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. Introduction: Endogeneity Strikes, You Have No Good IV at Hand—"Then Use the Lagged Term"

You are studying the effect of FDI on economic growth. You know FDI is endogenous—countries with faster growth also attract more foreign capital, so Cov(FDI, ε) ≠ 0. You need an instrumental variable.

You skim a few papers and find a common practice—using the one-period lag of FDI as an instrumental variable. The reasoning: last year's FDI is highly correlated with this year's FDI (satisfying the relevance condition), and "last year's FDI should not be affected by this year's error shocks to economic growth" (satisfying the exogeneity condition).

You follow suit. The second-stage results of your 2SLS show: the coefficient on FDI is significantly positive. The reviewer glances at your IV strategy and asks one question:

"You are assuming $\text{Cov}(\text{FDI}_{t-1}, \varepsilon_t) = 0$ . But if economic shocks are persistent—this year's positive shock is correlated with last year's positive shock—can you still guarantee this assumption holds?"

You fall silent. Because if the error term exhibits serial correlation— $\varepsilon_t$ and $\varepsilon_{t-1}$ are not independent—then last year's FDI (which is correlated with $\varepsilon_{t-1}$ ) is naturally also correlated with this year's $\varepsilon_t$ . Your instrumental variable is no longer exogenous. Its "exogeneity" rests on an assumption you never tested and that most likely does not hold.

Core message: The validity of using lagged variables as instrumental variables hinges on a key assumption— $\text{Cov}(X_{t-1}, \varepsilon_t) = 0$ , i.e., the lagged X is uncorrelated with the current-period error. This assumption is equivalent to requiring that the error term exhibits no serial correlation. If economic shocks are persistent (as they are in most macro and micro panels), this assumption fails—the lagged term is not a valid instrument. More generally, the validity of lagged variables as IVs is not a "yes or no" question but a continuous spectrum—from "clearly invalid" (strong serial correlation) to "marginally acceptable" (shocks are transitory, X is a slowly adjusting stock variable) to "valid and correct" (within the System GMM framework, exploiting the system moment conditions of both the differenced and level equations).

II. The Basic Logic of Using Lagged Variables as IVs—Why Is It Intuitively Appealing?

2.1 Setup

Your structural equation is:

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

where $X_t$ is endogenous: $\text{Cov}(X_t, \varepsilon_t) \neq 0$ . You need an instrumental variable $Z$ satisfying two conditions:

Relevance: $\text{Cov}(Z, X_t) \neq 0$
Exogeneity: $\text{Cov}(Z, \varepsilon_t) = 0$

You propose $Z = X_{t-1}$ (the one-period lag of X).

Relevance: $X_{t-1}$ and $X_t$ are usually highly correlated—most economic variables (GDP, investment, employment, income) exhibit persistence. This condition is easy to satisfy.

Exogeneity: $X_{t-1}$ is a variable from the "past." It was determined in period $t-1$ —and $\varepsilon_t$ is the shock in period $t$ . From a temporal perspective, variables from the "past" cannot be affected by shocks from the "future." Logically, this sounds reasonable.

2.2 The Hidden Premise of This Logic

$X_{t-1}$ is not affected by $\varepsilon_t$ —this is mathematically guaranteed (causation cannot flow backward in time). But IV exogeneity requires more than just " $X_{t-1}$ is not caused by $\varepsilon_t$ "—it requires $\text{Cov}(X_{t-1}, \varepsilon_t) = 0$ . The two are not the same.

Why is there a possibility that $\text{Cov}(X_{t-1}, \varepsilon_t) \neq 0$ ?

Because $X_{t-1}$ is determined in period $t-1$ —it is affected by $\varepsilon_{t-1}$ . If $\varepsilon_{t-1}$ and $\varepsilon_t$ are correlated (i.e., the error term exhibits serial correlation), then $X_{t-1}$ (through $\varepsilon_{t-1}$ ) is indirectly correlated with $\varepsilon_t$ .

