数据范儿_范老师-AI时代科研统计新范式-计量科研AI导航

I. Opening: Two Unrelated Variables—an R² = 0.95 and a t-statistic = 15.3

Granger and Newbold conducted a famous simulation experiment in 1974. They generated two completely independent random walk processes:

$Y_t = Y_{t-1} + u_t, \quad X_t = X_{t-1} + v_t$

where $u_t$ and $v_t$ are independent standard normal random numbers. There is no relationship whatsoever between Y and X—one is purely randomly generated, and so is the other. The data-generating process designed no connection between them.

They then regressed Y on X. By rights, β̂ should be approximately 0, R² should be approximately 0, and the t-statistic should be insignificant.

But the actual result was: out of 100 repeated simulations, 78 regressions produced "significant" results with p < 0.05. R² frequently exceeded 0.7, and the DW statistic frequently fell below 0.5.

Two completely independent, entirely unrelated variables—merely because they both "persist over time" (today's value is highly correlated with yesterday's)—were judged by OLS to be "significantly correlated." This is spurious regression.

Core message: Spurious regression occurs when two (or more) nonstationary time series are placed directly into an OLS regression—because each has its own "drift" that accumulates over time (random walk or trend), they can exhibit seemingly highly significant, but in reality entirely spurious, correlations. The classic signature is "three highs and two lows"—R² is very high, t-statistics are very high, F-statistics are very high, but the DW statistic is very low, and the coefficients typically lack a plausible magnitude. The consequence of spurious regression is that you mistake two completely unrelated variables for being "significantly correlated"—your paper's conclusion rests on a statistical illusion. The core remedy is: test whether variables are stationary before regression (unit root tests) → if not stationary, difference them to achieve stationarity, or test whether they are cointegrated (a long-run equilibrium relationship).

II. What Exactly Is Spurious Regression?—It's Not "You Failed to Control for Variables," It's "The Data Are Deceiving You"

2.1 Spurious Regression ≠ Omitted Variable Bias

Omitted variable bias means you left out a variable that should have been included—this is a model specification problem. Spurious regression means you included all the right variables—but because of the time series properties of these variables (nonstationarity), OLS's standard inference becomes unreliable.

In spurious regression, even if you omit no variables, even if Cov(X, ε) = 0 holds, even if the sample size tends to infinity—your t-test will still incorrectly reject the null hypothesis that "β = 0." This is not a small-sample problem—it is also wrong in an asymptotic sense.

2.2 Why Do Nonstationary Series Produce Spurious Regression?—A Simplified Intuition

Consider two independent random walks:

$Y_t = Y_{t-1} + u_t, \quad X_t = X_{t-1} + v_t$

The value of Y at time t is the accumulation of all past shocks $u_1, u_2, \dots, u_t$ . X is likewise the accumulation of its own past shocks. Each of these two accumulation processes drifts randomly through time.

If Y has accumulated several positive $u$ 's over the past few periods, Y will exhibit an "upward trend"—even though no deterministic trend is driving it at all.
If X happens to have also accumulated several positive $v$ 's, X will also exhibit an upward trend.

Both are rising—OLS sees two rising series and assumes they are positively correlated. But Y's rise comes from the accumulation of $u$ , and X's rise comes from the accumulation of $v$ —and these two sets of shocks are completely independent. This is the essence of spurious regression: OLS misreads the "individual trends" of two independent accumulation processes as "correlation with each other."

2.3 The Formal Reason for Spurious Regression—Variance Diverges Over Time

In nonstationary series, the variance of the variable diverges as time t grows ( $\text{Var}(Y_t) \to \infty$ as $t \to \infty$ ). This means OLS's asymptotic theory no longer applies—the standard laws of large numbers and central limit theorems assume finite variance. When variance is infinite, the t-statistic no longer converges to a t-distribution—it diverges to infinity, meaning that even when β = 0, you are increasingly likely to reject H₀.

