Durbin-Watson Test Interpretation: How to Read Your Regression Output

Durbin-Watson test interpretation is the one regression check that trips up more dissertation students than almost anything else I have seen in twelve years of guiding people through their statistics chapters. You run the test, SPSS or Stata gives you a number like 1.42, and then what. Most guides just tell you “close to 2 is good” and leave you stuck the moment your value isn’t close to 2.

I want to fix that here. This is the full explanation I actually give my dissertation students, not the textbook version.

What Is the Durbin-Watson Test?

The Durbin-Watson test checks whether the residuals (errors) in your regression model are correlated with each other in sequence. It comes from a landmark 1950 paper by James Durbin and Geoffrey Watson, published in the journal Biometrika, and it remains the most commonly reported autocorrelation check in regression output across SPSS, Stata, EViews, and Excel.

What It Actually Measures

In plain terms, it asks whether today’s error looks like yesterday’s error. If your residuals follow a pattern instead of behaving randomly, one of the core assumptions of ordinary least squares regression breaks down. That quietly damages your standard errors and your p-values.

Why It Matters for Your Results

If autocorrelation is present and you ignore it, your regression can look statistically significant when it really isn’t. I have reviewed dissertation drafts where a “significant” predictor disappeared completely once the autocorrelation was corrected. That is not a small issue to discover during a viva.

Durbin-Watson Test Interpretation: The Quick Answer

Definition: The Durbin-Watson statistic is a number between 0 and 4 that tests for autocorrelation in regression residuals. A value near 2 suggests none, values below 2 suggest positive autocorrelation, and values above 2 suggest negative autocorrelation.

DW Value Interpretation Table

DW Value	What It Means
0 to 1.5	Positive autocorrelation likely
1.5 to 2.5	No serious autocorrelation
2.5 to 4	Negative autocorrelation likely

What Counts as a “Good” DW Value

This is the version students actually need, not the theoretical version most websites publish. A lot of the guides I looked at while researching this piece stop right here and call it done. That is where they go wrong. This rough range does not replace the proper dL/dU bounds test, and using it alone in a dissertation methodology chapter will get flagged by a careful examiner.

Positive vs Negative Autocorrelation Explained

Understanding positive vs negative autocorrelation properly will save you from misreading your output.

Positive Autocorrelation (DW Below 2)

This is the one you will see most often, especially with time series data like sales figures, stock prices, or monthly economic indicators. A positive error tends to be followed by another positive error. Think of a regression predicting monthly revenue where an unmeasured seasonal factor keeps nudging every month’s error in the same direction.

Negative Autocorrelation (DW Above 2)

This is rarer in practice. A positive error is typically followed by a negative one here, so the residuals swing back and forth. I see this occasionally in overcorrected models or in data with an alternating structure, such as odd and even survey batches processed differently.

Common Causes

1. An omitted variable that changes gradually over time
2. A wrong functional form, such as fitting a straight line to curved data
3. Data collected in a time or spatial order without accounting for that order

The dL and dU Critical Bounds (Inconclusive Zone)

This is the part that separates a properly written dissertation from one that gets sent back for revisions.

The Durbin-Watson statistic does not have a single fixed cutoff. Durbin and Watson themselves gave two bounds, dL and dU, based on your sample size and number of predictors. You compare your DW value against these bounds, not against “2” alone.

• DW below dL: Positive autocorrelation confirmed
• DW between dL and dU: The test is inconclusive, and you genuinely cannot decide either way from this test alone
• DW above dU: No evidence of positive autocorrelation

Here is what that looks like with real numbers from the published critical value tables, at the 5% significance level:

Sample Size (n)	k = 1 predictor	k = 2 predictors	k = 3 predictors
30	dL 1.35, dU 1.49	dL 1.28, dU 1.57	dL 1.21, dU 1.65
50	dL 1.50, dU 1.59	dL 1.46, dU 1.63	dL 1.42, dU 1.67

Most students never learn this zone exists because their software just prints one number and stops. If your value lands here, do not force an interpretation. Report it honestly and back it up with a residual plot, or run the VIF calculator and Breusch-Pagan test calculator alongside it. A clean DW result means very little if your other regression assumptions are shaky.

Using a Durbin-Watson Calculator

Looking up dL and dU by hand from a printed table is tedious, and it is easy to get wrong with an unusual sample size. I built a Durbin-Watson calculator for exactly this reason. Enter your DW statistic, sample size, and number of predictors, and it gives you the correct bounds and a plain-language verdict instead of a table you have to decode yourself. If you need diagnostics beyond this one test, you can also browse the full calculator hub for every regression check I have built.

Running and Reading the Durbin-Watson Test by Software

The steps differ slightly depending on your software.

