Please cite or reference this post if you use it for publication purposes: Wiegand, J.P. (October 9, 2015). Bayesian SEM (BSEM) Application and Example [Blog post]. Retrieved from
https://justinwiegand.com/blog/?p=331
Sounds good, right?
Bayesian priors allow cross-loadings and residual covariances of SEM’s to vary a small degree (i.e., replace exact zeros with approximate zeros from informative, small-variance priors) and be evaluated (see Asparouhov, Muthén, & Morin, 2015; Muthén, & Asparouhov 2012). The researcher can thereby discover whether 0 cross-loadings likely exist in the “true” population model, given their data, and refine their model accordingly.
I recently had the opportunity to apply this technique to a bi-factor model for a new scale that had previously only been subjected to a traditional CFA. The model had strong theoretical support, but fit indices and some loadings were not supportive of the general factor as theorized. These conditions provide a perfect opportunity to use Bayesian CFA (BCFA) to refine the model for cross-validation.
MPlus 7 was used to specify and test a model where cross-loadings were assigned normally distributed priors with 0 means and variances of .01. The technique is not difficult for those familiar with CFA in MPlus. Working from a typical CFA, the researcher need only:
- Specify “Bayes” as your estimator (shown with option recommendations):
ANALYSIS: ESTIMATOR = BAYES; !Uses two independent MCMC chains
PROCESSORS = 2; !To speed up computations if you have 2 processors
FBITERATIONS=15000; !Minimum number of Markov Chain iterations
- Include all cross-loadings for a factor below the actual loadings (in our case, the specific factors in a bi-factor model).
- Thus for a data set with 10 items (y1-y10), and two factors (f1 and f2) with loadings y1-y3 and y4-y10 respectively, go from:
f1 BY y1-y3;
f2 BY y4-y10;
to
f1 BY y1-y3
y4-y10; !Cross-loadings
f2 BY y4-y10
y1-y3; !Cross-loadings
- Next, assign labels to the cross-loadings so a Bayesian prior can be specified for each. To do so, expand the syntax with cross-loadings as follows (cross-loadings can be labeled however desired–I am using “xlam” to refer to lamda’s [loadings] of the cross-loadings [“x” for cross]):
f1 BY y1-y3
y4-y10 (f1xlam1-f1xlam7); !Cross-loadings (with assigned labels)
f2 BY y4-y10
y1-y3(f2xlam1-f2xlam3); !Cross-loadings (with assigned labels)
- Finally, model priors are specified by defining all the cross-loadings’ distribution, mean, and variance. A statement is simply added at the end of the “Model” command as follows (for cross-loading priors normally distributed with mean 0 and variances of 0.01):
MODEL PRIORS:
f1xlam1-f2xlam3~N(0,0.01); !You can list across factors if ordered
Naturally, identification and metric setting needs to be addressed first (for example, MPlus sets the first loading of a factor to one by default, but you may want to free that and instead set the variance of each factor to 1 instead). Yet, hopefully this illustrates the ease at which cross-loadings can be assigned Bayesian priors. For those interested in assigning residual covariances prior distributions the concept is very similar. A full example is given in the Asparouhov, Muthén, and Morin (2015) article.
Now for the beauty (i.e., application) of BCFA. Cross-loadings were assigned a normal prior distribution with mean 0 and variance of 0.01. Were the actual loadings 0 given our data? The prior and posterior distributions for each loading can be viewed to verify this (choose “Plot” > “View Plots” in MPlus) or the confidence intervals for the cross-loadings can be examined in the MPlus output. Below is an example of a cross-loading for an item I analyzed:
First, examine the prior distribution (made up of 15,000 iterations):
bcfa2prior
This looks right, the mean is essentially 0 and variance (Std Dev squared) 0.01 as specified.
Now, did the distribution change given our data? Check the posterior distribution to find out (or the confidence intervals for the cross-loadings in the MPlus output):
bcfa2post
It appears the distribution did change. Given our data, the posterior distribution does not include 0 (for a 95% confidence interval). This suggests that the specific cross-loading should be freed.
