I have a data set with population information for 6 age groups. In theory, I wish to know if a specific two of these age groups are different from each other.

My first question is: Can I subset the data so that I only test one population against the other and completely ignore the other age groups? I am only interested in the difference between the specific two age groups. The data comes from one sample, and I created the age groups.

My second question is: In testing for significant differences in populations through different variables in the data set, I have conducted Kolmogorov-Smirnov and Wilcoxon rank sum tests on the variables that are obviously not normally distributed. There are some variables which are highly right-skewed, and a log transformation, confirmed by a Q-Q plot makes the data approximately normal. These tests have come out as significant, but I wish to know if there is a way to conduct an analysis on how much of a difference there is in each variable. Performing a t-test on the log transformed data doesn't seem to make sense because the geometric mean does not really carry the same meaning. Is it appropriate to conduct a regression with a continuous variable as the response and the categorical age variable as the predictor? In this case, I am not testing prediction power, but just wish to know the value of the coefficients. Is this method appropriate and how do I justify it if it is?



1 Answer 1


Regarding your first question, it depends whether or not you came to the data set with an a priori intention to focus on those two age sets (e.g. if the hypothesis you came in with was to compare differences between kids before and after they enter high school - or something like that). If you did have that kind of a priori hypothesis, then I don't see why focusing your analysis on those two groupings is a problem. However, if you split the data set into six yourself and then noticed after the fact that the comparison between two of them was interesting while the other pair-wise comparisons weren't, then no, you can't do that: that would be data dredging (aka data mining aka a fishing expedition...).

In this case, if you created the groups yourself and you were only interested in 2 of them, why did you create 6 groups in the first place?

  • $\begingroup$ The research question is very broad, so I came to the data set with the hypothesis already constructed; however, not all the data fits into the two groups that are of interest to me. Rather than eliminating the rest of the data, I just split the data into six groups in case I wanted to do any further analysis. The narrow hypothesis I have come up with here certainly is not the only hypothesis testing I will be doing. I was just wondering if estimation would be possible for this data, as well. $\endgroup$
    – dc3
    Commented Oct 28, 2015 at 21:18

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.