ttests for paired data with different number of trials in each pair

[Elaina Johns-Wolfe]

Hello,

I am trying to conduct a paired t-test on quasi-experimental data where matched (on sex, income, etc.) pairs of subjects who differ by race experience a different number of trials, meaning that individuals’ scores in each pair are the means across the varying numbers of trials. So the data look like:

pair no. trials_1 trials_2 xbar_1 xbar_2

1 15 10 4.0 2.0
2 3 8 2.9 5.5

etc.

My thinking is that in testing for overall differences in means between group 1 and group 2 I would want to weight case 1 more than case 2. So basically I want to account for different within-pair and between-pair variances.

Stata’s ttest command does not allow for aweights, and since I want to maintain the paired structure of the study, reshaping the data to long format and regressing xbar on a dummy variable for group 1 vs. group 2 with aweights is not a complete solution. Any suggestions?

Thanks,
Elaina

[Clyde Schechter]

If you reshape your data to long, you can regress the variable of interest on the group distinction and also include the pair number (as a factor variable) as a term in the regression. That's completely equivalent to a paired ttest in the unweighted situation. And -regress-, of course, supports aweights.

[Elaina Johns-Wolfe]

If you reshape your data to long, you can regress the variable of interest on the group distinction and also include the pair number (as a factor variable) as a term in the regression. That's completely equivalent to a paired ttest in the unweighted situation. And -regress-, of course, supports aweights.

Hi Clyde,

Thank you for the help. It looks like that method is going to work!