h</mi><mo>(</mo><mi>Y</mi><mo>,</mo><mi>C</mi><mo>,</mo><mi>ν</mi><mo>,</mo><mi>u</mi><mo>|</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><mi>θ</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ h\left(Y,C,\nu, u|{Y}_{obs},{C}_{obs},\theta \right) $$</annotation></semantics></math> </ephtml> by importance sampling via AGHQ, and estimated the HLM given the MI. These approaches within the maximum‐likelihood framework, however, are suboptimal given small sample sizes.</p> <p>We now explain our contribution to the literature. Unlike existing samplers that sample missing values of a cluster‐level covariate having a nonlinear effect by a metropolis algorithm via a proposal normal density with a constant variance [[<reflink idref="bib21" id="ref26">21</reflink>], [<reflink idref="bib23" id="ref27">23</reflink>]], our Gibbs sampler samples missing data and parameters from their exact posterior distributions, which are derived via selected PKREs. Therefore, our Gibbs sampler is guaranteed to be compatible and, thus, to produce unbiased estimation of an analytic HLM with nonlinear effects of cluster‐level covariates. Either one or both of the interactive covariates may be partially observed and partially observed covariates are continuous. The results of the simulation study below reveal that our Gibbs sampler is particularly effective with comparatively small sample sizes.</p> <p>Our motivating application is the analysis of racially discordant patient–physician medical interactions, focusing on pairs of patients and physicians from different racial groups during office visits. These visits were videotaped and coded to produce four repeated measurements of the outcome: a positive valence score measuring the physician's facial expression. Of interest are the main and interaction effects of continuous physician's implicit and explicit prejudices on the outcome that may be appropriately analyzed using a HLM. Three challenges motivated us to develop our Gibbs sampler. First, valence score is missing 20% of the values, and each of the two key physician covariates has 16% missing data at the cluster level. Next, the effects of partially observed physician prejudices include an interaction effect. Finally, due to the COVID restrictions, only 37 encounters were recorded from the small number of 6 physicians and 37 patients, posing a formidable challenge to the efficient and unbiased estimation of the HLM.</p> <p>The rest of the paper is organized as follows. Section 2 introduces our general HLM with the nonlinear effects of cluster‐level covariates. Section 3 presents our Gibbs sampler. Section 4 evaluates our sampler by comparing our estimators with those by existing methods in simulated analyses. Section 5 illustrates analysis of real data from the racially discordant patient–physician encounters using our Gibbs sampler. Lastly, section 6 discusses limitations and future extensions of our approach.</p> <hd id="AN0185452270-3">Model</hd> <p>Our interest focuses on a two‐level HLM 1 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0022" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>Yij=β0+βCTCj+βXTXij+∑s=1pβXCsTXijCsj+∑s=1p−1∑t=s+1pβCCstCsjCtj+uj+eij</mrow><annotation encoding="application/x-tex">$$ {Y}_{ij}\kern0.5em ={\beta}_0+{\beta}_C^T{C}_j+{\beta}_X^T{X}_{ij}+\sum \limits_{s=1}^p{\beta}_{XCs}^T{X}_{ij}{C}_{sj}+\sum \limits_{s=1}^{p-1}\sum \limits_{t=s+1}^p{\beta}_{CCst}{C}_{sj}{C}_{tj}+{u}_j+{e}_{ij} $$</annotation></semantics></math> </ephtml> where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0023" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> is the outcome variable, <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0024" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>=</mo><msup><mrow><mo>[</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mi>...</mi><msub><mrow><mi>C</mi></mrow><mrow><mi>p</mi><mi>j</mi></mrow></msub><mo>]</mo></mrow><mrow><mi>T</mi></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {C}_j={\left[{C}_{1j}\dots {C}_{pj}\right]}^T $$</annotation></semantics></math> </ephtml> is a <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0025" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi></mrow><annotation encoding="application/x-tex">$$ p $$</annotation></semantics></math> </ephtml> ‐by‐1 vector of partially observed cluster‐level continuous covariates having main effects <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0026" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>C</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_C $$</annotation></semantics></math> </ephtml> , and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0027" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>=</mo><mo>[</mo><msubsup><mrow><mi>x</mi></mrow><mrow><mn>1</mn><mi>i</mi><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>x</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mo>]</mo></mrow><annotation encoding="application/x-tex">$$ {X}_{ij}=\left[{x}_{1 ij}^T\kern3.0235pt {x}_{2j}^T\right] $$</annotation></semantics></math> </ephtml> is a <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0028" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>q</mi></mrow><annotation encoding="application/x-tex">$$ q $$</annotation></semantics></math> </ephtml> ‐by‐1 vector of fully observed covariates having main effects <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0029" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi></mrow></msub><mo>=</mo><msup><mrow><mo>[</mo><msubsup><mrow><mi>β</mi></mrow><mrow><mi>x</mi><mn>1</mn></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>β</mi></mrow><mrow><mi>x</mi><mn>2</mn></mrow><mrow><mi>T</mi></mrow></msubsup><mo>]</mo></mrow><mrow><mi>T</mi></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {\beta}_X={\left[{\beta}_{x1}^T\kern3.0235pt {\beta}_{x2}^T\right]}^T $$</annotation></semantics></math> </ephtml> for lower‐level or level‐1 covariates <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0030" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>x</mi></mrow><mrow><mn>1</mn><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {x}_{1 ij} $$</annotation></semantics></math> </ephtml> and cluster‐level or level‐2 covariates <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0031" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>x</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {x}_{2j} $$</annotation></semantics></math> </ephtml> . In addition, <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0032" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi><mi>s</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_{XCs} $$</annotation></semantics></math> </ephtml> is the q‐by‐1 interaction effects of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0033" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {X}_{ij} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0034" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>s</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{sj} $$</annotation></semantics></math> </ephtml> , whereas <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0035" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>C</mi><mi>C</mi><mi>s</mi><mi>t</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_{CCst} $$</annotation></semantics></math> </ephtml> represents the scalar interaction effect of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0036" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>s</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{sj} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0037" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>t</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{tj} $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0038" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>s</mi><mo>≤</mo><mi>t</mi></mrow><annotation encoding="application/x-tex">$$ s\le t $$</annotation></semantics></math> </ephtml> . Lastly, a level‐2 unit‐specific random effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0039" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>u</mi><mi>j</mi></msub><mo>∼</mo><mi>N</mi><mo form="prefix">(</mo><mn>0</mn><mo separator="true">,</mo><mi>τ</mi><mo form="postfix">)</mo></mrow><annotation encoding="application/x-tex">$$ {u}_j\sim N\left(0,\tau \right) $$</annotation></semantics></math> </ephtml> and a level‐1 unit specific random effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0040" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>e</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>0</mn><mo>,</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {e}_{ij}\sim N\left(0,{\sigma}^2\right) $$</annotation></semantics></math> </ephtml> are independent, and a level‐1 unit <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0041" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>i</mi></mrow><annotation encoding="application/x-tex">$$ i $$</annotation></semantics></math> </ephtml> is nested within a level‐2 cluster <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0042" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>j</mi></mrow><annotation encoding="application/x-tex">$$ j $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0043" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>i</mi><mo>=</mo><mn>1</mn><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>n</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ i=1,\dots, {n}_j $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0044" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>j</mi><mo>=</mo><mn>1</mn><mo>,</mo><mi>...</mi><mo>,</mo><mi>J</mi></mrow><annotation encoding="application/x-tex">$$ j=1,\dots, J $$</annotation></semantics></math> </ephtml> . Here, the covariates <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0045" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_j $$</annotation></semantics></math> </ephtml> and the outcome <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0046" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> may be partially observed.</p> <hd id="AN0185452270-4">Compatible Gibbs Sampler</hd> <p>This section explains our Gibbs sampler based on the exact posteriors of parameters and missing data, unlike existing Bayesian approaches using rejection sampling of missing data by the Metropolis algorithm [[<reflink idref="bib21" id="ref28">21</reflink>], [<reflink idref="bib23" id="ref29">23</reflink>]]. To handle missing data efficiently, we assume the joint normal distribution of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0047" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_j $$</annotation></semantics></math> </ephtml> conditional on known covariates to write our Bayesian joint distribution of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0048" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">Y</mi><mo>=</mo><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mn>11</mn></mrow></msub><mo>,</mo><msub><mrow><mi>Y</mi></mrow><mrow><mn>12</mn></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>Y</mi></mrow><mrow><msub><mrow><mi>n</mi></mrow><mrow><mi>J</mi></mrow></msub><mi>J</mi></mrow></msub><mo>)</mo><mo>,</mo><mi mathvariant="bold">C</mi><mo>=</mo><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>J</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{Y}=\left({Y}_{11},{Y}_{12},\dots, {Y}_{n_JJ}\right),\mathbf{C}=\left({C}_1,{C}_2,\dots, {C}_J\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0049" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>θ</mi></mrow><annotation encoding="application/x-tex">$$ \theta $$</annotation></semantics></math> </ephtml> given <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0050" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">X</mi><mo>=</mo><mo>(</mo><msub><mrow><mi>X</mi></mrow><mrow><mn>11</mn></mrow></msub><mo>,</mo><msub><mrow><mi>X</mi></mrow><mrow><mn>12</mn></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>X</mi></mrow><mrow><msub><mrow><mi>n</mi></mrow><mrow><mi>J</mi></mrow></msub><mi>J</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{X}=\left({X}_{11},{X}_{12},\dots, {X}_{n_JJ}\right) $$</annotation></semantics></math> </ephtml></p> <p>2 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0051" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>f(Y,C,θ|X)=∏j=1J∏i=1njf(Yij|Cj,Xij,uj,β,σ2)f(uj|τ)f(Cj|x2j,α,T)p(θ)</mrow><annotation encoding="application/x-tex">$$ f\left(\mathbf{Y},\mathbf{C},\theta |\mathbf{X}\right)\kern0.5em =\prod \limits_{j=1}^J\prod \limits_{i=1}^{n_j}f\left({Y}_{ij}|{C}_j,{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2\right)f\left({u}_j|\tau \right)f\left({C}_j|{x}_{2j},\alpha, T\right)p\left(\theta \right) $$</annotation></semantics></math> </ephtml> where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0052" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><mi>θ</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left(\theta \right) $$</annotation></semantics></math> </ephtml> is the prior distribution of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0053" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>θ</mi><mo>=</mo><mo>(</mo><mi mathvariant="bold-italic">β</mi><mo>,</mo><mi>τ</mi><mo>,</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>,</mo><mi>α</mi><mo>,</mo><mi>T</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \theta =\left(\boldsymbol{\beta}, \tau, {\sigma}^2,\alpha, T\right) $$</annotation></semantics></math> </ephtml> , as specified in the Gibbs sampler steps below, with <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0054" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold-italic">β</mi><mo>=</mo><mo>(</mo><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>C</mi></mrow></msub><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi></mrow></msub><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi></mrow></msub><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>C</mi><mi>C</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \boldsymbol{\beta} =\left({\beta}_0,{\beta}_C,{\beta}_X,{\beta}_{XC},{\beta}_{CC}\right) $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0055" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi></mrow></msub><mo>=</mo><mo>(</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi><mn>1</mn></mrow></msub><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi><mn>2</mn></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi><mi>p</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {\beta}_{XC}=\left({\beta}_{XC1},{\beta}_{XC2},\dots, {\beta}_{XC p}\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0056" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>β</mi><mrow><mi>C</mi><mi>C</mi></mrow></msub><mo>=</mo><mo form="prefix">(</mo><msub><mi>β</mi><mrow><mi>C</mi><mi>C</mi><mn>12</mn></mrow></msub><mo separator="true">,</mo><mi>...</mi><mo separator="true">,</mo><msub><mi>β</mi><mrow><mi>C</mi><mi>C</mi><mn>1</mn><mi>p</mi></mrow></msub><mo separator="true">,</mo><msub><mi>β</mi><mrow><mi>C</mi><mi>C</mi><mn>23</mn></mrow></msub><mi>...</mi><mo separator="true">,</mo><msub><mi>β</mi><mrow><mi>C</mi><mi>C</mi><mn>2</mn><mi>p</mi></mrow></msub><mo separator="true">,</mo><mi>...</mi><mo separator="true">,</mo><msub><mi>β</mi><mrow><mi>C</mi><mi>C</mi><mo form="prefix">(</mo><mi>p</mi><mo>−</mo><mn>1</mn><mo form="postfix">)</mo><mi>p</mi></mrow></msub><mo form="postfix">)</mo><mo separator="true">.