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: A Phenomenon You Have Noticed but No One Has Explained

Open the empirical section of any economics journal, scan the regression tables, and you will observe a pattern—

The coefficients are almost all small decimals. 0.08 (returns to education), 0.32 (labor supply elasticity), -0.15 (semi-elasticity of investment to tax rates), 0.07, 0.23, 0.11... The tables are filled with these small numbers between -1 and 1.

You rarely see regression coefficients of 30, or -500, or 2700.

In theory, this should not be the case.

The meaning of a regression coefficient is "when X changes by one unit, Y changes by β units." If X is years of education (unit: years) and Y is annual income (unit: yuan), then β = 2000 is perfectly reasonable—one more year of schooling, 2000 more yuan per year. Why does this coefficient never appear in papers?

If you have ever wondered about this—this article tells you that regression coefficients concentrated between -1 and 1 is not a law of nature, but a deliberate design choice by researchers. It is related to taking logarithms, variable scaling, standardization, and the behavioral regularities of human beings in the social sciences.

Key message: You do not see β = 2000 in papers, not because "returns to education are small," but because researchers have turned β = 2000 into β = 0.08—before running the regression, they have already "packaged" the coefficient into the readable interval [-1, 1] through log transformations, unit changes, or standardization.

2. Root Cause 1: Taking Logarithms—Converting "How Many Yuan" into "How Many Percent"

2.1 Semi-Log Models: The Most Widely Used Transformation

Suppose the true wage equation is:

$\text{wage} = 1000 + 2000 \times \text{edu} + \varepsilon$

β₁ = 2000. One more year of schooling, wages increase by 2000 yuan.

This coefficient has two problems:

First, it is difficult to compare across groups. 2000 yuan means something completely different to someone earning 3000 per month versus someone earning 30000 per month. The coefficient 2000 alone cannot tell you whether "this effect is large"—you need to know the baseline wage.

Second, it is difficult to compare across studies. American researchers use dollars and obtain a coefficient of 3000 (dollars); Chinese researchers use yuan and obtain a coefficient of 2000 (yuan). Whose effect is larger? You cannot tell directly from the magnitude of the coefficients.

Solution: Take the logarithm of Y.

$\ln(\text{wage}) = \beta_0 + \beta_1 \cdot \text{edu} + \varepsilon$

When Y is logged but X is not, the meaning of β₁ becomes: when X changes by one unit, Y changes by approximately (β₁ × 100)%.

The original meaning of β₁ = 2000—"one more year of schooling, wages increase by 2000 yuan"—is replaced by β₁ = 0.08—"one more year of schooling, wages increase by 8%."

This is the semi-elasticity: it is a percentage, naturally less than 1 (because most social factors are unlikely to double or halve wages within a single year).

2.2 Log-Log Models: Elasticities Are Almost Naturally Between -1 and 1

When X is also logged, β₁ becomes an elasticity:

$\ln(Y) = \beta_0 + \beta_1 \ln(X) + \varepsilon$

The meaning of β₁: when X changes by 1%, Y changes by β₁%.

In economics, most elasticities do indeed lie between -1 and 1. The labor supply elasticity is about 0.1–0.3, and the price elasticity of gasoline demand is about -0.2 to -0.3. Why are elasticities often less than 1?

Because elasticity < 1 (inelastic) is very common in human behavior. If wages rise by 10%, you will not immediately reduce your working hours by 10% (you might even increase them—the income effect may dominate the substitution effect). If gasoline prices rise by 10%, you will not immediately reduce your gasoline consumption by 10%—because your commute route and vehicle are difficult to change in the short run. Most short-run behavioral adjustments are smaller in magnitude than the stimulus, so elasticities are naturally less than 1.

But this is not a theorem. The demand elasticity for luxury goods can exceed 1 (elastic), and substitution elasticities for certain specific goods can far exceed 1. But for the relationships that most social scientists care about—returns to education, the effect of institutions on growth, the effect of income on health—elasticities are indeed on the order of a few tenths.