$X_{t-1} = \cdots + \gamma \varepsilon_{t-1} + \cdots$ $\varepsilon_t = \rho \varepsilon_{t-1} + \nu_t$

If $\rho \neq 0$ (errors are serially correlated), then $\text{Cov}(X_{t-1}, \varepsilon_t) = \gamma \rho \cdot \text{Var}(\varepsilon_{t-1}) \neq 0$ → the lagged variable is no longer a valid instrumental variable.

III. When Is It Justified?—A Spectrum from "Absolutely Not" to "Broadly Acceptable"

3.1 Case One: Clearly Invalid—Strong Serial Correlation in the Error Term

In most macroeconomic time series, economic shocks are highly persistent. A positive productivity shock this year (εₜ > 0) very likely means last year was also a positive shock (εₜ₋₁ > 0). In this case:

$\text{Cov}(\text{FDI}_{t-1}, \varepsilon_t) > 0$

Using lagged FDI as an IV, you are essentially using a variable positively correlated with εₜ to "replace" another variable also positively correlated with εₜ (FDIₜ). This is not a genuine instrumental variable—you are merely shifting the bias from one variable to another, and you have no independent source of variation to identify the causal effect.

Worse, if X is also highly persistent (i.e., $X_t \approx X_{t-1}$ ), then $X_{t-1}$ and $X_t$ are almost the same thing—you are essentially using X as its own instrument. An instrumental variable should provide variation that is "independent of the endogenous variable and exogenous to the error term"—a variable that is nearly identical to the endogenous variable and contaminated by the same serially correlated error provides no new independent variation.

Judgment criterion: If your regression errors exhibit significant serial correlation in an AR(1) test, the exogeneity of the lagged variable as an IV does not hold.

3.2 Case Two: Marginally Acceptable—Economic Shocks Are Transitory, X Is a Stock Variable

In certain micro settings, the shocks affecting current-period Y may be transitory and non-persistent. For example:

Weather shocks: This year's rainfall only affects this year's agricultural output. Last year's rainfall is likely uncorrelated with this year's rainfall, and last year's FDI investment decisions (made under last year's weather conditions) are uncorrelated with this year's weather shocks. In this case, $\varepsilon_t$ and $\varepsilon_{t-1}$ may be approximately uncorrelated → $\text{Cov}(X_{t-1}, \varepsilon_t) \approx 0$ .
X is a "slowly adjusting stock variable": For example, "firm capital stock." The capital stock is accumulated from past investments; this year's adjustment changes only a small fraction. $K_{t-1}$ and $K_t$ are highly correlated (good instrument), but $K_{t-1}$ is barely affected by $\varepsilon_t$ (because this year's shock εₜ can only change a small amount of this year's investment, with negligible impact on the stock itself).

But even in these cases, "lagged variable as IV" is still not a compelling IV strategy—at best, it serves as a baseline model that is "slightly better than OLS" and can be reported in robustness checks. If you claim this is your only identification strategy, reviewers will typically not be convinced.

3.3 Case Three: Valid Within the Correct Framework—System GMM

In dynamic panel models (models with a lagged dependent variable):

$Y_{it} = \beta_0 + \rho Y_{i,t-1} + \beta_1 X_{it} + u_i + \varepsilon_{it}$

$Y_{i,t-1}$ is naturally endogenous—because it is correlated with the unit fixed effects $u_i$ ( $u_i$ affects both $Y_{i,t-1}$ and $Y_{it}$ ). In this case, using lagged variables as IVs is no longer an ad hoc "backdoor trick"—it is a core component of formally derived and widely accepted system estimation methods.

Arellano-Bond (Difference GMM): For the differenced equation $\Delta Y_{it} = \rho \Delta Y_{i,t-1} + \beta_1 \Delta X_{it} + \Delta \varepsilon_{it}$ , use $Y_{i,t-2}, Y_{i,t-3}, \dots$ (earlier lagged levels) as instruments for $\Delta Y_{i,t-1}$ . These earlier lags are uncorrelated with $\Delta \varepsilon_{it}$ under the assumption that "errors are not serially correlated."

