I’m working on a measurement model in lavaan with both latent and observed variables. My model includes two categorical predictors: gender (male/female) and family business background (yes/no). When I run the CFA or SEM, I get warnings about not being able to compute standard errors and the latent variable covariance matrix not being positive definite.
Here’s a simplified version of my model:
model <- '
# Measurement model
Attitude =~ A1 + A2 + A3
Motivation =~ M1 + M2 + M3
Performance =~ P1 + P2 + P3
# Structural model
Performance ~ Attitude + Motivation + Gender + FamilyBusiness
'
fit <- cfa(model, data = mydata)
The model works fine without the categorical predictors. Should I recode them as numeric (0/1) instead of factors? Or is there a better way to include categorical variables in lavaan?
I’ve looked at other posts but couldn’t find a solution to this specific issue. Any advice would be really helpful!
I’ve encountered similar issues with categorical predictors in lavaan. One approach that worked for me was using dummy coding instead of treating them as factors or ordered variables. For gender, you could create a dummy variable (e.g., 1 for female, 0 for male) and for family business background, another dummy (1 for yes, 0 for no).
Then modify your model to use these dummy variables:
This approach often resolves the standard error computation issues and positive definite matrix problems. It also makes interpretation straightforward, as the coefficients represent the effect of being in one category versus the reference category.
I’ve been tinkering with lavaan models myself, and I totally get your frustration with those pesky categorical predictors. Have you considered using effect coding instead of dummy coding? It’s a neat trick that might just do the trick!
Here’s the gist: instead of 0 and 1, you use -1 and 1. So for gender, you could do -1 for male and 1 for female, and for family business, -1 for no and 1 for yes. It’s like giving lavaan a balanced diet of categories!
The cool thing about effect coding is it can help with model stability and interpretation. Plus, it keeps your reference group as the overall mean, which is pretty handy.
Just curious, have you looked into interaction effects between your categorical and continuous predictors? Sometimes those can reveal some interesting patterns!
What do you think about giving effect coding a shot? I’d love to hear if it works out for you or if you find another solution. Keep us posted!
it might help lavaan handle them better. also, check if ur sample size is big enough for the model complexity. sometimes that can cause issues w/ standard errors n stuff.
