In Empirical Research, How Do You Identify and Diagnose Autocorrelation? What Impact Does It Have on OLS? How Should You Adjust?

I. Opening: Your t-statistics Look Great—But Your Residuals Are "Colluding"

You ran a time series regression—using annual macro data from 1980–2020 to study "the effect of money supply on inflation." Your model is:

$\text{Inflation}_t = \beta_0 + \beta_1 \cdot \text{M2Growth}_t + \beta_2 \cdot \text{GDPGrowth}_t + \varepsilon_t$

You obtained significant coefficients, sensible signs, and reasonable economic interpretations. The abstract is already written—"Money growth has a significant positive effect on inflation: a 1 percentage point increase in M2 growth raises the inflation rate by 0.34 percentage points."

Then your advisor asks you to produce a plot—a time series plot of the residuals. You casually generate:

reg inflation m2growth gdpgrowth
predict e, residual
twoway (line e year), yline(0)

You notice something is off with the residuals—positive residuals this year tend to be followed by positive residuals; negative residuals this year tend to be followed by negative residuals. They do not look like "random noise"—they look like a group of friends walking together, positives sticking with positives, negatives sticking with negatives.

Your heart sinks—your t-statistics may have been inflated by underestimated standard errors.

Core message: Autocorrelation (also called serial correlation) means that the error terms are not independent across time (or space, or any ordered dimension)—today's error is correlated with yesterday's error. In the classical OLS assumptions, $\text{Cov}(\varepsilon_t, \varepsilon_{t-k}) = 0$ for any k ≠ 0. If this assumption is violated—if the errors "collude" over time—the OLS coefficient estimates remain unbiased and consistent (provided strict exogeneity holds), but OLS is no longer BLUE (best linear unbiased estimator): the OLS formula for standard errors becomes biased (typically downward—standard errors are underestimated), t-statistics are inflated, and significance tests are no longer reliable. There are three routes to identifying autocorrelation—first use your eyes (residual time series plots and autocorrelation plots), then use formal tests (the Durbin-Watson test for AR(1), the Breusch-Godfrey LM test for higher orders, and the Ljung-Box Q test for general forms). There are also three layers of solutions: if you only care about "correct inference"—Newey-West HAC standard errors (heteroskedasticity and autocorrelation consistent) are the standard answer; if you also care about "estimation efficiency"—feasible generalized least squares FGLS (Cochrane-Orcutt or Prais-Winsten) can extract additional information from the error structure, producing narrower confidence intervals; if you suspect the root cause of autocorrelation is model misspecification—adding lagged dependent variables, adding omitted trend variables, or rethinking the dynamic structure of your model—this is the fundamental solution.

II. What Exactly Is Autocorrelation?—Not Correlation in Y, but "Collusion" Among Errors

2.1 Autocorrelation ≠ Autocorrelation in Y

Time series variables $Y_t$ are naturally autocorrelated—this year's GDP is highly correlated with last year's GDP. This is not a problem; it is a feature of the data itself. The task of a regression model is to use X to explain the autocorrelation in Y—if your model is correctly specified, X should capture the persistence of Y over time, and the remaining errors should be "clean" and serially uncorrelated.

The problem of autocorrelation is that X does not fully explain away the persistence in Y—the remaining "persistence" leaks into the error term, creating correlation among the error terms.

2.2 Formal Definition

For the time series regression model $Y_t = \beta_0 + \beta_1 X_t + \varepsilon_t$, first-order autocorrelation (AR(1)) in the error term is:

$\varepsilon_t = \rho \cdot \varepsilon_{t-1} + u_t, \quad u_t \sim \text{WN}(0, \sigma^2_u), \quad |\rho| < 1$

• If $\rho > 0$: positive autocorrelation—today's positive error tends to follow yesterday's positive error ("inertia"—residuals "stick together").
• If $\rho < 0$: negative autocorrelation—today's positive error tends to follow yesterday's negative error ("overcorrection"—residuals oscillate back and forth, rare but possible in differenced data).
• If $\rho = 0$: no autocorrelation—errors are independent.