Blundell-Bond (System GMM): On top of the differenced equation, additionally exploit the level equation—using $\Delta Y_{i,t-1}, \Delta Y_{i,t-2}, \dots$ (lagged differences) as instruments for $Y_{i,t-1}$ in the level equation. System GMM is more efficient than Difference GMM—especially when Y is highly persistent (ρ close to 1), where lagged levels are weak instruments and lagged differences provide additional identifying power.

Within the System GMM framework, using lagged variables as IVs is valid and correct—provided you use these moment conditions properly and report the relevant diagnostic tests:

Hansen/Sargan overidentification test: Tests the joint exogeneity of all instruments. p > 0.05 → cannot reject the null hypothesis that the instruments are exogenous.
AR(2) test: Tests whether the differenced error term exhibits second-order serial correlation. p > 0.05 → cannot reject the null of no second-order autocorrelation (if AR(2) is present, $Y_{i,t-2}$ is no longer a valid IV).
Number of instruments: If the number of instruments approaches or exceeds the number of units N → a warning sign of overfitting. Typically, the number of IVs should be less than N.

IV. When Is It Absolutely Not Acceptable?—Three Red Flags

4.1 Red Flag One: You Have Only One Lagged IV and No Diagnostic Tests

Using only $X_{t-1}$ as an IV means the model is exactly identified—you cannot test the exogeneity of this instrument. Overidentification tests (Sargan/Hansen) require "number of instruments > number of endogenous variables"—when exactly identified, you cannot conduct any statistical test of IV exogeneity.

An exactly identified IV strategy that cannot be tested, combined with a serial-uncorrelatedness assumption that most likely fails, combined with an instrument that is nearly identical to the endogenous variable—this is almost circular reasoning. Your IV is essentially "a slightly older version of X"—if the new X is endogenous, why would the old X suddenly be exogenous?

4.2 Red Flag Two: X Is a "Flow Variable" Rather Than a "Stock Variable"

Stock variables (e.g., capital stock, cumulative years of education): Today's value contains a large amount of information accumulated from the past. $X_{t-1}$ and $X_t$ are highly correlated, but $X_{t-1}$ is less contaminated by the current shock εₜ.
Flow variables (e.g., current-year R&D expenditure, current-year advertising spending, current-year FDI inflows): Today's value is correlated with last year's value, but the correlation is far weaker than for stock variables. Moreover, flow variables are themselves the result of current-period decisions— $X_{t-1}$ is equally an "endogenous decision"—just made one period earlier. If decisions in both periods are driven by the same persistent factor (such as "management quality" or "market prospects"), $X_{t-1}$ and εₜ remain correlated.

Flow variable + lagged term as IV = red flag.

4.3 Red Flag Three: Your Panel Has Very Short T, but You Claim to Have Solved Endogeneity Using "Internal IVs"

In short panels (T = 3 or 4), the number of lagged IVs you can use is extremely limited, and the difference between the lagged and current terms is very small. If you claim that using $X_{i,t-1}$ as an IV for $X_{it}$ has "solved endogeneity," the reviewer only needs to ask—"What exactly is the independent exogenous variation in the difference between a variable and its own lag?"—and you will be unable to answer.

V. If You Should Not Use Lagged Variables, What Should You Use?—A Priority Ranking of Alternatives

5.1 First Priority: Search for a Genuine "External" Instrumental Variable

A good instrumental variable should come from "outside the model"—it should not merely be a lagged version of the model's endogenous variable. The source of variation in an external IV is independent of the data-generating process of the endogenous variable:

Institutional/policy shocks: Compulsory schooling law reforms → changes in years of education
Natural/geographic characteristics: Rainfall → agricultural output; distance to port → trade volume
Historical events: Colonial history → contemporary institutional quality
Randomized experiments/lotteries: Vietnam draft lottery → veteran status

Finding an external IV requires thinking at the level of research design, not data manipulation. When planning your study, first ask yourself: "What factor caused some people to receive more X, while being unrelated to all other factors affecting Y?"—rather than "What variable in my data can I use as an IV?"

