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: When you put "education" into your wage regression, is it really exogenous?

Suppose you run the most classic wage equation:

\text{Wage} = \beta_0 + \beta_1 \cdot \text{Education} + \beta_2 \cdot \text{Experience} + \varepsilon

You estimate it with OLS and obtain a return to education of $\hat{\beta}_1 = 0.08$ , with a standard error of 0.005, significant at the one percent level. Everything looks perfect.

But your advisor reads it and asks one question:

"What do you think is inside ε?"

You think about it—ε is the error term, containing all factors that affect wages but that you "did not put into the regression." Ability is certainly in there. High-ability people earn more (ability → wage ↑), and high-ability people typically receive more education (ability → education ↑). This implies:

$\text{Cov}(\text{Education}, \varepsilon) > 0$

In other words, your key explanatory variable "education" is correlated with the error term. The OLS assumption that Cov(X, ε) = 0 is violated. Your $\hat{\beta}_1 = 0.08$ contains, in part, not the causal effect of education but the confounding effect of ability—education is endogenous.

Core message: The distinction between exogenous and endogenous variables boils down to one sentence—whether the explanatory variable is correlated with the error term. Cov(X, ε) = 0 → exogenous; Cov(X, ε) ≠ 0 → endogenous. This is not an inherent property of the variable, but an assumption relative to a specific model and a specific error term. The same variable can be exogenous in Model A and entirely endogenous in Model B.

II. Precise definitions of endogenous and exogenous variables—one condition determines everything

2.1 Exogeneity

In the classical linear regression model, exogeneity is the core condition that makes OLS unbiased:

$\mathbb{E}[\varepsilon_i \mid X_{1i}, X_{2i}, \dots, X_{ki}] = 0$

That is: conditional on the values of all independent variables, the conditional expectation of the error term is zero. This means the error term is uncorrelated with any of the independent variables— $\text{Cov}(X_j, \varepsilon) = 0$ for all j.

An independent variable satisfying this condition is an exogenous explanatory variable.

Exogeneity implies that variation in X is determined "outside the model," unaffected by Y or by any confounding factors omitted in ε. Changes in X are as if driven by an "external hand"—independent of the factors you cannot observe.

2.2 Endogeneity

If an explanatory variable is correlated with the error term— $\text{Cov}(X_j, \varepsilon) \neq 0$ —it is an endogenous explanatory variable.

The literal meaning of endogenous is "generated from within"—the value of Xⱼ is partly influenced or determined by the "inside" of the model (i.e., factors contained in ε or Y itself). When X is endogenous:

OLS is no longer unbiased: $\mathbb{E}[\hat{\beta}_j] \neq \beta_j$
OLS is no longer consistent: even as n → ∞, $\hat{\beta}_j$ does not converge to the true $\beta_j$
Causal interpretation fails: $\hat{\beta}_j$ can no longer be interpreted as "the causal effect of a one-unit change in Xⱼ on Y"

2.3 A table to see the difference at a glance

Dimension	Exogenous explanatory variable	Endogenous explanatory variable
Definition	Cov(X, ε) = 0	Cov(X, ε) ≠ 0
Is OLS unbiased?	✅ Yes	❌ No
Is OLS consistent?	✅ Yes	❌ No (bias does not disappear as n grows)
Can it be interpreted causally?	✅ Yes, after controlling for confounders	❌ No—the coefficient is "contaminated"
"Status" in the model	Determined outside the model	Partly determined by factors inside the model (Y or ε)
Needs to be addressed?	No special treatment needed	Requires identification strategies such as instrumental variables, panel methods, etc.

III. The relationship between the two—not a classification, but an assumption

3.1 Exogenous/endogenous is an "assumption," not an "inherent label" of the variable

This is the easiest thing for beginners to confuse. The variable "education" itself is neither an "exogenous variable" nor an "endogenous variable." It is judged as exogenous or endogenous only within a specific model, relative to a specific error term.

In Wage ~ Education + ε₁, ε₁ contains "ability"—and ability is correlated with education → education is endogenous.
In Wage ~ Education + Ability + ε₂, you have controlled for ability, and ε₂ no longer contains ability. If there are no other factors in ε₂ correlated with Education → education is exogenous.

