数据范儿_范老师-AI时代科研统计新范式-计量科研AI导航

1. Introduction: The Textbook Says OLS Is Unbiased, So Isn't My 0.08 Correct?

Beginners in econometrics often feel a strange sense of reassurance after learning about the unbiasedness of OLS.

The textbook presents the Gauss-Markov theorem: under five classical assumptions, the OLS estimator is unbiased — $\mathbb{E}[\hat{\beta}_1] = \beta_1$ .

Then you open Stata, run a regression, and obtain $\hat{\beta}_1 = 0.08$ .

You naturally think: "Since OLS is unbiased, my 0.08 should be the true value — or at least not far from it, right?"

No. Absolutely not.

Unbiasedness provides no guarantee whatsoever about the specific 0.08 in your hands. It does not promise that your 0.08 is close to the true value. It does not even promise that your 0.08 is closer to the true value than a purely random guess. It only promises this — if you could repeat the entire study an infinite number of times, drawing a new sample from the population and re-estimating each time, then the average of these infinitely many estimates would equal the true value.

But you cannot. You have only one sample and one estimate. And unbiasedness says nothing about whether that single estimate is accurate.

Core message: Unbiasedness is a property of the estimator (the formula), not of the estimate (the specific number). It describes "this formula will not systematically deviate from the target in the long run," not "your single arrow will hit the bullseye." The gap between these two is the starting point of all statistical inference logic.

2. The Precise Meaning of Unbiasedness — In Mathematics and Intuition

2.1 Mathematical Definition

$\mathbb{E}[\hat{\beta}_1] = \beta_1$

Here, the expectation $\mathbb{E}[\cdot]$ is taken over all possible random samples. It concerns the average behavior under repeated sampling — and has nothing to do with the specific sample in your hands or the specific value 0.08.

2.2 Understanding Through a Thought Experiment

You have a population (e.g., the education years and wages of all workers in China). The true β₁ is 0.075.

Now you conduct a thought experiment:

Time 1: Randomly draw 1,000 people from the population and run OLS. You obtain $\hat{\beta}_1^{(1)} = 0.081$ .
Time 2: Randomly draw another 1,000 people and run OLS. You obtain $\hat{\beta}_1^{(2)} = 0.069$ .
Time 3: Draw again, obtaining $\hat{\beta}_1^{(3)} = 0.083$ .
……
Time 10,000: You obtain $\hat{\beta}_1^{(10000)} = 0.073$ .

Now, take the average of these 10,000 $\hat{\beta}_1$ values. You will find that this average is approximately 0.075 — that is, the true β₁.

This is the entire meaning of unbiasedness: the "long-run average" of the estimation method is not wrong. The 0.083 from the third sample was too high, and the 0.069 from the second was too low, but they cancel each other out — on average, it is accurate.

2.3 But You Said "Time 3" — And I Only Have One Sample

You have only one sample. You are "Time 3" — your estimate is 0.083, higher than the true value of 0.075. Your estimate deviates from the true value — even though the estimation method you used is completely unbiased.

Unbiasedness does not promise that "your deviation this time is small," let alone "whether your deviation this time is upward or downward." You cannot infer any definitive conclusion about the 0.08 in your hands from unbiasedness.

3. The Bullseye Analogy — Classic but Often Misunderstood

Imagine a shooter aiming at a bullseye.

An unbiased shooter: Fires 1,000 times; the bullet holes are evenly scattered around the bullseye — symmetric in all directions, with the average point of impact exactly at the center. But any single shot could hit the bullseye dead center or land as far out as the 3-ring.
A biased shooter: Fires 1,000 times; the bullet holes systematically cluster toward the upper right of the bullseye — the average point of impact is at the 2-ring position. He might occasionally hit the bullseye, but "in the long run" he drifts to the upper right.

An unbiased shooter ≠ every shot hits the bullseye. He is merely "aiming correctly" in the sense that "if you let him fire infinitely many shots." As an observer who has seen only one of his shots, you cannot determine from whether that shot hit the bullseye whether the shooter's aim is correct.

There is an extended version of this analogy — two shooters are both unbiased, but one has tightly clustered bullet holes (small variance) while the other has loosely scattered ones (large variance). The former is a "precise unbiased estimator" — OLS's BLUE property under Gauss-Markov conditions precisely guarantees that it has the smallest variance among all linear unbiased estimators. The latter, while also unbiased, has estimates that can deviate from the true value by a much larger magnitude in any single sample.

