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: Two Variables, One Measured in "Yuan," the Other in "Years"—How Do You Compare Their Relative Impacts?

After running your regression, your Stata output shows two coefficients:

Coefficient on income (annual household income, in yuan) = 0.0003
Coefficient on education (years of schooling, in years) = 0.0800

Which variable has a "larger" effect on wages? You cannot compare them directly—because the effect of a 1-yuan change in income and the effect of a 1-year change in education are fundamentally incommensurable in terms of units. It is like saying "1 kilometer is bigger than 1 hour"—they are not measuring the same kind of thing at all.

Faced with this problem, your first reaction might be—"Why not just standardize all the variables? Transform everything to have mean 0 and standard deviation 1, and then the coefficients become comparable, right?"

This intuition is only one-third correct. Standardization does solve the problem of "comparability across units," but it simultaneously introduces three new problems: the meaning of the coefficients changes, the inference framework changes, and for certain variables, the economic interpretation after standardization becomes difficult to convey to readers. This is precisely why, in empirical research, you far more often see "taking ln" rather than "standardization."

Key takeaway: Standardization (Z-score) transforms all variables into "unit-free" numbers with mean 0 and standard deviation 1—the regression coefficient becomes "for each one-standard-deviation increase in X, Y increases by so many standard deviations." This is useful when comparing the relative importance of different variables within the same model, but the cost is that the coefficient no longer carries the economic meaning in its original units ("one more year of schooling increases wages by X yuan" becomes "one more standard deviation of schooling increases wages by X standard deviations"). Taking ln takes a different path: it does not eliminate unit differences but rather converts "absolute changes" into "relative changes" (percentages), turning coefficients into elasticities and semi-elasticities—this both preserves economic intuition and offers additional benefits in handling skewed distributions and heteroskedasticity. In empirical work, "taking ln" is far more common than "standardization," precisely for this reason.

II. What Does Standardization Actually Do?

2.1 The Z-score Operation

For a variable X, standardization involves only one step:

$Z_i = \frac{X_i - \bar{X}}{s_X}$

where $\bar{X}$ is the sample mean of X and $s_X$ is the sample standard deviation of X.

The standardized Z has two properties:

$\bar{Z} = 0$ (mean of zero)
$s_Z = 1$ (standard deviation of 1)

The essence of standardization is "changing the ruler"—you take income originally measured in "yuan," education measured in "years," and age measured in "years," and convert them all to a unified ruler measured in "how many standard deviations from the mean."

2.2 Regression After Standardization—What Do the Coefficients Become?

Original regression:

$Y = \beta_0 + \beta_1 X + \varepsilon$

After standardization (all variables transformed via Z-score):

$\frac{Y - \bar{Y}}{s_Y} = \beta_1^* \cdot \frac{X - \bar{X}}{s_X} + \varepsilon^*$

Note: the intercept term disappears in the standardized model (because all variables have mean zero, the regression line must pass through the origin).

The relationship between the standardized coefficient $\beta_1^*$ (also denoted as the Beta coefficient in some literature) and the original coefficient $\beta_1$ is:

$\beta_1^* = \beta_1 \cdot \frac{s_X}{s_Y}$

That is, the standardized coefficient = original coefficient × (standard deviation of X / standard deviation of Y).

Interpretation: $\beta_1^* = 0.3$ means—for each one-standard-deviation increase in X, Y is predicted to increase by 0.3 standard deviations. It no longer measures "the effect of a one-unit change in X on Y" but rather "the effect of a one-standard-deviation 'typical magnitude of variation' in X on Y."

III. Can You Standardize All Data in Empirical Work?—Yes, But You Need to Know What You Are Losing

3.1 When Is Standardization Useful?

Scenario 1: Comparing the "relative importance" of different variables within the same model.

This is the core use of standardized coefficients. If you have a model with 10 independent variables whose units are all over the place (yuan, years, percentages, dummy variables), and you want to know which of these 10 variables has the "largest impact on Y"—standardized coefficients provide a common scale that makes such comparisons possible.

Scenario 2: You need to construct interaction terms, and the variables have vastly different scales.

When constructing an interaction term $X_1 \times X_2$ , if X₁ is income (magnitude 10⁴—10⁶) and X₂ is education (magnitude 0—20), the interaction term will be dominated by income—this can create serious numerical computation problems and multicollinearity. Standardizing before constructing the interaction term puts both variables on a similar scale, which can alleviate this issue.

Scenario 3: Certain machine learning models require features to be on a consistent scale.