This can be generalized to p-th order autocorrelation (AR(p)):

$\varepsilon_t = \rho_1 \varepsilon_{t-1} + \rho_2 \varepsilon_{t-2} + \dots + \rho_p \varepsilon_{t-p} + u_t$

2.3 Core Intuition—"Information Is Released Gradually in the Error Term"

The classical OLS assumption (spherical errors) holds that each $\varepsilon_t$ is entirely new, independent, and unrelated to errors from any other period—all past information has been fully absorbed by this point. The existence of autocorrelation means that information is not absorbed all at once—past shocks "spill over" into the future. Like a spring that, when pressed down, does not immediately return to its original position but oscillates several times before settling at equilibrium—your error term still carries the "inertia" of the past several periods.

III. When Does Autocorrelation Arise?—Four Classic Generating Mechanisms

3.1 Cause One: Inertia in Economic Variables—The Most Common Source

Most macroeconomic time series have natural persistence—GDP, consumption, investment, price levels—their values today are highly correlated with their values yesterday. If your model does not fully explain this persistence—for example, you only use "interest rates" to explain "investment," but investment decisions have strong inertia (last year's investment plans affect this year's investment execution)—the omitted inertia enters the error term, making $\varepsilon_t$ and $\varepsilon_{t-1}$ positively correlated.

Symptoms: The residual time series plot shows slow, continuous fluctuations—positive stretches followed by positive stretches, negative stretches followed by negative stretches.

3.2 Cause Two: Omitting Variables with Autocorrelated Structure

If your model omits a variable—and that omitted variable is itself autocorrelated over time—its autocorrelation will manifest through the error term. For example, you study "the effect of crime rates on housing prices" but do not control for "community income levels." Community income is a slowly changing, highly autocorrelated variable—its persistence will flow into the error term, producing positive autocorrelation.

This is the most important source of autocorrelation—because it means autocorrelation is not a "technical problem" but a signal of "model misspecification." Before considering Newey-West adjustments, you should first ask yourself: Have I omitted a variable that should be included and that is persistent over time?

3.3 Cause Three: Incorrect Dynamic Specification—Missing Lagged Terms

The true data-generating process may include lagged effects:

$Y_t = \beta_0 + \beta_1 X_t + \beta_2 X_{t-1} + \beta_3 Y_{t-1} + \varepsilon_t$

But you only estimated a static model:

$Y_t = \beta_0 + \beta_1 X_t + \varepsilon_t$

The effects of $X_{t-1}$ and $Y_{t-1}$ are omitted—but each is correlated with time—their omission creates systematic temporal patterns in the errors. This type of autocorrelation is telling you: your model is too "narrow"—it only looks at the current period, but causal relationships in the real world have temporal depth.

3.4 Cause Four: Data Operations—Interpolation, Smoothing, and Averaging

• Interpolation: The raw data are census data collected every five years, and you use linear interpolation to fill in the intervening four years—the filled values are weighted averages of surrounding observations, inevitably creating artificial smoothness and autocorrelation.
• Moving Average Smoothing: You apply a three-period moving average to the data to "remove noise"—in effect, you are implanting an MA(2) structure into the error term.
• Overlapping Data: You use "the average GDP growth rate over the next five years" as the dependent variable—adjacent observations share four years of data, so the errors are naturally correlated.

These operations "compile" autocorrelation into your error structure at the moment the data are generated. If you have done any of the above, you hardly need to test—you already know autocorrelation exists.

IV. How to Identify and Diagnose Autocorrelation?—With Your Eyes, With Formulas, With Tests

4.1 Step One: Use Your Eyes—Residual Time Series Plot

Plot the residuals against time. This is the most intuitive step and should never be skipped.

reg y x1 x2
predict e, residual
twoway (line e time), yline(0)

What to look for:

• Do the residuals bounce "randomly" above and below the zero line? (Good—no autocorrelation)
• Do the residuals show continuous "same-sign stretches"—a positive stretch, then a negative stretch, each spanning several periods? (Bad—positive autocorrelation)
• Do the residuals alternate signs frequently—positive, negative, positive, negative, positive, negative? (Bad—negative autocorrelation, possibly in differenced data)

Limitation of the residual time series plot: When autocorrelation is not so "obvious" (e.g., $\rho$ = 0.3), the naked eye struggles to judge. You need formal tests to supplement.

4.2 Step Two: Use Formulas—Autocorrelation Plot (ACF)

The autocorrelation function (ACF) displays the correlation coefficients between the residuals and their own lagged values:

$\hat{\rho}_k = \text{Corr}(\hat{\varepsilon}_t, \hat{\varepsilon}_{t-k})$

predict e, residual
ac e, lags(12)          // ACF plot, 12 lags

What to look for:

• If the ACF is high at k = 1 (e.g., > 0.5 or beyond the confidence band) and then decays slowly—a classic AR(1) positive autocorrelation pattern.
• If only k = 1 or k = 1, 2 are significant and then quickly drop to near zero—low-order MA or AR.
• If the ACF at all k values falls within the confidence band—no significant autocorrelation.