III. Under What Conditions Does Spurious Regression Easily Occur?

3.1 Case One: Both Are Random Walks (I(1) Processes)—The Most Classic Scenario

Any series that can be written as $X_t = X_{t-1} + \varepsilon_t$ is a random walk—it is nonstationary (specifically, I(1)—integrated of order one).

Stock prices, exchange rates, GDP levels, CPI levels—these are typically I(1) processes.
When two I(1) variables are placed in the same regression, even if they are completely independent, there is a high probability of producing spurious significance.

This is the most common and most dangerous scenario for spurious regression—because many variables in economics are naturally I(1).

3.2 Case Two: Both Contain Deterministic Time Trends

Even if variables are stationary (I(0)), if they both contain a deterministic time trend (such as $X_t = \alpha + \delta t + \varepsilon_t$ ), the correlation between the two trends can produce spurious regression.

Solution: Include a linear time trend t in the regression—once the trend is partialled out, only the correlation between the two stationary residuals is examined. This is much easier to handle than the I(1) case.

3.3 Case Three: One I(1) + One I(0)—Mixed Nonstationarity

Regressing an I(1) Y on an I(0) X is also invalid—a regression between one nonstationary variable and one stationary variable does not have standard asymptotic properties. However, this combination is less common—typically if Y is I(1), researchers will naturally focus on X variables that are also I(1).

3.4 Case Four: Long Panels (Large T) Macro Panel Data

In micro panels (large N, small T), T is typically only 3–10 periods—even if variables are nonstationary, T is too short for nonstationarity to cause severe distortion to inference (unit root tests also lack power in short T). Spurious regression is not a primary concern in micro panels.

But in macro panels (such as cross-country panels with T = 50–60 years), T is sufficiently long that nonstationarity problems surface. In this case, traditional panel fixed effects estimation may suffer from spurious regression—you need to conduct panel unit root tests (such as the IPS test, LLC test) on the variables for each country before regression.

IV. What Signs Appear After Spurious Regression Occurs?—The Classic Fingerprint of "Three Highs and Two Lows"

4.1 R² Is Abnormally High—But DW Is Abnormally Low

Granger and Newbold found that the typical fingerprint of spurious regression is:

R² is very high (often > 0.7, sometimes even > 0.9)—because the individual time persistence of the two series is mistaken for correlation between them.
The DW (Durbin-Watson) statistic is very low (often < 0.5, sometimes even < 0.2)—because the residuals are also nonstationary (if two I(1) series are not cointegrated, their linear combination is still I(1)). The strong persistence of residuals (today's residual is highly correlated with yesterday's) drives DW → 0.

If your regression has R² > DW (i.e., $R^2$ is much larger than the DW statistic), this is a classic alarm for spurious regression.

Granger and Newbold proposed a rule of thumb: If $R^2 > DW$ , you should seriously suspect the possibility of spurious regression.

4.2 t-Statistics and F-Statistics Are Abnormally High—"Looks" Too Significant

For two completely independent variables, the t-statistic should follow a t-distribution. But in spurious regression, the probability that the absolute value of the t-statistic greatly exceeds 2 is far higher than 5%—it diverges to infinity. So you will almost always obtain "highly significant" coefficients—but this "significance" is spurious.

4.3 The Magnitude and Sign of Coefficients Lack Economic Plausibility

Coefficients from spurious regression may:

Have implausibly large magnitudes ("a 1 yuan increase in GDP raises the stock index by 500 points")
Have signs that contradict any economic theory
Reverse signs across different subsamples (positive correlation in the first 20 years, negative in the last 20—the "drift direction" of the series happened to change)

An "implausible" coefficient is a soft but important signal—if the magnitude or sign of β̂ makes you frown, don't rush to write "this is a new finding of this paper"—first run a unit root test.

V. How to Avoid Spurious Regression?—A Complete Path from Testing to Treatment

5.1 Step One: Test the Stationarity of Variables—Unit Root Tests

Before regression, conduct unit root tests on each time series variable. The most commonly used test is the ADF test (Augmented Dickey-Fuller Test).

