An article by Laura Stapleton came out in the most recent issue of the SEM journal which shows how to use SAS in conjunction with Mplus and replicate weights. We will eventually be adding this to Mplus. In the meantime, you can see that article.
I'm using the ECLS-K data and am confused after reading the above paper and the others by Tihomir on analyzing multilevel complex samples in Mplus -- about which weights to use for these data. The file for the longitudinal data comes with both PSU and Strata weights, and full sample weights and Jackknife2 replicate weights for the kids.
Since I want to take the unequal selection probabilities of the schools into account for my longitudinal analysis, I'm thinking that I should set up a growth model multivariate style, but as a multilevel model to take PSUs and strata into account.
So, do I use the Taylor linearization weights for the clusters and strata for the multilevel part, as well as the full sample and JK2 replicate weights for the kids?
Dan Li posted on Wednesday, June 10, 2015 - 8:27 pm
Hi Dr. Muthen,
I am using PISA data to do a SEM. PISA use Fay's replicate weights. I read your paper "Asparouhov, T., & Muthen, B. O. (2010). Resampling methods in Mplus for complex survey data. Los Angeles: Muthen & Muthen", you provide syntax for using replicate weights:
Hello, If I only have the weight (and not strata and cluster) for my data, and I want to get bias-corrected bootstrapped confidence intervals for indirect effects in a mediation model, is it correct to do the following under the analysis and output commands?
Is repse=bootstrap needed under analysis? Or can replicate weights not be calculated when we do not have strata and cluster?
Dear Dr Muthen, I am using PISA data to do SEM with one dichotomous variable and several continuous variables. I use WLSMV method of estimation. How can I get fit indices for the model? can I also test indirect effects in mediational part of the model? Thanks, Sylwia
1. Probably assuming the CLUSTER = b_weight; STATIFICATION = c_weight; really specify the clusters and the strata. These are not weights.
2 and 3. You can remove BOOTSTRAP = 2000; REPSE = BOOTSTRAP; and you will get fit indices. The bootstrap method concerns standard errors. It doesn't change the point estimates. The bootstrap version of chi-square (Bollen-Stine) is not available yet in these complex sampling settings.
For a follow-up question regarding your response on my Q2 and Q3. If I am testing indirect effects in my path model, can I keep these two commands (BOOTSTRAP = 2000; REPSE = BOOTSTRAP; ) in my syntax so that I can get a bootstrap estimation for indirect effects?
I am looking for a statistical software to conduct a multilevel analysis (three-levels, dependent variable = binary 0/1) using replicate weights (complex survey design) as well as plausible values (measurement error) and multiply imputed data - is there a possibility to run such an analysis (combining all these features) in MPlus? (to me it seems that using multilevel commands in MPlus require stratification and PSU indicators and do not allow for replicate weights in conjunction with a cluster variable for the hierarchical levels?)
Thank you Dr. Asparouhov, for your very quick response with the Mplus command! It works great! Problem Solved! Thank you!
shonnslc posted on Monday, November 25, 2019 - 6:55 am
As far as I know, Mplus doesn't provide fit indices when replicate weights are used. Then, how can I evaluate my SEM model(global fit) in this case? Also, reviewers are going to ask for the information if SEM is employed. Any suggestions? Thank you so much.
Replicate weights based test of fit is not available. Include as much of the sampling information as you have: sampling weights etc. (excluding the replicates weights) and run type=complex to get fit indices and chi-square. I suppose if you compare the replicate weights SE and the type=complex SE you can make an approximate design effect of some kind which can be used to attempt to correct the test of fit but you are on your own with that. If the SE are not very different I would read that as endorsement that the test of fit and fit indices of type=complex are good enough.
Dear Mplus-Team, I want to use Mplus to generate replicate weights, but not as in the user guide for a factor model, but for an entire data set containing about 200 mainly ordinal items. Similar to Pisa, where they provide brr for the whole data set. How could this be done in Mplus. As I understand, the generation of repweights is independet from the analysed model. Only sampling weights, strata and clustering information are needed to calculate the weights. Thus, could I use an abritrary model in conjunction with an Id-variable to merge the repweights to the whole data set?