Hi! I noticed that mPlus offers two alternative approaches to modeling when the measurements are not independent. One approach is multilevel modelling (TYPE=TWOLEVEL) and the other is complex and handles the problem of non-independence of observations in a different way (TYPE=COMPLEX). Multilevel modelling is very popular, but the second approach seems less frequent. What do you think about the second approach in cases when you have no reason to model random effects and you just want to account for the non-independence of observations. Please find below a passage from mPlus manual where the authors write about the second approach:
"Complex survey data refers to data obtained by stratification, cluster sampling and/or sampling with an unequal probability of selection. Complex survey data are also referred to as multilevel or hierarchical data. For an overview, see Muthén and Satorra (1995). There are two approaches to the analysis of complex survey data in Mplus. One approach is to compute standard errors and a chi-square test of model fit taking into account stratification, non-independence of observations due to cluster sampling, and/or unequal probability of selection. Subpopulation analysis is also available. With sampling weights, parameters are estimated by maximizing a weighted loglikelihood function. Standard error computations use a sandwich estimator. This approach can be obtained by specifying TYPE=COMPLEX in the ANALYSIS command in conjunction with the STRATIFICATION, CLUSTER, WEIGHT, and/or SUBPOPULATION options of the VARIABLE command. Observed outcome variables can be continuous, censored, binary, ordered categorical (ordinal), unordered categorical (nominal), counts, or combinations of these variable types. The implementation of these methods in Mplus is discussed in Asparouhov (2005, 2006) and Asparouhov and Muthén (2005, 2006a)."
I wonder whether anyone has already tried it before and what are your impressions. Thank you!
Gary King argues that using robust SE's (type is complex) is like taking a canary down the mine; if the robust SE's give different results , there is something wrong with the model, it is not OK to interpret it . (If the canary dies there is something wrong, you need to get out of there!). For what it is worth I agree with him and thus I think it is better to have an explicit model (e.g. multilevel model) that does take the dependency into account and not just treat as a nuisance to be corrected.
http://gking.harvard.edu/publications/how-robust-standard-errors-expose-methodological-problems-they-do-not-fix
With the robust approach the standard errors are inflated but you get no idea of the relative size of the level 2 and level 1 variance, it is a technical fix and not about substantive interpretation, and the approach does not really extend to more complex models with complex variation
I think it's a sandwich estimator, and as Muthen & Muthen write, many researchers believe it's adequate to use it instead of full multilevel model structure whenever you have a multilevel structure in your data but are not interested in effects on the different levels/cross-level interactions. Very Relevant paper showing that "type is complex"-estimators often work better than going multilevel:
http://psycnet.apa.org/fulltext/2016-22467-001.html
Peter Edelsbrunner Thank you for your response. Could you please provide a valid link to the paper or just the title, because the link does not work for me. Thanks!
Yes, please prove a link to this paper if possible, or perhaps the authors and title?
Sorry Meghan Crabtree Lukasz Dominik Kaczmarek this should work Article On the Unnecessary Ubiquity of Hierarchical Linear Modeling
Peter, this paper is awesome and concludes the discussion! Thanks!
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics