**Summary:** This paper investigates the estimation and inference of cointegrated regression models in panel data, comparing the performance of three estimators: OLS, FMOLS, and DOLS. The study highlights the limitations of the OLS estimator in finite samples, showing it has significant bias. The FMOLS estimator does not consistently outperform OLS, while the DOLS estimator performs better than both. Monte Carlo simulations demonstrate that DOLS has the smallest bias and provides more accurate estimates, especially in heterogeneous panels. The paper also discusses the asymptotic properties of these estimators, showing that DOLS has a normal limiting distribution. The results suggest that DOLS is the most reliable estimator for cointegrated panel data, particularly when dealing with non-normal errors and heterogeneous panels. The study concludes that DOLS is the best estimator overall, although its t-statistic may have downward bias in non-normal error distributions. The paper also notes that the performance of FMOLS is limited in homogeneous panels due to its reliance on kernel estimators for the asymptotic covariance matrix.**Summary:** This paper investigates the estimation and inference of cointegrated regression models in panel data, comparing the performance of three estimators: OLS, FMOLS, and DOLS. The study highlights the limitations of the OLS estimator in finite samples, showing it has significant bias. The FMOLS estimator does not consistently outperform OLS, while the DOLS estimator performs better than both. Monte Carlo simulations demonstrate that DOLS has the smallest bias and provides more accurate estimates, especially in heterogeneous panels. The paper also discusses the asymptotic properties of these estimators, showing that DOLS has a normal limiting distribution. The results suggest that DOLS is the most reliable estimator for cointegrated panel data, particularly when dealing with non-normal errors and heterogeneous panels. The study concludes that DOLS is the best estimator overall, although its t-statistic may have downward bias in non-normal error distributions. The paper also notes that the performance of FMOLS is limited in homogeneous panels due to its reliance on kernel estimators for the asymptotic covariance matrix.