The ACF complements the Durbin-Watson test—DW only tests AR(1), while the ACF plot allows you to see the full picture of autocorrelation at all orders.

4.3 Step Three: Formal Tests

4.3.1 Durbin-Watson Test—The Most Classic, But Only for AR(1)

The construction of the DW statistic:

$\text{DW} = \frac{\sum_{t=2}^T (\hat{\varepsilon}_t - \hat{\varepsilon}_{t-1})^2}{\sum_{t=1}^T \hat{\varepsilon}_t^2} \approx 2(1 - \hat{\rho})$

• DW ≈ 2 → no autocorrelation ($\hat{\rho} \approx 0$).
• DW < 2 → positive autocorrelation ($\hat{\rho} > 0$). The closer DW is to 0, the stronger the positive autocorrelation.
• DW > 2 → negative autocorrelation ($\hat{\rho} < 0$). The closer DW is to 4, the stronger the negative autocorrelation.

reg y x1 x2
estat dwatson    // reports DW directly after regression

Three important limitations of the DW test (also the most common pitfalls for beginners):

1. DW can only test first-order autocorrelation AR(1)—if your errors follow AR(2) or MA(1), DW may completely fail to detect it.
2. DW is biased in models containing lagged dependent variables—in regressions with $Y_{t-1}$ as a regressor, DW tends to be close to 2 (no autocorrelation)—even when the true error autocorrelation may be severe. In these models, use Durbin's h test or the Breusch-Godfrey LM test.
3. DW has an "inconclusive region"—the distribution of the DW statistic depends on the X matrix, so DW critical values have an "uncertainty interval" ($d_L$ to $d_U$). If your DW falls in this interval, you cannot make a definitive judgment.

DW rule of thumb (for reference only, not a substitute for formal judgment): For sample sizes T ≈ 30–50, DW < 1.5 or > 2.5 should raise concern. For large T, DW < 1.8 or > 2.2 warrants caution.

4.3.2 Breusch-Godfrey LM Test—More General and More Recommended

The BG test overcomes all three limitations of DW:

• It can test autocorrelation of any order (AR(p) or MA(p)).
• It remains valid in models with lagged dependent variables.
• It does not depend on the X matrix and provides a clear p-value (no "inconclusive region").

Principle: Regress the residuals $\hat{\varepsilon}_t$ on the original regressors X and p lags of the residuals $\hat{\varepsilon}_{t-1}, \dots, \hat{\varepsilon}_{t-p}$ in an auxiliary regression. Test the null hypothesis $H_0$: all coefficients on the lagged residuals are zero (no autocorrelation up to order p).

reg y x1 x2
estat bgodfrey, lags(1)    // test AR(1)
estat bgodfrey, lags(4)    // test up to AR(4)

• p > 0.05 → cannot reject the null of "no autocorrelation."
• p < 0.05 → reject "no autocorrelation"; significant autocorrelation exists.

Recommended lag orders in practice:

• Annual data: lags(1) or lags(2)
• Quarterly data: lags(4)—seasonality may make errors four periods apart correlated
• Monthly data: lags(12)—same seasonal logic

4.3.3 Ljung-Box Q Test—A Standard Tool in Time Series Modeling

The null hypothesis of the Ljung-Box Q test: $\rho_1 = \rho_2 = \dots = \rho_k = 0$ (the first k autocorrelations are all zero). Unlike DW, it is not limited to AR(1), and it provides a joint test p-value across multiple lag orders.

predict e, residual
wntestq e, lags(8)    // Q test, testing whether the first 8 autocorrelations are jointly zero

Choosing between BG and Q tests:

• BG is better suited for regression diagnostics (its auxiliary regression directly includes the original regressors).
• Q is better suited for residual diagnostics in pure time series modeling (white noise tests for ARIMA model residuals).
• In a regression context, BG is recommended as the primary tool—it is also robust in models with lagged dependent variables.

