Lewbel Instrumental variable

[Abhinaya Kanakasabai]

Hi, I ran:

ivreg2 lnGDPper lag1_lnGDPper lag2_lnGDPper lag3_lnGDPper lag4_lnGDPper lnGDP_trans (high_frac_dem = distcr cen_lat), robust cluster (country)

The final output says, "Instrumented: high_frac_dem, Included instruments: lag1_lnGDPper lag2_lnGDPper lag3_lnGDPper lag4_lnGDPper lnGDP_trans, Excluded instruments: distcr cen_lat

The instruments I've used here are only distcr and cen_lat. The other variables which stata says included instruments are my control variables. can someone please tell me why it calls the control variables as included instruments? I'm quite concerned if I'm getting the right estimates as expected. are the controls used as instruments here?

Thanks in advance

[Jeff Wooldridge]

The message is expected because the control variables act as their own instruments. In particular, all exogenous variables appear in the first stages for distcr cen_lat. You can see this using the "first" option.

[Abhinaya Kanakasabai]

Oh got it! Thank you so much for the clarification and the wonderful book you've written for us, Jeff!

[Abhinaya Kanakasabai]

Hi again! I have a panel data of 177 countries from 1960 to 2010. I ran the following code:

qui ivreg2 lnGDPper lag1_lnGDPper lag2_lnGDPper lag3_lnGDPper lag4_lnGDPper lnGDP_trans (low_frac_dem high_frac_dem = frac_iv1 frac_iv2 frac_iv3 frac_iv4) i.country i.year, robust cluster(country)

This code works perfectly and gives me significant results as long as I don't include the country and the year fixed effects. When I do include them, I get insignificant results with the following warning:

I tried the partial option as well. I'm quite concerned about the reliability of my IV estimates and how helpful the reported standard errors are.

Thank you!

[Maxence Morlet]

This might help:

ivreghdfe

[Maxence Morlet]

From ssc install

ssc install ivreghdfe

[Jeff Wooldridge]

You definitely don't want to even see the cluster-robust standard errors/t statistics for the country fixed effects. They're meaningless. The others should be fine. To make sure, take Maxence's advince. Or use xtivregress after xtset country year and use the FE option. You include i.year as you did above. With T = 50 and N = 177, there is an issue of how reliable clustering will be. You might also want to think about the decision to include lags of your y variable.

[George Ford]

ivreghdfe will handle your problem with ease..

Including the lags of the DV can be problem because they too are endogenous. You'll need the Arellano-Bover/Blundell-Bond type estimator (xtdpdsys).

I'd keep it simple at first.

[Jeff Wooldridge]

With large T, the bias from including lags may be quite small. The lags are predetermined -- not contemporaneously endogenous. Under restrictions on the time dependence, the bias is on the order of 1/T. Still, I'm not sure I would start by including lags.

[Abhinaya Kanakasabai]

Thank you so much, Maxence, Jeff, and George! I removed robust & cluster and replaced ivreg2 with ivreghdfe. The warning disappeared now. However, when I tried xtivreg, I'm getting the following error:

"independent variables are collinear with the panel variable country"

Thanks again!

[Maxence Morlet]

Thank you so much, Maxence, Jeff, and George! I removed robust & cluster and replaced ivreg2 with ivreghdfe. The warning disappeared now. However, when I tried xtivreg, I'm getting the following error:

"independent variables are collinear with the panel variable country"

Thanks again!

Do these specific variables vary within each country, over time?

If not, do you have coarser panel indicators, such as continent or larger region? e.g. Western Europe, North America, etc.?

[Abhinaya Kanakasabai]

Maxence Morlet Thank you for your reply! Yes, these values do change across countries and years. I also have panel indicators - region

[Abhinaya Kanakasabai]

Maxence Morlet Sincere apologies, I missed the thread. Thank you for your reply! Yes, these values do change across countries and years. I also have panel indicators - region

[Maxence Morlet]

In this case, this Papke and Wooldridge (2022) paper may be of interest to you. It has really helped me in my research, and it is perhaps applicable in this case to determine whether you need country fixed effects, or whether region fixed effects are sufficient (countries must be perfectly nested within regions if I'm not wrong): https://link.springer.com/article/10...81-022-02337-y.

[Abhinaya Kanakasabai]

Maxence Morlet thank you so much for this! It is really helpful!