3
$\begingroup$

I understand that the fixed effects estimator in a panel model (say, individuals, $i$ across years, $t$) can be understood either as a including a dummy for each $i$ or running OLS on the time demean-ed data. My question is whether the estimate from the FE model (that is, the within estimator) is equivalent to the average of the estimates from running OLS on each individual separately. Consider the following two approaches:

$y_{it} = constant + \beta x_{it} + v_i + u_{it}$

$y_t = constant + \alpha x_t + e_t$ for all $i \in (1, 2, ... N) $

The second equation gives us an $\alpha^i$ for each individual and my question is whether $\beta = \frac{1}{N} \sum_i^N \alpha^i$

$\endgroup$
  • 1
    $\begingroup$ Shouldn't there be a $x_{i,t}$ somewhere in the second equation? How does the second equation give you an $\alpha^i$? It doesn't seem to contain a value of $\alpha^i$ at all. $\endgroup$ – BKay Mar 4 '16 at 15:34
  • $\begingroup$ As for BKay's concern, I would suggest rephrasing the second equation to $y_{it} = constant_i + \alpha^i x_{it} + e_{it}$. $\endgroup$ – chan1142 Oct 27 '16 at 6:38
3
$\begingroup$

I asked exactly the same question on math.stackexchange:

https://math.stackexchange.com/questions/1470490/fixed-effects-estimation

In short, the answer is yes, it can be viewed as running separate OLS regressions- the weights, however, are not arbitrary. It is a weighted average of the separate OLS regressions.

$\endgroup$
1
$\begingroup$

I believe the second approach you suggest is equivalent to using an OLS on the whole sample and with dummy variables for each person, country or whatever. Therefore, the answer to your question should be yes.

$\endgroup$
1
$\begingroup$

Let $\hat\alpha^i = \left[\sum_t (x_{it}-\bar{x}_i)^2 \right]^{-1} \sum_t (x_{it}-\bar{x}_i) (y_{it}-\bar{y}_i)$, estimator from the individual OLS regression. Let $\hat\beta$ be the FE estimator using the panel data. Then, math gives the identity $$ \hat\beta = \sum_{i=1}^N w_i \hat\alpha^i,\quad w_i = \frac{\sum_t (x_{it}-\bar{x}_i)^2}{\sum_{j=1}^N \sum_t (x_{jt}-\bar{x}_j)^2}. $$ This confirms ChinG's answer. (Note that the mean-group estimator is $N^{-1} \sum_{i=1}^N \hat\alpha^i$, which is different from the FE estimator.)

$\endgroup$
  • $\begingroup$ +1. The form of $w_i$ also suggests that the FE estimator gives more weight to individuals/groups that have larger variation in $X$. $\endgroup$ – Paul Oct 27 '16 at 17:14

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.