4.4 Comprehensive Diagnostic Strategy—A Recommended Path

After running the regression:
│
├── 1. Plot the residual time series ← A first step that is never wrong, 30 seconds, builds intuition
│
├── 2. Plot the ACF (12 lags) ← Understand the order and pattern of autocorrelation
│
├── 3. Formal tests:
│   ├── If the model has no lagged Y → DW test (traditional reporting requirement) + BG test (order p, depending on data frequency)
│   └── If the model has lagged Y → BG test (do NOT use DW!)
│
└── 4. If autocorrelation is found → first ask "why"—is a variable omitted? Is dynamic structure omitted? Or was it introduced by data operations?
    ├── If the root cause is model specification → fix the model (add lagged terms, add omitted variables)
    └── If the root cause cannot be eliminated or is uncertain → report Newey-West HAC standard errors

V. What Impact Does Autocorrelation Have on OLS?—Coefficients Are Fine, But Your Significance May Be Inflated

5.1 OLS Coefficient Estimates: Still Unbiased and Consistent (Under Strict Exogeneity)

This is a blind spot for many. Under error autocorrelation, if the strict exogeneity condition $\mathbb{E}[\varepsilon_t \mid X_1, \dots, X_T] = 0$ still holds—the OLS β̂ remains unbiased and consistent. Autocorrelation does not cause β̂ to deviate from the true β. Your coefficient signs and magnitudes are correct in expectation.

But note that "strict exogeneity" is a stronger condition than "contemporaneous exogeneity"—it requires that $\varepsilon_t$ is uncorrelated not only with contemporaneous X but also with past and future X. If your model includes a lagged dependent variable $Y_{t-1}$, this condition is automatically violated—the combination of autocorrelation and a lagged dependent variable makes OLS not only inefficient but also inconsistent. This is the core reason dynamic panels use GMM rather than OLS.

5.2 Standard Errors: Underestimated—This Is the Fatal Damage

The OLS variance formula under the spherical error assumption:

$\text{Var}_{\text{OLS}}(\hat{\beta}) = \sigma^2 (X'X)^{-1}$

But when autocorrelation exists, the true variance should be:

$\text{Var}_{\text{true}}(\hat{\beta}) = (X'X)^{-1} X' \Omega X (X'X)^{-1}$

where $\Omega$ is no longer $\sigma^2 I$—its off-diagonal elements are nonzero ($\Omega_{t,s} = \text{Cov}(\varepsilon_t, \varepsilon_s) \neq 0$).

When positive autocorrelation exists—the most common case—the OLS formula typically underestimates the true sampling variability of the standard errors. Your t-statistic = β̂ / SE(β̂)—the denominator is underestimated, the t-statistic is inflated, and insignificant results may appear significant.

The magnitude of underestimation depends on:

• The size of $\rho$ (the stronger the autocorrelation, the more severe the underestimation).
• The persistence of X itself (the more persistent X is, the more severe the underestimation—because autocorrelation in X and autocorrelation in ε "resonate").

5.3 OLS Is No Longer BLUE—You Can Do Better

The existence of autocorrelation means that past errors contain information useful for predicting the current error. If you know $\varepsilon_t$ and $\varepsilon_{t-1}$ are positively correlated, and you observe $\hat{\varepsilon}_{t-1}$ is positive—you know $\varepsilon_t$ is likely positive too. OLS wastes this information—while generalized least squares (GLS) exploits it, producing smaller standard errors than OLS.

So the problem of autocorrelation is essentially two-layered:

1. Inference layer: The OLS standard error formula is wrong—your significance tests are unreliable (fix with Newey-West SE).
2. Efficiency layer: OLS is not the best linear unbiased estimator—there are more precise estimators than OLS (fix with FGLS).

VI. How Should You Adjust When Autocorrelation Is Present?—Three Layers of Solutions

6.1 Solution One: Newey-West HAC Standard Errors—Most Common and Safest

The heteroskedasticity and autocorrelation consistent (HAC) standard errors proposed by Newey and West (1987) are a generalization of White's robust standard errors to the time series setting. They are robust to both heteroskedasticity and autocorrelation.

Principle: When estimating $\text{Var}(\hat{\beta})$, do not assume errors are independent. Instead, use a "kernel" to impose a decaying weighting structure on the error covariance matrix—the further the lag order, the smaller the weight. By choosing an appropriate lag truncation parameter, you control "how many periods of error correlation are allowed."

