2
$\begingroup$

I have two dependent variables that I want to predict. One is a normally-distributed continuous variable, and the other is a binary categorical variable (0 or 1). They're moderately correlated so that the binary variable is 1 for higher values of the continuous variable.

The theoretical question I'm interested in is whether the binary variable is just a coarser-grained version of the continuous variable (i.e. they can be predicted by the same model), or whether they're best predicted by two different models.

I have a large number of IVs that should predict both variables. My first thought was just to fit a linear and a logistic regression for the continuous and binary variables, respectively, and see whether the same set of IVs predicts both of the DVs, or whether they're a different set. That seems ad-hoc to me and I was interested in knowing if there's a better approach I can take.

$\endgroup$
6
  • 2
    $\begingroup$ I would fit a mediation model with the continuous version of the measure mediating the relationship between the predictor and he categorical version. If the categorical version is a coarser measure of the continuous version, then there should be no indirect effect between the predictor and the categorical outcome. You can do this in Lavaan in R. $\endgroup$ Jan 20, 2015 at 23:56
  • 2
    $\begingroup$ Here are some details. lavaan.ugent.be/tutorial/mediation.html $\endgroup$ Jan 20, 2015 at 23:57
  • 1
    $\begingroup$ @JeremyMiles: Your comments are perhaps worth turning into an answer. $\endgroup$ Jan 21, 2015 at 8:10
  • $\begingroup$ Yes, very helpful - thank you! Your comments also led me to these papers on Bayesian mediation analyses: ncbi.nlm.nih.gov/pubmed/24903686 (with R package BayesMed) and ncbi.nlm.nih.gov/pmc/articles/PMC2885293 which were helpful since I need to use mixed-effects. $\endgroup$ Jan 22, 2015 at 21:30
  • $\begingroup$ I know it's been almost a decade, but I agree with Richard; @JeremyMiles do you want to turn your two comments into an answer? $\endgroup$
    – Peter Flom
    Jan 6 at 12:03

1 Answer 1

1
$\begingroup$

[Copied from comment]

I would fit a mediation model with the continuous version of the measure mediating the relationship between the predictor and the categorical version. If the categorical version is a coarser measure of the continuous version, then there should be no indirect effect between the predictor and the categorical outcome.

You can do this using the Lavaan in package in R.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.