Methods such as LASSO, Ridge, neural networks, and K-means clustering are sensitive to the scale of variables—without standardization, variables with large magnitudes will dominate the penalty term or distance calculations. In these contexts, standardization is a standard preprocessing step, not an option.

3.2 What Do You Lose After Standardization?

Loss 1: The economic meaning of coefficients—from "marginal effect" to "standard-deviation effect."

This is the biggest cost of standardization. The original coefficient has an intuitive meaning that can be taken directly into policy discussions:

"Each additional year of education increases wages by 8%."

The standardized coefficient means:

"Each one-standard-deviation increase in education increases wages by 0.25 standard deviations."

Can you write in the "policy implications" section of your paper, "We recommend increasing the level of education by one standard deviation"? Policymakers need a translation—"How many years of education is one standard deviation? How much money is 0.25 standard deviations of wages?" When your readers or policymakers cannot intuitively understand your coefficients, the impact of your research is constrained.

Loss 2: The unit-free nature of coefficients makes them incomparable across different samples.

Standardization uses the sample mean and standard deviation. This means:

You run a standardized regression using CGSS 2015 data → $\beta^* = 0.3$
Your classmate runs the same standardized regression using CFPS 2018 data → $\beta^* = 0.2$

These 0.3 and 0.2 cannot be directly compared. This is because the standard deviations of education and wages may differ across the two samples—the standardized coefficient confounds "the magnitude of the effect" with "the degree of dispersion of the variable in the sample." Original coefficients are comparable across samples; standardized coefficients are not fully comparable.

Loss 3: Dummy variables lose their original interpretive meaning after standardization.

The variable you should least standardize is the dummy variable. A variable that only takes values 0/1—such as "gender (female = 1)"—has a standard deviation of $\sqrt{p(1-p)}$ , where p is the proportion of females in the sample. After standardization, "female" becomes a quantity whose value depends on the proportion of females in the sample—this has no economic meaning whatsoever. For dummy variables, either do not standardize them, or standardize only Y and the continuous X variables while keeping the 0/1 coding of the dummy variables.

Loss 4: Standardization masks the absolute magnitude of the effect.

If Y = wages (yuan), X₁ = education (years), $\beta_1 = 2000$ —this number immediately tells you "one more year of schooling means 2000 more yuan per year." This information is meaningful in policy discussions. After standardization, $\beta_1^* = 0.15$ —you cannot recover the 2000 yuan from this number unless the reader additionally knows the standard deviations of wages and education in the sample.

IV. Why Do Researchers Prefer Taking ln Over Standardization in Empirical Work?—Three Hard Reasons

4.1 Reason 1: ln Converts "Absolute Magnitudes" into "Percentage Changes"—Economic Meaning Preserved

Taking logs does not eliminate units, but it changes the interpretation of the coefficient:

Log-log model (ln Y ~ ln X): $\beta_1$ is the elasticity—for each 1% increase in X, Y increases by β₁%. This interpretation does not depend on the original units of X and Y and can be compared across different studies. You do not need to know what "one standard deviation" is.
Semi-log model (ln Y ~ X): $\beta_1$ is approximately the semi-elasticity—for each one-unit increase in X, Y increases by approximately β₁ × 100%.
Log-linear model (Y ~ ln X): $\beta_1$ means that for each 1-unit increase in ln X (i.e., X itself increases by approximately 172%), Y increases by β₁ units. A more commonly used approximation: for each 1% increase in X, Y increases by β₁/100 units.

Key comparison:

Operation	Meaning of the coefficient	Comparable across studies?	Policy communication friendly?
No transformation	For each 1-unit change in X, Y changes by β units	❌ Not comparable due to different units	⚠️ Depends on units
Standardization	For each 1-standard-deviation change in X, Y changes by β* standard deviations	⚠️ Depends on s_X and s_Y of each sample	❌ "One standard deviation" is difficult for policy communication
Taking ln	For each 1% change in X, Y changes by β% (elasticity)	✅ Unit-free, comparable across studies	✅ "1% change" is easy to understand

"Taking ln" puts coefficients into the language of percentage changes—this is the native language of economics. "Each additional year of education increases wages by 8%" is far easier for reviewers, readers, and policymakers to understand than "each one-standard-deviation increase in education increases wages by 0.25 standard deviations."