OLS's "efficiency" (minimum variance) guarantees that among all unbiased shooters, OLS has the most concentrated bullet-hole distribution. But it does not guarantee that your single shot will hit the bullseye.

4. Why Can't We Say "OLS Is Unbiased, So My Estimate Is Correct"?

4.1 The Estimate ≠ The Estimator

As detailed above: the estimator is the formula (a random variable); the estimate is the number computed by applying the formula to your specific sample (a fixed value).

Unbiasedness is a property of the formula. The 0.08 in your hands is not "unbiased" — it is just a number, just as 3.14 is not "unbiased." You can say that the method that produced it is "unbiased," but you cannot say that this number is "unbiased."

4.2 Sampling Error Always Exists

Even if your sample is a perfect simple random sample, your model specification is completely correct, and all classical assumptions are satisfied — your n=1,000 is not the population. Due to the randomness of sampling, your 1,000 people happen to differ slightly from the next 1,000 people, so your 0.08 happens to differ slightly from the 0.076 you would get by drawing another 1,000 people.

Sampling error is not a "mistake" but a natural consequence of the act of "random sampling." Unbiasedness guarantees that this error averages to zero in the long run, but it does not guarantee that the error is zero in any particular sample.

4.3 The CIA (Conditional Independence Assumption) Itself May Not Hold

Unbiasedness is, mathematically, a theorem "given the classical assumptions." If the zero conditional mean assumption ( $\mathbb{E}[\varepsilon|X] = 0$ ) among the classical assumptions does not hold — that is, if there is omitted variable bias — then even with a perfectly random sample, the OLS estimator is not unbiased.

At this point you face a double uncertainty: not only may your specific estimate deviate from the true value (sampling error), but your estimation method itself will not average to the true value in the long run (bias).

But even in the first scenario of pure "sampling error" — where all assumptions hold perfectly — your single estimate can still deviate wildly.

5. An Underappreciated Fact: Unbiased Estimators Can Produce "Wildly Wrong" Estimates

5.1 Unbiased Estimators in Small Samples

Suppose the true β₁ = 0. You run OLS with a sample of n=10. OLS is unbiased — if you repeated the sampling one million times, the average of $\hat{\beta}_1$ would be 0.

But when n=10, the sampling distribution of $\hat{\beta}_1$ is very dispersed. From this distribution, you might draw $\hat{\beta}_1 = 2.3$ , or you might draw $\hat{\beta}_1 = -1.7$ . Each specific sample can produce an estimate far from 0 — but because the positive and negative values cancel out, the long-run average is still 0.

Unbiasedness provides no probabilistic guarantee about the "distance from the true value" for any single estimate. That distance is measured by the standard error — and the standard error is large in small samples.

5.2 In Large Samples — Even Then, "This Sample's Luck" Still Matters

Even with n=100,000 — where the standard error is already extremely small — your specific estimate can still deviate from the true value due to "bad luck with this particular sample." The only difference is that the magnitude of this deviation is constrained to a very small range by the standard error.

This is precisely the entire purpose of the confidence interval: it acknowledges that you have only one sample and one estimate — and therefore you cannot assert that "my estimate equals the true value." What you can say is — "if the classical assumptions hold, then under repeated sampling, confidence intervals constructed by the same method will contain the true value 95% of the time. My specific interval, at the 95% level, provides a reliable range."

6. Unbiasedness, Consistency, and the Progressive Relationship of "What You Can Believe"

Beyond unbiasedness, econometrics textbooks discuss another property: consistency.

$\hat{\beta}_1 \xrightarrow{p} \beta_1 \quad \text{as } n \to \infty$

Consistency states that as the sample size tends to infinity, the estimator converges in probability to the true value. In other words, when the sample size is sufficiently large, the probability of a large deviation between the estimate and the true value can be made arbitrarily close to 0.

The difference between consistency and unbiasedness:

Unbiasedness is a small-sample property — for any fixed n, $\mathbb{E}[\hat{\beta}_1] = \beta_1$ .
Consistency is a large-sample property — as $n \to \infty$ , $\hat{\beta}_1$ is almost surely near the true value.