* Newey-West HAC standard errors (default lag order)
newey y x1 x2, lag(4)    // manually specify the truncation parameter
 
* Or use after any regression (with ivreg2 or heteroskedasticity commands)
reg y x1 x2
* ivreg2 also supports automatic BW lag selection

Choice of truncation parameter: A common rule of thumb is $m = \lfloor 4 \cdot (T/100)^{2/9} \rfloor$ (Newey-West's automatic selection formula), which is about 3 or 4 for $T$ = 50. Stata's newey command defaults to $m = \lfloor T^{1/3} \rfloor$, which is about 4 or 5 for T = 100.

Advantages of Newey-West:

• Does not change β̂—your coefficient estimates remain unchanged.
• Corrects inference—standard errors become larger (under positive autocorrelation), t-statistics become smaller, confidence intervals become wider, and inference becomes more conservative—more honest.
• Does not require assuming a specific form of autocorrelation (AR(1), AR(2), MA(1), etc.)—nonparametric correction.

Limitations of Newey-West:

• Only corrects inference; does not improve efficiency—β̂ is still OLS (not BLUE).
• With small sample sizes (T < 30), HAC standard errors perform worse in finite samples than parametric methods (FGLS).

When to use Newey-West?

• You are unsure of the exact form of autocorrelation (a common pragmatic choice).
• You only want robust inference and do not seek efficiency gains.
• Your sample size T is sufficiently large (≥ 50).
• Journal reviewers ask you to "do something about autocorrelation"—this is the most readily accepted approach.

6.2 Solution Two: Feasible Generalized Least Squares (FGLS)—Correcting Autocorrelation and Improving Efficiency

If you are willing to assume that autocorrelation has a specific parametric structure (the most common being AR(1): $\varepsilon_t = \rho \varepsilon_{t-1} + u_t$), you can use generalized least squares (GLS).

The core idea of GLS (using AR(1) as an example): Use $\rho$ to transform the data so that the transformed errors become white noise.

$\tilde{Y}_t = Y_t - \rho Y_{t-1}, \quad \tilde{X}_t = X_t - \rho X_{t-1}$

The transformed model is:

$\tilde{Y}_t = \beta_0 (1 - \rho) + \beta_1 \tilde{X}_t + u_t$

where $u_t = \varepsilon_t - \rho \varepsilon_{t-1}$ is white noise—all OLS assumptions are satisfied. Running OLS on the transformed data is GLS.

But $\rho$ is unknown—so in practice we use feasible generalized least squares (FGLS), which first estimates $\hat{\rho}$ and then uses $\hat{\rho}$ to transform the data.

Cochrane-Orcutt iterative method:

1. Run OLS and obtain residuals $\hat{\varepsilon}_t$.
2. Regress $\hat{\varepsilon}_t$ on $\hat{\varepsilon}_{t-1}$ to estimate $\hat{\rho}$.
3. Use $\hat{\rho}$ to transform the data and run OLS on the transformed data.
4. Obtain new residuals from the new regression and re-estimate $\hat{\rho}$.
5. Repeat steps 3–4 until $\hat{\rho}$ converges.

* Cochrane-Orcutt iterative method
prais y x1 x2              // Stata's prais command uses Prais-Winsten (retains the first observation)
prais y x1 x2, ssesearch   // use grid search to optimize ρ

Prais-Winsten vs Cochrane-Orcutt:

• Cochrane-Orcutt differences the t = 1 observation, but $Y_1 - \rho Y_0$ requires $Y_0$, which does not exist—the first observation is dropped. This is wasteful in small T.
• Prais-Winsten retains the first observation—through a special transformation $Y_1 \sqrt{1-\rho^2}$—and is more efficient when T is small. Stata's prais command uses Prais-Winsten by default.

Advantages of FGLS:

• Not only corrects standard errors but also improves estimation efficiency (the variance of $\hat{\beta}_{\text{GLS}}$ ≤ the variance of $\hat{\beta}_{\text{OLS}}$).
• When T is small, finite-sample performance is typically better than HAC standard errors.

Limitations of FGLS:

• Requires correct specification of the form of autocorrelation—if the true autocorrelation is AR(2) but you assume AR(1), FGLS remains inconsistent.
• The estimate of $\rho$ has sampling error—this error propagates into the estimation of β. With large T this is not a problem ($\hat{\rho}$ is consistent), but with small T, the efficiency advantage of FGLS may be offset by estimation error in $\hat{\rho}$.

When to use FGLS?

