Analysis of variance for a dummy dependent variable

[Mohammadreza Arabi]

Hello all,
My question might seem trivial but it's what I have been searching about for several days. I intend to do some descriptive statistics analysis on a dummy variable and study whether the mean of this variable varies by different characteristics (categorical and continuous ones) of the sample. Due to what I found, in the case of a continuous dependent variable it can be done using "anova" command in Stata, however, for a dummy, I should use chi2 option in "tabulate" command. But I was wondering what is the equivalent of two-way or three-way anova in this situation?

thank you so much in advance...

[Joseph Coveney]

You can start by typing the following at Stata's command line

help logit

[Mohammadreza Arabi]

Thanks a lot Joseph,

But is there any other possibility to achieve this without doing regression?

[Carlo Lazzaro]

Mohammedreza:
as your interested in interactions, too, I suspect there's no alternative to -logit-, as Joseph pointed out.

[Marcos Almeida]

I intend to do some descriptive statistics analysis on a dummy variable and study whether the mean of this variable varies by different characteristics (categorical and continuous ones) of the sample

We are not supposed to work with the mean of a binary variable, strictly speaking. Instead, we get the proportion as well as the absolute numbers.

[Nick Cox]

As the mean of a binary variable and a proportion are one and the same, I don't get Marcos' message here. I don't rule out anova dogmatically here; just watch out that it may be really hard to get conditional distributions that are even roughly normal.

[Marcos Almeida]

I fully agree with Nick in #6. I just wished to underline that, being a variable binary, the "mean" value we get is the proportion.

[Carlo Lazzaro]

Dear All,
and to make things messier,

...proportion is the natural estimate of the probability of an event.

(https://global.oup.com/academic/prod...cc=it&lang=en&, page 86).

[Maarten Buis]

But is there any other possibility to achieve this without doing regression?

Remember that ANOVA is just another way of doing regression. Internally Stata quietly uses -_regress- to estimate the model, and than plays around with the results to create the output expected from an ANOVA model. So I do not see why you want to avoid regression.

If you only had catogical variables, you could try log-linear analysis as the analogue of ANOVA. (see: http://www.stata.com/meeting/uk15/ab..._loglinear.pdf ) However, using logit is probably much easier in your case.

[Mohammadreza Arabi]

Remember that ANOVA is just another way of doing regression. Internally Stata quietly uses -_regress- to estimate the model, and than plays around with the results to create the output expected from an ANOVA model. So I do not see why you want to avoid regression.

If you only had catogical variables, you could try log-linear analysis as the analogue of ANOVA. (see: http://www.stata.com/meeting/uk15/ab..._loglinear.pdf ) However, using logit is probably much easier in your case.

Thanks you so much for the comprehensive explanation. Yes, I read about how ANOVA works, but from the beginning I was looking for some equivalent to t-test of the means in my case. Buy the way, as many approved, I am going to do it with logit.