I'm doing a report of the differences between women and men in their performance on a test with variables like: age, education, race and of course genre. I have two options:

  1. doing a linear regression of the performance over age, education and all my variables to check if being a women makes an statistically significant difference.

  2. Presenting exhausting descriptive statistics to see the differences between men and women. This descriptive statistics are like some kind of three in which I start by setting groups of age, for each group I separate different levels of education, then for each of those smaller groups I separate race and finally genre. At the end I can do a t test to check if the performance on the test is different for women (compared to men) in each small group.

I'm struggling on deciding how to approach to this problem because:

  • I feel that doing the second approach is the same as doing the first one because I feel like "controling" for all variables and then checking if there is a different effect of my variable of interest in each group.
  • Nonetheless, my variable of interest can have a different effect in each of the groups as compared with just one possible observable effect with the regression method.
  • Besides that, following the second method is easier to explain to the directives. Is easier to say that "white, educated, old women perform better on the test" than "ceteris paribus, women perform better on the test".

What thoughts do you have on my situation?. What would you do and why?. What differences do you find between these approaches?

Thank you!

  • $\begingroup$ Pl. restructure your question title and modify body to state your objective that backs your question $\endgroup$
    – user10619
    May 12, 2016 at 16:45
  • $\begingroup$ Hum Ok I will, I just need some time because I got really busy at work now... I'll try to do it at night $\endgroup$ May 12, 2016 at 17:40

2 Answers 2


Go with the regression. Don't go with cross-tabulate-everything-by-everything, because this opens you up to excessive Type I error. If your interest is narrowly about a "gender effect," interact gender with your other predictors for a start, like this, assuming R syntax will be helpful:

mod<-lm(outcome~(covariate1+covariate2+covariate3)*gender), data=data))

Of course, base R will appropriately include the necessary additive terms, with a specification like the above. If you're not using R, the approach is the same; just the syntax is different.

But as with all linear models with interaction terms, remember that you have to look at the marginal effect. That's the partial derivative of your outcome with respect to your predictor. The 'effects' package is helpful. So to get you moving in R, try:


See Understanding Interaction Models, which has 3,000+ cites for a more complete discussion. And no, I am not Brambor, Clark or Golder$\dots$

  • $\begingroup$ Thank you @user105360. I just did the regression as you and Loris proposed me!... it was really good because I found the same insights I found with the cross-tabulation plus some other conclusions that I didn't see because it's really difficult to cross tabulate everything. I'm just wondering: why do you say I'm more exposed to commit Type I error while cross tabulated? $\endgroup$ May 13, 2016 at 13:49
  • $\begingroup$ Let's say you do just one cross-tabulation and and a chi-square test, then the chances you get a false positive (find a 'statistically significant' difference that's not really there) are pretty small. But if you do this over and over and over again, the probability of getting a false positive is much higher. On Wikipedia, this is discussed as multiple comparisons problem. $\endgroup$
    – 5ayat
    May 13, 2016 at 16:50
  • $\begingroup$ @105360 what is co-variate 1, 2 , 3 ? pl.explain taking into account proposed independent variables. Further indicate definition and formula for computing such covariates. $\endgroup$
    – user10619
    May 15, 2016 at 10:34
  • 1
    $\begingroup$ @subhashc.davar Covariate is another word for independent variable. $\endgroup$
    – 5ayat
    May 15, 2016 at 12:09

The second method will lead to a slew of confusing pairwise comparisons. The main concern you express about multiple regression is "my variable of interest can have a different effect in each of the groups", but you can model that (and test that assumption) by specifying interactions between variables.

  • $\begingroup$ Oh I didn't think about that!. Thanks for the idea! $\endgroup$ May 12, 2016 at 16:02

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.