</mo></mrow><annotation encoding="application/x-tex">$$ {\beta}_{CC}=\left({\beta}_{CC12},\dots, {\beta}_{CC1p},{\beta}_{CC23}\dots, {\beta}_{CC2p},\dots, {\beta}_{CC\left(p-1\right)p}\right). $$</annotation></semantics></math> </ephtml> Here <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0057" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>f</mi><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>|</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>,</mo><msub><mrow><mi>u</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><mi mathvariant="bold-italic">β</mi><mo>,</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ f\left({Y}_{ij}|{C}_j,{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0058" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>f</mi><mo>(</mo><msub><mrow><mi>u</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>|</mo><mi>τ</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ f\left({u}_j|\tau \right) $$</annotation></semantics></math> </ephtml> are normal densities from the HLM (<reflink idref="bib1" id="ref30">1</reflink>), and 3 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0059" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>f(Cj|x2j)∼NWα=Ip⊗1x2jTα,T</mrow><annotation encoding="application/x-tex">$$ f\left({C}_j|{x}_{2j}\right)\sim N\left[ W\alpha =\left({I}_p\otimes \left[1\kern2.56804pt {x}_{2j}^T\right]\right)\alpha, T\right] $$</annotation></semantics></math> </ephtml> for a vector <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0060" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>α</mi></mrow><annotation encoding="application/x-tex">$$ \alpha $$</annotation></semantics></math> </ephtml> of fixed effects, a <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0061" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo form="prefix">×</mo><mi>p</mi></mrow><annotation encoding="application/x-tex">$$ p\times p $$</annotation></semantics></math> </ephtml> identity matrix <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0062" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>I</mi></mrow><mrow><mi>p</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {I}_p $$</annotation></semantics></math> </ephtml> , a kronecker product <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0063" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>A</mi><mo>⊗</mo><mi>B</mi></mrow><annotation encoding="application/x-tex">$$ A\otimes B $$</annotation></semantics></math> </ephtml> multiplying matrix B to each element of matrix A, and a <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0064" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo form="prefix">×</mo><mi>p</mi></mrow><annotation encoding="application/x-tex">$$ p\times p $$</annotation></semantics></math> </ephtml> variance‐covariance matrix <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0065" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>T</mi></mrow><annotation encoding="application/x-tex">$$ T $$</annotation></semantics></math> </ephtml> . To derive the Gibbs sampler, we partition complete data <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0066" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">Y</mi><mo>=</mo><mo>(</mo><msub><mrow><mi mathvariant="bold">Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi mathvariant="bold">Y</mi></mrow><mrow><mi>m</mi><mi>i</mi><mi>s</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{Y}=\left({\mathbf{Y}}_{obs},{\mathbf{Y}}_{mis}\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0067" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">C</mi><mo>=</mo><mo>(</mo><msub><mrow><mi mathvariant="bold">C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi mathvariant="bold">C</mi></mrow><mrow><mi>m</mi><mi>i</mi><mi>s</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{C}=\left({\mathbf{C}}_{obs},{\mathbf{C}}_{mis}\right) $$</annotation></semantics></math> </ephtml> into observed <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0068" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><msub><mrow><mi mathvariant="bold">Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi mathvariant="bold">C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left({\mathbf{Y}}_{obs},{\mathbf{C}}_{obs}\right) $$</annotation></semantics></math> </ephtml> and missing <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0069" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><msub><mrow><mi mathvariant="bold">Y</mi></mrow><mrow><mi>m</mi><mi>i</mi><mi>s</mi></mrow></msub><mo>,</mo><msub><mrow><mi mathvariant="bold">C</mi></mrow><mrow><mi>m</mi><mi>i</mi><mi>s</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left({\mathbf{Y}}_{mis},{\mathbf{C}}_{mis}\right) $$</annotation></semantics></math> </ephtml> .</p> <hd id="AN0185452270-5">Exact Posterior Distributions of Cmis$$ {C}_{mis} $$</hd> <p>Let <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0071" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><mi>A</mi><mo>|</mo><mo>·</mo><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left(A|\cdotp \right) $$</annotation></semantics></math> </ephtml> denote the posterior, or exact posterior, distribution of A given all other unknowns. To find the key posterior distribution <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0072" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub><mo>|</mo><mo>·</mo><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left({C}_{kj}|\cdotp \right) $$</annotation></semantics></math> </ephtml> for a missing element <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0073" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{kj} $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0074" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_j $$</annotation></semantics></math> </ephtml> , we first derive a bivariate normal conditional distribution from Equations (<reflink idref="bib1" id="ref31">1</reflink>) and (<reflink idref="bib3" id="ref32">3</reflink>) 4 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0075" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>YijCkjC(−k)j,Xij,uj,β,σ2,α,T∼Nμ1ij+μ2ijMk|(−k)Mk|(−k),μ2ij2Tk|(−k)+σ2μ2ijTk|(−k)μ2ijTk|(−k)Tk|(−k)</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}& \left[\begin{array}{c}{Y}_{ij}\\ {}{C}_{kj}\end{array}\right]\left|{C}_{\left(-k\right)j},{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2,\alpha, T\right.\\ {}& \kern1em \sim N\left(\left[\begin{array}{c}{\mu}_{1 ij}+{\mu}_{2 ij}{M}_{k\mid \left(-k\right)}\\ {}{M}_{k\mid \left(-k\right)}\end{array}\right],\left[\begin{array}{cc}{\mu}_{2 ij}^2{T}_{k\mid \left(-k\right)}+{\sigma}^2& {\mu}_{2 ij}{T}_{k\mid \left(-k\right)}\\ {}{\mu}_{2 ij}{T}_{k\mid \left(-k\right)}& {T}_{k\mid \left(-k\right)}\end{array}\right]\right)\end{array}} $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0076" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo><mi>j</mi></mrow></msub><mo>=</mo><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mi>k</mi><mo form="prefix">−</mo><mn>1</mn><mo>)</mo><mi>j</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mi>k</mi><mo>+</mo><mn>1</mn><mo>)</mo><mi>j</mi></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>p</mi><mi>j</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {C}_{\left(-k\right)j}=\left({C}_{1j},\dots, {C}_{\left(k-1\right)j},{C}_{\left(k+1\right)j},\dots, {C}_{pj}\right) $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0077" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>M</mi></mrow><mrow><mi>k</mi><mo>|</mo><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow></msub><mo>=</mo><mi>E</mi><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub><mo>|</mo><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo><mi>j</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {M}_{k\mid \left(-k\right)}=E\left({C}_{kj}|{C}_{\left(-k\right)j}\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0078" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>T</mi></mrow><mrow><mi>k</mi><mo>|</mo><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow></msub><mo>=</mo><mi>v</mi><mi>a</mi><mi>r</mi><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub><mo>|</mo><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo><mi>j</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {T}_{k\mid \left(-k\right)}=\mathit{\operatorname{var}}\left({C}_{kj}|{C}_{\left(-k\right)j}\right) $$</annotation></semantics></math> </ephtml> . Shin and Raudenbush [[<reflink idref="bib18" id="ref33">18</reflink>]] referred to missing values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0079" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{\left(-k\right)j} $$</annotation></semantics></math> </ephtml> as provisionally known random effects. The conditional mean <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0080" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>E</mi><mo>(</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>|</mo><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo><mi>j</mi></mrow></msub><mo>,</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>,</mo><msub><mrow><mi>u</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ E\left({Y}_{ij}|{C}_{\left(-k\right)j},{X}_{ij},{u}_j\right) $$</annotation></semantics></math> </ephtml> has two parts, <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0081" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>μ</mi></mrow><mrow><mn>1</mn><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\mu}_{1 ij} $$</annotation></semantics></math> </ephtml> excluding and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0082" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>μ</mi></mrow><mrow><mn>2</mn><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\mu}_{2 ij} $$</annotation></semantics></math> </ephtml> including <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0083" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{kj} $$</annotation></semantics></math> </ephtml> , to facilitate derivation of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0084" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub><mo>|</mo><mo>·</mo><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left({C}_{kj}|\cdotp \right) $$</annotation></semantics></math> </ephtml> : <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0085" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>μ1ij=β0+βC(−k)TC(−k)j+βXTXij +∑s≠k,s≥1pβXCsTXijCsj+∑s≠k,s≥1p−1∑t≠k,t≥spβCCstCsjCtj+uj,μ2ij=βCk+βXCkTXij+∑s=1p−1∑t=s+1pβCCstCtjI(s=k) +βCCstCsjI(t=k)</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}{\mu}_{1 ij}& ={\beta}_0+{\beta}_{C\left(-k\right)}^T{C}_{\left(-k\right)j}+{\beta}_X^T{X}_{ij}\\ {}& \kern1em +\sum \limits_{s\ne k,s\ge 1}^p{\beta}_{XCs}^T{X}_{ij}{C}_{sj}+\sum \limits_{s\ne k,s\ge 1}^{p-1}\sum \limits_{t\ne k,t\ge s}^p{\beta}_{CCst}{C}_{sj}{C}_{tj}+{u}_j,\\ {}{\mu}_{2 ij}& =\left({\beta}_{C_k}+{\beta}_{X{C}_k}^T{X}_{ij}+\sum \limits_{s=1}^{p-1}\sum \limits_{t=s+1}^p\left({\beta}_{CCst}{C}_{tj}I\left(s=k\right)\right.\right.\\ {}& \kern1em \left.\left.+{\beta}_{CCst}{C}_{sj}I\left(t=k\right)\right)\right)\end{array}} $$</annotation></semantics></math> </ephtml> for an indicator function <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0086" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>I</mi><mo>(</mo><mi>B</mi><mo>)</mo><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">$$ I(B)=1 $$</annotation></semantics></math> </ephtml> if condition <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0087" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>B</mi></mrow><annotation encoding="application/x-tex">$$ B $$</annotation></semantics></math> </ephtml> is true and 0 otherwise, <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0088" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>C</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_C $$</annotation></semantics></math> </ephtml> partitioned into the coefficients <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0089" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi></mrow></msub></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_{C_k} $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0090" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{kj} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0091" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mi>C</mi><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_{C\left(-k\right)} $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0092" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{\left(-k\right)j} $$</annotation></semantics></math> </ephtml> , and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0093" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>s</mi><mo>≠</mo><mi>k</mi><mo>,</mo><mi>s</mi><mo>≥</mo><mi>a</mi></mrow><mrow><mi>p</mi></mrow></msubsup><msub><mrow><mi>A</mi></mrow><mrow><mi>s</mi></mrow></msub><mo>=</mo><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>s</mi><mo>=</mo><mi>a</mi></mrow><mrow><mi>k</mi><mo form="prefix">−</mo><mn>1</mn></mrow></msubsup><msub><mrow><mi>A</mi></mrow><mrow><mi>s</mi></mrow></msub><mo>+</mo><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>s</mi><mo>=</mo><mi>k</mi><mo>+</mo><mn>1</mn></mrow><mrow><mi>p</mi></mrow></msubsup><msub><mrow><mi>A</mi></mrow><mrow><mi>s</mi></mrow></msub><mo>.</mo></mrow><annotation encoding="application/x-tex">$$ {\sum}_{s\ne k,s\ge a}^p{A}_s={\sum}_{s=a}^{k-1}{A}_s+{\sum}_{s=k+1}^p{A}_s. $$</annotation></semantics></math> </ephtml></p> <p>The joint distribution (<reflink idref="bib2" id="ref34">2</reflink>) then implies 5 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0094" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(Ckj|·)∝∏i=1njf(Yij|Cj,Xij,uj,β,σ2)×f(Ckj|C(−k)j,x2j,α,T)∼N(M˜kj,Δkj−1)</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}p\left({C}_{kj}|\cdotp \right)& \propto \prod \limits_{i=1}^{n_j}f\left({Y}_{ij}|{C}_j,{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2\right)\\ {}& \times f\left({C}_{kj}|{C}_{\left(-k\right)j},{x}_{2j},\alpha, T\right)\sim N\left({\tilde{M}}_{kj},{\Delta}_{kj}^{-1}\right)\end{array}} $$</annotation></semantics></math> </ephtml> where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0095" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mover accent="true"><mrow><mi>M</mi></mrow><mo>˜</mo></mover></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub><mo>=</mo><msub><mrow><mi>M</mi></mrow><mrow><mi>k</mi><mo>|</mo><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow></msub><mo>+</mo><msubsup><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow><mrow><mo form="prefix">−</mo><mn>1</mn></mrow></msubsup><msup><mrow><mi>σ</mi></mrow><mrow><mo form="prefix">−</mo><mn>2</mn></mrow></msup><msub><mrow><mi>μ</mi></mrow><mrow><mn>2</mn><mi>i</mi><mi>j</mi></mrow></msub><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mrow><msub><mrow><mi>n</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow></msubsup><mfenced separators="" open="(" close=")"><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo form="prefix">−</mo><mo>(</mo><msub><mrow><mi>μ</mi></mrow><mrow><mn>1</mn><mi>i</mi><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>μ</mi></mrow><mrow><mn>2</mn><mi>i</mi><mi>j</mi></mrow></msub><msub><mrow><mi>M</mi></mrow><mrow><mi>k</mi><mo>|</mo><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow></msub><mo>)</mo></mrow></mfenced></mrow><annotation encoding="application/x-tex">$$ {\tilde{M}}_{kj}={M}_{k\mid \left(-k\right)}+{\Delta}_{kj}^{-1}{\sigma}^{-2}{\mu}_{2 ij}{\sum}_{i=1}^{n_j}\left({Y}_{ij}-\left({\mu}_{1 ij}+{\mu}_{2 ij}{M}_{k\mid \left(-k\right)}\right)\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0096" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub><mo>=</mo><msubsup><mrow><mi>T</mi></mrow><mrow><mi>k</mi><mo>|</mo><mo>(</mo><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow><mrow><mo form="prefix">−</mo><mn>1</mn></mrow></msubsup><mo>+</mo><msub><mrow><mi>n</mi></mrow><mrow><mi>j</mi></mrow></msub><msubsup><mrow><mi>μ</mi></mrow><mrow><mn>2</mn><mi>i</mi><mi>j</mi></mrow><mrow><mn>2</mn></mrow></msubsup><msup><mrow><mi>σ</mi></mrow><mrow><mo form="prefix">−</mo><mn>2</mn></mrow></msup><mo>.</mo></mrow><annotation encoding="application/x-tex">$$ {\Delta}_{kj}={T}_{k\mid \left(-k\right)}^{-1}+{n}_j{\mu}_{2 ij}^2{\sigma}^{-2}. $$</annotation></semantics></math> </ephtml></p> <hd id="AN0185452270-6">Gibbs Sampler Steps Based on Exact Posterior Distributions</hd> <p>The joint distribution (<reflink idref="bib2" id="ref35">2</reflink>) implies the posterior distributions of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0097" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">u</mi><mo>=</mo><mo>(</mo><msub><mrow><mi>u</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>,</mo><mi>...</mi><mo>,</mo><msub><mrow><mi>u</mi></mrow><mrow><mi>J</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{u}=\left({u}_1,\dots, {u}_J\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0098" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>θ</mi></mrow><annotation encoding="application/x-tex">$$ \theta $$</annotation></semantics></math> </ephtml> : 6 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0099" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(uj|·)∝∏i=1njf(Yij|Cj,Xij,uj,β,σ2)f(uj|τ), p(τ|·)∝∏j=1Jf(uj|τ)p(τ), p(β|·)∝∏j=1J∏i=1njf(Yij|Cj,Xij,uj,β,σ2)f(β), p(σ2|·)∝∏j=1J∏i=1njf(Yij|Cj,Xij,uj,β,σ2)p(σ2), p(α|·)∝∏j=1Jf(Cj|x2j,α,T)p(α), p(T|·)∝∏j=1Jf(Cj|x2j,α,T)p(T)</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}& p\left({u}_j|\cdotp \right)\propto \prod \limits_{i=1}^{n_j}f\left({Y}_{ij}|{C}_j,{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2\right)f\left({u}_j|\tau \right),\\ {}& p\left(\tau |\cdotp \right)\propto \prod \limits_{j=1}^Jf\left({u}_j|\tau \right)p\left(\tau \right),\\ {}& p\left(\boldsymbol{\beta} |\cdotp \right)\propto \prod \limits_{j=1}^J\prod \limits_{i=1}^{n_j}f\left({Y}_{ij}|{C}_j,{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2\right)f\left(\boldsymbol{\beta} \right),\\ {}& p\left({\sigma}^2|\cdotp \right)\propto \prod \limits_{j=1}^J\prod \limits_{i=1}^{n_j}f\left({Y}_{ij}|{C}_j,{X}_{ij},{u}_j,\boldsymbol{\beta}, {\sigma}^2\right)p\left({\sigma}^2\right),\\ {}& p\left(\alpha |\cdotp \right)\propto \prod \limits_{j=1}^Jf\left({C}_j|{x}_{2j},\alpha, T\right)p\left(\alpha \right),\\ {}& p\left(T|\cdotp \right)\propto \prod \limits_{j=1}^Jf\left({C}_j|{x}_{2j},\alpha, T\right)p(T)\end{array}} $$</annotation></semantics></math> </ephtml> where we assume inverse gamma priors <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0100" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><mi>τ</mi><mo>)</mo><mo>∼</mo><mi>I</mi><mi>G</mi><mo>(</mo><msub><mrow><mi>α</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mn>1</mn><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mn>0</mn><mo>.