The same variable "education" switches its "endogenous/exogenous" identity across different models. The difference lies not in the variable itself, but in what you have controlled for and what remains in ε.

3.2 Exogenous and endogenous form a "spectrum," not a "binary opposition"

Strictly speaking, in empirical research, there is a gray zone between "fully exogenous" and "severely endogenous":

Strict exogeneity: Cov(X, ε) = 0, and X is uncorrelated with any functional form of ε.
Contemporaneous exogeneity: Only requires Cov(Xₜ, εₜ) = 0, but allows Xₜ to be correlated with past ε. More common in time series.
Weak endogeneity: Cov(X, ε) is small but nonzero. In many micro-evaluations, this "mild endogeneity" may not completely overturn core conclusions, but it needs to be discussed in robustness checks.
Severe endogeneity: Cov(X, ε) is of a magnitude that cannot be ignored, such that the OLS bias is large enough to change the sign or economic meaning of the coefficient.

Summary of the relationship: Exogenous and endogenous are not two "types" of variables—they are two possible states of the same variable under a specific model specification, relative to a specific error term. One of the core tasks of empirical researchers is to determine where the X they care about lies on this spectrum in the baseline model, and based on that, decide whether identification strategies such as IV, panel methods, or DID are needed.

IV. In empirical analysis, when do you need to worry about endogeneity?

4.1 Signals—if any of the following situations arises, endogeneity is a problem you must confront

Situation 1: You care about the causal effect of X on Y (not just prediction).

If you are doing causal identification—"Does education raise wages?" "Does the minimum wage reduce employment?" "Do tax incentives promote R&D investment?"—then endogeneity is the first threat you must rule out. An endogenous key explanatory variable means what you estimate is not the causal effect, but the causal effect plus confounding bias.

If you are only doing prediction—"Use these characteristics to forecast next year's GDP growth rate"—the importance of endogeneity drops considerably. In pure prediction tasks, as long as predictive accuracy is high and generalization is good, whether the coefficients are biased is not the primary concern.

Situation 2: You are using observational data rather than experimental data.

In randomized controlled trials (RCTs), the treatment variable is randomly assigned → the treatment is independent of any potential outcomes and confounders → the treatment is naturally exogenous. Endogeneity is not a problem in RCTs.

But in observational data—CGSS, CFPS, CHFS, listed company annual reports, macro statistical yearbooks—no variable is randomly assigned. Education, income, policy treatments, regional characteristics… almost every variable you put in a regression can be endogenous in observational data to unobserved confounders. Using observational data with the goal of causal inference → endogeneity is a hurdle you cannot bypass.

Situation 3: The X you care about is the result of a "choice" or "decision."

Individuals choose how much education to obtain, firms choose which province to invest in, governments choose whether to implement a policy—these are all outcomes of self-selection. The agents making choices decide based on some unobservable factors, and these unobservable factors typically also affect Y. In data generated by choice behavior, endogeneity is the default state; exogeneity is the exception that needs to be argued.

Situation 4: The literature has already extensively discussed the endogeneity of the X you care about.

If you study the return to education—classic papers such as Card (1999), Angrist & Krueger (1991) have thoroughly discussed the "endogeneity of education." If you work on this topic without addressing endogeneity, reviewers will almost certainly question you. Classic variables where endogeneity is widely recognized in the literature include:

Education (ability bias)
Income, prices, wages (simultaneity bias)
Policy treatment variables (self-selection bias)
Health measures and medical expenditure (reverse causality)
Firm-level R&D expenditure, advertising spending (omitted variables such as management quality)

Situation 5: Your OLS results differ greatly from instrumental variable or panel method results.

If you first run OLS → coefficient is 0.15. Then you find an instrumental variable and run 2SLS → coefficient becomes 0.06. The difference between the two (0.09) reflects the endogeneity bias in OLS. If you find that OLS and other identification strategies give very different estimates—do not rush to "pick the significant one to report." Instead, confront the existence of endogeneity and its magnitude.

