Durbin-Watson Test for Autocorrelated Residuals: Reading the Statistic and the Inconclusive Zone

Autocorrelation in the residuals does not bias your regression coefficients — it invalidates the standard errors computed from them. Under the usual OLS assumptions, the coefficient estimates stay unbiased and consistent even when residuals are serially correlated (as long as the regressors themselves are strictly exogenous and none of them is a lagged version of the dependent variable). What breaks is the formula OLS uses to compute standard errors, which assumes the errors are independent of each other. When they are not, that formula is wrong — usually too small under the positive autocorrelation that shows up constantly in time-ordered data — which inflates t-statistics, narrows confidence intervals, and makes p-values look more convincing than the data actually support. The Durbin-Watson statistic is the standard first check for this problem, and reading it correctly means understanding not just the 0–4 scale but the inconclusive zone built into the test itself.

What Autocorrelated Residuals Actually Invalidate

It is worth being precise about which parts of a regression output autocorrelation touches, because the intuitive assumption — that autocorrelated residuals mean the model itself is wrong — overstates the problem in one place and understates it in another.

• It does not bias the coefficients. Given strictly exogenous regressors (and no lagged dependent variable in the regressor set — see below), OLS coefficient estimates remain unbiased and consistent regardless of serial correlation in the errors. The point estimates are, on average, still centered on the true population values.
• It does invalidate the standard OLS standard-error formula. That formula is derived under the Gauss-Markov assumption that errors are uncorrelated with each other (spherical errors). Autocorrelation violates that assumption directly, so the reported standard errors, t-statistics, F-statistics, and confidence intervals are no longer trustworthy as computed.
• It usually means your standard errors are too small. With the positive autocorrelation that dominates real time-series and panel data (a shock this period tends to persist into next period), the conventional OLS standard error formula typically understates the true sampling variability of the coefficients — which inflates significance and increases the false-positive (Type I error) rate on hypothesis tests built from those coefficients.
• OLS is no longer efficient. Even with unbiased coefficients and a corrected standard-error formula, plain OLS stops being BLUE (Best Linear Unbiased Estimator) once the errors are correlated — a generalized least squares (GLS) estimator that accounts for the autocorrelation structure would produce more precise estimates from the same data.

In practice, this means a significant Durbin-Watson test is not a reason to distrust your coefficients or throw out the model. It is a reason to distrust the standard errors, confidence intervals, and p-values computed alongside those coefficients — and to fix that specific piece before reporting them. This is the same shape of problem as multicollinearity, where the diagnosis targets precision, not the point estimate itself — but the fix is different, because multicollinearity is a property of the predictors and autocorrelation is a property of the errors.

The Durbin-Watson Statistic: What It Measures

The Durbin-Watson statistic, introduced by James Durbin and Geoffrey Watson in a pair of 1950 and 1951 papers, tests specifically for first-order autocorrelation — correlation between a residual and the residual immediately before it. It is computed from the regression’s own residuals:

d = Σ(e_t − e_t−1)²  /  Σe_t²

where e_t is the residual at observation t, summed over the ordered sequence of observations. The statistic ranges from 0 to 4, and for large samples it is approximately related to the first-order autocorrelation coefficient ρ by d ≈ 2(1 − ρ):

• d ≈ 2 — little to no first-order autocorrelation (ρ ≈ 0).
• d approaching 0 — strong positive autocorrelation (ρ approaching 1): residuals track each other from one observation to the next, the common pattern in time series with a trend, momentum, or an omitted slow-moving variable.
• d approaching 4 — strong negative autocorrelation (ρ approaching −1): residuals alternate sign from one observation to the next, a less common pattern that can show up after over-differencing a series.

A value close to 2 is reassuring but not, by itself, a formal test result — deciding how far from 2 is far enough to matter is exactly what the critical-value tables below are for.