3. Root Cause 2: Scaling of Variable Units—Making Coefficients "Readable"

Even without taking logarithms, researchers control the magnitude of coefficients through the choice of variable units.

One example explains everything. You want to study "the effect of local fiscal expenditure on economic growth." You have two choices:

Choice A: X = fiscal expenditure (yuan), Y = GDP (yuan). The estimated β₁ might be 1.5—spending 1 yuan increases GDP by 1.5 yuan.
Choice B: X = fiscal expenditure (100 million yuan), Y = GDP (100 million yuan). The estimated β₁ is still 1.5—spending 100 million yuan increases GDP by 150 million yuan.
Choice C: X = fiscal expenditure (trillion yuan), Y = GDP (trillion yuan). The estimated β₁ is still 1.5.

You see, as long as X and Y are scaled by the same proportion, the regression coefficient does not change. Researchers naturally have an incentive to choose units that make the numbers "look good."

But more commonly: researchers enlarge the unit of X within a reasonable range, forcing the coefficient to become smaller. For example:

Instead of "age (years)," use "age/10"—the coefficient then becomes "how much Y changes for every 10 additional years," and the magnitude naturally goes from "over ten" to "around one."
Instead of "GDP (yuan)," use "GDP (thousand yuan)" or "take the logarithm."
Firm size is not measured by "number of employees (persons)" but by "number of employees (thousands)."

When researchers choose variable units, there is an unwritten aesthetic standard: keep the absolute value of coefficients between 0.01 and 1, so that the numbers in the table look clean, readable, and do not require scientific notation. A table full of 2748.32 and 0.000034 is very difficult to read—but after switching to appropriate units, the coefficients become 2.75 and 0.34, and everything is much cleaner.

4. Root Cause 3: Standardization and the "Gravity of Correlation Coefficients"

Recall a key formula derived earlier (in simple regression):

$\hat{\beta}_1 = r_{XY} \times \frac{s_Y}{s_X}$

Now, if the researcher happens to scale X and Y to roughly the same degree of dispersion (i.e., $s_Y \approx s_X$ ), then:

$\hat{\beta}_1 \approx r_{XY}$

And the correlation coefficient $r_{XY}$ is naturally bounded between [-1, 1]. So—when you study the effect of a one-standard-deviation change in X on Y, the magnitude of this effect is "naturally" close to the magnitude of a correlation coefficient.

This is precisely the intuition behind standardized regression. The standardized regression coefficient $\beta^*$ measures "how many standard deviations Y changes when X changes by one standard deviation." If the standard deviations of X and Y are similar (which often happens with carefully constructed variables), the standardized and unstandardized coefficients are close in magnitude.

This explains why many regression coefficients lie between -1 and 1—because researchers, by choosing appropriate variable measures, inadvertently make $s_Y / s_X$ approximately equal to 1, thereby bringing regression coefficients close to the magnitude of correlation coefficients.

This is also an underappreciated explanation. The phenomena studied in the social sciences—wages, academic performance, health status, voting behavior, entrepreneurial success—are not dominated by any single factor.

Wages are influenced by dozens of factors: education, experience, ability, industry, region, family background, social networks, luck... Taken individually, the "partial effect" of each factor on wages cannot possibly be very large. If one factor could multiply Y severalfold, where would that leave all the other factors?

This is not a mathematical constraint but an empirical fact: in complex systems of human behavior, the partial effect of a single variable typically falls in the "moderate but real" range.

This is also why econometrics distinguishes between "economic significance" and "statistical significance." A β₁ = 0.02 may be highly statistically significant (p < 0.001), but whether it is economically important depends on—what does 0.02 represent? If it is the coefficient of education on log(wage), 0.02 means each additional year of education increases earnings by 2%—which is actually quite substantial (4 years of college = 8.2% higher wages, with compounding effects). But if it is the effect of "using one more social media app on well-being (standardized 0-100 scale)," 0.02 means it is almost negligible.

