Examine effect over time

[Kathrine Weinreich]

I want to examine the effect of several xvar (gender, age, education) on the evaluation on partyleader, and if this effect changes depending on the leaders gender, age and education.

I have a dataset consisting of several choice studies (in the years 1994, 1998, 2001, 2005, 2007 and 2011) all using the same variable in measuring both xvar and yvar.

I would like to somehow cluster my data, so I will be able to se the effect over time, and if this effect is changing over time. So fx. the effect when there is a male respondent and a male party leader, and if this effect is larger or smaller in 1994 compared to 2011. Also fx. if the effect is larger if the respondent is female and the leader is female compared to if the respondent is male and the leader is male.


My main question is how I will be able to cluster og make this OLS reg between my xvar (gender, age and education) and my yvar (leader evaluation) across time and vary in different kinds of characteristics of the respondent and the leader.

[Clyde Schechter]

In general terms, this is what interaction terms do in regression models. So you would include interactions between your xvars and the different kinds of characteristics of the respondent and the leader. If you are not familiar with interaction terms in regressions, you should consult a basic textbook on regression before you proceed. If you are familiar with them conceptually but need help with coding, run -help fvvarlist- for information about how to represent interactions in Stata. Also, to facilitate the interpretation, you will want to learn about Stata's wonderful -margins- command. The most lucid explanation of it I know of is the excellent Richard Williams' https://www3.nd.edu/~rwilliam/stats/Margins01.pdf.

It is unclear to me from your description whether you have longitudinal (panel) data, or if these surveys are cross-sections involving different respondents. If the former, you can't use OLS--you must use a model that is suitable for longitudinal data.

Finally, a caution. It sounds like you are planning on looking at a large number of interactions. This is usually a bad idea. Compared to main effects, the detection of interactions is substantially underpowered--typically the required sample size is about four times as large as is needed for main effects. When you run underpowered analyses, the results are noisy. If you run multiple hypothesis tests in noisy data, you will generate a lot of Type I errors, and you won't be able to distinguish those from findings of actual importance if there are any. So unless you have a really huge data set, be quite cautious here.