4.2 Reason 2: ln Pulls Right-Skewed Distributions Toward Symmetry—Also Addressing Heteroskedasticity

Many economic variables are naturally right-skewed—a few people have extremely high incomes, a few firms are extremely large, a few cities have extremely large populations. Under such distributions:

OLS is sensitive to extreme values—a few high-income earners can pull the entire regression line off course.
The variance of the error term may increase with X (heteroskedasticity).

Taking ln "compresses" the extreme values on the right—the gap between 1000 and 10000 is 9000 on the original scale, but on the ln scale it is ln(10000) − ln(1000) = 9.21 − 6.91 ≈ 2.30. This compression makes the distribution more symmetric, reduces the influence of extreme values, and typically alleviates heteroskedasticity.

Standardization cannot achieve this. Standardization only changes the unit of measurement (from "yuan" to "number of standard deviations"); it does not change the shape of the distribution—if the original data are right-skewed, they remain right-skewed after standardization. Standardization is "changing the ruler"; ln is "reshaping the distribution."

4.3 Reason 3: Taking ln Preserves Disciplinary Tradition and Communicability

In labor economics, the Mincer wage equation uses ln(wage) as the dependent variable—half a century of literature has done so. If you use ln(wage), your coefficient (return to education of approximately 0.06—0.10) can be directly compared with the existing literature. If you standardize it, your coefficient becomes "0.2 standard deviations"—no one knows whether this number is large or small relative to the literature.

In macroeconomics, GDP, consumption, and investment almost always enter regressions in ln form. In industrial organization research, firm size, R&D expenditure, and advertising expenditure are almost always logged. Taking ln allows your results to be positioned and compared within disciplinary traditions—this is the infrastructure of scientific communication, not a mere technical convenience.

V. Standardization vs. Taking ln vs. Doing Nothing—A Decision Table

Your goal	Recommended operation	Why?
Compare the relative importance of different variables within the same model	Standardize (report Beta coefficients as supplementary)	Standardized coefficients provide a common scale—but not as a replacement for original coefficients; rather, as supplementary information in an appendix or footnote
Interpret elasticities or semi-elasticities (X changes by 1%, Y changes by X%)	Take ln	Direct economic meaning, comparable across studies
Address right-skewed distributions + heteroskedasticity	Take ln	Equivalent to reshaping the distribution and stabilizing the variance; standardization cannot do this
X and Y already have intuitive economic units (e.g., "years," "%")	Do nothing	The original unit of the coefficient is "years"—you can directly say "one more year of schooling"
X and Y have abstract units or no natural interpretation (e.g., psychological scale scores)	Standardize or do nothing	The original units have no intuitive meaning anyway; standardization is a reasonable choice
Construct interaction terms or polynomial terms	Center first (standardization is sometimes acceptable)	Reduces collinearity after construction, and centering preserves more of the original unit information than standardization
Feature preprocessing for machine learning models	Standardize	Methods such as LASSO and neural networks require features on similar scales
Dummy variables	Never standardize	Standardized dummy variables lose their meaning; their 0/1 coding is the most informationally complete form

VI. An Overlooked but Important Question: Does Inference Change After Standardization?

6.1 Standardization Does Not Affect t-Statistics or p-Values

If you apply a linear standardization transformation (Z-score) to all variables (X and Y), the t-statistics and p-values from the standardized regression are identical to those from the original regression. This is because standardization is merely a linear transformation:

$\hat{\beta}^* = \hat{\beta} \cdot \frac{s_X}{s_Y}, \quad \text{SE}(\hat{\beta}^*) = \text{SE}(\hat{\beta}) \cdot \frac{s_X}{s_Y}$

Both the numerator and denominator are multiplied by $s_X/s_Y$ → the t-value is unchanged. Standardization does not affect your statistical inference—it only changes the scale and interpretation of the coefficients.

6.2 But Standardized Coefficients Cannot Be Used for "Cross-Sample F-tests"

If you have two samples (e.g., males and females) and you want to test whether "the effect of education on wages differs between men and women"—you should test $H_0: \beta_{\text{male}} = \beta_{\text{female}}$ , using the original coefficients (both are on the same "yuan/year" scale and thus comparable).

If you use standardized coefficients, the difference between $\beta^*_{\text{male}}$ and $\beta^*_{\text{female}}$ reflects not only differences in the effect but also confounds differences in the standard deviations of wages and education between the male and female samples—you effectively cannot distinguish between "different effects" and "different dispersions."