6. When Do Regression Coefficients Exceed -1 to 1?

Of course they do. Coefficients easily exceed ±1 in the following situations:

Case 1: Y is not logged, and the units of Y and X differ greatly.

If you use Y = annual income (yuan) and X = years of education, β₁ might be 2000, 5000, or even 8000. This is not "wrong"—it simply means the researcher did not take logarithms. In earlier empirical literature, this specification was common. But contemporary empirical research typically takes logarithms—so what you mostly see is 0.08 rather than 2000.

Case 2: X is a dummy variable (0/1).

If Y is a continuous variable with a large magnitude (such as income) and X is gender (female = 1), β₁ can be -8000 (women earn 8000 yuan less than men on average). This is common in the literature that does not log Y. After taking logarithms (ln(wage)), this coefficient becomes about -0.15 (women earn about 15% less than men).

Case 3: The range of X is very small.

If X = "years since policy implementation" (range 0–5) and Y = GDP growth rate (%), β₁ might be 1.5, indicating that the growth effect of the policy accumulates and strengthens year by year. Because the range of X is extremely narrow, the coefficient becomes larger.

Case 4: Both X and Y are ratios or percentages.

For example, X = tax rate (between 0 and 1) and Y = investment rate (percentage of GDP). β₁ might be -2.3, indicating that when the tax rate falls from 20% to 19% (i.e., X changes by 0.01), the investment rate rises by 0.023 percentage points. Here the coefficient is large because the range of X is extremely small.

All of these are "correct"—they are merely visual differences arising from different unit specifications and have no direct bearing on whether the effect "exists" or is "statistically significant."

7. Summary: β Between -1 and 1 Is Not a Theorem—It Is a Convention

When you flip through regression tables and see decimals everywhere, what you are actually seeing is a set of "coefficient aesthetics" that researchers collectively follow:

Take logarithms whenever possible → coefficients become elasticities or semi-elasticities → naturally near [-1, 1].
Choose variable units carefully → avoid large numbers, keep the magnitude consistent across columns.
Understand the complexity of social systems → accept that the partial effect of a single factor will not be so large as to overwhelm all other factors in the system.

But you must remember—if a researcher interprets a coefficient of 0.08 as "a small effect" while forgetting that they used log(wage) as the dependent variable (0.08 = 8%, which is actually not small), then they have been fooled by their own "coefficient aesthetics."

When reading regression tables, do not let the digits after the decimal point distract you—first ask about the units of the independent variable, then the form of the dependent variable (whether it is logged), and only then judge the magnitude of the effect.

One-sentence conclusion:

"Regression coefficients concentrated between -1 and 1 is not because social laws only permit effects this small—but because researchers, through log transformations, unit changes, and standardization, have 'translated' all effects into the same easy-to-understand scale. Your job is not to marvel at how 'small' they are, but to reconstruct what the number actually means in reality."

8. Presentation Suggestions for Bilibili/WeChat Official Account

Bilibili video: For the core demonstration, we suggest a "transformation animation"—the center of the screen shows β₁ = 2000, then the researcher takes the logarithm of Y, the number shrinks and morphs into β₁ = 0.08. Voiceover: "The same effect, different expressions." Then show that if you change units (yuan → ten thousand yuan), the number continues to transform.
WeChat article: The four causes should be presented as progressive sections, with a one-sentence summary as the subtitle of each chapter. The concepts of semi-elasticity and elasticity can be emphasized with boxed definitions. The argument that elasticity is "naturally less than 1" should be accompanied by classic supply-demand curve diagrams.
Recommended titles:
- Main title: 《Why Are Most Regression Coefficients in Empirical Papers Around Zero-Point-Something?》
- Alternative title: 《β = 0.08 and β = 2000 May Be the Same Effect—It Depends on Whether You Take the Log》
Key quote:

"Regression coefficients concentrated between -1 and 1 is not a theorem—it is a convention. Taking logarithms, changing units, standardizing—researchers translate all effects onto the same scale. Your job is not to marvel at how 'small' they are, but to reconstruct their true meaning."