I'm planning to do a structural equation model (SEM) in Mplus where high school math motivation (continuous) and gender (dichotomous) would predict attainment in an undergraduate degree in mathematical STEM, life science, or non-STEM (nominal).
A multinomial logistic regression comes to mind here, but it doesn't produce the usual model fit statistics. I would prefer to conduct an SEM with 2 dummy variables as the outcome (third group is the reference group) but I'm having problems with convergence even after increasing iterations to 100,000 and wanted to check if this is the right analysis to do?
Hi Li
What you want to do is not possible as a SEM model because you have a categorical outcome variable. SEM, as generally thought of, is a linear model and you will not be able to predict categorical outcomes in this way - this probably why Mplus is complaining a lot.
Are you using dummy variables because "...it (Mplus) doesn't produce the usual model fit statistics"? You can't get the SEM indices such as TLI/CFI etc in a logistic-type model, but you can work out if the model is good or not. For example, you could fit the multinomial model twice - first with all the parameters fixed to zero, and then with all the paths being estimated. You can then compare the models using BIC, AIC or do a 2*LL test (using the difference in # parameters as the df).
Cheers
Mark
It seems your objective is to see whether the achievement as measured by ? in three undergraduate programs differs in males and female students and more motivated students vs less motivated students. You can categorize the motivation into two/three categories and use two- Analysis of variance using gender and motivation as factors.
Why are you not considering their high school GPA and/or SAT scores? Do you think they are not relevant? I would suggest you check the literature on it. The
students in these groups are likely to differ in these. Think about it, if you want to use either of them, you can use Analysis of Covariance using that as a covariate.
By the way, there is some research evidence to suggest that grit is a better predictor than these for achievement.
Hi Harinder Mahajan
My objective is to examine how latent continuous factors (math motivation: success expectancies, intrinsic value, importance value and costs) predict whether participants graduated with a math science, life science or non-STEM undergraduate degree (categorical DV) over and above high school math achievement scores, for men and women.
Thank you for your suggestion, but I am using an existing longitudinal dataset (Time 1: 1995) and thus do not have access to other variables such as grit.
I've consulted with others in an Mplus forum and have decided to go with a multinomial logistic regression considering the nature of the dependent variables.
Best wishes,
Lili
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling