ZIP vs linear model
Message/Author
 Carl-Etienne Juneau posted on Friday, November 06, 2015 - 12:22 pm
I have an outcome that is skewed to the left. To improve its distribution, I have used its square root.

A colleague suggested I use Poisson or zero-inflated poisson models instead.

How would this choice of modelling impact results?
 Bengt O. Muthen posted on Friday, November 06, 2015 - 1:14 pm
You need to have a count outcome for that, so non-negative integer values.
 Carl-Etienne Juneau posted on Saturday, November 07, 2015 - 5:09 pm
Dear Dr. Muthén,

Yes, I am aware of this.

I can recode my outcome so it has only non-negative integer values. And use zero-inflated poisson models.

Or I can use its square root. And use "normal" (linear?) models instead.

For an outcome that's skewed to the left, how would this choice of modelling impact results?

Would the estimates of coefficients be more accurate with Poisson models?

Would this affect model fit in any way?
 Bengt O. Muthen posted on Monday, November 09, 2015 - 10:11 am
You may want to ask these general modeling questions on SEMNET.
 Carl-Etienne Juneau posted on Monday, November 09, 2015 - 11:00 am
Thanks for the tip! Will try SEMNET.
 Jennie Jester posted on Thursday, April 12, 2018 - 1:42 pm
I had a reviewer ask me for an effect size for a zero-inflated Poisson model, can you tell me how I can get this?

Here is the model I used:
analysis:type = complex;
model:
fag_revt6c on behunct4 cohort1 cohort2 cohort3 ;
fag_revt6c#1 on behunct4 cohort1 cohort2 cohort3 ;

Thanks,

Jennie
 Bengt O. Muthen posted on Thursday, April 12, 2018 - 5:34 pm
I don't think there is an established E.S. metric for counts. Conventional ES has to do with a standardized mean difference for 2 groups for a continuous outcome and you don't have a continuous outcome. One possibility that is relevant because count distributions are characterized by probabilities is to look at group differences in probabilities for e.g. count = 0. See Chapter 6 of our RMA book.

You can look at count mean differences but that is more abstract.