4.2 A decision flowchart

What is your research objective?
├── Predict Y → Endogeneity is not the primary concern
└── Estimate the causal effect of X on Y
    ├── Is the data from a randomized experiment?
    │   └── Yes → X is naturally exogenous; endogeneity is not a problem
    └── Is the data observational?
        └── Yes → Endogeneity must be taken seriously
            ├── Is X the result of choice/decisions? → Almost certainly endogenous
            ├── Does the literature recognize endogeneity in X? → Must be addressed
            └── Do you have omitted confounders? → If correlated with X → endogenous

V. What types of variables are generally exogenous?

The answer is: In causal inference with observational data, truly "naturally exogenous variables" are far rarer than you think. Here are several categories of variables widely accepted as exogenous (or at least arguable as exogenous) in empirical work:

5.1 Variables that are almost always exogenous

(1) Naturally random events—weather, natural disasters, geographic features.

Rainfall, temperature, earthquakes, typhoon paths—these are not determined by any individual's choices or decisions. When studying "the effect of rainfall on agricultural output," rainfall is exogenous. This is also why "weather" is so popular as an instrumental variable in empirical work—its exogeneity requires almost no argument.

But note: if the subjects of study can choose to "move to places with more rainfall" (e.g., farmer migration), then there may still be self-selection in "individual-level rainfall exposure."

(2) Biological characteristics—age, sex, year of birth, genetic traits.

You cannot choose your year of birth or your sex. When studying individual-level labor market outcomes, these variables are generally considered exogenous.

But note: age = year of birth + survey year. If your sample spans multiple years, age and cohort effects are confounded—age can be exogenous (you cannot choose your birth year), but the coefficient on age may be contaminated by cohort effects or period effects. Exogenous ≠ the coefficient has a clear causal interpretation.

(3) Institutional rules and historical events—compulsory schooling laws, military draft lotteries, colonial boundaries.

Angrist & Krueger (1991) used "quarter of birth × compulsory schooling law" as an instrument for education. Compulsory schooling laws are set by the government; individuals cannot choose them—they are exogenous. The same logic applies to various "natural experiments": policy changes, institutional discontinuities, historical boundaries.

(4) Random numbers or lottery outcomes.

The Vietnam War draft lottery (Angrist, 1990), license plate restrictions based on the last digit—exogeneity generated by genuinely random processes is the cleanest and requires the least argumentation.

(5) Lagged variables (under specific conditions).

In time series or panel data, predetermined variables—i.e., X at period t-1—can be exogenous (relative to the contemporaneous εₜ) under the assumption that "the contemporaneous εₜ is not serially correlated." But this is not permanent exogeneity—it is merely a weaker condition than contemporaneous X.

5.2 Variables that "look exogenous but may actually be endogenous"—common traps

Variable	Why it looks exogenous	Why it may be endogenous
Policy treatment variable (a province implements a policy)	The policy is set by the government; individuals cannot choose	Policies are not randomly implemented—the government chose the provinces that "need it most" or have the "best conditions" to go first
Education level	Education is completed before entering the labor market	Ability affects both education and wages (omitted variable); wage expectations reversely affect education decisions (reverse causality)
Firm size	Size looks like an "objective characteristic" of the firm	Size is the result of past firm performance—and past performance is correlated with current performance
Regional dummy variables	Region is fixed; you cannot choose where you are born	But you can choose which region to move to—migration decisions are endogenous to economic opportunities
Parental characteristics (e.g., mother's education)	You cannot choose your mother	But mother's education is highly correlated with your own education → if you control for own education but not mother's education, own education may still be endogenous due to omitted family background

5.3 Arguing exogeneity—it is not "say it and it is so"

In a paper, you cannot simply write "we assume X is exogenous" and be done. You need to argue why your X satisfies exogeneity in this particular empirical setting. Common argumentation strategies:

Randomized experiment: "Treatment was randomly assigned, therefore Cov(Treatment, ε) = 0."
Institutional features: "The policy was implemented uniformly nationwide; the timing of implementation across provinces was determined centrally, unrelated to provincial economic conditions."
Predeterminedness argument: "X is determined at period t-1, while εₜ is the shock at period t—if shocks are not serially correlated, then Cov(Xₜ₋₁, εₜ) = 0."
Controlling for many confounders: "After controlling for A, B, and C, we have reason to believe that factors remaining in the error term are uncorrelated with X." (This is the weakest argument, but also the most common—it relies on the untestable assumption that "you have controlled for enough variables.")

