OLS回归异方差判定咨询：残差图与检验结果矛盾时的判断

Great question—this is a super common point of confusion when validating OLS assumptions, and it’s smart that you’re paying attention to both statistical tests and visual diagnostics!

The short answer is: No, you can’t definitively conclude there’s no heteroskedasticity just because the Breusch-Pagan/Cook-Weisberg test isn’t significant, especially when your residual vs. fitted plot has unusual features. Here’s why, and what to do next:

1. The Limitations of the BP Test #

The BP test is a specific tool—it’s designed to detect heteroskedasticity that’s linearly related to your fitted values (or the covariates you specify in the test). It misses a lot of other heteroskedasticity patterns:

• If the variance of residuals changes in a non-linear way with fitted values (e.g., a U-shape, or variance spikes only at high/low fitted values), the BP test (which tests for a linear relationship) might not pick it up.
• It’s also sensitive to sample size: with small samples, the test might lack power to detect real heteroskedasticity (a Type II error).
• If heteroskedasticity is tied to a specific subgroup (like your x variable’s small/medium/large levels, rather than overall fitted values), the BP test might not flag it unless you explicitly include x in the test specification.

2. Why the Residual Plot Matters More Than Any Single Test #

Visual diagnostics like your rvfplot are critical because they capture unstructured or complex patterns that formal tests can’t. If you’re seeing weird behavior in the lowess curve or residual spread (e.g., residuals suddenly widening in one section, or the lowess line deviating sharply from flat), that’s a red flag—even if the BP test is non-significant.

For example: maybe the variance of residuals is stable across most fitted values, but spikes when x = large and age is above 60. The BP test (which looks at overall fitted values) won’t catch that, but your residual plot (if you slice it by those subgroups) will show it clearly.

3. Next Steps to Investigate #

Don’t stop at the BP test—dig into that residual plot and run additional checks:

• Slice the residuals by your categorical variable x: Make separate residual plots for small/medium/large groups. Do you see differences in residual variance across these groups?
• Run a White test: This is a more flexible test that doesn’t assume heteroskedasticity follows a linear pattern. It tests for heteroskedasticity related to fitted values, squared fitted values, and interactions—better at catching non-linear patterns.
• Check for model misspecification: Sometimes odd residual patterns aren’t heteroskedasticity at all—they’re a sign your model is missing a key term (like a quadratic term for age, or an interaction between x and gender). If your model doesn’t capture the true relationship between predictors and y, residuals can show strange patterns that look like heteroskedasticity but are actually misfit.
• Look for outliers/influential points: A few extreme points can distort both the BP test and the residual plot. Use Cook’s distance or studentized residuals to check if any points are pulling the model off track.

Final Takeaway #

Statistical tests are tools, not final answers. The BP test can rule out linear heteroskedasticity related to fitted values, but it can’t confirm there’s no heteroskedasticity of any kind. When your residual plot shows unusual features, you owe it to yourself to dig deeper—don’t just rely on the test result to declare no heteroskedasticity.

内容的提问来源于stack exchange，提问作者bobmcpop