[Charlie Hammond]

Hi Joao Santos Silva, I tested my independent variables with xtunitroot and found that several of them are non-stationary. I have read in posts on this forum that some fields ignore stationarity while others pay attention to it.

I am interested in avoiding spurious regression!

Do you have any recommendations on how I should handle this? There is differencing, but then that muddies the interpretation right? I have not been able to find much useful information on this forum regarding this subject.

Thanks,

Charlie

[Joao Santos Silva]

Dear Charlie Hammond,

That is not really my area, but I would not use xtunitroot in this context. Because your N is so small, I would used dfgls series by series and try to get a detailed picture of the problem. Also, your T is large, but is the span of the data large enough to meaningfully test for unit roots? Having 1000 observations covering 1 second or 100 years is not the same thing.

Best wishes,

Joao

[Charlie Hammond]

Hi Joao Santos Silva, I have N = 128 and I have up to 10 independent variables, so that would mean going through 1,280 individual time series, is this what you intend?

The time index is 943 consecutive daily observations.

Say I find stationarity using dfgls, are there any resources you can point me to? I have not been able to much. I am perusing through [1] right now and it seems to be somewhat relevant.

[1] Matyas L, ed. The Econometrics of Multi-Dimensional Panels. Vol 50. Cham: Springer International Publishing; 2017. doi:10.1007/978-3-319-60783-2

[Joao Santos Silva]

Dear Charlie Hammond,

In #29, you said that your N was 10-20, that is why I suggested going 1 by 1. Anyway, with 3 years of data, it may not make sense to look at stationarity, but it will depend on your field, so I will not comment.

Best wishes,

Joao

[Charlie Hammond]

Joao Santos Silva, you are right I did!! I did not know what N meant in that post. Now I know N is the number of entities in the panel. Forgive my confusion!

[Charlie Hammond]

Joao Santos Silva: The Stata documentation for xtpoisson says that the fe option gives the conditional fixed effects, I cannot find any information on how this differs from the fixed effects included in, say, "xtreg, fe". Do you have any advice or resources that could clarify this?

I would also like to run one of my research questions by you: we are investigating how long it takes for an effect of an independent variable to influence the dependent variable and we are doing this by lagging or doing rolling window calculations on the independent variables and including them in the regression. We'd like to "optimize" this regression with 5-fold cross validation (CV), but the issue is what is the objective function value? With "xtreg, fe", we use the test-set average (from 5-fold CV) mean square error, but in this case you said predictions are not possible because the "predict, xb" command gives predictions along an index rather than time. You also said the deviance is not a good idea. You mentioned using F tests to see whether a variable can be dropped or not, but what if two different lags of the same variable indicate inclusion in the model? How can we decide between the two? Maximize the F test?

Thanks!

Charlie

[Joao Santos Silva]

Dear Charlie Hammond,

About the terminology, I am not the best person here to comment on it, but I believe that Stata refers to conditional fixed effects when the estimator "conditions out" the fixed effects rather than using a dummy for each unit/individual. In the case of linear models and Poisson regression, there is no difference between the two approaches, but in the logit these are very different approaches.

On your second question, if the F test indicates that both lags should be included, you may want to consider both. Anyway, because you have large T, you can simply use the poisson command (or ppml or ppmlhdfe) with a dummy for each individual (always use clustered standard errors). With these estimators you can get the predictions and use the mean square error in the CV.

Best wishes,

Joao