Is the reported statistic a p-value?

No. A separate valid reference calculation is needed for a p-value.

Does a value near two guarantee independence?

No; it mainly addresses adjacent linear correlation.

Can I compute it after sorting residuals?

That changes the question and is inappropriate for assessing the original temporal ordering.

AI Glossary / Statistics and probability / Intermediate

Durbin-Watson statistic

Definition

The Durbin-Watson statistic measures how strongly successive regression residuals differ relative to their overall squared size. It is commonly used to investigate first-order serial correlation in a regression's errors. Its value lies between zero and four when the residual sum of squares is positive. Values near two are broadly associated with little lag-one correlation, smaller values with positive correlation and larger values with negative correlation. These are diagnostic interpretations, not universal significance cutoffs or proof that a model is valid.

What it means

A model can fit the average level reasonably while repeatedly underpredicting for several periods and then overpredicting for several periods. Such runs suggest that residuals carry temporal information the model has not captured. Durbin-Watson summarizes one aspect of that pattern. The residuals must be in their meaningful observation order, usually time order. Sorting them by size would manufacture a pattern unrelated to the original regression. The statistic concerns adjacent residuals and does not measure every possible form of dependence or forecast inadequacy.

How it works

After fitting the relevant regression, calculate d = sum from t = 2 to n of (e_t - e_(t-1)) squared, divided by the sum from t = 1 to n of e_t squared. Small successive changes relative to total residual variation produce a small statistic. Alternating signs often produce larger changes and a larger statistic. A common approximation is d approximately 2(1 - r), where r describes lag-one residual autocorrelation; finite-sample endpoint terms prevent treating this as a universal exact identity. Formal testing depends on the regression design and appropriate reference distribution. Classical bounds procedures may yield an inconclusive region, and standard Durbin-Watson inference is unsuitable when lagged dependent variables appear among the regressors.

Three practical examples

Demand modeling: inspect whether errors from a regression arrive in sustained positive or negative runs.
Process analysis: compare the statistic with a residual time plot before deciding whether temporal structure needs investigation.
Forecast review: identify when an automated report provides only a statistic rather than a valid significance test.

Think of it this way

Think of walking along a line of residual heights and measuring the size of each step. Smooth stretches have small steps even when the heights are far from zero. Rapid alternation creates large steps. The measure depends on the order of the walk, not just the collection of heights.

Worked example

Take hypothetical ordered residuals -2, -1, 0, 1 and 2, such as centered observations from an intercept-only fit. Their squared total is 4 + 1 + 0 + 1 + 4 = ten. Each successive difference is one, so the numerator is four and d = 4/10 = 0.4. Now arrange the same values as -2, 1, 0, -1 and 2. The squared successive differences are nine, one, one and nine, giving d = 20/10 = two. Both sequences have the same mean, variance and residual magnitudes, but different adjacency patterns. This demonstrates order sensitivity; it is not permission to reorder real time-series residuals. Neither five-value calculation supplies a significance conclusion without the applicable model and testing assumptions.

Illustrative case study

Fictional case: Willow Planning reviews a monthly demand regression whose residual chart shows long runs above and below zero. A low Durbin-Watson statistic prompts investigation of omitted seasonal structure and model dynamics. Analysts then evaluate revised models on held-out periods and use diagnostics suited to their regressors. They do not repeatedly add predictors merely to push the statistic toward two. This invented case illustrates a diagnostic trigger, without claiming that one particular omitted variable caused the pattern or that a desirable statistic guarantees accurate forecasts.

Common misconceptions and mistakes

Do not interpret two as proof of independent errors: higher-order or nonlinear dependence can remain. Do not apply a fixed rule such as every value below 1.5 is statistically significant regardless of sample size and design. Irregular time gaps and panel boundaries complicate adjacency and require deliberate handling. A lagged response predictor invalidates ordinary Durbin-Watson testing assumptions, even though the arithmetic statistic can still be computed. If every residual is zero, the denominator is zero and the statistic is undefined. It is not the Durbin rank test for incomplete blocks.

Business use

Businesses should use this statistic as part of a model-review record that includes residual plots, temporal structure and the actual decision task. Serial dependence can affect conventional uncertainty calculations and signal opportunities to improve forecasting, but the appropriate response depends on the model. Record observation ordering and any gaps or aggregation. If a diagnostic leads to model changes, evaluate their consequences outside the fitting sample. A single convenient number should not replace checking whether errors behave adequately across the time horizons and operating conditions that matter.

Check your understanding

Why did the worked example's statistic change while its squared residual total stayed ten?

Answer: Reordering changed successive differences in the numerator, while the denominator depends only on the residual values.

Related terms

Autocorrelation
regression residual
lag
Breusch-Godfrey test
residual diagnostics
time-series regression
Durbin h test

Learn next

compare residual time plots and autocorrelation functions, then study which serial-correlation tests remain appropriate when a regression contains lagged dependent variables.

Autocorrelation

Sources

Sources: Hyndman, Glossary of Forecasting Terms: https://robjhyndman.com/mwh3/FG4.pdf ; statsmodels, durbin_watson, formula and approximation: https://www.statsmodels.org/v0.13.5/generated/statsmodels.stats.stattools.durbin_watson.html ; SAS, Testing for Autocorrelation, lagged-dependent-variable limitation: https://support.sas.com/documentation/cdl/en/etsug/68148/HTML/default/etsug_autoreg_gettingstarted07.htm ; SAS forecasting text, bounds discussion: https://www.sas.com/storefront/aux/en/spforetimeseries/57275_excerpt.pdf .

Previous: Durbin test Next: Durbin - Wu - Hausman test