Reading the 0–4 Scale and the Inconclusive-Zone Problem

Durbin and Watson could not derive a single critical value for d, because its exact sampling distribution depends on the specific matrix of regressor values in your data — not just on sample size and number of predictors, the way a t- or F-distribution’s critical value does. Their solution, published as statistical tables, was to bound the true (data-dependent) critical value between two figures: a lower bound d_L and an upper bound d_U, both indexed by sample size, number of predictors, and significance level. The decision rule for positive autocorrelation is:

• d < d_L — reject the null of no autocorrelation; positive autocorrelation is present.
• d > d_U — fail to reject the null; no evidence of positive autocorrelation at this significance level.
• d_L ≤ d ≤ d_U — the test is inconclusive. The true critical value for your specific data could fall anywhere in that gap, and the bounds alone cannot say which side of it your statistic lands on.

The same logic applies symmetrically for negative autocorrelation, testing 4 − d against the same d_L/d_U bounds. The inconclusive zone is not a flaw in how a particular study ran the test — it is a structural feature of the Durbin-Watson test itself, and a genuinely common outcome, especially in smaller samples where d_L and d_U sit further apart. When your statistic lands there, the honest options are: report it as inconclusive and move to a test that does not have this gap, or gather more data (the bounds narrow as sample size grows).

The test that closes the gap is the Breusch-Godfrey test (developed independently by Breusch and Godfrey in 1978), a Lagrange-multiplier test for autocorrelation with no inconclusive region. It also generalizes to higher-order autocorrelation (correlation at lag 2, 3, and beyond, not just lag 1) and remains valid when a lagged dependent variable appears among the regressors — both real limitations of the plain Durbin-Watson test, covered next. In practice, many researchers now run Breusch-Godfrey directly rather than starting with Durbin-Watson, precisely to avoid the inconclusive-zone problem; Durbin-Watson remains worth knowing because it is still the statistic most statistical software reports by default and the one most reviewers expect to see referenced.

The Lagged-Dependent-Variable Trap

The Durbin-Watson test assumes none of the regressors is a lagged value of the dependent variable itself. If the model includes Y_t−1 as a predictor of Y_t — a common and often necessary specification in time-series work — the Durbin-Watson statistic is biased toward 2, understating the true degree of autocorrelation and making the test unreliable exactly when autocorrelation is most likely to be present. For that specification, the standard alternative is Durbin’s h-test (also developed by Durbin, as a direct fix for this case) or, in larger samples, the Breusch-Godfrey test above, which stays valid with lagged dependent variables among the regressors. Checking whether a lagged Y is in the model is a five-second check worth doing before trusting a Durbin-Watson result at all.

When It Matters — and When It Doesn’t

Autocorrelation is fundamentally about ordering: it asks whether the residual at one position in a sequence is correlated with the residual at the adjacent position. That framing determines where the test is meaningful and where it is not.

• It matters for time-series data — regressions on data observed sequentially over time (monthly sales, daily prices, annual outcomes), where a real, meaningful adjacency exists between observation t and observation t−1. This is the setting the test was built for, and where omitted trends, seasonality, or momentum routinely produce exactly the positive autocorrelation the test flags.
• It matters for panel/longitudinal data — repeated observations of the same units over time carry the same ordering concern within each unit, though the plain Durbin-Watson statistic is not the right tool there. Panel data needs a panel-specific serial-correlation test (for example, the Wooldridge test for autocorrelation in panel data) that accounts for the panel structure rather than treating the pooled data as one long time series.
• It does not meaningfully apply to cross-sectional data — a dataset of independent units observed at a single point in time (survey respondents, patients, firms in one year) has no natural ordering for “adjacent” to refer to. Running Durbin-Watson on cross-sectional data does not error out — it silently computes a statistic based on whatever row order the dataset happens to be in, which is an artifact of data entry or sorting, not a real dependency between observations. A cross-sectional dataset sorted by region, by collection date, or by any variable correlated with the outcome can produce a spuriously “significant” Durbin-Watson result that reflects the sort order, not genuine serial dependence. If you must report it for a cross-sectional model (a reviewer asked, or a template requires it), shuffle the row order first, or better, drop it in favor of checking residual plots and independence assumptions through the study design itself (independent sampling, clustering-aware standard errors) rather than a test built for sequential data.