VI. Three classic sources of endogeneity—why is Cov(X, ε) ≠ 0?

Having understood what exogenous and endogenous mean, a natural question is: What causes a variable to become endogenous? All endogeneity problems can be traced back to one of the following three mechanisms (or their combinations):

6.1 Omitted variables

This is the most common source of endogeneity. If there is a variable Z that affects both X and Y, and you do not control for it—Z enters ε. Because Z is correlated with X, Cov(X, ε) ≠ 0.

$\text{True model: } Y = \beta_0 + \beta_1 X + \beta_2 Z + \nu \quad (\text{Cov}(X,\nu)=0)$ $\text{What you run: } Y = \tilde{\beta}_0 + \tilde{\beta}_1 X + \varepsilon \quad (\varepsilon = \beta_2 Z + \nu)$

Because Z is correlated with X, Cov(X, ε) = β₂·Cov(X, Z) ≠ 0.

Classic example: Ability Z affects both education X and wages Y, but you cannot measure ability. Education becomes endogenous.

6.2 Measurement error

If the X* you observe is not equal to the true X, but rather X* = X + ν (where ν is random noise), then:

The true Y is generated by X: Y = β₀ + β₁X + ε
You put X* into the regression: Y = β₀ + β₁(X* − ν) + ε = β₀ + β₁X* + (ε − β₁ν)

The new error term (ε − β₁ν) is correlated with X* (because X* = X + ν, and ν appears in both X* and the error term) → measurement error makes the variable endogenous.

Under the classical "random measurement error" assumption, this type of endogeneity leads to attenuation bias— $\hat{\beta}_1$ is pulled toward zero.

6.3 Simultaneity / reverse causality

If X affects Y, but Y also reversely affects X—you cannot tell which is the cause and which is the effect.

$\text{Demand: } Q = \alpha_0 + \alpha_1 P + \varepsilon_1$ $\text{Supply: } Q = \beta_0 + \beta_1 P + \varepsilon_2$

In equilibrium, price P and quantity Q are simultaneously determined by both equations. P is correlated with ε₁ (because P reflects supply shocks ε₂ in equilibrium) → P is endogenous in the demand equation.

Classic example: When studying "the effect of crime rates on police numbers," more police may reduce crime (X → Y), but cities with high crime rates may hire more police (Y → X). Causality runs in both directions simultaneously, making police numbers endogenous.

6.4 Quick reference table for the three sources

Source	Mechanism for Cov(X, ε) ≠ 0	Classic example	Direction of OLS bias
Omitted variables	Omitted confounder Z enters both X and Y	Ability (education → wage)	Depends on the direction of correlation between Z and X, Y
Measurement error	Measurement noise in X enters both X* and the error term	Underreporting of income in surveys	Pulled toward zero in the classical case (attenuation bias)
Reverse causality	Y reversely affects X	Price and quantity, crime and police	Depends on the relative magnitude of the two-directional effects

VII. How are these two concepts typically examined in master's and doctoral exams?

7.1 Exam formats

In master's (intermediate econometrics) and doctoral (advanced econometrics) exams, "endogeneity" and "exogeneity" are among the most central examination points, typically appearing in the following three forms:

Form 1: Conceptual distinction questions (short answer / terminology explanation).

"Please explain the difference between endogenous and exogenous explanatory variables. Why does endogeneity cause OLS estimates to be inconsistent?"

This type of question tests precise understanding of definitions and mechanisms. You need to write:

Exogeneity: $\mathbb{E}[\varepsilon \mid X] = 0$ → OLS is unbiased and consistent
Endogeneity: $\text{Cov}(X, \varepsilon) \neq 0$ → OLS is biased and inconsistent
Use the plim formula for OLS to prove: $\text{plim } \hat{\beta}_1 = \beta_1 + \frac{\text{Cov}(X,\varepsilon)}{\text{Var}(X)} \neq \beta_1$

Form 2: Applied analysis questions (given a scenario, judge endogeneity and propose solutions).

"A researcher uses OLS to estimate the 'effect of health on wages,' controlling for age, sex, and education. Please analyze: could the health variable be endogenous? If so, which mechanism does the endogeneity come from? How would you address it?"