</mo><mn>5</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left(\tau \right)\sim IG\left({\alpha}_0=1,{\beta}_0=0.5\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0101" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo><mo>∼</mo><mi>I</mi><mi>G</mi><mo>(</mo><msub><mrow><mi>α</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mn>1</mn><mo>,</mo><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mn>0</mn><mo>.</mo><mn>5</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left({\sigma}^2\right)\sim IG\left({\alpha}_0=1,{\beta}_0=0.5\right) $$</annotation></semantics></math> </ephtml> , an inverse wishart <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0102" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><mi>T</mi><mo>)</mo><mo>∼</mo><mi>I</mi><mi>W</mi><mo>(</mo><msub><mrow><mi>V</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>,</mo><msubsup><mrow><mi>S</mi></mrow><mrow><mn>0</mn></mrow><mrow><mo form="prefix">−</mo><mn>1</mn></mrow></msubsup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p(T)\sim IW\left({V}_0,{S}_0^{-1}\right) $$</annotation></semantics></math> </ephtml> with <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0103" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>V</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mi>p</mi><mo>+</mo><mn>2</mn></mrow><annotation encoding="application/x-tex">$$ {V}_0=p+2 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0104" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>S</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mover accent="true"><mrow><mi>T</mi></mrow><mo>^</mo></mover></mrow><annotation encoding="application/x-tex">$$ {S}_0=\hat{T} $$</annotation></semantics></math> </ephtml> (an estimated variance‐covariance matrix using complete cases), and noninformative priors <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0105" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><mi>β</mi><mo>)</mo><mo>=</mo><mi>p</mi><mo>(</mo><mi>α</mi><mo>)</mo><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">$$ p\left(\beta \right)=p\left(\alpha \right)=1 $$</annotation></semantics></math> </ephtml> , following Schafer and Yucel [[<reflink idref="bib6" id="ref36">6</reflink>]], Hoff [[<reflink idref="bib27" id="ref37">27</reflink>]] and Enders et al. [[<reflink idref="bib21" id="ref38">21</reflink>]].</p> <p>At cycle <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0106" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>t</mi></mrow><annotation encoding="application/x-tex">$$ t $$</annotation></semantics></math> </ephtml> , we update the current values of parameters <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0107" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>θ</mi><mo>=</mo><msup><mrow><mi>θ</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo form="prefix">−</mo><mn>1</mn><mo>)</mo></mrow></msup></mrow><annotation encoding="application/x-tex">$$ \theta ={\theta}^{\left(t-1\right)} $$</annotation></semantics></math> </ephtml> and completed data <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0108" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">Y</mi><mo>=</mo><mo>(</mo><msub><mrow><mi mathvariant="bold">Y</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msubsup><mrow><mi mathvariant="bold">Y</mi></mrow><mrow><mi>m</mi><mi>i</mi><mi>s</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo form="prefix">−</mo><mn>1</mn><mo>)</mo></mrow></msubsup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{Y}=\left({\mathbf{Y}}_{obs},{\mathbf{Y}}_{mis}^{\left(t-1\right)}\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0109" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold">C</mi><mo>=</mo><mo>(</mo><msub><mrow><mi mathvariant="bold">C</mi></mrow><mrow><mi>o</mi><mi>b</mi><mi>s</mi></mrow></msub><mo>,</mo><msubsup><mrow><mi mathvariant="bold">C</mi></mrow><mrow><mi>m</mi><mi>i</mi><mi>s</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo form="prefix">−</mo><mn>1</mn><mo>)</mo></mrow></msubsup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \mathbf{C}=\left({\mathbf{C}}_{obs},{\mathbf{C}}_{mis}^{\left(t-1\right)}\right) $$</annotation></semantics></math> </ephtml> from cycle <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0110" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>t</mi><mo form="prefix">−</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">$$ t-1 $$</annotation></semantics></math> </ephtml> in eight steps, as follows:</p> <p></p> <ulist> <item> Step 1: Sample <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0111" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msubsup><mrow><mi>u</mi></mrow><mrow><mi>j</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msubsup></mrow><annotation encoding="application/x-tex">$$ {u}_j^{(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0112" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(uj|Y=Y(t−1),θ=θ(t−1))∼N(Δj−1σ−2∑i=1nj(Yij−XijTβ),Δj−1)</mrow><annotation encoding="application/x-tex">$$ \kern0.5em p\left({u}_j|\mathbf{Y}={\mathbf{Y}}^{\left(t-1\right)},\theta ={\theta}^{\left(t-1\right)}\right)\sim N\left({\Delta}_j^{-1}{\sigma}^{-2}\sum \limits_{i=1}^{n_j}\left({Y}_{ij}-{\mathbf{X}}_{ij}^T\boldsymbol{\beta} \right),{\Delta}_j^{-1}\right) $$</annotation></semantics></math> </ephtml> where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0113" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msubsup><mrow><mi mathvariant="bold">X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mo>=</mo><mfenced separators="" open="[" close="]"><mrow><mn>1</mn><mspace width="3.0235pt" /><msubsup><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>X</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mo>⊗</mo><msubsup><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msup><mrow><mtext>vech</mtext></mrow><mrow><mi>T</mi></mrow></msup><mfenced separators="" open="(" close=")"><mrow><msub><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mn>1</mn><mo>)</mo><mi>j</mi></mrow></msub><msubsup><mrow><mi>C</mi></mrow><mrow><mo>(</mo><mo form="prefix">−</mo><mi>p</mi><mo>)</mo><mi>j</mi></mrow><mrow><mi>T</mi></mrow></msubsup></mrow></mfenced></mrow></mfenced></mrow><annotation encoding="application/x-tex">$$ {\mathbf{X}}_{ij}^T=\left[1\kern3.0235pt {C}_j^T\kern3.0235pt {X}_{ij}^T\kern3.0235pt {X}_{ij}^T\otimes {C}_j^T\kern3.0235pt {\mathrm{vech}}^T\left({C}_{\left(-1\right)j}{C}_{\left(-p\right)j}^T\right)\right] $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0114" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi mathvariant="bold-italic">β</mi><mo>=</mo><msup><mrow><mfenced separators="" open="[" close="]"><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub><mspace width="3.0235pt" /><msubsup><mrow><mi>β</mi></mrow><mrow><mi>C</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>β</mi></mrow><mrow><mi>X</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>β</mi></mrow><mrow><mi>X</mi><mi>C</mi></mrow><mrow><mi>T</mi></mrow></msubsup><mspace width="3.0235pt" /><msubsup><mrow><mi>β</mi></mrow><mrow><mi>C</mi><mi>C</mi></mrow><mrow><mi>T</mi></mrow></msubsup></mrow></mfenced></mrow><mrow><mi>T</mi></mrow></msup></mrow><annotation encoding="application/x-tex">$$ \boldsymbol{\beta} ={\left[{\beta}_0\kern3.0235pt {\beta}_C^T\kern3.0235pt {\beta}_X^T\kern3.0235pt {\beta}_{XC}^T\kern3.0235pt {\beta}_{CC}^T\right]}^T $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0115" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>=</mo><msub><mrow><mi>n</mi></mrow><mrow><mi>j</mi></mrow></msub><msup><mrow><mi>σ</mi></mrow><mrow><mo form="prefix">−</mo><mn>2</mn></mrow></msup><mo>+</mo><msup><mrow><mi>τ</mi></mrow><mrow><mo form="prefix">−</mo><mn>1</mn></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {\Delta}_j={n}_j{\sigma}^{-2}+{\tau}^{-1} $$</annotation></semantics></math> </ephtml> . Here, the operator <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0116" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mtext>vech</mtext><mo>(</mo><mo>·</mo><mo>)</mo><mo>:</mo><msup><mrow><mi>ℝ</mi></mrow><mrow><mi>r</mi><mo form="prefix">×</mo><mi>r</mi></mrow></msup><mo>→</mo><msup><mrow><mi>ℝ</mi></mrow><mrow><mi>r</mi><mo>(</mo><mi>r</mi><mo>+</mo><mn>1</mn><mo>)</mo><mo>/</mo><mn>2</mn></mrow></msup></mrow><annotation encoding="application/x-tex">$$ \mathrm{vech}\left(\cdotp \right):{\mathbb{R}}^{r\times r}\to {\mathbb{R}}^{r\left(r+1\right)/2} $$</annotation></semantics></math> </ephtml> stacks the upper triangular entries of an <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0117" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>r</mi><mo form="prefix">×</mo><mi>r</mi></mrow><annotation encoding="application/x-tex">$$ r\times r $$</annotation></semantics></math> </ephtml> square matrix into a column vector.</item> <p></p> <item> Step 2: Draw <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0118" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>τ</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {\tau}^{(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0119" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(τ|u=u(t),θ=θ(t−1))∼IGJ2+α0,∑i=1njuj22+1β0−1.</mrow><annotation encoding="application/x-tex">$$ p\left(\tau |\mathbf{u}={\mathbf{u}}^{(t)},\theta ={\theta}^{\left(t-1\right)}\right)\sim IG\left(\frac{J}{2}+{\alpha}_0,{\left[\frac{\sum_{i=1}^{n_j}{u}_j^2}{2}+\frac{1}{\beta_0}\right]}^{-1}\right). $$</annotation></semantics></math> </ephtml></item> <p></p> <item> Step 3: Draw <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0120" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi mathvariant="bold-italic">β</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {\boldsymbol{\beta}}^{(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0121" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(β|Y=Y(t−1),u=u(t),σ2(t−1)) ∼N∑j=1J∑i=1njXijXijT−1∑j=1J∑i=1njXij(Yij−uj), σ2∑j=1J∑i=1njXijXijT−1.</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}& p\left(\boldsymbol{\beta} |\mathbf{Y}={\mathbf{Y}}^{\left(t-1\right)},\mathbf{u}={\mathbf{u}}^{(t)},{\sigma}^{2\left(t-1\right)}\right)\\ {}& \kern1em \sim N\left({\left(\sum \limits_{j=1}^J\sum \limits_{i=1}^{n_j}{\mathbf{X}}_{ij}{\mathbf{X}}_{ij}^T\right)}^{-1}\sum \limits_{j=1}^J\sum \limits_{i=1}^{n_j}{\mathbf{X}}_{ij}\left({Y}_{ij}-{u}_j\right),\right.\\ {}& \kern1em \left.{\sigma}^2{\left(\sum \limits_{j=1}^J\sum \limits_{i=1}^{n_j}{\mathbf{X}}_{ij}{\mathbf{X}}_{ij}^T\right)}^{-1}\right).\end{array}} $$</annotation></semantics></math> </ephtml></item> <p></p> <item> Step 4: Draw <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0122" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {\sigma}^{2(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0123" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(σ2|Y=Y(t−1),u=u(t),β=β(t))∼IGN2+α0,∑j=1J∑i=1njeij22+1β0−1</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}\hfill & p\left({\sigma}^2|\mathbf{Y}={\mathbf{Y}}^{\left(t-1\right)},\mathbf{u}={\mathbf{u}}^{(t)},\boldsymbol{\beta} ={\boldsymbol{\beta}}^{(t)}\right)\\ {}\hfill & \kern1em \sim IG\left(\frac{N}{2}+{\alpha}_0,{\left[\frac{\sum_{j=1}^J{\sum}_{i=1}^{n_j}{e}_{ij}^2}{2}+\frac{1}{\beta_0}\right]}^{-1}\right)\end{array}} $$</annotation></semantics></math> </ephtml> where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0124" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>N</mi><mo>=</mo><msubsup><mrow><mo>∑</mo></mrow><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>J</mi></mrow></msubsup><msub><mrow><mi>n</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ N={\sum}_{j=1}^J{n}_j $$</annotation></semantics></math> </ephtml> .</item> <p></p> <item> Step 5: For <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0125" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> missing, impute <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0126" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msubsup><mrow><mi>e</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msubsup></mrow><annotation encoding="application/x-tex">$$ {e}_{ij}^{(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0127" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi><mo>(</mo><msub><mrow><mi>e</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>|</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>=</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msup><mo>)</mo><mo>∼</mo><mi>N</mi><mo>(</mo><mn>0</mn><mo>,</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ p\left({e}_{ij}|{\sigma}^2={\sigma}^{2(t)}\right)\sim N\left(0,{\sigma}^2\right) $$</annotation></semantics></math> </ephtml> and set <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0128" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msubsup><mi>Y</mi><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mo form="prefix">(</mo><mi>t</mi><mo form="postfix">)</mo></mrow></msubsup><mo>=</mo><msubsup><mi mathvariant="bold">X</mi><mrow><mi>i</mi><mi>j</mi></mrow><mi>T</mi></msubsup><msup><mi mathvariant="bold-italic">β</mi><mrow><mo form="prefix">(</mo><mi>t</mi><mo form="postfix">)</mo></mrow></msup><mo>+</mo><msubsup><mi>u</mi><mi>j</mi><mrow><mo form="prefix">(</mo><mi>t</mi><mo form="postfix">)</mo></mrow></msubsup><mo>+</mo><msubsup><mi>e</mi><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mo form="prefix">(</mo><mi>t</mi><mo form="postfix">)</mo></mrow></msubsup><mo separator="true">.</mo></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij}^{(t)}={\mathbf{X}}_{ij}^T{\boldsymbol{\beta}}^{(t)}+{u}_j^{(t)}+{e}_{ij}^{(t)}. $$</annotation></semantics></math> </ephtml></item> <p></p> <item> Step 6: Draw <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0129" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>α</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {\alpha}^{(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0130" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(α|C=C(t−1),T=T(t−1)) ∼N∑j=1JWTT−1W−1∑j=1JWTT−1Cj,∑j=1JWTT−1W−1.</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}& p\left(\alpha |\mathbf{C}={\mathbf{C}}^{\left(t-1\right)},T={T}^{\left(t-1\right)}\right)\\ {}& \kern1em \sim N\left({\left(\sum \limits_{j=1}^J{W}^T{T}^{-1}W\right)}^{-1}\sum \limits_{j=1}^J{W}^T{T}^{-1}{C}_j,{\left(\sum \limits_{j=1}^J{W}^T{T}^{-1}W\right)}^{-1}\right).\end{array}} $$</annotation></semantics></math> </ephtml></item> <p></p> <item> Step 7: Draw <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0131" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>T</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msup></mrow><annotation encoding="application/x-tex">$$ {T}^{(t)} $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0132" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(T|C=C(t−1),α=α(t)) ∼IWV0+J,S0+∑j=1J(Cj−Wα)(Cj−Wα)T−1.</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}& p\left(T|\mathbf{C}={\mathbf{C}}^{\left(t-1\right)},\alpha ={\alpha}^{(t)}\right)\\ {}& \kern1em \sim IW\left({V}_0+J,{\left({S}_0+\sum \limits_{j=1}^J\left({C}_j- W\alpha \right){\left({C}_j- W\alpha \right)}^T\right)}^{-1}\right).\end{array}} $$</annotation></semantics></math> </ephtml></item> <p></p> <item> Step 8: Repeat the following sub‐step as many times as the number of missing values in <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0133" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_j $$</annotation></semantics></math> </ephtml> . For each <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0134" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{kj} $$</annotation></semantics></math> </ephtml> missing, define <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0135" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>C(−k)j(t)=(C1(t),...,Ck−1(t),Ck+1(t−1),...,Cp(t−1))</mrow><annotation encoding="application/x-tex">$$ {C}_{\left(-k\right)j}^{(t)}=\left({C}_1^{(t)},\dots, {C}_{k-1}^{(t)},{C}_{k+1}^{\left(t-1\right)},\dots, {C}_p^{\left(t-1\right)}\right) $$</annotation></semantics></math> </ephtml> that consists of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0136" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>k</mi><mo form="prefix">−</mo><mn>1</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(k-1\right) $$</annotation></semantics></math> </ephtml> observed or imputed missing values at cycle <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0137" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>t</mi></mrow><annotation encoding="application/x-tex">$$ t $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0138" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>p</mi><mo form="prefix">−</mo><mi>k</mi><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(p-k\right) $$</annotation></semantics></math> </ephtml> observed or imputed values at cycle <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0139" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>t</mi><mo form="prefix">−</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">$$ t-1 $$</annotation></semantics></math> </ephtml> . We compose a bivariate distribution (<reflink idref="bib4" id="ref39">4</reflink>) given the imputed missing values or PKREs of <ephtml> <math altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0140" display="inline" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msubsup><mi>C</mi><mrow><mo form="prefix">(</mo><mo>−</mo><mi>k</mi><mo form="postfix">)</mo><mi>j</mi></mrow><mrow><mo form="prefix">(</mo><mi>t</mi><mo form="postfix">)</mo></mrow></msubsup></mrow><annotation encoding="application/x-tex">$$ {C}_{\left(-k\right)j}^{(t)} $$</annotation></semantics></math> </ephtml> : <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0141" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>f(Yij,Ckj|Y=Y(t),C(−k)j=C(−k)j(t),uj=uj(t),θ=θ(t))</mrow><annotation encoding="application/x-tex">$$ f\left({Y}_{ij},{C}_{kj}|\mathbf{Y}={\mathbf{Y}}^{(t)},{C}_{\left(-k\right)j}={C}_{\left(-k\right)j}^{(t)},{u}_j={u}_j^{(t)},\theta ={\theta}^{(t)}\right) $$</annotation></semantics></math> </ephtml> to draw <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0142" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mi>k</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{kj} $$</annotation></semantics></math> </ephtml> from the implied posterior distribution (<reflink idref="bib5" id="ref40">5</reflink>): <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0143" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>p(Ckj|Y=Y(t),C(−k)j=C(−k)j(t),uj=uj(t),θ=θ(t)).