Consider the following data generating process. For simplicity, assume all quantities are scalars with $Y_{it}$, $X_{it}$ and $G_i$ observable for all $i\in \{1..N\}$ and $t\in \{1..T\}$ where $N\gg T$ (but lets assume away dynamic panel bias). Denote the process as:

\begin{align} Y_{it} &=a_i+\beta_i(G_i) X_{it-1}+\varepsilon_{it}\\ &s.t.\\ \beta_i(G_i) &= \delta+\gamma G_i+\eta_i\\ &and\\ E_i[X_{it}\varepsilon_{it}] &=0\;\forall\; i\in \{1...N\}\\ E[G_{i}\eta_{i}] &= 0 \end{align}

To estimate the model, I run the first regression for all i using OLS and then use the estimated coefficients to run the second regression.

My question is as follows: Is there a single-pass OLS panel-regression equivalent of the above two pass procedure? If so, what is it? If not, is there a general classification of these types of problems so I can do further research?

My initial instinct was to run the following regression: \begin{align} Y_{it} &=a_i+\gamma G_i X_{it-1}+\delta X_{it-1}+\epsilon_{it} \end{align} but this didn't seem to reproduce the same coefficients (it's possible I have a coding error, or some wonky issues with missing data, but I've been working with it long enough that I'm starting to doubt that).

As far as the standard errors on $\gamma$ goes, I was thinking a GMM or block-bootstrap approach might make sense.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.