1. SPSS: Analyze, then Regression, then Linear, then Statistics, and tick Durbin-Watson. It appears in the Model Summary table.
2. Stata: Run your regression, then type estat dwatson. Stata prints the statistic directly.
3. EViews: The DW statistic appears automatically at the bottom of every regression output table.
4. Excel: There is no built-in function. Calculate residuals manually, then divide the sum of squared differences between consecutive residuals by the sum of squared residuals.

If you are working in Stata for your dissertation and want someone to check your full model output rather than just this one statistic, my Stata tutoring sessions cover exactly this kind of diagnostic review.

Regression Assumptions: Where Durbin-Watson Fits In

Durbin-Watson checks only one of several regression assumptions. It tests independence of errors, but it says nothing about multicollinearity or non-constant variance. I have reviewed too many chapters where a student proudly reported a perfect DW value while a VIF above 10 sat two tables below it, completely unaddressed.

Treat the full regression assumptions checklist as one package: independence through Durbin-Watson, constant variance through Breusch-Pagan, and low multicollinearity through VIF. Passing one check and skipping the rest does not make your model sound.

Autocorrelation Is Not the Same as Heteroscedasticity

Students often mix these two up. Autocorrelation is about errors correlating with each other in sequence, which Durbin-Watson tests. Heteroscedasticity is about errors having unequal variance across observations, which is a completely different problem tested by the Breusch-Pagan test. A model can pass one and fail the other, so check both.

When Durbin-Watson Doesn’t Apply

This is the mistake I catch most often in dissertation reviews.

Panel or Cross-Sectional Data

Panel data means multiple subjects, such as firms or countries, observed repeatedly over time. The standard Durbin-Watson test is not built for that structure. Use the Wooldridge test for autocorrelation instead, which checks the same independence assumption but accounts for the panel structure properly. Most online explainers built around single time series examples skip this warning entirely, which is exactly the gap that gets panel-data dissertations sent back.

Models With a Lagged Dependent Variable

If your regression includes a lagged value of your outcome variable as a predictor, Durbin-Watson is biased toward 2 and will understate autocorrelation. Use Durbin’s h-statistic instead, a version of the test built specifically to correct for that bias.

Higher-Order Autocorrelation: The Breusch-Godfrey Test

Durbin-Watson only ever tests for correlation between consecutive residuals, known as first-order or lag-one autocorrelation. If you suspect a pattern that repeats every few periods rather than every single one, such as a quarterly cycle, you need the Breusch-Godfrey test, which checks multiple lags at once. A lot of pages mention Breusch-Godfrey as a footnote without explaining when to actually reach for it, which defeats the point of naming it at all.

What to Do If You Detect Autocorrelation

Detecting the problem is only half the job. Here is what I actually recommend to students, in the order I reach for each option.

1. Add a lagged version of the dependent or an independent variable to capture the missing time dynamics
2. Switch to Newey-West standard errors, a correction that keeps your coefficients unchanged but recalculates the standard errors so your significance tests stay valid despite the autocorrelation
3. Apply the Cochrane-Orcutt or Prais-Winsten correction, two closely related methods that transform the data to remove the autocorrelation before re-estimating the model, most useful when the correlation is strong and persistent
4. Re-check your model specification, since autocorrelation is often a symptom of a missing variable rather than a standalone problem

How to Report Durbin-Watson in Your Dissertation

Sample Reporting Sentence

Examiners want the number, the comparison, and the conclusion in one sentence. A version I give students often looks like this:

“The Durbin-Watson statistic was 1.84 (n = 120, k = 3), which falls above the upper bound dU of 1.75 at the 5% significance level, indicating no evidence of positive autocorrelation in the residuals.”

That single sentence does more for your credibility than three paragraphs of vague explanation.

A Quick Case Study

Here is a pattern I see often enough with dissertation students that it is worth walking through as a typical example, not a one-off. A student studying the effect of digital ad spend on weekly online sales for a retail brand ran her regression on monthly data spanning three years. Her Durbin-Watson value came out at 1.12, a strong signal of positive autocorrelation, most likely from an unmeasured seasonal pattern in online shopping behaviour.

We added a one-period lag of sales as a predictor and reran the model. The DW value moved to 1.89, comfortably inside the acceptable range. The coefficient for ad spend, which had looked significant before the fix, dropped in significance once the correction was applied. That single check changed her entire discussion chapter.

Frequently Asked Questions

---

If your dissertation methodology chapter needs a full regression assumptions review rather than just this one test, my statistics tutoring sessions are built for exactly that.

Author: Siddharth Gupta is an independent statistics researcher and dissertation analytics consultant with 12 years of experience helping graduate students across the US, UK, and India get through their regression diagnostics chapters, working across Stata, SPSS, R, and EViews. Connect on LinkedIn.