H₀: A unit root exists (i.e., the variable is a nonstationary I(1) process). H₁: The variable is stationary.

dfuller y, lags(4)       // ADF test
dfuller y, trend lags(4) // ADF with deterministic trend

p > 0.05 → cannot reject unit root → the variable may be I(1) → needs careful handling.
p < 0.05 → reject unit root → the variable is stationary (I(0)) → lower risk of spurious regression.

Limitations of the ADF test: Low power, especially when T is small or when the series is close to a unit root but slightly stationary—it can easily fail to reject H₀ (high Type II error rate). In practice, if the ADF test shows "cannot reject unit root" and your T is not sufficiently large—treating the variable as I(1) and using it in differenced form is generally the safer strategy.

Panel unit root tests (for macro panels):

* IPS test (allows different autoregressive coefficients across individuals)
ipshin y, lags(2)
 
* LLC test (assumes all individuals share the same autoregressive coefficient)
levinlin y, lags(2)

5.2 Step Two: If All Variables Are I(1) → Two Paths

Path A (simple and safe): Take the first difference of the variables—transforming I(1) variables into I(0) variables.

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

After differencing, $\Delta Y_t$ and $\Delta X_t$ are both stationary (assuming the original series are I(1)). The risk of spurious regression is eliminated. Cost: The interpretation of coefficients changes—β₁ is now the effect of the change in X on the change in Y—not a long-run relationship between levels, but a short-run relationship between differences. If what you care about happens to be short-run dynamics—this is exactly what you need. But if you care about the long-run equilibrium relationship—the differenced model cannot provide it.

Path B (long-run relationship): Test for cointegration—whether there exists a stationary linear combination of the two I(1) variables.

If Y and X are both I(1), but there exists a β such that $Y_t - \beta X_t$ is I(0) (stationary) → Y and X have a cointegrating relationship—that is, although they each drift, the "gap" between them is stationary and will revert to equilibrium in the long run.

Engle-Granger two-step method: First run OLS (even if it might be spurious), then conduct an ADF test on the residuals. If the residuals are stationary → cointegration exists → the original OLS is "superconsistent" (coefficient estimates converge to the true values at a faster rate than usual).
Johansen test: A more general multivariate cointegration test (allowing for multiple cointegrating vectors).

* Engle-Granger
reg y x
predict e, residual
dfuller e, lags(4)    // Test whether residuals are stationary (note: critical values differ from standard ADF!)
 
* Johansen
vecrank y x, lags(4)  // Determine cointegrating rank

If cointegration exists—the original levels regression is not spurious. There is a genuine long-run equilibrium relationship between the two nonstationary series, and OLS can consistently estimate this equilibrium relationship (in fact, OLS is superconsistent in cointegrated systems—it converges faster than the usual $\sqrt{T}$ rate).

If cointegration does not exist—the levels regression is spurious. You must use the differenced model.

5.3 Step Three: Standard Reporting in Empirical Papers

Report ADF test results for all time series variables (including lag order selection and whether a trend term is included).
If variables are I(1) → report cointegration test results (Engle-Granger or Johansen).
If cointegration exists → use levels as the baseline regression ("long-run equilibrium relationship"); report the differenced model (short-run dynamics) in robustness checks.
If cointegration does not exist → use differences as the baseline regression (or report an error correction model ECM).

VI. Spurious Regression and Panel Data—Macro Panels Require Caution, Micro Panels Usually Do Not

6.1 Micro Panels (Large N, Small T): Spurious Regression Is Usually Not a Problem

When T = 3–10, even if variables are I(1), the distortion of nonstationarity to inference is relatively limited in finite samples—unit root tests lack power in short T, and the severity of spurious regression is far less than in long-T settings. In typical micro panel empirical work (N > 1000, T < 10), spurious regression is not the first thing you need to worry about.