This simple illustration is just one way BCFA (and BSEM) adds value for scale and model development. I refer you to the linked articles for more. Let me know if you’ve had a chance to apply the technique and what you learned.
References
1559346
IU7X3CWB,N2WSKQJC
items
1
apa
0
author
asc
331
https://justinwiegand.com/blog/wp-content/plugins/zotpress/
%7B%22status%22%3A%22success%22%2C%22updateneeded%22%3Afalse%2C%22instance%22%3A%22zotpress-e587a33fc29c030a0a97084e2c4486c0%22%2C%22meta%22%3A%7B%22request_last%22%3A0%2C%22request_next%22%3A0%2C%22used_cache%22%3Atrue%7D%2C%22data%22%3A%5B%7B%22key%22%3A%22IU7X3CWB%22%2C%22library%22%3A%7B%22id%22%3A1559346%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Asparouhov%20et%20al.%22%2C%22parsedDate%22%3A%222015-06-30%22%2C%22numChildren%22%3A2%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20padding-left%3A%201em%3B%20text-indent%3A-1em%3B%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%3EAsparouhov%2C%20T.%2C%20Muth%26%23xE9%3Bn%2C%20B.%2C%20%26amp%3B%20Morin%2C%20A.%20J.%20S.%20%282015%29.%20Bayesian%20Structural%20Equation%20Modeling%20With%20Cross-Loadings%20and%20Residual%20Covariances%20Comments%20on%20Stromeyer%20et%20al.%20%3Ci%3EJournal%20of%20Management%3C%5C%2Fi%3E%2C%200149206315591075.%20%3Ca%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.1177%5C%2F0149206315591075%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.1177%5C%2F0149206315591075%3C%5C%2Fa%3E%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22journalArticle%22%2C%22title%22%3A%22Bayesian%20Structural%20Equation%20Modeling%20With%20Cross-Loadings%20and%20Residual%20Covariances%20Comments%20on%20Stromeyer%20et%20al.%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tihomir%22%2C%22lastName%22%3A%22Asparouhov%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Bengt%22%2C%22lastName%22%3A%22Muth%5Cu00e9n%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alexandre%20J.%20S.%22%2C%22lastName%22%3A%22Morin%22%7D%5D%2C%22abstractNote%22%3A%22A%20recent%20article%20in%20the%20Journal%20of%20Management%20gives%20a%20critique%20of%20a%20Bayesian%20approach%20to%20factor%20analysis%20proposed%20in%20Psychological%20Methods.%20This%20commentary%20responds%20to%20the%20authors%5Cu2019%20critique%20by%20clarifying%20key%20issues%2C%20especially%20the%20use%20of%20priors%20for%20residual%20covariances.%20A%20discussion%20is%20also%20presented%20of%20cross-loadings%20and%20model%20selection%20tools.%20Simulated%20data%20are%20used%20to%20illustrate%20the%20ideas.%20A%20reanalysis%20of%20the%20example%20used%20by%20the%20authors%20reveals%20a%20superior%20model%20overlooked%20by%20the%20authors.%22%2C%22date%22%3A%222015-06-30%22%2C%22language%22%3A%22en%22%2C%22DOI%22%3A%2210.1177%5C%2F0149206315591075%22%2C%22ISSN%22%3A%220149-2063%2C%201557-1211%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Fjom.sagepub.com%5C%2Fcontent%5C%2Fearly%5C%2F2015%5C%2F06%5C%2F25%5C%2F0149206315591075%22%2C%22collections%22%3A%5B%225A94Z44M%22%5D%2C%22dateModified%22%3A%222015-09-25T20%3A08%3A10Z%22%7D%7D%2C%7B%22key%22%3A%22N2WSKQJC%22%2C%22library%22%3A%7B%22id%22%3A1559346%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Muth%5Cu00e9n%20and%20Asparouhov%22%2C%22parsedDate%22%3A%222012-09%22%2C%22numChildren%22%3A4%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20padding-left%3A%201em%3B%20text-indent%3A-1em%3B%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%3EMuth%26%23xE9%3Bn%2C%20B.