Message/Author 

Joseph posted on Monday, May 12, 2014  9:05 am



Dear Drs Muthen, I have data from a study which collected data across 7 time points (waves) across the same 588 respondents. I would like to run a model to assess the relationship between 2 categorical variables while accounting for clustering at respondent level. Based on the examples in chapter 9, I have written the script below (see below), I am hoping you could comment on this to let me know if it appears sensible: Variable: Names are ID wave CENTRE topic3g_w st_topic3g_w sat_meanw invw impw st_sat_meanw st_sa3gw sat3gw st_invw CODEP; Missing are all (9999) ; USEVARIABLES ARE topic3g_w invw wave; CATEGORICAL ARE invw; ! BETWEEN = topic3g_w wave; WITHIN = topic3g_w wave; CLUSTER = ID; DATA IMPUTATION: impute = topic3g_w (c) invw (c) wave (c); Analysis: TYPE=TWOLEVEL; ESTIMATOR= WLSMV; Model: %WITHIN% invw ON topic3g_w wave; 


You need TYPE=RANDOM to specify a long format growth model. This is not available with WLSMV. I suggest specifying your growth model using the wide format. See the examples in Chapter 6. This is also a more flexible model as the residual variances are not held equal over time. 

Joseph posted on Tuesday, May 13, 2014  2:08 am



Hi Linda, Thank you for your reply. After some more reading I noticed that the MLR estimator used by default in my original script actually uses a logistic model while the WLSMV a probit model. I'd actually prefer a logistic solution. I am mostly interested in assessing if there is an association between satisfaction and topic selected rather than modelling growth. However, is it wrong to use the multilevel modelling above (after specifying TYPE=TWOLEVEL RANDOM)? Thanks again. 


The model you specify has no need for TYPE=RANDOM because you have no random effect except for a random intercept which does not require TYPE=RANDOM. See Examples 9.1 and 9.2 in the user's guide. The first is a random intercept model and the second includes a random slope and random intercept. 

Back to top 