6.2 Macro Panels (Large N, Large T): The Risk of Spurious Regression in Long Panels Is Real

When T reaches 30–60 (such as cross-country panels, or Chinese provincial panels from 1978–2020), the problem of spurious regression begins to emerge. In this case, your paper should:

Conduct panel unit root tests on key variables (IPS, LLC)
If variables are I(1) → conduct panel cointegration tests (Pedroni, Westerlund)
Report differenced GMM or system GMM results as robustness checks

* Panel cointegration (Pedroni)
xtcointtest pedroni y x, trend
 
* Panel cointegration (Westerlund, allows heterogeneous panels and cross-sectional dependence)
xtcointtest westerlund y x

VII. Common Misconceptions

7.1 Misconception One: High R² + Significant t = Successful Regression

The core lesson of spurious regression is precisely that high R² and significant t do not mean the regression was successful. In nonstationary data, high R² and high t-values are precisely the typical fingerprint of spurious regression, not evidence of a good model. The rule of thumb "R² > DW" is worth checking every time you run a time series regression.

7.2 Misconception Two: Differencing Solves Everything—Losing Long-Run Information

Differencing solves the spurious regression problem—but at the cost of moving from "relationships between levels" to "relationships between changes." If the theory concerns the long-run equilibrium between X and Y (such as "the long-run ratio of GDP to consumption"), the differenced model cannot answer this question. Before differencing, test for cointegration—perhaps you don't need to difference because you have a genuine long-run relationship that can be estimated.

7.3 Misconception Three: "My Data Are Not Time Series—They're Panel Data, So Spurious Regression Can't Happen"

If the time dimension T in panel data is sufficiently long (macro panels), spurious regression can still occur. Moreover, fixed effects models cannot solve the spurious regression problem—fixed effects only remove individual-specific time-invariant means, but the nonstationarity of variables (the I(1) property) may persist after within transformation.

VIII. Summary

Five core pieces of knowledge about spurious regression:

The essence of spurious regression: Two independent nonstationary processes, each drifting randomly—OLS interprets the coincidence of their individual drifts as correlation with each other. This is not a model specification problem—it is the time series structure of the data deceiving OLS's inference framework.
The scenario where it most easily occurs: Both variables are I(1) processes (random walks) with no cointegrating relationship. GDP, stock prices, exchange rates, consumption levels—these common macroeconomic variables are almost all I(1).
The classic fingerprint—"R² > DW": High R², high t-values, and a low DW statistic are the triple alarm for spurious regression. If your regression's R² is much larger than DW—before getting excited, run a unit root test.
The core path to avoidance: Test stationarity (ADF) → if not stationary → difference (short-run change relationships) or test for cointegration (long-run equilibrium relationships). If cointegration exists → the levels regression is valid (and superconsistent). If cointegration does not exist → the levels regression is spurious and differencing is required.
Micro panels (large N, small T) usually need not worry about spurious regression—macro panels (large N, large T) do. The length of T determines whether nonstationarity has "time" to produce substantive distortion to inference.

One-sentence conclusion:

"Spurious regression is the most insidious trap that time series data sets for you—it gives you a beautiful R², a three-star t-statistic, and a correlation that seems intuitively plausible but simply does not exist in the data-generating process. It's not that your model is misspecified—it's that your variables are each wandering through time and happened to drift in similar directions during your sample period. There is only one way to avoid it: before you run your regression, ask your variables—'Are you stationary, or are you adrift?' If they are all adrift, you must either make them stop (differencing), or prove that although they drift, their drifting is tethered together by an invisible rope (cointegration)."

