Statisticians call the least squares estimator BLUE — best linear unbiased estimate — when errors are normally distributed. But political science often uses this method with non-normal data, increasing risk of unreliable estimates.
This article argues that BLUE's requirement for normality doesn't apply to real-world politics where skewed error distributions are common. It shows how standard practices can mislead researchers by being sensitive to unusual observations.
What We Did:
We combined theoretical analysis with Monte Carlo simulations, demonstrating the pitfalls when data deviates from normal assumptions.
Alternative Approaches Needed:
Researchers should use robust estimators or variable transformations. These methods handle skewed errors better and provide more reliable results in political science contexts.
Communicating Findings:
Effective detection strategies are crucial for summarizing these issues and conveying the influence of specific data points to fellow researchers.
| When BLUE Is Not Best: Non-Normal Errors and the Linear Model. was authored by Carlisle Rainey and Daniel Baissa. It was published by Cambridge in PSR&M in 2020. |