• T is not very large (30 ≤ T ≤ 100).
• You have reason to believe the autocorrelation structure is AR(1) (very common in annual macro data).
• You care not only about correct inference but also about estimation precision.

6.3 Solution Three: Fix the Model Itself—Addressing Autocorrelation at the Root

Autocorrelation is often a signal of model misspecification—not technical noise that needs to be "tolerated" or "corrected." Before reaching for Newey-West or FGLS, you should first check:

6.3.1 Add omitted lagged terms

If your model is:

$Y_t = \beta_0 + \beta_1 X_t + \varepsilon_t$

and $\varepsilon_t$ exhibits AR(1) positive autocorrelation—this suggests that Y's response to X is not instantaneous but delayed. Consider extending to:

$Y_t = \beta_0 + \beta_1 X_t + \beta_2 X_{t-1} + \varepsilon_t$

or adding a lag of Y itself (partial adjustment model):

$Y_t = \beta_0 + \beta_1 X_t + \beta_2 Y_{t-1} + \varepsilon_t$

Economic intuition for adding $Y_{t-1}$: Current Y is affected not only by current X but also by "where Y was yesterday"—because economic variables have adjustment costs and inertia. If you add $Y_{t-1}$, your model transitions from a static to a dynamic model—closer to the real-world adjustment process.

Note: After adding $Y_{t-1}$, the DW test is no longer valid—use the BG test or Durbin's h test instead.

6.3.2 Add an omitted time trend

If the residuals show a slow upward or downward trend—this suggests you may have omitted a deterministic time trend. Adding a linear time trend t or year dummies (year fixed effects) to the model may eliminate trend-induced autocorrelation in the residuals.

6.3.3 Check your data operations

If you interpolated, smoothed with moving averages, or used overlapping samples—autocorrelation is something you created yourself. Consider returning to the original, unsmoothed data. If smoothing is necessary—then directly acknowledge and model the MA structure of the errors in your model.

6.4 Framework for Choosing Among Solutions

After detecting autocorrelation:
│
├── First ask: Is the model specification correct?
│   ├── Did I omit a lag of Y (dynamic effects)?
│   ├── Did I omit an X that is itself autocorrelated over time?
│   ├── Did I omit a time trend?
│   └── Did I artificially smooth the data?
│
├── If you can eliminate autocorrelation by fixing the model:
│   → Fix the model. This is the most fundamental solution—you will obtain a more correct model, not just "more robust" standard errors.
│
├── If autocorrelation persists after fixing the model (or you are unsure what to add):
│   │
│   ├── T is large (> 50) + you only want to ensure correct inference:
│   │   → Newey-West HAC standard errors (report "NW-adjusted t-statistics")
│   │
│   ├── T is moderate (30–100) + you also want to improve efficiency + autocorrelation appears to be AR(1):
│   │   → FGLS / Prais-Winsten
│   │
│   └── T is small (< 30) + autocorrelation:
│       → Your sample size is too small for any autocorrelation correction. Prioritize fixing the model—or honestly acknowledge that "given the limited sample size, this paper does not formally correct for potential autocorrelation but reports BG test results for the reader's judgment."

VII. Autocorrelation vs Heteroskedasticity—Two Frequently Confused Violations of OLS Assumptions

Both can coexist: Your errors can be both heteroskedastic and autocorrelated—for example, in stock return data, volatility is larger during financial crises (heteroskedasticity), and volatility itself is persistent (today's volatility is correlated with yesterday's—autocorrelation). Newey-West HAC standard errors correct for both simultaneously—it provides comprehensive immunity to heteroskedasticity and autocorrelation.

VIII. Autocorrelation in Panel Data—Not Just a Time Series Problem

Panel data also have a time dimension and therefore may exhibit autocorrelation. But autocorrelation in panels has different "levels":

8.1 Two Levels of Autocorrelation in Panels

• Serial correlation within individuals: $\varepsilon_{it}$ and $\varepsilon_{i,t-1}$ for the same individual i are correlated. This is equivalent to having T time series autocorrelation problems across N individuals. In short T panels, the impact of this autocorrelation is limited; in long T panels, it needs to be addressed.
• Cross-sectional correlation: Contemporaneous errors across different individuals are correlated—$\text{Cov}(\varepsilon_{it}, \varepsilon_{jt}) \neq 0$. For example, "in the year of the financial crisis, the unexplained performance declines of all firms are correlated."