This type of question tests the ability to identify the source of endogeneity and propose solutions. Your answer should:

Judge endogeneity: Health could be endogenous—omitted variables (childhood nutrition affects both adult health and wages), reverse causality (high-wage individuals can afford better medical care → wages affect health), measurement error (self-reported health contains subjective error).
Propose solutions:
- Instrumental variables: find a variable that affects health but does not directly affect wages (e.g., "whether you had a major childhood illness," "birth weight"; economists often use "mother's education" or "health status at age 16" as IVs for health).
- Panel data fixed effects: if endogeneity comes from time-invariant omitted variables (e.g., childhood nutrition), panel fixed effects can eliminate it.
- Natural experiments: exploit exogenous shocks from health policies.

Form 3: Mathematical derivation questions.

"Suppose the true model is Yᵢ = β₀ + β₁Xᵢ + εᵢ, where Cov(Xᵢ, εᵢ) = σ_{Xε} ≠ 0. Please derive the probability limit of the OLS estimator $\hat{\beta}_1$ and explain why $\hat{\beta}_1$ is inconsistent."

This is a standard doctoral-level exam question. The key derivation steps:

$\hat{\beta}_1 = \frac{\sum (X_i - \bar{X})(Y_i - \bar{Y})}{\sum (X_i - \bar{X})^2} = \beta_1 + \frac{\sum (X_i - \bar{X})\varepsilon_i}{\sum (X_i - \bar{X})^2}$

$\text{plim } \hat{\beta}_1 = \beta_1 + \frac{\text{Cov}(X,\varepsilon)}{\text{Var}(X)}$

Because $\text{Cov}(X, \varepsilon) \neq 0$ , the second term is nonzero → $\hat{\beta}_1$ is inconsistent.

7.2 Summary of high-frequency exam points

Exam point	Master's	Doctoral	Typical question format
Definition of exogeneity	✅ Required	✅ Required	"What is the exogeneity assumption? Why is it key to OLS unbiasedness?"
Three sources of endogeneity	✅ Required	✅ Required	"List and illustrate with examples the three sources of endogeneity"
Formula for omitted variable bias	✅ Common	✅ Required	"Derive the direction and magnitude of omitted variable bias"
plim derivation of OLS inconsistency	⭕ Sometimes	✅ Required	"Prove that OLS is inconsistent when Cov(X,ε) ≠ 0"
Idea of the Hausman test	⭕ Sometimes	✅ Common	"How does the Hausman test test for endogeneity?"
Two conditions for instrumental variables	✅ Common	✅ Required	"What two conditions must a valid instrument satisfy?"
Two-stage least squares (2SLS) with IVs	✅ Common	✅ Required	"Write out the two stages of the 2SLS regression equations"
Weak instruments problem	❌ Rare	✅ Common	"What are weak instruments? Why is F > 10 an empirical threshold?"
Overidentification test (Sargan/J test)	❌ Rare	✅ Common	"Under what conditions can you test the exogeneity of instruments?"
How panel fixed effects mitigate endogeneity	⭕ Sometimes	✅ Common	"Which type of endogeneity can fixed effects models solve? Which type can they not solve?"

7.3 Answering techniques

Technique 1: In conceptual distinction questions, be sure to distinguish between "exogeneity is an assumption of OLS" and "exogeneity is a property of the variable itself." The former is correct, the latter is wrong—this is precisely the grading point instructors use to distinguish "true understanding" from "rote memorization."

Technique 2: In applied analysis questions, do not just say "use instrumental variables." You need to state what specific instrument you would use, and why it satisfies the two conditions of relevance and exogeneity. A good IV answer is: variable name + relevance argument + exclusion restriction argument.

Technique 3: Endogeneity = Cov(X, ε) ≠ 0. This inequality is the logical starting point for all your answers in the exam room. Whether deriving, giving examples, or proposing solutions, always start from this inequality—"Why endogenous? Because Cov ≠ 0 → Why can 2SLS solve it? Because it finds a Z that is uncorrelated with ε but correlated with X → Why can panel fixed effects partially solve it? Because it eliminates the time-invariant component of ε."

