A Problem with Time Dummies in Pooled Time Series Model

[Nihat Mugurtay]

I am running a pooled time series model in STATA 17, the standard errors are clustered around different units. I have a concern about my model and prediction.

One of my independent variables is time-invariant, which is the Muslim majority variable. Also, all the variables are logged and lagged. Instead of fixed effects, I thought using a pooled-time series with necessary tools will be better.

I included lagged-DV as usual. But, I am not sure if I need to add time- dummies. Because I used lagged-DV as an independent variable. I thought I may inflate my coefficients with time-dummies (even if it might be unnecessary).

Best,

reg dv L.dv L.c.a i.muslim L.lnb L.logn L.logdacoda L.loggdpconstant L.logtrade L.concor, vce(cluster id)

However, I may use a different model with time dummies.

reg dv L.dv L.c.a i.muslim L.lnb L.logn L.logdacoda L.loggdpconstant L.logtrade L.concor i.time, vce(cluster id)

[Jared Greathouse]

Why lag your outcome? Use time dummies if you think there's unit stable, time variant confounding that you don't observe.

[Nihat Mugurtay]

Without taking Lagged DV as an independent variable (in time series cross-sectional models), we cannot capture the effects of other independent variables properly. I think Lagged-DV plays a role similar to year dummies. However, I think it might be unnecessary to use both a lagged-DV and year dummy together. I know we can use year dummies without a lagged DV like this:

reg dv  L.c.a i.muslim L.lnb L.logn L.logdacoda L.loggdpconstant L.logtrade L.concor i.year, vce(cluster id)

However, I wonder if I can use this:

  reg dv lagged dv  L.c.a i.muslim L.lnb L.logn L.logdacoda L.loggdpconstant L.logtrade L.concor i.year, vce(cluster id)

[Jared Greathouse]

Without taking Lagged DV as an independent variable (in time series cross-sectional models), we cannot capture the effects of other independent variables properly.

No that's not true. At least, it isn't so cut and dry.

I pretty much only do panel econometrics and causal inference, and there's no such rule that essentially says you always have to lag your outcome and include it as a predictor or the interpretation of your IVs are flawed. As far as I'm aware, you lag your dv if you wanna account for (I think) serial correlation or maybe other issues, but it isn't some catch all rule.

[Joro Kolev]

OP does not quite explain what he is trying to achieve here, and what is the problem that he faces.

Did the regression brake down when you included full set of time dummies plus the lagged value of the dependant variable?

What do you want to measure, the effect of a Muslim majority (a time constant variable) on a time varying outcome?

[Nihat Mugurtay]

Joro Kolev

Yes, My DV is time-variant, and I have two time-invariant independent variables in my model: "a binary religion variable" and "log of distance". This binary religion variable is essential to measure one of my hypotheses.

I am interested in time effects, instead of country effects.