IX. Presentation Suggestions for Bilibili/WeChat Official Account

Bilibili video: It is recommended to use "two drunk men" as the narrative metaphor running through the entire piece. Opening: two drunk men—one named "GDP" and one named "Stock Index." They each start from the origin and wander randomly without purpose. No one guides them toward the same direction—but by pure coincidence, at midnight (the end of the sample period), the two happen to both arrive at the upper-right corner of the block. A researcher (OLS) standing under a streetlight sees the two men's positions at the endpoint and excitedly announces: "The walking paths of these two men are highly significantly correlated—R² = 0.95!" Narration: "They are each just drunk on their own, but you only see their starting and ending points—you have captured enough length in time for their random drifts to create an illusion of 'correlation' at the endpoint." Act One "The fingerprint of spurious regression": The screen pops up a regression output table—R² = 0.93, t = 12.5, DW = 0.23. The three numbers flash in sequence—R² high (green → turns red), t high (green → turns red), DW extremely low (red). A large headline presses down—"R² > DW—the classic fingerprint of spurious regression." Act Two "Why it happens": Animation of two sets of shocks— $u_t$ and $v_t$ are each generated independently, but after they accumulate into $Y_t$ and $X_t$ , the two happen to exhibit similar directions during the sample period. In the panel— $Y_t$ and $X_t$ are both moving upward, but not because they are related—because their respective shocks have produced similar drifts in accumulation. Narration: "Y's rise comes from the accumulation of its own shocks—X's rise also comes from the accumulation of its own shocks. The two waves of shocks are completely independent, but the accumulated paths, in a finite sample, can easily look correlated." Act Three "How to avoid it": A three-step path—Step One: ADF test (visual: a checkpoint where variables are labeled with a green "stationary I(0)" tag or a red "nonstationary I(1)" tag as they pass through). Step Two: Differencing (visual: I(1) variables enter a "differencing machine"—what comes out are ΔY and ΔX, with labels turning green I(0), annotated "short-run change relationship"). Step Three: Cointegration test (visual: an elastic rope appears between the two I(1) variables—they drift but are tethered together, and the length of the rope (the linear combination) is stationary. An annotation beside it reads "long-run equilibrium relationship—this is not spurious regression").
WeChat Official Account: The "three highs and two lows" fingerprint of spurious regression (high R², high t, high F, low DW, implausible coefficient magnitudes) should be made into a diagnostic card. The decision path of ADF test → differencing vs. cointegration should be made into a flowchart infographic as the core. The three cases of I(0) vs. I(1) vs. cointegration (two horizontal lines vs. two random walks vs. two tethered random walks) should each be paired with a diagram. Granger-Newbold's $R^2 > DW$ rule of thumb should be made into a highlight box. The Engle-Granger two-step method (OLS residuals + ADF) should be made into a step card. The comparison of spurious regression risk between micro panels and macro panels should be made into a clear comparison table.
Recommended titles:
- Main title: 《What Is Spurious Regression?—When Two Completely Unrelated Variables Are Judged by OLS as "Significantly Correlated"》
- Alternative title: 《R² = 0.95, t = 15.3—This Is Not a Good Paper, This Is a Spurious Regression》
- New media title: 《The "Perfect Correlation" of Two Independent Random Walks—How Spurious Regression Deceives Your Significance Tests》
Key quotes:

"Spurious regression is not that your model is wrong—it's that your variables are all adrift in time, and OLS interprets their drifting paths as following each other. Two drunk men each walk their own way, but if you only look at their starting and ending positions—you would think they were traveling together."

"R² > DW—this is the most concise diagnostic rule for spurious regression left by Granger and Newbold. A high R² tells you 'these two variables are highly correlated'; a low DW tells you 'the residuals are still drifting—their relationship is not stable.' Both pieces of contradictory information present at once—this is the fingerprint of spurious regression."

"Differencing makes drifting variables stop—you no longer ask 'the relationship between GDP levels and the stock index,' but rather 'the relationship between changes in GDP and changes in the stock index.' The cost is losing long-run information. Cointegration lets you keep levels—provided you can prove that although both variables drift, their drifting is tethered by an invisible long-run equilibrium rope."

"In time series, you should not trust any regression that has not passed a unit root test—just as you would not trust the driving of a driver you do not know whether they are drunk. Stationary variables are sober—they fluctuate around a mean. Nonstationary variables are drunk—they can drift anywhere, including toward the same direction as another equally drunk variable."