Statistics for testing the difference in mean difference?

[Sabrina Gong]

Hello:

I have two groups of observations, group A and B. Within each group, I have a special event to split observations into news=0 (pre-event) and news=1 (post-event). Now I have the mean differences within each group, DIFF1=groupA/new=1-groupA/new=0, and DIFF2=groupB/new=1-groupB/new=0, how can I compare the two differences in means (difference between DIFF1 and DIFF2)? I am thinking to calculate T-statistics but is that appropriate here? I don't think ttest in Stata works. Any suggestion will be highly appreciated.

[Clyde Schechter]

The specific commands depend on the structure of your data. If the people/firms/whatever entity the unit of observation is, in group A are separate from those in group B, and if those in the news = 0 and news = 1 are also separate, then the simplest way would be:

regress outcome i.group##i.news

where outcome is the variable whose mean difference in differences you want to estimate, group indicates groups A and B, and news is 0 or 1. The coefficient of group#news is the estimated mean difference in differences, and its associated statistics (t-tests, confidence intervals, etc.) come with it.

If you have panel data, that is, if the same entities are observed in both news = 0 and news = 1 conditions, then you need to use a model that accounts for the lack of independence among observations. If the variable id identifies observations on the same entities, then you would do something like:

xtset id
xtreg outcome i.group##i.news, // fe or re option

If you are not already familiar with the -regress- and -xtreg- commands, I suggest you familiarize yourself with them through the corresponding chapters in the user's manual. The examples in both of those chapters are quite helpful. Also read the section on factor variable notation in the online help (-help fvvarlist-) so you understand both these commands and the output they generate.