A few more things about panel data estimation techniques. Fixed-effects versus a dummy variable for each observation: These techniques are equivalent. They give exactly the same estimates of coefficients, and exactly the same standard errors, and exactly the same results when testing hypotheses. The dummy-variable technique is computationally intensive, but it automatically gives you estimates of the c i values. This might be useful if you’re interested in knowing the distribution of c i . It’s also useful when you want to account for two (or more) different sources of fixed effects. (Interestingly, while these techniques are exactly equivalent, they yield different R 2 values—another example of why you can’t interpret an R 2 coefficient on its own.) Fixed-effects versus first-differences: Both estimation techniques offer unbiased results whenever cov(xit ,c i ) ≠ 0 , as long as sequential exogeneity holds. In many cases, there’s little reason to prefer one over the other. Some considerations: 1. We haven’t discussed any trends in the uit unobservables. Sometimes, luck in one year might correlate with luck in other years: if this year you have a job that pays unusually well, then next year you’ve got a good chance of being in the same job with high wages. Sometimes not: if this year you have high income because you won the lottery, then next year you go back to being average. This issue of “serial correlation” is complex. Both FE and FD rely on some assumptions, and I’m going to avoid going into detail about those. Firstdifferences might be more appropriate when the randomness tends to be correlated over time, while fixed-effects might be more appropriate when randomness tends to dissipate completely between periods. 2. When “strict exogeneity” fails, both FE and FD are biased. This serves as a useful check of this assumption. Estimate the model both ways, and compare the estimated coefficients. If strict exogeneity is satisfied, then you expect to get similar results. If there is a violation, then both estimators are biased, but in different ways. This will cause FE and FD to look very different. 3. When strict exogeneity fails, it’s usually easier to remedy the problem using FD than FE. Constant terms: If the model is yit = β1 + β 2 xit + … , then you’re not going to recover β1 by doing regression with the fixed-effects or first-differences transformed data. They cancel out of the transformations. If you capture a constant term, it reflects the time trend. When strict exogeneity fails: All of the techniques result in biased estimates, and we have to resort to instrumental variables to correct the problem. We can combine the IV technique with panel data techniques. However, we still need an instrument that’s uncorrelated with the error term—or with the transformed error term, when we’re doing panel data transformations. This is another reason why we might favor one estimation technique. Fixed-effects requires that the instrument zit is uncorrelated with u it = uit − ui = uit − T1 ∑ uis . Usually, this requires that zit is uncorrelated with every single uis : past, present, or future. First-differences, with instrumental variables, is unbiased if the instrument is zit is unrelated to ∆ uit = uit − uit −1 . This is a small difference, but sometimes it comes in handy: it requires that the instrument zit is uncorrelated with just uit and uit −1 . In some cases, that means that we can use xit − 2 as our instrumental for Δxit . It’s likely that the past xit − 2 is correlated with xit −1 and xit (and thus, with Δxit ) but not with future uit or uit −1 , so not with Δuit . This situation is sometimes described as “sequential exogeneity”.
© Copyright 2026 Paperzz