5.2 Second Priority: Panel Data Methods—Fixed Effects + Quasi-Natural Experiments

If your data are panel data and the endogeneity mainly comes from time-invariant omitted variables (such as ability, geographic location, institutional traditions), fixed effects models can directly eliminate this source of endogeneity—no instrumental variable needed.

If your endogeneity comes from time-varying confounders or reverse causality, and you have some quasi-natural experiment (such as a staggered policy rollout), you can use:

DID + event study: Exploit variation in policy timing and geographic coverage to identify causal effects.
Bartik instruments (shift-share IV): Use the interaction of "initial shares × aggregate growth rates" as an IV—it provides "external" variation within a panel structure.

5.3 Third Priority: If You Must Use Lagged Terms—System GMM with Full Diagnostic Reporting

If you truly have no external IV and your data have a panel structure (large N, moderate T), System GMM may be your best option. But using System GMM is not an excuse to be lazy—you must fully report in your paper:

The p-value of the Hansen/Sargan test
The p-value of the AR(2) test
The number of instruments (and whether it exceeds N)
The Difference-in-Hansen test (testing the validity of the additional moment conditions in System GMM)
As robustness: the coefficient range from OLS (upper bound) and fixed effects (lower bound), showing whether the GMM estimate falls within this reasonable range

VI. A Self-Checklist

Before deciding to "use lagged variables as IVs," ask yourself six questions:

Is my endogenous variable a stock or a flow variable? → Stock: marginally acceptable. Flow: basically not acceptable.
In my data, does the error term exhibit serial correlation? → Test it. If yes → the exogeneity of the lagged IV does not hold.
Am I using only one lagged term as my IV? → Exactly identified, exogeneity cannot be tested. You need at least two IVs to conduct an overidentification test.
Do I have panel data? → If yes, consider System GMM with full diagnostics. If no, a lagged IV is almost never convincing.
Have I searched for an external IV? → If not—go look first. The lagged term is a "last resort"—but it cannot be the "only resort."
Is my T very short (≤3)? → If yes, the difference between the lagged and current terms is too small, and the IV has almost no independent identifying power.

VII. Summary

Five core takeaways on using lagged variables as IVs:

The exogeneity assumption of a lagged variable = no serial correlation in the error term. If shocks are persistent (as they are in most economic data), the lagged variable is no longer exogenous—it is contaminated by the previous period's error, and the previous period's error is correlated with the current period's error.
The problem with lagged variables as IVs is not "relevance" (usually strong enough) but that "exogeneity" cannot be tested. When exactly identified, you cannot conduct any statistical test of IV exogeneity—you can only "assume" it holds.
Stock variables are more suitable than flow variables for using lagged terms as IVs. Stock variables contain substantial information from the past and are less contaminated by current shocks. Flow variables are the result of current-period decisions—the one-period lag is equally an endogenous decision.
System GMM is the correct framework for using lagged IVs under certain conditions. But it requires a panel structure, extensive diagnostic testing, and is highly sensitive to the number of instruments and serial correlation assumptions. It is not a "universal lagged-term processor."
Before searching for "an IV in the data," search for "an IV in the research design." A good instrumental variable comes from institutional, natural, historical, or random variation outside the model—not merely from the previous period of a variable you already have.

One-sentence conclusion:

"Using a lagged variable as an IV essentially says: 'I don't know what exogenous factor causes X to vary—but I know last year's X is similar to this year's X, and it "should" be uncorrelated with this year's error.' The first statement is a fact; the second is an assumption you have never tested and that most likely does not hold. When you have exact identification, cannot test exogeneity, and the error may exhibit serial correlation—you are not using an instrumental variable; you are using an older version of the same endogenous variable, dressed in IV clothing and pretending to be exogenous."