</mrow><annotation encoding="application/x-tex">$$ p\left({C}_{kj}|\mathbf{Y}={\mathbf{Y}}^{(t)},{C}_{\left(-k\right)j}={C}_{\left(-k\right)j}^{(t)},{u}_j={u}_j^{(t)},\theta ={\theta}^{(t)}\right). $$</annotation></semantics></math> </ephtml></item> </ulist> <hd id="AN0185452270-7">Simulation Study</hd> <p>We now assess our compatible Gibbs sampler based on exact posterior distributions (GSExact) using simulated data under four different scenarios: i) correctly specified distributions of the response, covariates and missing data mechanism; ii) misspecified distribution of a covariate; iii) violation of the MAR assumption; and iv) controlling for multiple partially observed interaction terms. The latter three cases evaluate GSExact in terms of robust estimation against a misspecified covariate distribution, a violated missing data mechanism, and additional partially observed nonlinear terms, respectively. The simulated data from HLM (<reflink idref="bib1" id="ref41">1</reflink>) will closely resemble real data that we analyze in the next section in terms of correlations, sample sizes and missing rates. In each case, we compare our estimators with those from: 1) the lme4 package in R [[<reflink idref="bib28" id="ref42">28</reflink>]] that estimates maximum‐likelihood estimates given complete data (CDML) and 2) software Blimp [[<reflink idref="bib29" id="ref43">29</reflink>]] that implements the Gibbs Sampler of Enders et al. [[<reflink idref="bib21" id="ref44">21</reflink>]]. We do not consider FCS, as it has been shown to be incompatible and produce biased estimates due to nonlinear effects of covariates MAR [[<reflink idref="bib14" id="ref45">14</reflink>], [<reflink idref="bib16" id="ref46">16</reflink>], [<reflink idref="bib21" id="ref47">21</reflink>]]. However, we compare GSEsxact with imputations by predictive mean matching in terms of robust estimation in the Appendix. Because CDML estimators are estimated from complete data, whereas others are based on data MAR, a good method will produce estimates near CDML counterparts.</p> <hd id="AN0185452270-8">Correctly Specified Distributions of the Response, Covariates, and MAR Mechanism</hd> <p>We consider two cases where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0144" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>n</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ {n}_j=4 $$</annotation></semantics></math> </ephtml> units are nested within each of 1) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0145" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>J</mi><mo>=</mo><mn>36</mn></mrow><annotation encoding="application/x-tex">$$ J=36 $$</annotation></semantics></math> </ephtml> clusters or a small sample and 2) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0146" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>J</mi><mo>=</mo><mn>200</mn></mrow><annotation encoding="application/x-tex">$$ J=200 $$</annotation></semantics></math> </ephtml> clusters or a large sample. We validate the correct execution of the R code that implements GSExact given the large sample size and compare our estimators with the competing ones given the small and large sample sizes.</p> <p>We simulate sequentially: i) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0147" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>2</mn><mo>,</mo><mn>1</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {X}_j\sim N\left(2,1\right) $$</annotation></semantics></math> </ephtml> ; ii) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0148" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>0</mn><mo>.</mo><mn>75</mn><mo>+</mo><mn>0</mn><mo>.</mo><mn>7</mn><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><mn>1</mn><mo>.</mo><mn>25</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {C}_{1j}\sim N\left(0.75+0.7{X}_j,1.25\right) $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0149" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mo form="prefix">−</mo><mn>0</mn><mo>.</mo><mn>5</mn><mo>+</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><mn>1</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {C}_{2j}\sim N\left(-0.5+{X}_j,1\right) $$</annotation></semantics></math> </ephtml> with <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0150" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>T</mi></mrow><mrow><mn>12</mn></mrow></msub><mo>=</mo><mi>c</mi><mi>o</mi><mi>v</mi><mo>(</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>)</mo><mo>=</mo><mo form="prefix">−</mo><mn>0</mn><mo>.</mo><mn>5</mn></mrow><annotation encoding="application/x-tex">$$ {T}_{12}=\mathit{\operatorname{cov}}\left({C}_{1j},{C}_{2j}\right)=-0.5 $$</annotation></semantics></math> </ephtml> in model (<reflink idref="bib3" id="ref48">3</reflink>); and iii) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0151" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>1</mn><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>,</mo><mi>τ</mi><mo>+</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij}\sim N\left(1+{C}_{1j}+{C}_{2j}+{X}_j+{C}_{1j}{C}_{2j},\tau +{\sigma}^2\right) $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0152" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ \tau =4 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0153" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>=</mo><mn>16</mn></mrow><annotation encoding="application/x-tex">$$ {\sigma}^2=16 $$</annotation></semantics></math> </ephtml> in HLM (<reflink idref="bib1" id="ref49">1</reflink>). The simulated coefficients of the HLM are all equal to 1 to facilitate the comparison.</p> <p>Next, we simulate the missing values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0154" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0155" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0156" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> by a MAR mechanism that depends on the fully known <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0157" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {X}_j $$</annotation></semantics></math> </ephtml> from <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0158" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>B</mi><mi>e</mi><mi>r</mi><mi>n</mi><mi>o</mi><mi>u</mi><mi>l</mi><mi>l</mi><mi>i</mi><mo>(</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ Bernoulli\left({p}_j\right) $$</annotation></semantics></math> </ephtml> where 7 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0159" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>logit(pj)∼N(c0+c1Xj,δ).</mrow><annotation encoding="application/x-tex">$$ logit\left({p}_j\right)\sim N\left({c}_0+{c}_1{X}_j,\delta \right). $$</annotation></semantics></math> </ephtml> We manipulate <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0160" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>c</mi></mrow><mrow><mn>0</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {c}_0 $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0161" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>c</mi></mrow><mrow><mn>1</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {c}_1 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0162" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>δ</mi></mrow><annotation encoding="application/x-tex">$$ \delta $$</annotation></semantics></math> </ephtml> to simulate 20% missing values for each of the following variables: <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0163" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> given <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0164" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>c</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mo form="prefix">−</mo><mn>1</mn><mo>.</mo><mn>9</mn><mo>,</mo><msub><mrow><mi>c</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>=</mo><mn>0</mn><mo>.</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">$$ {c}_0=-1.9,{c}_1=0.1 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0165" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>δ</mi><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">$$ \delta =1 $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0166" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> given <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0167" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>c</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mn>0</mn><mo>.</mo><mn>8</mn><mo>,</mo><msub><mrow><mi>c</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>=</mo><mo form="prefix">−</mo><mn>1</mn><mo>.</mo><mn>5</mn></mrow><annotation encoding="application/x-tex">$$ {c}_0=0.8,{c}_1=-1.5 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0168" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>δ</mi><mo>=</mo><mn>0</mn></mrow><annotation encoding="application/x-tex">$$ \delta =0 $$</annotation></semantics></math> </ephtml> , and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0169" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> given <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0170" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>c</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mo form="prefix">−</mo><mn>2</mn><mo>.</mo><mn>8</mn><mo>,</mo><msub><mrow><mi>c</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>=</mo><mn>0</mn><mo>.</mo><mn>5</mn></mrow><annotation encoding="application/x-tex">$$ {c}_0=-2.8,{c}_1=0.5 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0171" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>δ</mi><mo>=</mo><mn>0</mn></mrow><annotation encoding="application/x-tex">$$ \delta =0 $$</annotation></semantics></math> </ephtml> .</p> <p>We repeated simulating data and estimating the simulated HLM 1,000 times to compute the % bias equal to (estimated‐simulated) * 100 / simulated, average estimated standard error (ASE), empirical estimate of the true standard error (ESE) over the samples and the coverage probability (coverage) of each estimator. Both Blimp and GSExact are based on 2,500 burn‐in and 2,500 post burn‐in iterations. Given each simulated data set, we run two separate Gibbs sampler chains given different initial values to evaluate the convergence by Geweke's method [[<reflink idref="bib30" id="ref50">30</reflink>]] and the potential scale reduction factor (PSRF) [[<reflink idref="bib31" id="ref51">31</reflink>]].</p> <p>Table 1 summarizes the simulation results for the large sample size. Both GSExact and Blimp produce estimates close to those by CDML. All three approaches produce very accurate and precise estimates with biases <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0188" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo><</mo></mrow><annotation encoding="application/x-tex">$$ < $$</annotation></semantics></math> </ephtml> 2% except those of the <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0189" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> estimates by GSExact (‐2.5%) and Blimp (4.9%), ASEs close to ESEs, and coverage probabilities near the nominal 0.95. The standard errors by GSExact and Blimp are comparatively inflated to reflect extra uncertainty from missing data.</p> <p>1 TABLE Estimated % biases, ASEs, ESEs and coverages from 1,000 simulated large‐sample data sets (nj=4,J=200)$$ \left({n}_j=4,J=200\right) $$.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center" valign="top">Simulated</th><th align="center">CDML</th><th align="center">GSExact</th><th align="center">Blimp</th></tr><tr><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th>Coverage</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0173" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><mo>=</mo>4<annotation encoding="application/x-tex">$$ \tau =4 $$</annotation></semantics></math></p></td><td>0.6 (0.84)</td><td>0.85</td><td>0.95</td><td>−2.5 (1.07)</td><td>1.05</td><td>0.95</td><td>4.9 (1.13)</td><td>1.07</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0174" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<mo>=</mo>16<annotation encoding="application/x-tex">$$ {\sigma}^2=16 $$</annotation></semantics></math></p></td><td>0.1 (0.92)</td><td>0.88</td><td>0.95</td><td>0.8 (1.08)</td><td>1.03</td><td>0.96</td><td>0.5 (1.08)</td><td>1.02</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0175" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_0=1 $$</annotation></semantics></math></p></td><td>−1.7 (0.58)</td><td>0.58</td><td>0.95</td><td>0.2 (0.66)</td><td>0.65</td><td>0.95</td><td>0.1 (0.67)</td><td>0.65</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0176" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_1=1 $$</annotation></semantics></math></p></td><td>−0.6 (0.26)</td><td>0.26</td><td>0.95</td><td>0.0 (0.32)</td><td>0.32</td><td>0.94</td><td>−0.3 (0.32)</td><td>0.32</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0177" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_2=1 $$</annotation></semantics></math></p></td><td>1.0 (0.33)</td><td>0.34</td><td>0.94</td><td>−0.3 (0.39)</td><td>0.41</td><td>0.94</td><td>−0.7 (0.40)</td><td>0.41</td><td>0.94</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0178" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_3=1 $$</annotation></semantics></math></p></td><td>0.5 (0.37)</td><td>0.38</td><td>0.94</td><td>0.5 (0.45)</td><td>0.46</td><td>0.94</td><td>1.6 (0.46)</td><td>0.46</td><td>0.94</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0179" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_4=1 $$</annotation></semantics></math></p></td><td>−0.5 (0.11)</td><td>0.11</td><td>0.95</td><td>−0.1 (0.13)</td><td>0.13</td><td>0.94</td><td>−0.3 (0.14)</td><td>0.14</td><td>0.95</td></tr></tbody></table> </ephtml> </p> <p>Table 2 summarizes the results from the small sample simulation. Because the small sample size led to greater sampling variability compared with the large sample size, we ran 5,000 simulations to reduce this variability. The CDML estimates are consistently close to simulated values with bias <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0190" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo><</mo></mrow><annotation encoding="application/x-tex">$$ < $$</annotation></semantics></math> </ephtml> ‐3.5% in magnitude and small ASEs close to ESEs, achieving good coverages near the nominal 0.95. GSExact estimates are reasonably close to the CDML estimates with small biases <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0191" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo><</mo></mrow><annotation encoding="application/x-tex">$$ < $$</annotation></semantics></math> </ephtml> 2% in magnitude except two fixed effects. The estimator for the intercept <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0192" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math> </ephtml> has a comparatively large bias of 6.3%, which is 2.8% larger in magnitude than the ‐3.5% bias of the CDML counterpart. Additionally, the effect of the known covariate <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0193" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {X}_j $$</annotation></semantics></math> </ephtml> , on which the missing pattern depends, showed a ‐3.8% bias in the estimation. ASEs are small near ESEs with good coverages.</p> <p>2 TABLE Estimated % biases, ASEs, ESEs and coverages from 5,000 simulated small‐sample data sets (nj=4,J=36)$$ \left({n}_j=4,J=36\right) $$.