Pooled OLS

Pooled OLS runs a single OLS regression on all $N \times T$ panel observations, treating them as if they were $N \times T$ independent draws. It ignores the panel structure entirely: no unit dummies, no time dummies, no clustering. Pooled OLS conflates between and within variation, so its estimate of $\beta$ is contaminated by every time-invariant cross-unit confounder.

When to use

Pooled OLS is a useful baseline to show how much bias your panel structure introduces — but it should rarely be the headline estimate. PS_4 reports pooled OLS at −18% and then FE at −7%, making the bias from unobserved fixed state heterogeneity transparent. Use pooled OLS only when you can credibly argue that there are no unit-level unobservables correlated with the regressor (rare in practice).