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: Those Two Columns in Stata Output — Do You Only Look at p-values and Stars?

After running a regression, Stata output looks like this:

------------------------------------------------------------------------------
        wage | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
   education |   .0800000    .0100000     8.00   0.000      .060399    .0996001
  experience |   .0300000    .0080000     3.75   0.000      .014317    .0456831
        _cons |   1.500000    .2000000     7.50   0.000      1.10792     1.89208
------------------------------------------------------------------------------

Most people's eyes first land on the P>|t| column — p < 0.05 → happy, add stars, write it into the paper.

But those two columns on the right — [95% conf. interval] — have you ever looked at them carefully? It is not just "an interval"; mathematically, it is completely equivalent to the p-value you are staring at. Understanding this equivalence, you will find that for many judgments about "significance," looking at the interval is more intuitive and more informative than looking at the p-value.

Core message: A coefficient's 95% confidence interval says — "If I could repeatedly sample from the same population and run the same regression an infinite number of times, then across those repeated samples, 95% of the confidence intervals computed from those samples would cover the true coefficient value." Its relationship with the point estimate is: the interval is centered at the point estimate, extending roughly 2 standard errors to each side. And its relationship with the p-value is: whether the 95% confidence interval contains 0 and whether the p-value is less than 0.05 are mathematically two ways of stating the same judgment — they must either both hold or both fail.

2. What Does a Confidence Interval Actually Say? — A Concept Widely Misinterpreted

2.1 Correct Interpretation vs. Common Misinterpretation

❌ Misinterpretation (extremely common): "The true coefficient β has a 95% probability of falling within this interval."

This statement sounds natural, but it is incorrect within the frequentist framework. Why? Because in the frequentist framework (which is also the statistical framework adopted by Stata and almost all econometric software), the true β is a fixed, unknown constant — it is not a random variable. A fixed constant is either in the interval or not — there is no such thing as "a 95% probability of falling inside." Probability applies to random variables, and β is not one.

✅ Correct interpretation: "If I were to draw samples from the same population infinitely many times, run the same regression each time, and compute a 95% confidence interval each time — then among those infinitely many confidence intervals, approximately 95% would contain the true β."

The difference between these two statements may seem like splitting hairs, but it touches the foundation of frequentist inference: the confidence level (95%) describes the property of the "interval construction method," not the property of "any particular interval." The specific interval in your hands is an already-realized random outcome — it either covers β or it does not. You simply do not know which. The 95% is your degree of trust in the "long-run success rate" of this interval construction method.

2.2 An Analogy: Casting a Net

Imagine you are on a pitch-dark sea trying to catch a fixed fish (the true β). You have a net with a fixed radius (the confidence interval), where the radius is determined by the noise level in your data (the standard error). Each time you cast the net into the sea — the casting position is determined by your sample (the point estimate $\hat{\beta}$ is the center of the net), and the net's size is determined by your sample size and how well your model fits.

If you use a method that "catches the fish 95 times out of 100 casts," then the net you cast this time is a 95% confidence interval.

The fish's position itself is fixed. Whether this particular cast caught it — you do not know. But what you do know is: if you used this casting method your entire life, your probability of catching the fish would be 95%.

2.3 How Is [95% conf. interval] in Stata Output Computed?

Under the classical linear regression assumptions, the OLS estimator $\hat{\beta}_j$ follows a normal distribution (asymptotically normal in large samples):

\hat{\beta}_j \sim N\left(\beta_j, \text{Var}(\hat{\beta}_j)\right)

From this, the 95% confidence interval is constructed:

\text{95\% CI} = \hat{\beta}_j \pm t_{0.025, n-k-1} \times \text{SE}(\hat{\beta}_j)

where:

$\hat{\beta}_j$ is the point estimate (the Coefficient column in Stata output)
$\text{SE}(\hat{\beta}_j)$ is the standard error (the Std. err. column)
$t_{0.025, n-k-1}$ is the 2.5% critical value of the t-distribution. When the sample size is large (n − k − 1 > 120), this value is approximately 1.96

Therefore, when the sample size is sufficiently large, 95% CI ≈ point estimate ± 1.96 × standard error.

Take the education coefficient from the Stata output above as an example:

\hat{\beta} = 0.080, \quad \text{SE} = 0.010

\text{95\% CI} = 0.080 \pm 1.96 \times 0.010 = [0.0604, 0.0996]

This matches the Stata output exactly.