VIII. Presentation Suggestions for Bilibili/WeChat Official Account

Bilibili video: Use "time-travel cheating" as the narrative framework. Opening: two time points—last year (t−1) and this year (t). This year's X (FDI) and this year's Y (GDP) exhibit endogeneity—"Higher GDP also attracts more FDI; you can't tell which causes which." A researcher pulls out a "time machine remote control" and takes out last year's FDI: "I'll use last year's FDI as my instrumental variable—last year's investment is highly correlated with this year's investment, but last year's investment should not be affected by this year's GDP shock!" A green "IV VALID" stamp flashes on screen. But the narrator says: "Wait—what if economic shocks are persistent? Last year's good times and this year's good times are correlated. Last year's FDI contains last year's shock—and last year's shock carries over to this year." The green stamp cracks, revealing a red "Cov(Z, ε) ≠ 0." Then unfold in four acts: Act One "Why It's Tempting"—relevance (a tight scatter plot of last year's FDI vs. this year's FDI) + temporal logic (a timeline, the past cannot be reversed), two green checkmarks. Act Two "The Hidden Assumption"—two Greek letters enter: ρ (persistence of shocks) + γ (sensitivity of X to shocks). If both are nonzero → the exogeneity of the lagged IV is contaminated. Use two knobs to demonstrate—the higher the knobs, the more contaminated the exogeneity. Act Three "The Spectrum"—three scenarios shown in sequence: absolutely not acceptable (macro series, strongly persistent errors, lag = old version of the same thing), marginally OK (weather shocks, transitory and non-persistent errors), correct framework (System GMM equations, large N moderate T, full diagnostic test table). Act Four "What to Do"—a "ladder of IV choice": first rung → external IVs (rainfall, legal institutions, history); second rung → panel methods + quasi-experiments; third rung → System GMM + full reporting; fourth rung (dimmed) → lagged IV with exact identification.
WeChat Official Account: For the temporal logic chain of lagged variables as IVs ( $X_{t-1}$ ← $\varepsilon_{t-1}$ → $\varepsilon_t$ correlation transmission), suggest a causal path diagram. The three red flags (exact identification, flow variables, short panels) should be made into warning cards. The judgment spectrum for the three cases (invalid → marginal → valid) should be made into a horizontal gradient color bar. The priority ladder of IV choice should be made into a vertical flowchart. System GMM's "three-step diagnostic report" (Hansen/AR(2)/number of IVs) should be made into a checklist card. The six questions in the self-checklist should be made into Q&A cards.
Recommended titles:
- Main title: 《Using Lagged Variables as Instrumental Variables—Clever or Self-Deception?》
- Alternative title: 《 $\text{Cov}(X_{t-1}, \varepsilon_t) = 0$ —The One Assumption Lagged IVs Require but Most Easily Violate》
- New media title: 《Your "Lagged-Term Instrumental Variable"—May Just Be an Older Version of the Same Endogenous Variable》
Key quotes:

"Using a lagged variable as an IV is essentially making a copy of the same endogenous variable—just with the date stamp changed to t−1. If shocks are persistent, this old version and the new version are contaminated by the same persistent source—your IV is not exogenous; it is the younger version of the same variable."

"Exact identification + lagged IV + untested serial-uncorrelatedness assumption = you cannot distinguish between 'I have identified a causal effect' and 'I just ran an OLS with a time offset.'"

"Good instrumental variables come from outside the model—rainfall, laws, history, lotteries. Bad instrumental variables come from inside the model—the same variable, just one period earlier. The difference between the two is not whether the coefficients in the regression table look good, but whether you have found a source of variation truly independent of the endogenous process."

"System GMM is the legitimate path to using lagged terms as IVs under the right conditions—but it is not a panacea. It requires you to report the Hansen test, the AR(2) test, the number of instruments—it requires you to honestly demonstrate that your IV strategy withstands scrutiny. If you just throw X_{t-1} into ivreg2 and report the 2SLS results—you are not 'solving endogeneity'; you are 'borrowing endogeneity.'"