Technique 4: If the exam paper asks "evaluate the validity of the following instrumental variable"—the first thing you should do is not to think of an answer, but to silently recite the two conditions: relevance (Cov(Z, X) ≠ 0) and exogeneity (Cov(Z, ε) = 0). If either condition fails, the IV is invalid. Relevance can be tested (first-stage F-statistic); exogeneity is untestable in the just-identified case—this is an advanced exam point that instructors love to test.

VIII. Summary

Five core pieces of knowledge about endogenous and exogenous variables:

The distinction is just one condition: Cov(X, ε) = 0 → exogenous; Cov(X, ε) ≠ 0 → endogenous. This determines whether OLS is unbiased (exogenous) or biased (endogenous).
Exogenous/endogenous is not an inherent label of the variable, but an assumption relative to a specific error term in a specific model. The same "education" variable is exogenous after controlling for ability and endogenous when ability is not controlled.
In causal inference with observational data, endogeneity is the default state; exogeneity is the exception. As long as X is the result of a "choice" or "decision," you need to argue why it is exogenous—rather than assuming it is automatically exogenous.
Endogeneity has three classic sources: omitted variables (most common), measurement error (pulls toward zero in the classical case), and reverse causality (X and Y mutually determine each other). Any one of them is sufficient to make Cov(X, ε) ≠ 0.
In exams, the core of endogeneity/exogeneity is derivation—starting from Cov(X, ε) ≠ 0, derive OLS inconsistency, prove why IV works, and judge the source of endogeneity in a given scenario. Rote memorization of "education is endogenous" is not enough—you need to be able to explain "why it is endogenous, which mechanism it comes from, and how to solve it."

One-sentence conclusion:

"The difference between endogeneity and exogeneity is just one condition—whether Cov(X, ε) is zero. But behind this condition lies your understanding of the causal logic of your entire research design. A variable is exogenous in Model A and endogenous in Model B—the judgment lies between these two models. This is why econometrics is not just a set of tools, but a way of thinking."

IX. Presentation suggestions for Bilibili / WeChat Official Account

Bilibili video: It is recommended to use "the ε black box" as a visual metaphor running through the entire video. Opening: a black box labeled "ε" containing unobservable factors such as ability, motivation, and family background. Then a hand puts "education (X)" into the regression—but the variation in education is linked to "ability" inside the ε box (connect X and ε with a dashed line). Narration: "When there is this line between X and ε—your OLS estimate is no longer unbiased. This is endogeneity." Then develop in three parts: omitted variables (from the ε box, "ability" jumps out pointing to both X and Y) → measurement error (the value of X jitters between the "true value" and the "observed value") → reverse causality (Y shoots an arrow back at X). For the exam section, an animation of "the three endogeneity questions in the exam room" is recommended—question pops up → your answer logic unfolds → scoring points are highlighted.
WeChat Official Account: The comparison table of exogenous vs. endogenous definitions is recommended as a two-column infographic. The three sources of endogeneity (omitted variables, measurement error, reverse causality) should each be paired with a causal diagram (DAG) plus a classic example card. The classification of naturally exogenous variables (weather/biology/institution/random) can be made into a four-quadrant card layout. The master's vs. doctoral exam point comparison table is recommended as a vertical comparison chart. The plim derivation should be displayed in a standalone formula box.
Recommended titles:
- Main title: 《Endogenous vs. Exogenous Explanatory Variables: A Question of Whether Cov(X, ε) Is Zero》
- Alternative title: 《When Is Your X Endogenous?—From Identification to Solutions, Explained for Both Exams and Empirical Work》
- New media title: 《Why Is OLS Biased?—Where Endogeneity Comes From, How to Judge It, and How Exams Test It》
Key quotes:

"Exogenous and endogenous are not identity labels of variables, but the 'state' of a variable in a specific model, relative to a specific error term. The same education variable is exogenous when ability is controlled and endogenous when ability is not controlled. The judgment lies between these two models."

"In causal inference with observational data, endogeneity is the default state—exogeneity is the exception that you need to argue for."

"The three sources of endogeneity: omitted variables (there is a Z you do not see simultaneously manipulating X and Y), measurement error (the X you see is not the true X), and reverse causality (Y also affects X). Their common result: Cov(X, ε) ≠ 0 → OLS is biased and inconsistent."