[Clyde Schechter]

No, that's not quite it. You don't do separate regressions for each lag. You do them all at once (close to what you show in your third -xtpoisson- command above) and then do a simultaneous test of the relevant coefficients. The code shown here would include unlagged time, as well as lags 1 through 3 and forward times 1 and 2. You can modify to whichever lags and leads are actually relevant to the dynamics of your situation, of course.

xtpoisson nb_trial_y treated##(L(0/3).policy F(1/2).policy), fe

To do the simultaneous test on the lags or leads you will first have to run -xtpoisson, coeflegend- to see what names Stata has given to the corresponding terms and then list those as arguments in your -test- command.

[Spirae Raïs]

Thank you very much for your precise answer.

Last question I have is :

My average DID effect is 186 when I compute the margins as we discussed in #12. But what does it represent ? Is it an average number by disease for my treatment group? Is it %age points ? It is not clear for me how I should interpret this figure.

Thank you very much for your help.
Spirae

[Clyde Schechter]

That number is the expected difference between:
A. The increase in the average number of studies per disease in the "treated" group following policy change, and,
B. The increase in the average number of studies per disease in the "untreated" group following policy change.

The metric here is counts, not percentages or rates.

[Paola Bertolini]

Dear all,

I resume this old discussion with the hope of clarifying something that I had in the back of my mind for a while. I was estimating a simple nbreg, fe model when I started wondering about the sense of using the -margins- command in an nbreg, fe setting such the one discussed above. My question is: can we actually compute (average) marginal effects in a negative binomial or poisson fixed effect effect set-up, at all?

I am aware that marginal effects in non-linear models depend on the unknown fixed effects. See for example page 46/79 of Trivedi's review: http://www.stata.com/meeting/mexico1...ug_trivedi.pdf

In other contexts (for example, fractional response panels), a Mundlak-Chamberlain device can be adopted by exploting the ad-hoc properties of a given response function. In the fractional response (balanced) panel fe scenario of Papke and Woolridge (2008), identification of average partial effects follows from properties of the probit distribution.

In other words, I would like to know what is the -margin- command going to give me in the present context.

Thanks a lot. I apologize if my question does not make sense

[Clyde Schechter]

It's actually an excellent question!

As you point out, the logit, poisson, and nbreg estimators all condition out the fixed effects, so fixed effects cannot be estimated. That is why the predicted values calculated by -margins- are all conditional on "the fixed effect = 0." But as Jeff Wooldridge has pointed out on this Forum in several places, this criterion is itself poorly defined because its meaning can change with a change of scale! Consequently, I think that most, if not all, of us are now recommending -margins- not be used following these models, at least in most circumstances. In one of his recent posts Jeff Wooldridge has gone so far as to suggest that StataCorp disable the -margins- command after these.

The contrary advice earlier in this thread dates back about 5 years to a time when I, at least, was not fully aware of the difficulties and their implications.

[Paola Bertolini]

It's actually an excellent question!

As you point out, the logit, poisson, and nbreg estimators all condition out the fixed effects, so fixed effects cannot be estimated. That is why the predicted values calculated by -margins- are all conditional on "the fixed effect = 0." But as Jeff Wooldridge has pointed out on this Forum in several places, this criterion is itself poorly defined because its meaning can change with a change of scale! Consequently, I think that most, if not all, of us are now recommending -margins- not be used following these models, at least in most circumstances. In one of his recent posts Jeff Wooldridge has gone so far as to suggest that StataCorp disable the -margins- command after these.

The contrary advice earlier in this thread dates back about 5 years to a time when I, at least, was not fully aware of the difficulties and their implications.

Hey Clyde,

thanks a lot for the kind answer. I was about to edit my post to refer any interested reader into your and others' comments at this [discussion]. The label applied by stata on marginsplot on expected counts (which reads something along the line of "assuming fe = 0) provided some hint

I guess my subsequent question is inevitably: how to deal with coefficient interpretation in the context of a nbreg, fe set-up? Is there an alternative model you would recommend (e.g. quantile regression)?

thanks,

[Clyde Schechter]

Well, the exponentiated coefficients of a negative binomial (or Poisson) regression are estimates of the rate ratios. In my discipline, epidemiology, we work with those fairly often and are comfortable talking about that. But the model is not good for predicting expected outcomes for the reasons we are discussing here. Given your original citations of Mundlak and Wooldridge, I'll go out on a limb and guess that you are in econometrics or a related field. And I am very aware that in those fields, random effects models are viewed with extreme skepticism. In my epidemiology, we do not share that extreme skepticism. In part this is because are often analyzing randomized trials, so that the assumptions necessary for those estimates to be consistent are (if the study is properly implemented and the measurements are valid) met. And, unlike economists, we are often willing to go out on a limb and at least tentatively believe that our observational studies have sufficiently many and sufficiently good quality covariates to also resolve this difficulty. Bottom line is that in epidemiology we would generally go to a random effects model if we need to estimate expected outcome numbers or rates.

I don't know what would be considered the best alternative in an economic discipline. I'll stick my neck out here and hazard a guess that a correlated random-effects model (implemented as -xthybrid- by Francisco Perales and Reinhard Schunck, available from Stata Journal) would be acceptable. It supports the non-linear link functions we are discussing here. But it does not support factor-variable notation, so it is not possible to use -margins- in postestimation. That leaves you the task of writing code that emulates the calculations that -margins- would do. And, as I say, I don't know how acceptable this would be in the world of economics anyway.

[Paola Bertolini]

Well, the exponentiated coefficients of a negative binomial (or Poisson) regression are estimates of the rate ratios. In my discipline, epidemiology, we work with those fairly often and are comfortable talking about that. But the model is not good for predicting expected outcomes for the reasons we are discussing here. Given your original citations of Mundlak and Wooldridge, I'll go out on a limb and guess that you are in econometrics or a related field. And I am very aware that in those fields, random effects models are viewed with extreme skepticism. In my epidemiology, we do not share that extreme skepticism. In part this is because are often analyzing randomized trials, so that the assumptions necessary for those estimates to be consistent are (if the study is properly implemented and the measurements are valid) met. And, unlike economists, we are often willing to go out on a limb and at least tentatively believe that our observational studies have sufficiently many and sufficiently good quality covariates to also resolve this difficulty. Bottom line is that in epidemiology we would generally go to a random effects model if we need to estimate expected outcome numbers or rates.

I don't know what would be considered the best alternative in an economic discipline. I'll stick my neck out here and hazard a guess that a correlated random-effects model (implemented as -xthybrid- by Francisco Perales and Reinhard Schunck, available from Stata Journal) would be acceptable. It supports the non-linear link functions we are discussing here. But it does not support factor-variable notation, so it is not possible to use -margins- in postestimation. That leaves you the task of writing code that emulates the calculations that -margins- would do. And, as I say, I don't know how acceptable this would be in the world of economics anyway.

Hi Clyde,

Apologies for the delay in coming back, I was navigating the board getting stuck here and there (the learning surface is so non-convex!). First of all, you're totally right - my field exists at the periphery of the economics imperialism and as such (also due to the nature of the data under analysis) I suspect it would be easier to justify the use of a xtpois, fe regression. Relative to the interpretation of coefficient, I understand that getting percentage variations from rate ratios is straightforward, and universally very understandable, so it should work well.


However, I will look into the correlated random-effects model as it appears to be flexible and derivation of partial effects would not be problematic.

From this and previous discussions on the board, as well as a RESET test, I indeed decided to drop the nbreg approach and I am going to temptatively stick to xtpois, fe.

thanks again for the great help,