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center" valign="top">Simulated</th><th align="center">CDML</th><th align="center">GSExact</th><th align="center">Blimp</th></tr><tr><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th>Coverage</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0181" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><mo>=</mo>4<annotation encoding="application/x-tex">$$ \tau =4 $$</annotation></semantics></math></p></td><td>−0.2 (2.10)</td><td>2.07</td><td>0.96</td><td>−1.9 (2.54)</td><td>2.00</td><td>0.95</td><td>25.9 (3.57)</td><td>2.60</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0182" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<mo>=</mo>16<annotation encoding="application/x-tex">$$ {\sigma}^2=16 $$</annotation></semantics></math></p></td><td>0.2 (2.18)</td><td>2.20</td><td>0.95</td><td>1.9 (2.57)</td><td>2.49</td><td>0.94</td><td>2.5 (2.66)</td><td>2.51</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0183" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_0=1 $$</annotation></semantics></math></p></td><td>−3.5 (1.51)</td><td>1.55</td><td>0.95</td><td>6.3 (1.89)</td><td>1.84</td><td>0.97</td><td>4.1 (2.16)</td><td>1.90</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0184" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_1=1 $$</annotation></semantics></math></p></td><td>−0.8 (0.69)</td><td>0.69</td><td>0.94</td><td>−0.9 (0.90)</td><td>0.88</td><td>0.97</td><td>−4.7 (1.02)</td><td>0.90</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0185" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_2=1 $$</annotation></semantics></math></p></td><td>−0.3 (0.86)</td><td>0.87</td><td>0.95</td><td>−0.9 (1.13)</td><td>1.13</td><td>0.96</td><td>−5.9 (1.26)</td><td>1.11</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0186" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_3=1 $$</annotation></semantics></math></p></td><td>−1.1 (0.94)</td><td>0.94</td><td>0.94</td><td>−3.8 (1.19)</td><td>1.20</td><td>0.97</td><td>8.9 (1.31)</td><td>1.19</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0187" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<mo>=</mo>1<annotation encoding="application/x-tex">$$ {\beta}_4=1 $$</annotation></semantics></math></p></td><td>0.3 (0.30)</td><td>0.31</td><td>0.94</td><td>1.3 (0.40)</td><td>0.39</td><td>0.96</td><td>0.3 (0.44)</td><td>0.39</td><td>0.97</td></tr></tbody></table> </ephtml> </p> <p>Blimp produces comparatively large biases overall. In particular, the biases in the main effects ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0194" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>1</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0195" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>2</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0196" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>3</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math> </ephtml> ) are noticeably larger than those of GSExact, but the estimated interaction effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0197" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>4</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math> </ephtml> is very accurate. The ASEs are near ESEs, and overall modestly inflated compared with those of GSExact. As a result, the coverages are good near the nominal level. The level‐2 variance estimate of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0198" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> is biased upward by 25.9%, the largest bias observed, which is larger in magnitude than the GSExact counterpart bias of ‐1.9%. Additionally, the ASE of the <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0199" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> estimate is 3.57, which is 41% higher than the GSexact counterpart of 2.54. Therefore, small sample estimates differ noticeably between GSExact based on exact posterior distributions and Blimp based on a metropolis algorithm in sampling missing values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0200" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0201" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> from normal proposal densities with constant variances. Overall, GSExact and Blimp ASEs are again larger than the CDML counterparts, which reflects extra uncertainty due to missing data.</p> <p> <bold>Convergence.</bold> We assessed convergence to a stationary distribution by both Geweke's method [[<reflink idref="bib30" id="ref52">30</reflink>]] and PSRF [[<reflink idref="bib31" id="ref53">31</reflink>]]. Geweke's statistic tests if the two means of an estimator, typically from the first 20% and the last 50% of post burn‐in iterations, are equal within a single Markov chain by the two‐sample normal Z test. On the contrary, the PSRF compares the within‐chain and between‐chain variances from multiple Markov chains—two in our case—to evaluate convergence. A PSRF value <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0202" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo><</mo></mrow><annotation encoding="application/x-tex">$$ < $$</annotation></semantics></math> </ephtml> 1.1 indicates convergence [[<reflink idref="bib32" id="ref54">32</reflink>]]. We assess the convergence of GSExact as all 7 estimators of the HLM satisfying each convergence criterion and compare the results from both criteria.</p> <p>Our simulation revealed that for the large sample, only 64.5% out of the 1,000 simulations satisfied the Geweke's convergence criterion, whereas 99% met the PSRF criterion. Similarly, for the small sample, 66.7% satisfied the Geweke's criterion, whereas 99% met the PSRF criterion. The converged and nonconvergent estimates based on Geweke's criterion revealed virtually no differences in terms of % biases, ASEs, ESEs, and coverages.</p> <p>The Geweke's statistics seem to inflate the Type I error probability in multiple comparisons of the seven convergences. The rejection rates are quite close to the overall Type I error probability under the unrealistic assumption of independent estimators <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0205" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>1</mn><mo form="prefix">−</mo><mn>0</mn><mo>.</mo><mn>9</mn><msup><mrow><mn>5</mn></mrow><mrow><mn>7</mn></mrow></msup><mo>=</mo><mn>0</mn><mo>.</mo><mn>301</mn></mrow><annotation encoding="application/x-tex">$$ 1-0.9{5}^7=0.301 $$</annotation></semantics></math> </ephtml> , implying that it is easier to reject convergence by the Geweke's statistics under the null hypothesis of convergence. Furthermore, the Geweke's statistics rely on a single MCMC chain and, thus, is based on the within‐chain variance only. Consequently, given small sample sizes where ASEs are expected to be comparatively high, we particularly expect the nonconvergence rates by Geweke's criterion to be also high. Therefore, we used the PSRF criterion to assess the convergence in real analysis below.</p> <p>Additionally, we assessed convergence using trace plots from two chains with different initial values. Chain 1 initializes the missing data by predictive mean matching, using the MICE R package, whereas chain 2 imputes them by sample means. Next, both chains fit the HLM (<reflink idref="bib1" id="ref55">1</reflink>) to initialize the parameters. Each trace plot tracks the sampled values of a parameter from its posterior on the vertical axis, with iteration 1 through 5,000 on the horizontal axis. Figure 1 shows the trace plots for the level‐2 variance <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0206" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> and the interaction effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0207" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>4</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math> </ephtml> from one of the simulated small samples. See the Appendix for additional trace plots. Each plot exhibits evidence of convergence to a stable distribution with rapid random fluctuations around a stable horizontal band, no trends, and extensive overlap between chains, indicating exploration of the same posterior distribution.</p> <p> <img src="https://imageserver.ebscohost.com/img/embimages/rdk/2DZ/01may25/sim70051-fig-0001.jpg?ephost1=dGJyMNXb4kSepq84yOvqOLCmsE6epq5Srqa4SK6WxWXS" alt="sim70051-fig-0001.jpg" title="1 (a) Trace plot for the level‐2 variance, τ$$ \tau $$ and (b) trace plot for the interaction effect, β4$$ {\beta}_4 $$." /> </p> <p></p> <p>Although the trace plot for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0208" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> in Figure 1a oscillates randomly around its mean, it also exhibits some large spikes to reveal considerable uncertainly in the estimation of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0209" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> from the small number of 36 clusters and high missing rates. Additionally, these spikes reflect the fact that the posterior inverse gamma distribution of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0210" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> , which is right‐skewed and heavy tailed, is more likely to produce large sampled values compared with the symmetric normal posterior distributions of fixed effects.</p> <hd id="AN0185452270-10">A Misspecified Covariate Distribution</hd> <p>To investigate the robustness of our estimators against violations of the normality assumption for covariates, we simulate: 1) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0211" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>2</mn><mo>,</mo><mn>1</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {X}_j\sim N\left(2,1\right) $$</annotation></semantics></math> </ephtml> ; 2) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0212" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>∼</mo><mi>l</mi><mi>o</mi><mi>g</mi><mi>N</mi><mi>o</mi><mi>r</mi><mi>m</mi><mi>a</mi><mi>l</mi><mo>(</mo><mn>0</mn><mo>.</mo><mn>5</mn><mo>+</mo><mn>0</mn><mo>.</mo><mn>1</mn><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><mn>0</mn><mo>.</mo><mn>2</mn><mo>)</mo><mo>,</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>1</mn><mo>+</mo><mn>0</mn><mo>.</mo><mn>1</mn><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>+</mo><mn>0</mn><mo>.</mo><mn>3</mn><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><mn>1</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {C}_{1j}\sim logNormal\left(0.5+0.1{X}_j,0.2\right),{C}_{2j}\sim N\left(1+0.1{C}_{1j}+0.3{X}_j,1\right) $$</annotation></semantics></math> </ephtml> and 3) <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0213" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>1</mn><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>,</mo><mi>τ</mi><mo>+</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij}\sim N\left(1+{C}_{1j}+{C}_{2j}+{X}_j+{C}_{1j}{C}_{2j},\tau +{\sigma}^2\right) $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0214" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ \tau =4 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0215" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>=</mo><mn>16</mn></mrow><annotation encoding="application/x-tex">$$ {\sigma}^2=16 $$</annotation></semantics></math> </ephtml> . The parameters of the log‐normal distribution are set to generate <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0216" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_1 $$</annotation></semantics></math> </ephtml> with a mean of around 2 and variance near 1, and a skewness of 1.6. Therefore, the bivariate normality assumption of covariates in both GSExact and Blimp is violated. We again simulate 1,000 large and 5,000 small data sets, using the same missing rates as before.</p> <p>Table 3 summarizes the estimated HLM for the large‐sample case. Because the CDML estimation given complete data does not depend on the distributional assumption of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0217" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> , the CDML biases and ASEs near ESEs are small with good coverages. Both GSExact and Blimp result in reasonably small biases up to 4.4% in magnitude, except for the 8.1% and 10.7% biases in the intercept and the −5.6% and −6.5% biases in the effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0218" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>1</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0219" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> , respectively. They also produce reasonably small and similar ASEs and ESEs, with similar coverages near the nominal level. Overall, both GSExact and Blimp produce estimates that are quite robust against the violated normality assumption in the large‐sample scenario.</p> <p>3 TABLE Estimated % biases, ASEs, ESEs and coverages from 1,000 simulated large data sets (nj=4,J=200)$$ \left({n}_j=4,J=200\right) $$ with the normality assumption of C1j$$ {C}_{1j} $$ violated.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center" valign="top">Simulated</th><th align="center">CDML</th><th align="center">GSExact</th><th align="center">Blimp</th></tr><tr><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th>Coverage</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0222" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math></p> = 4</td><td>0.0 (0.84)</td><td>0.87</td><td>0.94</td><td>−3.2 (1.08)</td><td>1.13</td><td>0.93</td><td>4.4 (1.15)</td><td>1.17</td><td>0.93</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0223" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<annotation encoding="application/x-tex">$$ {\sigma}^2 $$</annotation></semantics></math></p> = 16</td><td>−0.2 (0.92)</td><td>0.92</td><td>0.95</td><td>0.5 (1.08)</td><td>1.08</td><td>0.95</td><td>0.2 (1.08)</td><td>1.07</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0224" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math></p> = 1</td><td>1.7 (0.99)</td><td>1.03</td><td>0.94</td><td>8.1 (1.17)</td><td>1.22</td><td>0.93</td><td>10.7 (1.19)</td><td>1.22</td><td>0.94</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0225" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math></p> = 1</td><td>−0.8 (0.45)</td><td>0.48</td><td>0.93</td><td>−5.6 (0.54)</td><td>0.58</td><td>0.92</td><td>−6.5 (0.55)</td><td>0.58</td><td>0.93</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0226" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math></p> = 1</td><td>0.0 (0.42)</td><td>0.44</td><td>0.93</td><td>−3.2 (0.51)</td><td>0.53</td><td>0.94</td><td>−3.9 (0.52)</td><td>0.53</td><td>0.94</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0227" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math></p> = 1</td><td>−0.3 (0.21)</td><td>0.21</td><td>0.96</td><td>−0.2 (0.25)</td><td>0.25</td><td>0.96</td><td>0.3 (0.26)</td><td>0.25</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0228" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math></p> = 1</td><td>0.3 (0.18)</td><td>0.20</td><td>0.94</td><td>2.5 (0.22)</td><td>0.24</td><td>0.92</td><td>2.5 (0.23)</td><td>0.24</td><td>0.93</td></tr></tbody></table> </ephtml> </p> <p>Table 4 summarizes the simulation results for the small sample size. Almost all biases, ASEs and ESEs are larger than those of the large‐sample results in Table 3. The CDML biases are still reasonably small up to −5.3% in magnitude with small ASEs near ESEs and good coverages. GSExact estimates result in biases of 10.6%, −9.0%, and ‐8.1% for the intercept <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0238" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math> </ephtml> , the effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0239" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>1</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0240" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> , and the effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0241" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>2</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0242" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> , respectively. Other estimates are reasonably accurate with biases up to 5%. ASEs are modestly distant from, but reasonably close to ESEs, with coverages near the nominal level. Blimp estimates produce higher biases than their GSExact counterparts overall. The biases for the intercept, the effect of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0243" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> , and the effects of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0244" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> are 23.9%, −11.9%, and −10.9%, respectively, which are noticeably higher than the GSExact counterparts. The <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0245" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> estimator is biased by 37% and is the least inaccurate. Other biases are below 4.3%. ASEs are generally higher than ESEs and their GSExact counterparts. The coverages are near the nominal level. Therefore, GSExact estimators appear more robust than their Blimp counterparts overall under our scenario involving the violated normality assumption of a covariate and a small sample size.</p> <p>4 TABLE Estimated % biases, ASEs, ESEs and coverages from 5,000 simulated data sets of the large sample size (nj=4,J=36)$$ \left({n}_j=4,J=36\right) $$ with the normality assumption of C1j$$ {C}_{1j} $$ violated.