Controlling for time trend in a fixed effects model

[Alex Kucharski]

Dear Statalisters,

as part of my dissertation I am running a fixed effects model investigating the relationship between the share of females in parliament and explanatory variables such as proportional representation, GDP, the share of females in the labour force etc.

I have 8 points in time for with I have observations (every 5 years between 1977 and 2012).

I want to control for a general global trend towards more gender equality that could have occurred in that time.

I have included a lagged variable for the share of females in parliament and dummies for each of the years but I am not sure if that's enough?

Thanks for you help!

[Clyde Schechter]

I must be missing something. You have 8 observations and you are fitting a model with more than 10 predictor variables? How is that going to work?

[bsc.j.j.w]

I agree with Clyde, but to answer your question. Why can't you just include time? That will take care of the "time effect", in this case the trend towards more equality. In addition, I am not sure whether it is sensible to include a lagged version of females in the parliament, because I think the number of females in the parliament is a Markov process. Please correct me if I am wrong.

[Alex Kucharski]

Apologies I didn't specify this clearly- I have 8 points in time (1977, 1982, 1987, 1992, 1997, 2002, 2007, 2012) for which I have 495 observations for 155 countries.

So in my fixed effects regression I would like to control for the global trend towards more females in parliament and I'm not sure what method is most appropriate.

Based on the data I think the number of females in parliament is strongly correlated with the number of females there were in the previous parliament as that reflects the position of women in the society, which is not random with every parliament.

Thanks for your help!

[Sebastian Kripfganz]

For global time trends you can easily control by adding a set of time dummies to your regression (or a linear time trend as a more parsimonious specification). Regarding the lagged dependent variable: If you have a short time dimension (only 8 observations) you have to deal properly with the short-T dynamic panel data bias (also known as the Nickell bias). The simple fixed effects estimator xtreg, fe yields biased estimates and is not appropriate here. GMM estimators, among others, qualify as an alternative. See for example the Stata manual on xtdpd, or the user-written command xtabond2.

[Sebastian Kripfganz]

As a side comment: You might also consider estimating a fractional response model because your dependent variable is a fraction between 0 and 1.