I wonder if I may have a similar issue to the one implied above; I'm using weights for a nationally representative sample which are both frequency and sampling weights combined in one number (i.e. they are all greater than 1 and most have fractions). I thought that perhaps I could get this to work in MPlus by using the weight as a sampling weight and specifying the number of observations (NOBSERVATIONS) as equal to the weighted population (my sample has about 2000 cases that represent about 200,000 individuals), but that does not seem to work. Are there any other solutions within Mplus?
Dear Mplus team, my question regards weighting data. I collected 3 measures in 3 different occasions for each participant. Time delay between the 3 measures was different among all participants.
This is the simplest example: •for participant 1 I collected var1, var2 and var3 the 1st July •for participant 2 I collected the var1 the 1st July, var2 the 1st August and var3 1st September
I assume that the more delay I have between the measures, the less strong is the relationship found between the measures. Therefore, I would like to weight data in order to give more importance to participant 1 that have the least time delay and less importance to participant 2 that has the biggest time delay.
Thus I have 3 times delay variables: number of days between var1-var2; var2-var3; var1-var3. For each variable I would give the most importance to time delay=0.
I have 2 questions: (1)How does Mplus weight data? Does it give more importance to bigger numbers? Because in that case I need to transform my times delay variables. (2)Can I use more than one variable to weight data? Because I need to weight the association between var1 and var2, var1 and var3 and var2 and var3 by the time delay.
1) we would simply maximize the weighted likelihood
Sum weight_i * log-likelihood_i
where weight_i and log-likelihood_i are the weight and the livelihood for individual i
2) I would discourage you to use weighting for your problem. The wights are meant to be inverse of probability of selection and this is what is being assumed while computing SE - the way standard errors are computed maters a lot with weights. There are two methods implemented in Mplus (frequency and sampling weights) neither one of which I would recommend. I would use this instead (you should look up example 5.23 in the user's guide for explanations)
The above model lets the relationship between the variables be stronger or weaker depending no how much time has elapsed between the observations. There are tons of variations and you can compare them using the BIC.
Standardized results are not available because the model estimated variance covariance is different for every subject (because the regression coefficients are subject specific as well). The standardized regression coefficients will also be subject specific but somewhat more complicated than the unstandardized. You have two options - standardize these by hand yourself or standardize the dependent variables before the analysis using "define: standardize v1 v2 v3". While the second option has some drawbacks I would not hesitate to use it in this situation.
You can use likelihood ratio test or BIC to evaluate model fit.