3. The Relationship Between Point Estimates and Confidence Intervals — Center + Radius = Interval

3.1 The Point Estimate Is "Your Best Guess from What You See"; the Interval Is "the Precision Range of That Guess"

Returning to the three core columns of information in the Stata output:

Output Column	Meaning	What It Tells You
`Coefficient` ( $\hat{\beta}$ )	Point estimate	"Based on this sample, my best guess for β is 0.08"
`Std. err.` (SE)	Standard error	"The precision of this guess — the smaller the standard error, the more precise the guess"
`[95% conf. interval]`	Confidence interval	"Given this precision, the true β likely falls between 0.060 and 0.100"

Confidence interval = point estimate ± critical value × standard error. The point estimate tells you the "direction" and "magnitude"; the confidence interval tells you "how much uncertainty surrounds this magnitude."

3.2 Identical Point Estimates, Completely Different Confidence Intervals

Suppose two researchers use different data:

Researcher A (large sample, n = 10,000): $\hat{\beta} = 0.080$ , SE = 0.005, 95% CI = [0.070, 0.090]
Researcher B (small sample, n = 100): $\hat{\beta} = 0.080$ , SE = 0.040, 95% CI = [0.002, 0.158]

The two point estimates are exactly the same — both are 0.080. But A's interval is narrow and precise, telling us that β is almost certainly between 0.07 and 0.09. B's interval is wide and vague — ranging from nearly zero (0.002) to very large (0.158), covering everything from a small effect to a large effect.

Looking only at the point estimate without the confidence interval is like looking only at the steering wheel direction without the fuel gauge — you have no idea how much confidence you should place in that estimate.

3.3 The Position of the Point Estimate Within the Interval

You may have noticed that the confidence interval in Stata output is symmetric around the point estimate — $\hat{\beta}$ sits exactly at the center of the interval. This is a property of OLS (and most classical estimation methods). However, in certain nonlinear models (such as marginal effects in logit, or bootstrap confidence intervals), the interval may be asymmetric. If you see an asymmetric confidence interval, do not be surprised — it simply means the distribution of the estimator is not symmetric.

4. Confidence Intervals and p-values — Two Faces of the Same Judgment

4.1 What Question Does the p-value Ask? What Question Does the Confidence Interval Ask?

The p-value asks: "If the true β equals 0 (the null hypothesis is true), what is the probability of observing a $\hat{\beta}$ this large (or larger)?"

If this probability is very small (p < 0.05) → "In a world where β = 0, this data is too 'extreme'" → reject the null hypothesis → conclusion: β is significantly different from 0.
If this probability is not small (p ≥ 0.05) → "In a world where β = 0, this data is not too unusual" → cannot reject the null hypothesis.

The confidence interval asks: "What range of values might the true β take?"

If this range does not contain 0 → "0 is not within the plausible range" → β is significantly different from 0.
If this range contains 0 → "0 is within the plausible range" → cannot rule out the possibility that β = 0.

4.2 Why Does 95% CI Not Containing 0 ⇔ p < 0.05? — A Mathematical Proof

This is an exact mathematical equivalence, not a coincidence.

Definition of the p-value (two-sided test, null hypothesis H₀: βⱼ = 0):

p = 2 \times P\left(t_{n-k-1} > \left|\frac{\hat{\beta}_j}{\text{SE}(\hat{\beta}_j)}\right|\right)

That is, the p-value is the two-tailed tail area of the t-distribution beyond $\pm |t|$ .

Condition for the 95% confidence interval to contain 0:

0 \in \left[\hat{\beta}_j - c \times \text{SE}, \hat{\beta}_j + c \times \text{SE}\right]

where $c = t_{0.025, n-k-1}$ . This is equivalent to:

-c \times \text{SE} < \hat{\beta}_j < c \times \text{SE}

That is:

\left|\frac{\hat{\beta}_j}{\text{SE}}\right| < c \quad \Leftrightarrow \quad |t| < c

Therefore:

	Interval contains 0	Interval does not contain 0
Condition	$	t
p-value	$p > 0.05$	$p < 0.05$
Conclusion	Cannot reject H₀	Reject H₀

Key point: $c = t_{0.025, n-k-1}$ is precisely defined as the critical value that makes the two-tailed area of the t-distribution exactly equal to 0.05. Therefore, $|t| > c$ and $p < 0.05$ are saying the same thing.

"The interval does not contain 0" ⇔ " $|t| >$ critical value" ⇔ "p < 0.05." These three statements are mathematically completely equivalent.