Fixing Autocorrelated Residuals

Once autocorrelation is confirmed, the practical response depends on what is causing it:

• Heteroskedasticity- and autocorrelation-consistent (HAC) standard errors — most commonly the Newey-West estimator — recompute the standard errors to be valid under autocorrelation without changing the coefficient estimates themselves. This is usually the simplest fix when the coefficients are fine and only the inference around them needs correcting.
• Respecify the model — if the autocorrelation is coming from a real, omitted dynamic (a trend, a lag structure, seasonality), adding the missing term directly addresses the cause rather than papering over the symptom in the standard errors.
• Generalized least squares (GLS) — the Cochrane-Orcutt and Prais-Winsten procedures transform the model to account for the estimated autocorrelation structure directly, restoring efficiency (BLUE) rather than only correcting standard errors.
• Cluster-robust standard errors — in panel settings, clustering by the panel unit is often the more direct fix than a time-series-style correction, since it addresses correlation within a unit’s residuals over time without needing to specify the exact autocorrelation structure.

Running the Test

Most general-purpose statistical software reports Durbin-Watson as an optional regression diagnostic rather than a default output. In SPSS, it is requested from the Statistics button in the Linear Regression dialog — see running multiple regression in SPSS for where that checkbox sits alongside the model’s other diagnostics. In Stata, it is a postestimation command, estat dwatson, run after regress — see regression in Stata for the broader postestimation workflow. In R, the lmtest package’s dwtest() function returns both the statistic and a p-value computed directly (bypassing the need to look up d_L/d_U tables by hand), and the same package’s bgtest() runs Breusch-Godfrey.

Frequently Asked Questions

What counts as a good Durbin-Watson value?

A value close to 2 indicates little first-order autocorrelation. There is no universal exact threshold, because the actual critical bounds (d_L, d_U) depend on sample size, number of predictors, and significance level — a value that is fine in a large sample can fall in the inconclusive or problem zone in a small one. Look up the bounds for your specific n and k rather than relying on a fixed rule of thumb like “below 1.5 is bad.”

Does autocorrelation bias regression coefficients?

No. Given strictly exogenous regressors and no lagged dependent variable in the model, OLS coefficients remain unbiased and consistent under autocorrelation. What is invalidated is the standard-error formula used to build confidence intervals and significance tests around those coefficients.

Can the Durbin-Watson test be used on cross-sectional data?

It can be computed, but the result is not meaningful in the way it is for time-series data, because cross-sectional observations have no natural ordering for “adjacent residuals” to refer to. A statistic computed on arbitrarily or non-randomly sorted cross-sectional data reflects the sort order, not genuine serial dependence.

What should I use instead when the Durbin-Watson result is inconclusive?

The Breusch-Godfrey test has no inconclusive region, generalizes to higher-order autocorrelation, and remains valid when a lagged dependent variable is among the regressors — all situations where Durbin-Watson is either uninformative or invalid.

Why does the Durbin-Watson statistic range from 0 to 4 instead of, say, −1 to 1?

It is built from squared differences between consecutive residuals rather than a direct correlation coefficient, which is why 2 (not 0) represents “no autocorrelation” and the scale runs from 0 (strong positive autocorrelation) to 4 (strong negative autocorrelation) rather than the more familiar −1-to-1 correlation scale.

For the broader set of quantitative diagnostics this sits alongside — checking predictors for multicollinearity, reading residual plots for homoscedasticity, and reporting results correctly once assumptions are checked (see how to report p-values and what statistical significance does and doesn’t mean) — see the Research Methods & Statistics hub.