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center" valign="top">Simulated</th><th align="center">CDML</th><th align="center">GSExact</th><th align="center">Blimp</th></tr><tr><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th>Coverage</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0231" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math></p> = 4</td><td>0.0 (2.10)</td><td>2.06</td><td>0.96</td><td>1.2 (2.65)</td><td>2.10</td><td>0.97</td><td>35.7 (3.88)</td><td>2.89</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0232" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<annotation encoding="application/x-tex">$$ {\sigma}^2 $$</annotation></semantics></math></p> = 16</td><td>−0.2 (2.17)</td><td>2.15</td><td>0.95</td><td>1.5 (2.56)</td><td>2.47</td><td>0.95</td><td>2.0 (2.66)</td><td>2.48</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0233" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math></p> = 1</td><td>−5.3 (2.67)</td><td>2.78</td><td>0.94</td><td>10.6 (3.62)</td><td>3.77</td><td>0.94</td><td>23.9 (3.98)</td><td>3.70</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0234" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math></p> = 1</td><td>3.8 (1.27)</td><td>1.33</td><td>0.94</td><td>−9.0 (1.75)</td><td>1.84</td><td>0.93</td><td>−11.9 (1.93)</td><td>1.81</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0235" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math></p> = 1</td><td>2.3 (1.17)</td><td>1.23</td><td>0.94</td><td>−8.1 (1.62)</td><td>1.69</td><td>0.93</td><td>−10.9 (1.78)</td><td>1.66</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0236" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math></p> = 1</td><td>−0.4 (0.54)</td><td>0.55</td><td>0.95</td><td>1.0 (0.70)</td><td>0.71</td><td>0.94</td><td>3.6 (0.77)</td><td>0.71</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0237" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math></p> = 1</td><td>−1.4 (0.54)</td><td>0.57</td><td>0.94</td><td>5.0 (0.76)</td><td>0.81</td><td>0.93</td><td>4.3 (0.83)</td><td>0.80</td><td>0.95</td></tr></tbody></table> </ephtml> </p> <hd id="AN0185452270-11">Violation of the MAR Assumption</hd> <p>To assess the robustness of our estimators against violations of the MAR assumption, we designed a simulation that incorporates both MAR and MNAR (missing not at random) mechanisms [[<reflink idref="bib5" id="ref56">5</reflink>]]. After simulating sample data as in Section 4.1, we simulate the missing values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0246" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0247" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> using a <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0248" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>B</mi><mi>e</mi><mi>r</mi><mi>n</mi><mi>o</mi><mi>u</mi><mi>l</mi><mi>l</mi><mi>i</mi><mo>(</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ Bernoulli\left({p}_j\right) $$</annotation></semantics></math> </ephtml> under a MNAR mechanism that depends on the values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0249" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> (before simulating missing values) where <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0250" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>logit(pj)=d0+d1C1j.</mrow><annotation encoding="application/x-tex">$$ logit\left({p}_j\right)={d}_0+{d}_1{C}_{1j}. $$</annotation></semantics></math> </ephtml> To simulate missing values, we set <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0251" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>d</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mo form="prefix">−</mo><mn>5</mn></mrow><annotation encoding="application/x-tex">$$ {d}_0=-5 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0252" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>d</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>=</mo><mn>1</mn><mo>.</mo><mn>3</mn></mrow><annotation encoding="application/x-tex">$$ {d}_1=1.3 $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0253" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> , and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0254" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>d</mi></mrow><mrow><mn>0</mn></mrow></msub><mo>=</mo><mo form="prefix">−</mo><mn>10</mn><mo>.</mo><mn>5</mn></mrow><annotation encoding="application/x-tex">$$ {d}_0=-10.5 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0255" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>d</mi></mrow><mrow><mn>1</mn></mrow></msub><mo>=</mo><mn>3</mn></mrow><annotation encoding="application/x-tex">$$ {d}_1=3 $$</annotation></semantics></math> </ephtml> for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0256" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> . These parameters are chosen to produce approximately 20% missing values for each covariate. This approach allows us to simulate data under a MNAR mechanism, as the missing pattern of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0257" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0258" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> depends on the missing values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0259" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> . We maintain the MAR mechanism (<reflink idref="bib7" id="ref57">7</reflink>) to simulate the missing values of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0260" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> , as described in Section 4.1. We evaluate our estimators under this combination of the MAR and MNAR mechanisms by simulating small sample sizes.</p> <p>Table 5 summarizes the resulting estimates. The CDML estimates remain unaffected by the violation, resulting in accurate and precise estimates. In contrast, the GSExact biases for the fixed effects range from 2.4% to 5.1% in magnitude, which are comparatively larger than the CDML counterparts. Additionally, bias for the random effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0261" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> is 4.7%, approximately four times larger than that of the CDML counterpart. The ASEs for GSExact are a slightly higher or lower than those of CDML, and are reasonably close to the ESEs. Coverages are near the nominal 0.95.</p> <p>5 TABLE Estimated % biases, ASEs, ESEs and coverages from 5,000 small‐sample simulations (nj=4,J=36)$$ \left({n}_j=4,J=36\right) $$ with the MNAR mechanism for C1j$$ {C}_{1j} $$ and C2j$$ {C}_{2j} $$ and the MAR mechanism for Yij$$ {Y}_{ij} $$.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center" valign="top">Simulated</th><th align="center">CDML</th><th align="center">GSExact</th><th align="center">Blimp</th></tr><tr><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th>Coverage</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0266" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math></p> = 4</td><td>1.0 (2.98)</td><td>2.12</td><td>0.95</td><td>4.7 (2.73)</td><td>2.21</td><td>0.97</td><td>39.4 (4.01)</td><td>3.0</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0267" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<annotation encoding="application/x-tex">$$ {\sigma}^2 $$</annotation></semantics></math></p> = 16</td><td>−0.2 (2.96)</td><td>2.17</td><td>0.96</td><td>1.6 (2.63)</td><td>2.49</td><td>0.95</td><td>2.1 (2.71)</td><td>2.50</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0268" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math></p> = 1</td><td>0.1 (1.51)</td><td>1.58</td><td>0.93</td><td>5.1 (1.91)</td><td>1.90</td><td>0.95</td><td>−5.4 (2.12)</td><td>1.89</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0269" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math></p> = 1</td><td>−0.9 (0.68)</td><td>0.72</td><td>0.94</td><td>−4.3 (1.00)</td><td>1.01</td><td>0.94</td><td>−1.4 (1.11)</td><td>1.00</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0270" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math></p> = 1</td><td>1.1 (0.87)</td><td>0.91</td><td>0.94</td><td>2.4 (1.13)</td><td>1.16</td><td>0.94</td><td>5.5 (1.25)</td><td>1.14</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0271" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math></p> = 1</td><td>0.7 (0.94)</td><td>0.96</td><td>0.94</td><td>2.6 (1.20)</td><td>1.21</td><td>0.94</td><td>10.1 (1.34)</td><td>1.20</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0272" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math></p> = 1</td><td>−0.2 (0.30)</td><td>0.31</td><td>0.95</td><td>4.0 (0.47)</td><td>0.47</td><td>0.94</td><td>−0.5 (0.52)</td><td>0.46</td><td>0.97'</td></tr></tbody></table> </ephtml> </p> <p>The Blimp biases for the variance <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0286" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> and the effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0287" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>3</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0288" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {X}_j $$</annotation></semantics></math> </ephtml> are 39.4% and 10.1%, larger than their GSExact counterparts (4.7% and 2.6%). Other biases are modest, up to 5.5%, and moderately larger or smaller than the GSExact counterparts. ASEs are reasonably close to ESEs, but noticeably larger than the GSExact ASEs, with coverages near the nominal level.</p> <p>Overall, GSExact demonstrated comparative robustness, with most biases up to 5.1% under the simulated MNAR scenario. The Blimp estimators of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0289" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> , the main effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0290" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>2</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0291" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> and the main effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0292" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>3</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0293" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {X}_j $$</annotation></semantics></math> </ephtml> were less robust than the GSExact estimators, whereas the remaining Blimp estimators exhibited comparable robustness to their GSExact counterparts when the MAR assumption was violated in small sample sizes. These findings, however, are based on a single MNAR mechanism we simulated. Investigating robust estimation under different violations of the MAR assumption is an important avenue for future research, which is beyond the scope of this paper.</p> <hd id="AN0185452270-12">Controlling for More Covariates or Nonlinear Terms</hd> <p>We now evaluate the performance of our estimators while controlling for additional covariates. In particular, we simulate data as in Section 4.1 with the HLM (<reflink idref="bib1" id="ref58">1</reflink>): <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0294" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>1</mn><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>+</mo><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub><mo>,</mo><mi>τ</mi><mo>+</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij}\sim N\left(1+{C}_{1j}+{C}_{2j}+{X}_j+{C}_{1j}{C}_{2j}+{C}_{1j}{X}_j+{C}_{2j}{X}_j,\tau +{\sigma}^2\right) $$</annotation></semantics></math> </ephtml> , where <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0295" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ \tau =4 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0296" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>=</mo><mn>16</mn></mrow><annotation encoding="application/x-tex">$$ {\sigma}^2=16 $$</annotation></semantics></math> </ephtml> . This model includes two additional interaction terms, <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0297" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j}{X}_j $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0298" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub><msub><mrow><mi>X</mi></mrow><mrow><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j}{X}_j $$</annotation></semantics></math> </ephtml> , which are partially observed. We then generate missing values in <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0299" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {Y}_{ij} $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0300" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>1</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{1j} $$</annotation></semantics></math> </ephtml> , and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0301" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>C</mi></mrow><mrow><mn>2</mn><mi>j</mi></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {C}_{2j} $$</annotation></semantics></math> </ephtml> based on the MAR mechanism (<reflink idref="bib7" id="ref59">7</reflink>). Again, we simulate the partially observed data with small sample sizes 5,000 times.</p> <p>Table 6 summarizes the estimates. The CDML estimates are accurate and precise as before, with modestly increased ASEs from those in Table 2 to reflect more parameters estimated. The GSExact estimators are biased by up to ‐3.6% in magnitude, maintaining the accuracy of those in Table 2 despite the addition of two partially observed nonlinear terms. ASEs and ESEs are slightly increased comparatively, reflecting increased uncertainty due to additional partially observed interaction terms in the HLM, and coverages are good near the nominal level.</p> <p>6 TABLE Estimated % biases, ASEs, ESEs, and coverages from 5,000 simulated small‐sample data sets (nj=4,J=36)$$ \left({n}_j=4,J=36\right) $$ with C1j$$ {C}_{1j} $$, C2j$$ {C}_{2j} $$, and Yij$$ {Y}_{ij} $$ MAR.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center" valign="top">Simulated</th><th align="center">CDML</th><th align="center">GSExact</th><th align="center">Blimp</th></tr><tr><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th align="center">Coverage</th><th align="center">%Bias (ASE)</th><th align="center">ESE</th><th>Coverage</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0277" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math></p> = 4</td><td>−0.3 (2.17)</td><td>2.17</td><td>0.96</td><td>−0.2 (2.73)</td><td>2.21</td><td>0.97</td><td>37.2 (4.01)</td><td>3.0</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0278" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<annotation encoding="application/x-tex">$$ {\sigma}^2 $$</annotation></semantics></math></p> = 16</td><td>0.1 (2.18)</td><td>2.17</td><td>0.95</td><td>1.6 (2.63)</td><td>2.49</td><td>0.95</td><td>2.3 (2.71)</td><td>2.50</td><td>0.95</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0279" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math></p> = 1</td><td>−1.5 (2.07)</td><td>2.21</td><td>0.94</td><td>−1.9 (1.91)</td><td>1.90</td><td>0.95</td><td>7.5 (2.12)</td><td>1.89</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0280" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math></p> = 1</td><td>1.6 (1.06)</td><td>1.10</td><td>0.94</td><td>2.2 (1.00)</td><td>1.01</td><td>0.94</td><td>−8.4 (1.11)</td><td>1.00</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0281" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math></p> = 1</td><td>−0.3 (1.23)</td><td>1.30</td><td>0.94</td><td>−3.6 (1.13)</td><td>1.16</td><td>0.94</td><td>−5.8 (1.25)</td><td>1.14</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0282" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math></p> = 1</td><td>−0.3 (1.71)</td><td>1.79</td><td>0.94</td><td>−0.3 (1.20)</td><td>1.21</td><td>0.94</td><td>5.3 (1.34)</td><td>1.20</td><td>0.96</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0283" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math></p> = 1</td><td>0.4 (0.48)</td><td>0.52</td><td>0.94</td><td>2.4 (0.47)</td><td>0.47</td><td>0.94</td><td>1.1 (0.52)</td><td>0.46</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0284" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>5<annotation encoding="application/x-tex">$$ {\beta}_5 $$</annotation></semantics></math></p> = 1</td><td>−0.6 (0.57)</td><td>0.62</td><td>0.94</td><td>−2.3 (0.47)</td><td>0.47</td><td>0.94</td><td>1.3 (0.52)</td><td>0.46</td><td>0.97</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0285" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>6<annotation encoding="application/x-tex">$$ {\beta}_6 $$</annotation></semantics></math></p> = 1</td><td>0.2 (0.42)</td><td>0.45</td><td>0.94</td><td>1.3 (0.47)</td><td>0.47</td><td>0.94</td><td>0.3 (0.52)</td><td>0.46</td><td>0.97</td></tr></tbody></table> </ephtml> </p> <p>The Blimp bias for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0302" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi></mrow><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math> </ephtml> is 37.2%, higher than the 25.9% bias in Table 2, whereas other biases change less noticeably. The Blimp biases are noticeably higher than those of the GSExact estimators, except for the biases for the interaction effects, which are lower than their GSExact counterparts. Blimp ASEs are also modestly inflated compared with the GSExact ASEs.</p> <p>These differences in accuracy and precision of GSExact and Blimp estimators highlight the importance of GSExact implementing exact posteriors and, thus, ensuring compatibility. In contrast, Blimp implements a Gibbs sampler using a Metropolis algorithm, which has not been shown to ensure compatibility. Confer Shin and Raudenbush [[<reflink idref="bib26" id="ref60">26</reflink>]] for a practical implication of compatibility.