4.3 Numerical Verification

Returning to the education coefficient in the Stata output:

\hat{\beta} = 0.080, \quad \text{SE} = 0.010, \quad t = 0.080/0.010 = 8.00

p = 2 \times P(t_{n-k-1} > 8.00) \approx 0.0000

95% CI = [0.0604, 0.0996] — does not contain 0. p < 0.001, far less than 0.05. Equivalent.

Now suppose another variable x_weak:

\hat{\beta} = 0.020, \quad \text{SE} = 0.015

t = 0.020/0.015 = 1.333

95\% \text{ CI} = 0.020 \pm 1.96 \times 0.015 = [-0.0094, 0.0494]

The interval contains 0. If n is large enough, the t critical value is approximately 1.96, and |1.333| < 1.96 → p > 0.05. Suppose n − k − 1 = 200, then p ≈ 0.184 > 0.05. Equivalent.

5. What Additional Information Does a Confidence Interval Give You Beyond a p-value?

5.1 The p-value Only Answers "Is It Zero?"; the Interval Answers "Where Might It Be?"

The p-value is a binary tool: p < 0.05 → significant, p ≥ 0.05 → not significant. It does not tell you how large the effect is, nor how precise the estimate is.

A confidence interval tells you three things simultaneously:

The direction of the effect (which side of zero is the entire interval on? All positive → sufficient to determine direction; crossing zero → direction uncertain)
The magnitude of the effect (the economic meaning of the interval endpoints — is the effect "large" or "small"?)
The precision of the estimate (how wide is the interval? Very wide → the data contains insufficient information)

Illustrative comparison:

Researcher A: $\hat{\beta} = 0.080$ , 95% CI = [0.060, 0.100], p < 0.001
Researcher C: $\hat{\beta} = 0.080$ , 95% CI = [0.008, 0.152], p = 0.030

Both researchers have the same point estimate, and both p-values are less than 0.05 (both significant). But A's interval precisely tells us the effect lies between 0.06 and 0.10 — a well-determined, moderately large effect. C's interval tells us — the effect could be 0.008 (tiny, almost no economic significance), or it could be 0.152 (very large). The p-value masks this enormous difference in precision; the confidence interval makes it immediately apparent.

5.2 What Does It Mean When an Interval Endpoint Is Close to Zero?

Suppose your key variable's coefficient has a 95% CI = [0.001, 0.159], p = 0.045. By the p < 0.05 standard → "significant."

But the lower endpoint of 0.001 means — the true effect could be 0.001, meaning a one-unit increase in X raises Y by only 0.1%. If the variable you are studying theoretically implies "there should be a substantial economic effect," then 0.001 is virtually indistinguishable from zero — your data can neither rule out a "zero effect" nor a "large effect."

In this case, "p < 0.05" is a conclusion that is technically valid but substantively weak. You should be candid in your paper: "The point estimate of the key coefficient is 0.080, but the 95% confidence interval is wide [0.001, 0.159], spanning effects that are economically negligible to substantively large — this indicates that the current sample contains insufficient information to precisely determine the magnitude of the effect."

5.3 The Interval Entirely on One Side of Zero and Far Away — "Not Just Statistically Significant, but Economically Significant"

If your 95% CI = [0.075, 0.085], and the standard deviation of the dependent variable wage is 0.5, then an effect of 0.075 to 0.085 (relative to the standard deviation) is approximately 0.15–0.17 standard deviations — a moderately small but stable and well-determined effect.

A confidence interval allows you to assess both "statistical significance" and "economic significance" simultaneously — a p-value can only answer the former.

6. Common Misuses of Confidence Intervals and Caveats

6.1 Misconception 1: Comparing Two Coefficients by Checking Whether Their Confidence Intervals Overlap

❌: "Variable A's 95% CI is [0.05, 0.15], and variable B's 95% CI is [0.10, 0.20]. The two intervals overlap → there is no significant difference between the coefficients of A and B."

This judgment may be wrong. Whether the difference between two coefficients is significant requires testing $H_0: \beta_A - \beta_B = 0$ , which requires $\text{Var}(\hat{\beta}_A - \hat{\beta}_B) = \text{Var}(\hat{\beta}_A) + \text{Var}(\hat{\beta}_B) - 2\text{Cov}(\hat{\beta}_A, \hat{\beta}_B)$ . Whether two individual confidence intervals overlap is not equivalent to this test — because the Cov term may not be zero.

