SEM post-estimation - standardized results

[Anna Warren]

Hello,

I am running a model to see if my predictor (PSAI) predicts my outcome (PTSD), and whether there is a significant interaction between PSAI and sex ('sexXpsai' is an interaction I manually calculated):

sem (ptsd <- psai mum_edu ethnicity sex sexXpsai ), method(mlmv) vce(robust) stand

To calculate the separate effects by sex, I have used lincom, e.g.

lincom psai + sexXpsai

As in my initial SEM I had asked for the standardized results, I re-ran the lincom with the 'estat stdize' option:

estat stdize: lincom psai+ sexXpsai

I am getting very different results from lincom depending on whether or not I standardize it. E.g.:
β=0.38 (-.34, 1.10), p=.297 (lincom non-standardised)
β=0.03 (-.001, .07), p=.057 (lincom standardised)


I would have expected the coefficients to be different, but I'm unsure why the p-value is so different.

I have tried running the SEM without vce(robust) and without method(mlmv) and I still get very different results.

The stata manual says lincom can be run after SEM, and that 'estat stdize' can be used with lincom, so I'm not sure why my results are so different.

I would appreciate any advice
All the best

[Clyde Schechter]

When you run standardized -sem- the calculations are actually done on the unstandardized variables, and then the results are transformed to reflect standardization of all variables. If you run -lincom- without -estat stdize:-, it just works with the non-standardized SEM results that Stata calculated but didn't show you. So that calculation is not appropriate to working with your standardized model. You must use -estat stdize:- with it to get results applicable to your standardized model.

Now, because you have an interaction term, the model you are estimating is actually incorrect: you have standardized the interaction term. But a standardized interaction term is not the interaction of the standardized variables. So your standardized model is not an interaction model and should not be interpreted as if it were one.

The above confusions are among the many reasons why standardized analyses are usually a bad idea. I don't know how your measure of PTSD is defined, and I don't know what the PSAI is. But I know that there are standard scales for measuring PTSD symptoms, such as the PCL (PTSD symptom checklist) which has a well-known scoring system and well-characterized cutpoints for probable PTSD. So standardizing such a measure serves no purpose other than to obfuscate the study findings because nobody but the author(s) will grasp what a 1 SD difference in PCL scores happens to be in the study sample. The same is even more clearly true of sex: standardizing it in an analysis is an exercise in obscurantism. If the PSAI also has a well-known scoring system, then it, too, can only be made more confusing by standardization.

[Anna Warren]

When you run standardized -sem- the calculations are actually done on the unstandardized variables, and then the results are transformed to reflect standardization of all variables. If you run -lincom- without -estat stdize:-, it just works with the non-standardized SEM results that Stata calculated but didn't show you. So that calculation is not appropriate to working with your standardized model. You must use -estat stdize:- with it to get results applicable to your standardized model.

Now, because you have an interaction term, the model you are estimating is actually incorrect: you have standardized the interaction term. But a standardized interaction term is not the interaction of the standardized variables. So your standardized model is not an interaction model and should not be interpreted as if it were one.

The above confusions are among the many reasons why standardized analyses are usually a bad idea. I don't know how your measure of PTSD is defined, and I don't know what the PSAI is. But I know that there are standard scales for measuring PTSD symptoms, such as the PCL (PTSD symptom checklist) which has a well-known scoring system and well-characterized cutpoints for probable PTSD. So standardizing such a measure serves no purpose other than to obfuscate the study findings because nobody but the author(s) will grasp what a 1 SD difference in PCL scores happens to be in the study sample. The same is even more clearly true of sex: standardizing it in an analysis is an exercise in obscurantism. If the PSAI also has a well-known scoring system, then it, too, can only be made more confusing by standardization.

Thanks so much for your help, that makes sense now!