非时间序列数据调用durbinWatsonTest()得到p-value=0是否可能？

Absolutely, this scenario is not only possible but entirely explainable once you break down what the Durbin-Watson (D-W) test actually does—regardless of whether your data is time-series or not.

Let’s start with the basics: The D-W test is designed to detect first-order autocorrelation in regression residuals, meaning it checks if each residual is correlated with the one immediately before it. It doesn’t care if your data is ordered by time, space, or even an arbitrary sequence—if the residuals have a meaningful linear relationship with their lagged counterparts, the test will pick it up.

In your case, the output shows a D-W statistic of ~1.31 (far below the baseline value of 2, which indicates no autocorrelation) and a p-value of 0. This happens because:

• A D-W statistic this low signals strong positive autocorrelation in the residuals.
• If your sample size is large enough, even a moderate autocorrelation (like your 0.34 value) will lead to an extremely small p-value—so small that it rounds to 0 in the output.

Why would non-time-series data have autocorrelated residuals? #

Even without a time component, there are common scenarios where this occurs:

• Implicit ordering: Your data might be sorted by a variable that introduces dependency (e.g., geographic proximity, product category, or survey response order). For example, if you sorted automobile data by engine size before running the regression, residuals for similar-sized cars might cluster together, creating autocorrelation.
• Omitted variables: If you left out a key predictor that has a systematic relationship with your outcome, the residuals could pick up that pattern and show autocorrelation.
• Spatial or clustered data: Data points that are grouped (e.g., cars from the same manufacturer, or observations from the same region) often have correlated residuals because they share unmeasured characteristics.

Recommendations to address this: #

• Check your data’s ordering: Verify if your dataset is sorted by any variable (intentional or accidental). If it is, try shuffling the data randomly and re-running the test to see if the autocorrelation disappears.
• Diagnose residual patterns: Plot the residuals against their lagged values, or use the acf() function in R to visualize the autocorrelation structure. This will confirm if the first-order correlation is truly driving the D-W result.
• Refine your regression model: Look for omitted variables or misspecified functional forms (e.g., adding polynomial terms or interaction effects) that could be causing the residual autocorrelation. Fixing these often eliminates the issue.
• Use robust standard errors: If the autocorrelation is unavoidable (e.g., with clustered data), switch to robust standard errors (like those from the sandwich package’s vcovCL() function) to get more reliable inference, even if the residuals are correlated.

内容的提问来源于stack exchange，提问作者SungwonAhn