The correct approach: use the test command (in Stata, test _b[varA] = _b[varB]).

6.2 Misconception 2: The Narrower the Interval, the Better — Ignoring Bias

A narrow confidence interval only means high precision (small standard error) — it does not mean the estimate is accurate (unbiased). If your model suffers from severe omitted variable bias, a small and narrow confidence interval surrounding a wrong value is actually more dangerous — because it gives you the illusion of being "precise and wrong."

Before focusing on confidence intervals, first confirm that your identification strategy is clean. Precision cannot substitute for unbiasedness.

6.3 Misconception 3: A 99% Confidence Interval Is "Better" Because It Is More "Conservative"

A 99% confidence interval is wider than a 95% interval because it requires a higher level of confidence. Neither is "better" — the choice depends on how much Type I error risk you are willing to bear (i.e., incorrectly rejecting H₀ when β = 0).

95% CI corresponds to α = 0.05 → when β = 0, you have a 5% probability that the computed interval does not contain 0.
99% CI corresponds to α = 0.01 → the interval is wider and "safer," but it also makes it more likely that you fail to reject H₀ when β ≠ 0 (increased Type II error risk).

In empirical economics, 95% is the conventional standard — you do not need to justify why you use 95%. But if you use 90% or 99%, you should provide a reason in your paper.

6.4 Caveat: Under Heteroskedasticity and Clustering, the Default CI Is Wrong

Stata's reg command assumes by default that the error terms are homoskedastic. If you have heteroskedasticity or within-cluster correlation, the default standard errors will be underestimated → the confidence interval will be narrower than the correct one → you will "over-reject" H₀.

Solution: Use robust standard errors (reg y x, robust) or clustered standard errors (reg y x, cluster(id)); Stata will automatically recompute the confidence interval based on the corrected standard errors.

7. How Are Confidence Intervals Examined in Master's and Doctoral Exams?

7.1 High-Frequency Exam Formats

Format 1: Given Stata output, interpret it.

"The table below reports the OLS estimates of the return to education. Please interpret the meaning of the education coefficient and explain the economic significance of its 95% confidence interval."

You need to answer:

Point estimate = 0.08 → each additional year of education increases wages by 8% on average.
95% CI = [0.06, 0.10] → the true return to education supported by the data is likely between 6% and 10%.
The interval does not contain 0 → p < 0.05 → the effect of education on wages is statistically significant at the 5% level.

Format 2: Conceptual distinction question.

"Please explain why the statement 'the true coefficient β has a 95% probability of falling within the confidence interval' is incorrect."

This is a favorite distinction question for instructors — testing whether you understand the core framework of frequentist inference. The correct answer: β is a fixed constant, not a random variable; 95% is the coverage probability of the construction method, not the probability that any particular interval contains β.

Format 3: Proof question.

"Please prove: in a two-sided t-test, the 95% confidence interval does not contain 0 if and only if p < 0.05."

Derivation path: 95% CI does not contain 0 ⇔ |t| > critical value ⇔ p < 0.05. Write out the three equivalences clearly.

Format 4: Given a coefficient and standard error, construct a confidence interval and perform a hypothesis test.

"A regression output gives $\hat{\beta}_1 = 0.05$ , $\text{SE} = 0.02$ , n = 200, k = 5. Construct the 95% confidence interval and determine whether β₁ is significantly different from 0 at the 5% significance level."

\text{95\% CI} = 0.05 \pm 1.96 \times 0.02 = [0.0108, 0.0892]

Does not contain 0 → significant at the 5% level.

7.2 Key Points for Answering

Distinguish between β and $\hat{\beta}$ : β is the true parameter (fixed and unknown), $\hat{\beta}$ is the estimator (a random variable). The confidence interval concerns β and is constructed based on $\hat{\beta}$ .
Remember the approximation formula: in large samples, 95% CI ≈ $\hat{\beta} \pm 1.96 \times \text{SE}$ ; 99% CI ≈ $\hat{\beta} \pm 2.58 \times \text{SE}$ .
If no t-distribution table is provided in the exam and the sample size is large — directly use the normal approximation with 1.96.
The proof of the equivalence between the interval and the p-value must be mastered — it is the bridge from "interval" to "hypothesis testing."

