Twitter user data - Fixed Effects Logit the right/only choice?

Question

I'm wondering whether a fixed-effects (logit) model fits the purpose of my study, and would like to get some feedback or ideas which other models could potentially be useful:

I have an unbalanced panel (daily observations, 3 months time period) with Twitter user data. Among the independent variables are

• number of followers
• is the user verified (binary)
• number of tweets on that day
• time between tweet
• ...

the dependent variable is binary (0,1).

As far as I understand, the fixed effects model does not consider the variance between the individuals. However, isn't the variance between the user what really matters here? Given that I suspect that there is much less variation for the independent variables within each user (e.g. the number of followers does not increase so much during the study period. Also whether or not a user is a verified user does not change).

Still, I understand the advantages of the fixed-effects model, given that there are likely individual time-constant differences between the users (e.g. intelligence, talent,...) which the FE model would control for. Is there a model that accounts for individual fixed effects but still considers between variance?

Also, in case I wanted to do a follow-up cluster analysis with the variables that turned out to be statistically significant in the regression, wouldn't I need to consider the between variance in the regression in the first step, because any cluster algorithm essentially tries to separate (between) users.

Thanks a lot for your help and ideas!

Answer 1

I'll reply to your first and second questions but I cannot help you with the last (the one about clustering).

1) Isn't the variance between the user what really matters here?

You are correct that FE uses only within observation variation (all that doesn't vary over time is dropped from the equation). This is because you want to get rid of unobserved heterogeneity to deal with omitted variable bias. That is the point of using FE. If there is little variation over time, it makes not much sense to use FE models, as it is less efficient than other panel data models.

2) Is there a model that accounts for individual fixed effects but still considers between variance?

As for models that use both within and between variation, there are RE (random effects) models. They model individual unobserved heterogeneity as individual random effects, with the trade-off that you have to assume that these effects are uncorrelated to the other predictors. You can augment them to allow for correlated random effects with what is called Mundlak-Chamberlain approach (Woodridge has a paper on how to use these models in unbalanced panels).

A bit of extra advice

Above all, though, you should think well about how your data meets the assumptions of the models you are trying to use. You are dealing with network data, and network data is not independently and identically distributed, mainly because who follows whom depends on who the other people are following (e.g. there is transitivity, reciprocity, etc) and because there are all sorts of spill over effects between individuals. These characteristics of network data are what makes it really interesting, but they also require specialised statistical models. Most panel models assume cross-sectional independence (i.e. that the series of observations for each subject are independent from the series of observations of the other subjects), and that is not met by network data.