</p> <hd id="AN0185452270-13">Analysis of Racially Discordant Patient–Physician Interactions</hd> <p>We analyze data from racially discordant medical encounters between patients and physicians by GSExact. Investigators video‐recorded physicians' behavioral and facial expressions during medical interactions with patients in office visits, and coded physicians' communication behaviors from the recordings. Each physician completed a baseline survey on demographics and other characteristics including implicit prejudice measured by the Implicit Association Test [[<reflink idref="bib34" id="ref61">34</reflink>]] and explicit prejudice measured by a symbolic racism test 2000 [[<reflink idref="bib36" id="ref62">36</reflink>]]. Each encounter lasted about 20 minutes, and the video‐taped physician's facial expression of each encounter was rated by a machine as a valence score at each of four consecutive time points. Due to the COVID‐19 restrictions, the investigators were able to recruit only 37 patients and 6 physicians, much less than planned.</p> <p>Of interest are the main and interaction effects of two physician characteristics, implicit prejudice( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0314" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>I</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ IPre $$</annotation></semantics></math> </ephtml> ) and explicit prejudice( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0315" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>E</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ EPre $$</annotation></semantics></math> </ephtml> ), on their communication behavior or a positive valence score ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0316" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>V</mi><mi>a</mi><mi>l</mi><mi>e</mi><mi>n</mi><mi>c</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ Valence $$</annotation></semantics></math> </ephtml> ) that measures the intensity of positive facial expression during the medical encounter. These variables are partially observed. We also analyze a fully known physician's Communication Training( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0317" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>C</mi><mi>T</mi></mrow><annotation encoding="application/x-tex">$$ CT $$</annotation></semantics></math> </ephtml> ) covariate as the time elapsed since the last training on communication skills. Because a positive valence score is expected to be higher at the beginning and end (e.g., when greeting) than mid time points of the encounter (e.g., when talking about health issues), we also control for dummy variables indicating the second to fourth measurement time points of the outcome Q2, Q3, and Q4 with the first time point as a reference. To improve the interpretability of the intercept and reduce potential collinearity between the covariates and the interaction term, we center the covariates at their sample means and write the model 8 <ephtml> <math display="block" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0318" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow>Valenceij=β0+β1IPrej+β2EPrej+β3CTj+β4IPrejEPrej+uj+β5Q2ij+β6Q3ij+β7Q4ij+eij</mrow><annotation encoding="application/x-tex">$$ {\displaystyle \begin{array}{ll}\hfill Valenc{e}_{ij}& ={\beta}_0+{\beta}_1 IPr{e}_j+{\beta}_2 EPr{e}_j+{\beta}_3C{T}_j+{\beta}_4 IPr{e}_j EPr{e}_j\\ {}\hfill & \kern1em +{u}_j+{\beta}_5{Q}_{2 ij}+{\beta}_6{Q}_{3 ij}+{\beta}_7{Q}_{4 ij}+{e}_{ij}\end{array}} $$</annotation></semantics></math> </ephtml> where occasion <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0319" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>i</mi></mrow><annotation encoding="application/x-tex">$$ i $$</annotation></semantics></math> </ephtml> is nested within the <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0320" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>j</mi><mtext>th</mtext></mrow><annotation encoding="application/x-tex">$$ j\mathrm{th} $$</annotation></semantics></math> </ephtml> patient–physician encounter for <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0321" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>i</mi><mo>=</mo><mn>1</mn><mo>,</mo><mi>...</mi><mo>,</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ i=1,\dots, 4 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0322" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>j</mi><mo>=</mo><mn>1</mn><mo>,</mo><mi>...</mi><mo>,</mo><mn>37</mn></mrow><annotation encoding="application/x-tex">$$ j=1,\dots, 37 $$</annotation></semantics></math> </ephtml> . Here <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0323" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>0</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math> </ephtml> is the expected positive valence score of encounters for physicians having average implicit prejudice ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0324" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>I</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ IPre $$</annotation></semantics></math> </ephtml> ), average explicit prejudice ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0325" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>E</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ EPre $$</annotation></semantics></math> </ephtml> ) and average level of communication training ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0326" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>C</mi><mi>T</mi></mrow><annotation encoding="application/x-tex">$$ CT $$</annotation></semantics></math> </ephtml> ) at occasion 1 ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0327" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>Q</mi></mrow><mrow><mn>2</mn></mrow></msub><mo>=</mo><msub><mrow><mi>Q</mi></mrow><mrow><mn>3</mn></mrow></msub><mo>=</mo><msub><mrow><mi>Q</mi></mrow><mrow><mn>4</mn></mrow></msub><mo>=</mo><mn>0</mn></mrow><annotation encoding="application/x-tex">$$ {Q}_2={Q}_3={Q}_4=0 $$</annotation></semantics></math> </ephtml> ). The fixed effects include the main effects <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0328" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>1</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0329" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>2</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0330" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>3</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0331" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>I</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ IPre $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0332" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>E</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ EPre $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0333" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>C</mi><mi>T</mi></mrow><annotation encoding="application/x-tex">$$ CT $$</annotation></semantics></math> </ephtml> , respectively; the interaction effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0334" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>4</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math> </ephtml> of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0335" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>I</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ IPre $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0336" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>E</mi><mi>P</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ EPre $$</annotation></semantics></math> </ephtml> ; and the mean differences <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0337" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>5</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_5 $$</annotation></semantics></math> </ephtml> , <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0338" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>6</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_6 $$</annotation></semantics></math> </ephtml> , and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0339" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>7</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_7 $$</annotation></semantics></math> </ephtml> from the mean positive valence of occasion one at occasions 2‐4, respectively, ceteris paribus. An encounter‐specific random effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0340" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>u</mi><mi>j</mi></msub><mo>∼</mo><mi>N</mi><mo form="prefix">(</mo><mn>0</mn><mo separator="true">,</mo><mi>τ</mi><mo form="postfix">)</mo></mrow><annotation encoding="application/x-tex">$$ {u}_j\sim N\left(0,\tau \right) $$</annotation></semantics></math> </ephtml> and an occasion‐specific random effect <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0341" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>e</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>∼</mo><mi>N</mi><mo>(</mo><mn>0</mn><mo>,</mo><msup><mrow><mi>σ</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ {e}_{ij}\sim N\left(0,{\sigma}^2\right) $$</annotation></semantics></math> </ephtml> are independent. The valence score is missing 20% of the values, and one physician failed to report both <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0342" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>I</mi><mi>p</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ Ipre $$</annotation></semantics></math> </ephtml> and <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0343" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>E</mi><mi>p</mi><mi>r</mi><mi>e</mi></mrow><annotation encoding="application/x-tex">$$ Epre $$</annotation></semantics></math> </ephtml> such that 16% of their values are missing, respectively.</p> <p>Table 7 shows the estimated HLM by GSexact. Each estimate and its associated standard error in parentheses are listed in column three followed by a 95% Bayesian credible interval in the last column. At a significance level 0.05, explicit prejudice is negatively associated with positive valence score ( <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0344" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow><mi>β</mi></mrow><mrow><mn>2</mn></mrow></msub></mrow><annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math> </ephtml> = −21.83, standard error(se)=11.22), and the mean outcome at occasion two is 9.58 (se=2.65) units less than that at occasion one, controlling for other covariates in the model. Neither large antagonistic interaction effect −78.24 (52.36) of implicit prejudice and explicit prejudice nor the effect −0.35 (5.62) of implicit prejudice is statistically significant, ceteris paribus. The intra‐cluster correlation coefficient 4.57/(4.57 + 80.89)=0.05 implies that only 5% of the remaining total variance in the positive valence score has to be explained at the encounter level. Therefore, high explicit prejudice, measured on the symbolic racism scale, is associated with low positive valence score during patient–physician medical encounters on average. Additionally, the mean positive valence score at occasion 2 drops significantly from the mean initial score at occasion 1, controlling for other covariates in the model. Finally, we used 2,500 burn‐in iterations and 2,500 post burn‐in iterations. The PSRF statistic for each parameter was less than 1.1, satisfying the convergence criterion. The R package GSExact, available at https://github.com/shind10/GSExact, provides all simulation codes from Section 4, as well as the estimation in Table 7.</p> <p>7 TABLE Estimated HLM (<reflink idref="bib8" id="ref63">8</reflink>) for analysis of racially discordant patient–physician encounters.</p> <p> <ephtml> <table><thead valign="bottom"><tr><th align="center">Parameters</th><th align="center">Covariates</th><th align="center">Estimates (se<sup>+</sup>)</th><th>CI<sup>++</sup> (2.5th %ile, 97.5th %ile)</th></tr></thead><tbody valign="top"><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0303" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>0<annotation encoding="application/x-tex">$$ {\beta}_0 $$</annotation></semantics></math></p></td><td>Intercept</td><td>84.46(2.42)*</td><td>(79.71, 107.21)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0304" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>1<annotation encoding="application/x-tex">$$ {\beta}_1 $$</annotation></semantics></math></p></td><td>IPrej</td><td>−0.35(5.62)</td><td>(−11.29, 10.72)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0305" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>2<annotation encoding="application/x-tex">$$ {\beta}_2 $$</annotation></semantics></math></p></td><td>EPrej</td><td>−21.83(11.22)*</td><td>(−44.61, −1.49)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0306" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>3<annotation encoding="application/x-tex">$$ {\beta}_3 $$</annotation></semantics></math></p></td><td>CT</td><td>−0.65(1.81)</td><td>(−3.99, 3.24)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0307" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>4<annotation encoding="application/x-tex">$$ {\beta}_4 $$</annotation></semantics></math></p></td><td>IPrej <p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0308" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mo form="prefix">×</mo><annotation encoding="application/x-tex">$$ \times $$</annotation></semantics></math></p> EPrej</td><td>−78.24(52.36)</td><td>(−188.84, 14.82)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0309" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>5<annotation encoding="application/x-tex">$$ {\beta}_5 $$</annotation></semantics></math></p></td><td>Q2</td><td>−9.58(2.65)*</td><td>(−14.99, −4.62)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0310" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>6<annotation encoding="application/x-tex">$$ {\beta}_6 $$</annotation></semantics></math></p></td><td>Q3</td><td>−4.25(2.59)</td><td>(−9.30, 0.68)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0311" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>β</mi>7<annotation encoding="application/x-tex">$$ {\beta}_7 $$</annotation></semantics></math></p></td><td>Q4</td><td>−2.14(2.47)</td><td>(−7.10, 2.66)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0312" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>τ</mi><annotation encoding="application/x-tex">$$ \tau $$</annotation></semantics></math></p></td><td>—</td><td>4.57(5.07)</td><td>(0.57, 19.59)</td></tr><tr><td><p><math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0313" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics xmlns=""><mi>σ</mi>2<annotation encoding="application/x-tex">$$ {\sigma}^2 $$</annotation></semantics></math></p></td><td>—</td><td>80.89(12.48)</td><td>(59.33, 107.21)</td></tr></tbody></table> </ephtml> </p> <p>1 <emph>Note:</emph> +: standard error; ++: a 95% credible interval; *: Significantly different from zero at a level 0.05.</p> <hd id="AN0185452270-14">Discussion</hd> <p>We estimated a HLM with the nonlinear effects of cluster‐level covariates from partially observed data assumed MAR. Our Bayesian estimation by a Gibbs sampler consists of exact posterior distributions, which are derived via selected provisionally known random effect to ensure compatibility with the HLM [[<reflink idref="bib18" id="ref64">18</reflink>], [<reflink idref="bib26" id="ref65">26</reflink>]]. Our simulation results show that the sampler produces reasonably accurate and precise estimates given sample sizes as small as <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0345" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>n</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ n=4 $$</annotation></semantics></math> </ephtml> units nested within <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0346" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>J</mi><mo>=</mo><mn>36</mn></mrow><annotation encoding="application/x-tex">$$ J=36 $$</annotation></semantics></math> </ephtml> clusters and given partially observed outcome and cluster‐level continuous covariates having interaction effects. Our estimators were as accurate and precise as those of competing methods given large sample sizes <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0347" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>n</mi><mo>=</mo><mn>4</mn><mo>,</mo><mi>J</mi><mo>=</mo><mn>200</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(n=4,J=200\right) $$</annotation></semantics></math> </ephtml> , and more accurate and precise than completing ones given small sample sizes <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0348" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>(</mo><mi>n</mi><mo>=</mo><mn>4</mn><mo>,</mo><mi>J</mi><mo>=</mo><mn>36</mn><mo>)</mo></mrow><annotation encoding="application/x-tex">$$ \left(n=4,J=36\right) $$</annotation></semantics></math> </ephtml> in our simulation study.</p> <p>We checked convergence to the posterior distribution by the PSRF and Geweke's criteria and found that the PSRF criterion was preferable to Geweke's criterion in our simulation scenarios. This preference was more apparent given additional partially observed covariates and smaller sample sizes.