Consistency is stronger than unbiasedness in the following sense: even if an estimator is biased in small samples ( $\mathbb{E}[\hat{\beta}_1] \neq \beta_1$ ), if the bias gradually disappears and the variance decays to zero as n increases, then it can still be trusted in sufficiently large samples. In contemporary microeconometrics, many estimation methods (such as instrumental variables 2SLS, propensity score matching) possess only large-sample properties (consistency) and do not have finite-sample unbiasedness.

7. So What Is "Unbiasedness" Actually Good For? — Value at Four Levels

Level 1: The minimum guarantee at the methodological level. Unbiasedness is the first line of defense for econometric methods — if an estimation method cannot even "aim at the bullseye in the long run," then you have no reason to trust its performance in any particular sample. Unbiasedness is the minimum threshold for entering the ranks of "acceptable methods."

Level 2: Eliminating systematic concerns. Unbiasedness removes the fear that "the direction of research will be systematically misled." If your estimation method is biased, you might repeatedly conduct 100 studies, and every single estimate would be too high — this is a terrifying cumulative consequence. Unbiasedness guarantees that the bias is zero in the long run — if policymakers conduct meta-analyses based on a large number of independent studies, the core findings of these studies will not fall into "collective bias."

Level 3: The core prerequisite for BLUE. "Efficiency" (minimum variance) is only meaningful within the set of unbiased estimators. The Gauss-Markov theorem defines OLS as having the smallest variance among "linear unbiased estimators" — if you abandon the unbiasedness requirement and pursue only minimum variance, your method could be a constant estimator (i.e., an estimator that always equals 3 — with zero variance but no relationship whatsoever to the true value). This is clearly undesirable.

Level 4: A theoretical benchmark — when it fails, you know where the problem lies. The classical assumptions for unbiasedness ( $\mathbb{E}[\varepsilon|X] = 0$ , homoskedasticity, no autocorrelation) can each be "diagnosed" one by one. When you discover in empirical work that a particular assumption may be violated, the derivation of unbiasedness tells you which direction of bias will appear and which method can remedy it — this itself is the greatest value of the unbiasedness framework for empirical researchers.

8. Summary

What you want to know…	Can unbiasedness answer it?	Then what can answer it?
"Does my 0.08 equal the true value?"	❌ No	Cannot be answered — the true value is forever unknown
"Is my 0.08 far from the true value?"	❌ No	Standard error + confidence interval
"Will this estimation method be biased in the long run?"	✅ Yes! Unbiasedness precisely answers this	—
"If I sample again, what would the estimate be?"	❌ No — but unbiasedness tells you the "expected value"	Sampling distribution (summarized by the standard error)
"With a larger sample, will my estimate approach the true value?"	❌ Unbiasedness does not guarantee this (but consistency does)	Consistency + asymptotic standard error

One-sentence conclusion:

"Unbiasedness is a check written to the 'long run' — if you could repeatedly draw samples an infinite number of times, the average answer from your estimation method would be correct. But you hold only one sample in your hands, and you must add the standard error and confidence interval to know roughly how far your reading this time is from the true value. The art of econometrics lies in finding a way to express the relationship between 'correct in the long run' and 'approximately within what range this time.'"

9. Presentation Suggestions for Bilibili / WeChat Official Account

Bilibili video: Consider using a "shooter aiming at a target" animation — two shooters, one unbiased (bullet holes surrounding the bullseye, with the average point of impact at the center) and one biased (bullet holes concentrated to the upper right). Then zoom in on a single frame of the unbiased shooter — "you only saw this one shot," and this shot happens to land outside the 3-ring. "Unbiased = correct in the long run, ≠ this shot hits the target." Finally, overlay an animation of OLS bullet holes narrowing as the sample size increases, visually demonstrating how the standard error shrinks as n grows.
WeChat Official Account: For the bullseye analogy, consider including two target diagrams (unbiased vs. biased). The distinction between "estimator vs. estimate" should maintain a consistent card style with the preceding text. The three-column table "What you want to know / Can unbiasedness answer it / What can answer it" is well suited for an infographic.
Recommended titles:
- Main title: 《OLS Is Unbiased, So Why Can My Estimate Still Be Wildly Wrong?》
- Alternative title: 《Unbiasedness Is a Check Written to the "Long Run" — But You Hold Only One Sample》
Key quote:

"Unbiasedness is a check written to the 'long run' — if you could repeatedly draw samples an infinite number of times, the average answer from your method would be correct. But you hold only one sample in your hands, and you must add the standard error and confidence interval to know roughly how reliable your answer this time is."