8. Summary

Five core takeaways about confidence intervals:

95% CI = point estimate ± critical value × standard error. In large samples, this is approximately $\hat{\beta} \pm 1.96 \times \text{SE}$ . It is centered at the point estimate, extending on each side by an "uncertainty boundary" determined by the standard error.
The confidence level is the "long-run success rate of the interval construction method," not "the probability that this particular interval contains β." β is a fixed constant, not a random variable. The 95% is your degree of trust in the method — if you cast the net 100 times using this method, roughly 95 times you will catch the fish.
95% CI does not contain 0 ⇔ p < 0.05. This is not an empirical regularity; it is a mathematical equivalence. The bridge between the two is the t-statistic: $|t| >$ critical value is simultaneously equivalent to both "p < 0.05" and "the interval does not contain 0."
A confidence interval gives you more information than a p-value: "effect magnitude" and "estimation precision." Looking only at the p-value is a binary judgment (significant/not significant); looking at the interval simultaneously tells you direction, magnitude, and precision — far more information than a single p-value.
A narrow interval ≠ an accurate estimate. A narrow interval means high precision, but if the model is biased (omitted variables, endogeneity), precision merely gives you more confidence around a wrong value — this is not a good thing.

One-sentence conclusion:

"The p-value asks you 'is it zero?'; the confidence interval tells you 'where might it be?' For the same judgment — 'not zero' — the p-value answers with a probability, and the interval answers with a range. But the interval does one more thing: it tells you that even after excluding zero, the true effect could be as small as an economically meaningless 0.001, or as large as a 0.15 that could change policy recommendations. This information, the p-value will never give you."

9. Presentation Suggestions for Bilibili/WeChat Official Account

Bilibili video: It is recommended to use the "casting a net" animation as the visual metaphor running through the entire piece. Opening: a fixed fish (the true β) in a pitch-dark sea. A hand casts a net (confidence interval) each time, with the center of the net being the point estimate and the radius being 1.96 × SE. Cast 100 times, roughly 95 times the fish is caught, 5 times it is not. Voiceover: "The 95% confidence interval does not mean the fish has a 95% probability of being in this net — the fish is fixed. The 95% means — if you cast the net this way your entire life, roughly 95% of your nets will catch this fish." Then transition to "interval vs. p-value": the screen shows two faces side by side — the left face is the p-value (only two expressions: 😊 significant / 😐 not significant), the right face is the confidence interval (showing the range and direction of the effect). Voiceover: "The p-value only tells you 'is it zero?'; the confidence interval tells you 'is it positive or negative, and roughly how large?'" Finally, use an interactive slider to demonstrate: drag the SE slider (standard error increases → interval widens → transitions from "does not contain 0" to "contains 0" → p-value changes from < 0.05 to > 0.05), visually showing the interconnected relationship among the standard error, the interval, and the p-value.
WeChat Official Account: The confidence interval formula should be displayed in a formula box. The correct vs. incorrect interpretations should be made into a two-column comparison card (left column ❌ common error, right column ✅ correct interpretation). The mathematical proof that CI not containing 0 ⇔ p < 0.05 should be presented as a logical chain (a three-step arrow diagram: "interval does not contain 0 → |t| > c → p < 0.05"). The information comparison between the interval and the p-value should be made into an infographic centerpiece (p-value: ❓direction? ❓magnitude? ❓precision? vs. confidence interval: ✅direction ✅magnitude ✅precision). The Stata output example should be paired with a real code block screenshot, annotating the correspondence among the five columns: Coefficient, SE, t, p, and CI. The four common misconceptions should be made into "misconception cards."
Recommended titles:
- Main title: 《What Exactly Is the Confidence Interval in Stata Regression Output? Why Is It the Same Thing as the p-value?》
- Alternative title: 《95% CI Not Containing 0 ⇔ p < 0.05 — Where Does This Equivalence Come From?》
- New media title: 《Don't Just Stare at the Stars: The Confidence Interval Tells You Three More Things Than the p-value》
Key quotes:

"The p-value asks you 'is it zero?'; the confidence interval tells you 'where might it be?' The former is a true/false question; the latter is a map."

"The 95% confidence interval not containing 0 and p < 0.05 are two faces of the same judgment — they are mathematically exactly equivalent, and the bridge between them is the t-statistic."

"The point estimate tells you the direction; the confidence interval tells you how certain that direction is. Looking only at the point estimate without the interval is like driving while looking only at the steering wheel direction without the fuel gauge — you do not know how much farther you can go."

"A narrow interval ≠ a correct estimate. A narrow confidence interval around a biased point estimate gives you the illusion of being 'precise and wrong' — this is more dangerous than a wide, honest confidence interval."