8.2 The Clustering Perspective—Why Autocorrelation in Panels Is Usually Handled with Clustering

The most common adjustment in panel data is not Newey-West (essentially a time series tool) but rather cluster-robust standard errors—clustering at the individual level.

$\text{Cluster}(\hat{\beta}) \text{ by } i: \quad \text{allows } \text{Cov}(\varepsilon_{it}, \varepsilon_{is}) \neq 0 \text{ for any } t \neq s$

Clustered standard errors allow errors for the same individual to be arbitrarily correlated across all time periods—a more flexible structure than AR(1) autocorrelation (no parametric form assumed, no assumption that correlation decays over time).

xtreg y x, fe cluster(id)   // cluster at the individual level
reghdfe y x, absorb(id year) vce(cluster id)  // more modern approach

In micro-panel empirical work (large N, small T), clustered standard errors are almost the only requirement—they simultaneously handle heteroskedasticity and autocorrelation in panels without requiring you to assume the serial correlation structure of the errors. Newey-West is less commonly used in panels—because the typical panel structure has large N and small T, and Newey-West's asymptotic theory performs poorly in such data (it assumes T is large).

IX. Common Misconceptions

9.1 Misconception One: "DW close to 2 → no autocorrelation → I do not need to worry"

DW can only test first-order autocorrelation AR(1). If your error structure is AR(2), MA(1), or seasonal autocorrelation (errors four periods apart are correlated, but errors one period apart are not), DW may be close to 2—yet your errors still exhibit autocorrelation. DW not significant ≠ no autocorrelation—it is only a sentinel for AR(1), not a pass for all autocorrelation. When reporting DW, also run at least one BG test (lags(2) or lags(4)).

9.2 Misconception Two: "Autocorrelation → OLS coefficients are biased → cannot be used"

Under strict exogeneity, OLS coefficients remain unbiased and consistent. What autocorrelation destroys is not unbiasedness but efficiency and the correctness of standard errors. Your coefficients are accurate in expectation—but your confidence intervals and t-tests are wrong in magnitude. This is exactly the same logic as how you treat heteroskedasticity—you do not discard OLS coefficients because of heteroskedasticity; you switch to robust standard errors. The same applies to autocorrelation.

But there is a critical exception to this logic: when the model includes a lagged dependent variable, the combination of autocorrelation and a lagged dependent variable makes OLS inconsistent—because strict exogeneity is violated ($Y_{t-1}$ is correlated with $\varepsilon_{t-1}$, and $\varepsilon_t$ is correlated with $\varepsilon_{t-1}$ → $Y_{t-1}$ is indirectly correlated with $\varepsilon_t$).

9.3 Misconception Three: "I used Newey-West → my standard errors are larger → the autocorrelation problem is solved"

Newey-West solves the inference problem—your t-tests are now more reliable. But if you also want to improve estimation efficiency, you should consider FGLS. Moreover, if the root cause of autocorrelation is model misspecification (omitting $Y_{t-1}$ or a time trend), Newey-West may simply be putting a band-aid on a misspecified model—the correct approach is to first fix the model specification, then use HAC standard errors as insurance.

9.4 Misconception Four: "Panel data do not need to worry about autocorrelation—fixed effects handle it automatically"

The within transformation in fixed effects models does remove individual-specific, time-invariant components—but it cannot eliminate autocorrelation in the error term. In fact, the within transformation may create negative autocorrelation in the errors—because demeaning creates correlation among errors of different individuals in the same period (if N is not large) and also creates negative correlation among errors of the same individual across periods (because each observation is subtracted from the individual mean).

This is why clustered standard errors are almost mandatory in panel empirical work—clustered standard errors do not assume errors are independent over time; they allow errors within each individual to be arbitrarily correlated. Fixed effects + clustered standard errors is the standard configuration in micro-panel empirical work—both are indispensable.

9.5 Misconception Five: "My data are cross-sectional—autocorrelation has nothing to do with me"

Autocorrelation does not only appear in the time dimension. Spatial autocorrelation—correlation among errors of neighboring regions, neighboring firms, or adjacent observations—is widespread in regional economics, urban economics, and economic geography. If your data have a "proximity" structure (geographic distance, social networks, supply chain linkages), you should consider the possibility of spatial autocorrelation. Moran's I test is the standard diagnostic tool for spatial autocorrelation.

X. Summary

Eight core takeaways about autocorrelation:

1. **Autocorrelation ≠