%2C%20%26amp%3B%20Asparouhov%2C%20T.%20%282012%29.%20Bayesian%20structural%20equation%20modeling%3A%20a%20more%20flexible%20representation%20of%20substantive%20theory.%20%3Ci%3EPsychological%20Methods%3C%5C%2Fi%3E%2C%20%3Ci%3E17%3C%5C%2Fi%3E%283%29%2C%20313%26%23x2013%3B335.%20%3Ca%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.1037%5C%2Fa0026802%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.1037%5C%2Fa0026802%3C%5C%2Fa%3E%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22journalArticle%22%2C%22title%22%3A%22Bayesian%20structural%20equation%20modeling%3A%20a%20more%20flexible%20representation%20of%20substantive%20theory%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Bengt%22%2C%22lastName%22%3A%22Muth%5Cu00e9n%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tihomir%22%2C%22lastName%22%3A%22Asparouhov%22%7D%5D%2C%22abstractNote%22%3A%22This%20article%20proposes%20a%20new%20approach%20to%20factor%20analysis%20and%20structural%20equation%20modeling%20using%20Bayesian%20analysis.%20The%20new%20approach%20replaces%20parameter%20specifications%20of%20exact%20zeros%20with%20approximate%20zeros%20based%20on%20informative%2C%20small-variance%20priors.%20It%20is%20argued%20that%20this%20produces%20an%20analysis%20that%20better%20reflects%20substantive%20theories.%20The%20proposed%20Bayesian%20approach%20is%20particularly%20beneficial%20in%20applications%20where%20parameters%20are%20added%20to%20a%20conventional%20model%20such%20that%20a%20nonidentified%20model%20is%20obtained%20if%20maximum-likelihood%20estimation%20is%20applied.%20This%20approach%20is%20useful%20for%20measurement%20aspects%20of%20latent%20variable%20modeling%2C%20such%20as%20with%20confirmatory%20factor%20analysis%2C%20and%20the%20measurement%20part%20of%20structural%20equation%20modeling.%20Two%20application%20areas%20are%20studied%2C%20cross-loadings%20and%20residual%20correlations%20in%20confirmatory%20factor%20analysis.%20An%20example%20using%20a%20full%20structural%20equation%20model%20is%20also%20presented%2C%20showing%20an%20efficient%20way%20to%20find%20model%20misspecification.%20The%20approach%20encompasses%203%20elements%3A%20model%20testing%20using%20posterior%20predictive%20checking%2C%20model%20estimation%2C%20and%20model%20modification.%20Monte%20Carlo%20simulations%20and%20real%20data%20are%20analyzed%20using%20Mplus.%20The%20real-data%20analyses%20use%20data%20from%20Holzinger%20and%20Swineford%27s%20%281939%29%20classic%20mental%20abilities%20study%2C%20Big%20Five%20personality%20factor%20data%20from%20a%20British%20survey%2C%20and%20science%20achievement%20data%20from%20the%20National%20Educational%20Longitudinal%20Study%20of%201988.%22%2C%22date%22%3A%22Sep%202012%22%2C%22language%22%3A%22eng%22%2C%22DOI%22%3A%2210.1037%5C%2Fa0026802%22%2C%22ISSN%22%3A%221939-1463%22%2C%22url%22%3A%22%22%2C%22collections%22%3A%5B%225A94Z44M%22%5D%2C%22dateModified%22%3A%222015-09-25T20%3A03%3A17Z%22%7D%7D%5D%7D
Asparouhov, T., Muthén, B., & Morin, A. J. S. (2015). Bayesian Structural Equation Modeling With Cross-Loadings and Residual Covariances Comments on Stromeyer et al.
Journal of Management, 0149206315591075.
https://doi.org/10.1177/0149206315591075
Muthén, B., & Asparouhov, T. (2012). Bayesian structural equation modeling: a more flexible representation of substantive theory.
Psychological Methods,
17(3), 313–335.
https://doi.org/10.1037/a0026802
It is very useful. Thank you.