Industry fixed effects in a dynamic panel

[Danielle leblanc]

Hi,

I want to include the industry fixed effects in a dynamic panel model based on the 2 digit sic code (S&P industry classification). I used xtabond2 and the model was running just fine: the model was

xtabond2 dsow    L.dsow     epu cfw leveragew f.growth2w faw sizew roaw dgs3mo gdpqg infq i.q_date, gmm(L.dsow,eq(both) collapse lag(1 1))  gmm(epu,eq(both collapse lag(1 1))  gmm(cfw,eq(both) collapse lag(2 2)) gmm(leveragew,eq(both) collapse lag(1 1)) gmm(leveragew,eq(lboth) collapse lag(1 1)) gmm(growth2w,eq(both) collapse lag(2 2)) gmm(faw,eq(both) collapse lag(1 1)) gmm(sizew,eq(both) collapse lag(2 2)) gmm(roaw,eq(both) collapse lag(1 1)) iv(i.L.q_date,eq(diff)) iv(i.L.q_date,eq(level))iv(L.dgs3mo) iv(L.gdpqg) iv(L.infq) twostep robust

When i trued to put the industry fixed effect using i.sic I started to receive an error that regressors outnumber instruments.So what should I put as instrumens for i.sic to solve the issue . In other words by adding i.sic to the regression as follows:

xtabond2 dsow    L.dsow     epu cfw leveragew f.growth2w faw sizew roaw dgs3mo gdpqg infq i.q_date

then what should I put as instruments for i.sic???? sic lagged values are the same as the current ones . So I am really puzzled

[Sebastian Kripfganz]

In the absence of appropriate instruments, you can only identify the coefficients of industry-fixed effects by assuming that they are uncorrelated with the unobserved "fixed effects". You will then need to instrument them by themselves, i.e. iv(i.sic, eq(level)). If the purpose of including the industry-fixed effects is merely to control for them, you do not need to include them at all into your model because all time-invariant effects are already accounted for by the unobserved fixed effects.

A few more comments on your specification:

It seems a bit weird to me that you are instrumenting the time dummies i.q_date by their time lag i.L.q_date. Time dummies are usually exogenous and can simply be instrumented by
themselves, similar to the industry-fixed effects. Note that there is a bug in xtabond2 that produces incorrect degrees of freedom (and therefore incorrect p-values) for the Sargan/Hansen overidentification tests of time dummies are specified with the factor notation, as some of those time dummies will be shown as omitted but xtabond2 still counts them as estimated coefficients. You would need to manually create separate time dummies for each period and make sure to include only non-omitted variables in the model specification. (The same applies to any other omitted variables or categories of factor variables.)

The iv() option of xtabond2 without the equation() suboption is probably not doing what you think it does. It does not create separate instruments for the first-differenced and the level model. If you are aware of this and know what this option is actually doing, nothing is wrong. Otherwise, you might want to specify the instruments separately for the first-differenced and the level model. Also, keep in mind that instruments specified with the iv() option for the level model must be uncorrelated with the unobserved "fixed effects".

More on GMM estimation of linear dynamic panel data models:
XTDPDGMM: new Stata command for efficient GMM estimation of linear (dynamic) panel models