Normality test for multiple imputations

[Irene Gómez]

Dear all,

I have a data set with multiple imputations performed. I want to perform linear and logistic regression to test the effectiveness of an intervention. Before of that, I would like to test the normality of variables that I am going to include in the regresion models. I want to test univariate and multivariate normality.

The normality comands are not officially supported by mi estimate so I tried the option cmdok but I could not perfom the analysis because an error occured. Could someone explain to me how can I test the normality with multiple imputation data.

I hava 2,471 observations aprox.

Thanks you so much in advance.

Petra.

[Carlo Lazzaro]

Petra (please read the FAQ about the highly welcomed habit to introduce ourselves with real given and family names. Thanks):
welcome to this forum.
The good news is that you do not need to check your variables for normality (regardless tye fact that your values are fully, partially observed or imputed).
In linear regression, normality is a (weak) requirement of the residual distribution.
I do not think that normality has even a role in logistic regression.

[daniel klein]

While Carlo is correct in stating that neither linear nor logistic regression assume normally distributed data, note that any assumptions that you might have regarding your analysis model(s) already apply at the imputation stage. It would strike me as rather inconsistent to check model assumptions only after imputation.

Best
Daniel

[Carlo Lazzaro]

Daniel gave, as usual, a wise recommendation.

[Irene Gómez]

Dear Carlo and Daniel,

Thanks you so much for your replies both of them are really useful for me I am doing the analysis of my PhD and my tessis advisor ask me to test normality before perform correlations and then regressions. Is becouse of that I want to test normality but with imputed data. Any recomendation?

Carlo, related to my name, you ara right, but my real name was occupate by another participants so I have no options. I will try again with a more realistic name.

Kind regards,
Irene.

[daniel klein]

[...] my tessis advisor ask me to test normality before perform correlations and then regressions. [...] Any recomendation?

Yes: Read up on the methods/models that you are using and form your own opinion. It is mostly math, so most of it is probably not really an opinion. Anyway, if you indeed come to the conclusion that you need to do those tests, get back, convince me, and I am happy to help with code. If instead, you come to the same conclusion as Carlo and I did, talk to your advisor, argue why you do not believe that such tests are necessary or make any sense and ask for his/her arguments to convince you otherwise.

I am aware that my advice might be a bit idealistic. However, even though I have still not yet finished my own Ph. D., I would expect a Ph. D. holer to be able to make an argument for their case and defend it. Do not blindly follow your advisor. That is not how science is supposed to work.

Others might well disagree; that's the point.

Best
Daniel

[Carlo Lazzaro]

Irene:
thanks for clarifying.
The core of the issue is why does your supervisor want you to test for normality before correlation and regression? (and, at the top of that, why performing correlation when you have planned a subsequent regression?)

PS: crossed in the cyberspace with Daniel's reply (on which I agree verbatim).