</p> <p>In our simulation, we also assessed robustness of our estimators under three cases: i) a misspecified distribution of a partially observed covariate; ii) violation of the MAR assumption; and iii) multiple interaction terms of partially observed covariates. The simulation results reveal that our estimators are reasonably robust against these violations given large and small sample sizes in our simulation scenarios. An importance avenue for future research is to conduct extensive simulations under different misspecified models and missing data mechanisms.</p> <p>Our simulation study is limited in the sense that the cluster size was fixed at <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0349" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>n</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">$$ n=4 $$</annotation></semantics></math> </ephtml> units, as analyzed in our real data analysis, while the number of <ephtml> <math display="inline" altimg="urn:x-wiley:02776715:media:sim70051:sim70051-math-0350" overflow="scroll" xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>J</mi></mrow><annotation encoding="application/x-tex">$$ J $$</annotation></semantics></math> </ephtml> clusters were varied. Valuable future research is to assess the impact of varied cluster sizes on our estimators.</p> <p>In near future, we aim to extend our sampler to efficiently handle partially observed discrete covariates or a mixture of partially observed discrete and continuous covariates having interaction effects at the cluster level. We also plan to extend our method to models where partially observed lower‐level covariates have nonlinear effects. See Shin and Raudenbush [[<reflink idref="bib18" id="ref66">18</reflink>], [<reflink idref="bib26" id="ref67">26</reflink>]] for ML estimation of a HLM for covariates MAR with nonlinear effects.</p> <hd id="AN0185452270-15">Conflicts of Interest</hd> <p>The authors declare no conflicts of interest.</p> <hd id="AN0185452270-16">Data Availability Statement</hd> <p>Research data are not shared.</p> <p>GRAPH: Supporting Information.</p> <ref id="AN0185452270-17"> <title> Footnotes </title> <blist> <bibl id="bib1" idref="ref1" type="bt">1</bibl> <bibtext> Funding: This study was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (R01DK112009), National Cancer Institute (R01CA263501), National Institute of Nursing Research (R01NR020030) and Institute of Education Sciences (R305D210022).</bibtext> </blist> </ref> <ref id="AN0185452270-18"> <title> References </title> <blist> <bibtext> S. W. Raudenbush and A. S. Bryk, "Hierarchical Linear Models: Applications and Data Analysis Methods," 2002.</bibtext> </blist> <blist> <bibl id="bib2" idref="ref34" type="bt">2</bibl> <bibtext> H. Goldstein, Multilevel Statistical Models (John Wiley & Sons, Incorporation, 2011).</bibtext> </blist> <blist> <bibl id="bib3" idref="ref2" type="bt">3</bibl> <bibtext> R. J. A. Little and D. B. Rubin, " Bayes and Multiple Imputation ," in Statistical Analysis With Missing Data (John Wiley & Sons, 2002), 200 – 220.</bibtext> </blist> <blist> <bibl id="bib4" idref="ref3" type="bt">4</bibl> <bibtext> D. B. Rubin, Multiple Imputation for Nonresponse in Surveys (John Wiley & Sons, Incorporation, 1987).</bibtext> </blist> <blist> <bibl id="bib5" idref="ref4" type="bt">5</bibl> <bibtext> D. B. Rubin, " Inference and Missing Data ," Biometrika 63, no. 3 (1976): 581 – 592.</bibtext> </blist> <blist> <bibl id="bib6" idref="ref5" type="bt">6</bibl> <bibtext> J. L. Schafer and R. M. Yucel, " Computational Strategies for Multivariate Linear Mixed‐Effects Models With Missing Values ," Journal of Computational and Graphical Statistics 11, no. 2 (2002): 437 – 457.</bibtext> </blist> <blist> <bibl id="bib7" idref="ref57" type="bt">7</bibl> <bibtext> Y. Shin and S. W. Raudenbush, " Just‐Identified Versus Overidentified Two‐Level Hierarchical Linear Models With Missing Data ," Biometrics 63, no. 4 (2007): 1262 – 1268.</bibtext> </blist> <blist> <bibl id="bib8" idref="ref6" type="bt">8</bibl> <bibtext> T. E. Raghunathan, J. M. Lepkowski, J. Van Hoewyk, and P. Solenberger, " A Multivariate Technique for Multiply Imputing Missing Values Using a Sequence of Regression Models ," Survey Methodology 27, no. 1 (2001): 85 – 96.</bibtext> </blist> <blist> <bibl id="bib9" type="bt">9</bibl> <bibtext> S. Van Buuren, " Multiple Imputation of Discrete and Continuous Data by Fully Conditional Specification ," Statistical Methods in Medical Research 16, no. 3 (2007): 219 – 242.</bibtext> </blist> <blist> <bibtext> S. Van Buuren and K. Groothuis‐Oudshoorn, " Mice: Multivariate Imputation by Chained Equations in R ," Journal of Statistical Software 45, no. 3 (2011): 1 – 67.</bibtext> </blist> <blist> <bibtext> S. Van Buuren, Flexible Imputation of Missing Data (CRC Press, 2018).</bibtext> </blist> <blist> <bibtext> B. C. Arnold and S. J. Press, " Compatible Conditional Distributions ," Journal of the American Statistical Association 84, no. 405 (1989): 152 – 156.</bibtext> </blist> <blist> <bibtext> B. C. Arnold, E. Castillo, and J. M. Sarabia, " Conditionally Specified Distributions: An Introduction (With Comments and a Rejoinder by the Authors) ," Statistical Science 16, no. 3 (2001): 249 – 274.</bibtext> </blist> <blist> <bibtext> J. Liu, A. Gelman, J. Hill, Y. S. Su, and J. Kropko, " On the Stationary Distribution of Iterative Imputations ," Biometrika 101, no. 1 (2014): 155 – 173.</bibtext> </blist> <blist> <bibtext> J. W. Bartlett, S. R. Seaman, I. R. White, J. R. Carpenter, and A. D. N. Initiative*, " Multiple Imputation of Covariates by Fully Conditional Specification: Accommodating the Substantive Model ," Statistical Methods in Medical Research 24, no. 4 (2015): 462 – 487.</bibtext> </blist> <blist> <bibtext> C. K. Enders, S. A. Mistler, and B. T. Keller, " Multilevel Multiple Imputation: A Review and Evaluation of Joint Modeling and Chained Equations Imputation ," Psychological Methods 21, no. 2 (2016): 222.</bibtext> </blist> <blist> <bibtext> S. Kim, C. A. Sugar, and T. R. Belin, " Evaluating Model‐Based Imputation Methods for Missing Covariates in Regression Models With Interactions ," Statistics in Medicine 34, no. 11 (2015): 1876 – 1888.</bibtext> </blist> <blist> <bibtext> Y. Shin and S. W. Raudenbush, " Maximum Likelihood Estimation of Hierarchical Linear Models From Incomplete Data: Random Coefficients, Statistical Interactions, and Measurement Error ," Journal of Computational and Graphical Statistics 33, no. 1 (2024): 112 – 125.</bibtext> </blist> <blist> <bibtext> J. Carpenter and M. Kenward, Multiple Imputation and Its Application (John Wiley & Sons Incorporation, 2013).</bibtext> </blist> <blist> <bibtext> N. S. Erler, D. Rizopoulos, R. Jv, V. W. Jaddoe, O. H. Franco, and E. M. Lesaffre, " Dealing With Missing Covariates in Epidemiologic Studies: A Comparison Between Multiple Imputation and a Full Bayesian Approach ," Statistics in Medicine 35, no. 17 (2016): 2955 – 2974.</bibtext> </blist> <blist> <bibtext> C. K. Enders, H. Du, and B. T. Keller, " A Model‐Based Imputation Procedure for Multilevel Regression Models With Random Coefficients, Interaction Effects, and Nonlinear Terms ," Psychological Methods 25, no. 1 (2020): 88.</bibtext> </blist> <blist> <bibtext> S. Kim, T. R. Belin, and C. A. Sugar, " Multiple Imputation With Non‐additively Related Variables: Joint‐Modeling and Approximations ," Statistical Methods in Medical Research 27, no. 6 (2018): 1683 – 1694.</bibtext> </blist> <blist> <bibtext> H. Goldstein, J. R. Carpenter, and W. J. Browne, " Fitting Multilevel Multivariate Models With Missing Data in Responses and Covariates That May Include Interactions and Non‐linear Terms ," Journal of the Royal Statistical Society: Series A (Statistics in Society) 177, no. 2 (2014): 553 – 564.</bibtext> </blist> <blist> <bibtext> N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, " Equation of State Calculations by Fast Computing Machines ," Journal of Chemical Physics 21, no. 6 (1953): 1087 – 1092.</bibtext> </blist> <blist> <bibtext> W. Hastings, " Monte Carlo Sampling Methods Using Markov Chains and Their Applications ," Biometrika 57, no. 1 (1970): 97 – 109.</bibtext> </blist> <blist> <bibtext> Y. Shin and S. W. Raudenbush, Multiple Imputation to Estimate Hierarchical Models From Data MAR: Latent Covariates (Random Coefficients and Statistical Interactions, 2024).</bibtext> </blist> <blist> <bibtext> P. D. Hoff, A First Course in Bayesian Statistical Methods (Springer, 2009).</bibtext> </blist> <blist> <bibtext> M. Douglas Bates, B. Bolker, and S. Walker, " Fitting Linear Mixed‐Effects Models Using lme4 ," Journal of Statistical Software 67, no. 1 (2015): 1 – 48.</bibtext> </blist> <blist> <bibtext> B. T. Keller and C. K. Enders, "Blimp User's Guide," 2022, <ulink href="http://www.appliedmissingdata.com/blimp">www.appliedmissingdata.com/blimp</ulink>.</bibtext> </blist> <blist> <bibtext> J. Geweke, Evaluating the Accuracy of Sampling‐Based Approaches to the Calculation of Posterior Moments. Technical Report 148 (Federal Reserve Bank of Minneapolis, 1991).</bibtext> </blist> <blist> <bibtext> S. P. Brooks and A. Gelman, " General Methods for Monitoring Convergence of Iterative Simulations ," Journal of Computational and Graphical Statistics 7, no. 4 (1998): 434 – 455.</bibtext> </blist> <blist> <bibtext> H. Du, Z. Ke, G. Jiang, and S. Huang, " The Performances of Gelman‐Rubin and Geweke's Convergence Diagnostics of Monte Carlo Markov Chains in Bayesian Analysis ," Journal of Behavioral Data Science 2, no. 2 (2022): 47 – 72.</bibtext> </blist> <blist> <bibtext> M. K. Cowles and B. P. Carlin, " Markov Chain Monte Carlo Convergence Diagnostics: A Comparative Review ," Journal of the American Statistical Association 91, no. 434 (1996): 883 – 904.</bibtext> </blist> <blist> <bibtext> A. G. Greenwald, D. E. McGhee, and J. L. Schwartz, " Measuring Individual Differences in Implicit Cognition: The Implicit Association Test ," Journal of Personality and Social Psychology 74, no. 6 (1998): 1464.</bibtext> </blist> <blist> <bibtext> B. A. Nosek, A. G. Greenwald, and M. R. Banaji, " Understanding and Using the Implicit Association Test: II. Method Variables and Construct Validity ," Personality and Social Psychology Bulletin 31, no. 2 (2005): 166 – 180.</bibtext> </blist> <blist> <bibtext> P. J. Henry and D. O. Sears, " The Symbolic Racism 2000 Scale ," Political Psychology 23, no. 2 (2002): 253 – 283.</bibtext> </blist> </ref> <aug> <p>By Dongho Shin; Yongyun Shin and Nao Hagiwara</p> <p>Reported by Author; Author; Author</p> </aug> <nolink nlid="nl1" bibid="bib10" firstref="ref7"></nolink> <nolink nlid="nl2" bibid="bib12" firstref="ref8"></nolink> <nolink nlid="nl3" bibid="bib14" firstref="ref9"></nolink> <nolink nlid="nl4" bibid="bib16" firstref="ref11"></nolink> <nolink nlid="nl5" bibid="bib17" firstref="ref12"></nolink> <nolink nlid="nl6" bibid="bib19" firstref="ref15"></nolink> <nolink nlid="nl7" bibid="bib21" firstref="ref16"></nolink> <nolink nlid="nl8" bibid="bib22" firstref="ref18"></nolink> <nolink nlid="nl9" bibid="bib23" firstref="ref20"></nolink> <nolink nlid="nl10" bibid="bib24" firstref="ref21"></nolink> <nolink nlid="nl11" bibid="bib18" firstref="ref24"></nolink> <nolink nlid="nl12" bibid="bib26" firstref="ref25"></nolink> <nolink nlid="nl13" bibid="bib27" firstref="ref37"></nolink> <nolink nlid="nl14" bibid="bib28" firstref="ref42"></nolink> <nolink nlid="nl15" bibid="bib29" firstref="ref43"></nolink> <nolink nlid="nl16" bibid="bib30" firstref="ref50"></nolink> <nolink nlid="nl17" bibid="bib31" firstref="ref51"></nolink> <nolink nlid="nl18" bibid="bib32" firstref="ref54"></nolink> <nolink nlid="nl19" bibid="bib34" firstref="ref61"></nolink> <nolink nlid="nl20" bibid="bib36" firstref="ref62"></nolink>
Header DbId: eric
DbLabel: ERIC
An: ED673559
AccessLevel: 3
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 0
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: Bayesian Estimation of Hierarchical Linear Models from Incomplete Data: Cluster-Level Interaction Effects and Small Sample Sizes
– Name: Language
  Label: Language
  Group: Lang
  Data: English
– Name: Author
  Label: Authors
  Group: Au
  Data: <searchLink fieldCode="AR" term="%22Dongho+Shin%22">Dongho Shin</searchLink> (ORCID <externalLink term="https://orcid.org/0009-0006-5754-6815">0009-0006-5754-6815</externalLink>)<br /><searchLink fieldCode="AR" term="%22Yongyun+Shin%22">Yongyun Shin</searchLink> (ORCID <externalLink term="https://orcid.org/0000-0003-0654-2620">0000-0003-0654-2620</externalLink>)<br /><searchLink fieldCode="AR" term="%22Nao+Hagiwara%22">Nao Hagiwara</searchLink> (ORCID <externalLink term="https://orcid.org/0000-0003-3933-8917">0000-0003-3933-8917</externalLink>)
– Name: TitleSource
  Label: Source
  Group: Src
  Data: <searchLink fieldCode="SO" term="%22Grantee+Submission%22"><i>Grantee Submission</i></searchLink>. 2025 44.
– Name: PeerReviewed
  Label: Peer Reviewed
  Group: SrcInfo
  Data: Y
– Name: Pages
  Label: Page Count
  Group: Src
  Data: 11
– Name: DatePubCY
  Label: Publication Date
  Group: Date
  Data: 2025
– Name: SourceSuprt
  Label: Sponsoring Agency
  Group: SrcSuprt
  Data: National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) (DHHS/NIH)<br />National Cancer Institute (NCI) (DHHS/NIH)<br />National Institute of Nursing Research (NINR) (DHHS/NIH)<br />Institute of Education Sciences (ED)
– Name: NumberContract
  Label: Contract Number
  Group: NumCntrct
  Data: R01DK112009<br />R01CA263501
– Name: TypeDocument
  Label: Document Type
  Group: TypDoc
  Data: Journal Articles<br />Reports - Research
– Name: Subject
  Label: Descriptors
  Group: Su
  Data: <searchLink fieldCode="DE" term="%22Bayesian+Statistics%22">Bayesian Statistics</searchLink><br /><searchLink fieldCode="DE" term="%22Hierarchical+Linear+Modeling%22">Hierarchical Linear Modeling</searchLink><br /><searchLink fieldCode="DE" term="%22Multivariate+Analysis%22">Multivariate Analysis</searchLink><br /><searchLink fieldCode="DE" term="%22Data+Analysis%22">Data Analysis</searchLink><br /><searchLink fieldCode="DE" term="%22Sample+Size%22">Sample Size</searchLink><br /><searchLink fieldCode="DE" term="%22Mathematical+Formulas%22">Mathematical Formulas</searchLink><br /><searchLink fieldCode="DE" term="%22Statistical+Distributions%22">Statistical Distributions</searchLink>
– Name: DOI
  Label: DOI
  Group: ID
  Data: 10.1002/sim.70051
– Name: Abstract
  Label: Abstract
  Group: Ab
  Data: We consider Bayesian estimation of a hierarchical linear model (HLM) from partially observed data, assumed to be missing at random, and small sample sizes. A vector of continuous covariates C includes cluster-level partially observed covariates with interaction effects. Due to small sample sizes from 37 patient-physician encounters repeatedly measured at four time points, maximum-likelihood estimation is suboptimal. Existing Gibbs samplers impute missing values of C by a Metropolis algorithm using proposal densities that have constant variances while the target posterior distributions have nonconstant variances. Therefore, these samplers may not ensure compatibility with the HLM and, as a result, may not guarantee unbiased estimation of the HLM. We introduce a compatible Gibbs sampler that imputes parameters and missing values directly from the exact posterior distributions. We apply our Gibbs sampler to the longitudinal patient-physician encounter data and compare our estimators with those from existing methods by simulation.
– Name: AbstractInfo
  Label: Abstractor
  Group: Ab
  Data: As Provided
– Name: CodeSource
  Label: IES Funded
  Group: SrcInfo
  Data: Yes
– Name: DateEntry
  Label: Entry Date
  Group: Date
  Data: 2025
– Name: AN
  Label: Accession Number
  Group: ID
  Data: ED673559
PLink https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=eric&AN=ED673559
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1002/sim.70051
    Languages:
      – Text: English
    PhysicalDescription:
      Pagination:
        PageCount: 11
    Subjects:
      – SubjectFull: Bayesian Statistics
        Type: general
      – SubjectFull: Hierarchical Linear Modeling
        Type: general
      – SubjectFull: Multivariate Analysis
        Type: general
      – SubjectFull: Data Analysis
        Type: general
      – SubjectFull: Sample Size
        Type: general
      – SubjectFull: Mathematical Formulas
        Type: general
      – SubjectFull: Statistical Distributions
        Type: general
    Titles:
      – TitleFull: Bayesian Estimation of Hierarchical Linear Models from Incomplete Data: Cluster-Level Interaction Effects and Small Sample Sizes
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Dongho Shin
      – PersonEntity:
          Name:
            NameFull: Yongyun Shin
      – PersonEntity:
          Name:
            NameFull: Nao Hagiwara
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 05
              Type: published
              Y: 2025
          Numbering:
            – Type: volume
              Value: 44
          Titles:
            – TitleFull: Grantee Submission
              Type: main
ResultId 1