6.3 The Intercept Disappears After Standardization—But This Does Not Affect Core Inference

The intercept in a standardized regression is always 0 (because $\bar{Y}^* = 0$ and $\bar{X}^* = 0$ )—there is nothing to worry about here; the intercept is not typically the object of interpretation in most empirical studies anyway.

VII. How Is This Tested in Exams?

7.1 Conceptual Distinction Between Standardization and Taking Logs

"A researcher wants to compare the relative impacts of education (years) and household income (yuan) on individual wages. What treatment should be applied to the variables? How should the coefficients be interpreted after treatment?"

Core answer:

Apply standardization (Z-score) to obtain Beta coefficients.
Interpretation of Beta coefficients: for each one-standard-deviation increase in X, Y increases by that many standard deviations.
Beta coefficients can be compared with each other—because they are all on the same "standard-deviation change" scale.
However, it must be noted: Beta coefficients cannot replace original coefficients—original coefficients preserve economic meaning.

7.2 Interpreting Coefficients After Taking ln

"Model: ln(wage) = β₀ + β₁·education + β₂·ln(experience) + ε. Please explain the economic meanings of β₁ and β₂."

β₁ (semi-elasticity): for each additional year of education, wage increases by approximately β₁ × 100% (exact formula: $(e^{\beta_1} - 1) \times 100\%$ , which approximates β₁ × 100% when |β₁| < 0.2).
β₂ (elasticity): for each 1% increase in experience, wage increases by β₂%.

7.3 Converting Between Standardized and Original Coefficients

"OLS estimation yields $\hat{Y} = 3 + 0.05 X_1 + 2000 X_2$ . Given $s_{X_1} = 4$ , $s_{X_2} = 5000$ , $s_Y = 15000$ . Calculate the standardized coefficients and compare the relative importance of X₁ and X₂."

$\beta_1^* = 0.05 \times \frac{4}{15000} \approx 0.0000133$

$\beta_2^* = 2000 \times \frac{5000}{15000} \approx 666.67$

Note: this result shows that standardized coefficients greatly "favor" variables with large standard deviations—X₂'s original coefficient looks large (2000) because its unit is small (the effect of a 1-yuan increase), while after standardization $\beta_2^*$ becomes unusually large because $s_{X_2}$ is large. This example precisely reveals—standardized coefficients do not measure "whether the effect is large," but rather "how large a standardized effect is produced by a one-standard-deviation change."

In actual exams, the purpose of this type of question is usually to point out that comparisons of standardized coefficients are only truly meaningful when $s_Y$ is the same (obviously, since it is the same Y) and when the $s_{X_j}$ values have economically meaningful comparability.

VIII. Summary

Five core takeaways about standardization:

Standardization = changing the ruler, not the distribution. Standardization transforms each variable into "how many standard deviations from the mean," but the shape of the distribution remains unchanged—right-skewed remains right-skewed, outliers remain outliers.
A standardized coefficient = for each 1-standard-deviation increase in X, Y increases by that many standard deviations. You lose the economic meaning in the original units ("one more year of schooling"), and in exchange you gain comparability across different variables within the same model.
Taking ln = changing the language, reshaping the distribution. ln converts "absolute changes" into "percentage changes" (elasticities/semi-elasticities), while simultaneously pulling right-skewed distributions toward symmetry and alleviating heteroskedasticity—none of these three things can be achieved by standardization.
In empirical work, the reason taking ln is more common than standardization—economic meaning, disciplinary tradition, and distributional handling: three hard reasons. The coefficient from taking ln is a "percentage change," which happens to be the most natural language in economics and the easiest for cross-study communication.
Never standardize dummy variables. The meaning of a standardized dummy variable depends on the proportion of 1s in the sample—this has no substantive meaning. For dummy variables, preserve the 0/1 coding.

One-sentence conclusion:

"Standardization liberates you from the Tower of Babel of units—but the new language it gives you is 'standard deviations,' a language your readers and policymakers do not speak. Taking ln gives you another, more universal language: 'percentage changes.' This is why, in empirical economics, ln almost always appears before standardization."

IX. Presentation Suggestions for Bilibili/WeChat Official Account

Bilibili video: Suggested three acts. Act 1 "The Tower of Units": show three variables on screen—annual household income (magnitude 10⁵ yuan), years of education (magnitude 10 years), age (magnitude 10 years)—each with a ruler of different scale. Narration: "How do you compare these coefficients? You cannot—because they use different rulers." Act 2 "Two Keys": split-screen showing two approaches—on the left, the "key of standardization" replaces the three rulers with one unified ruler (the standard-deviation ruler), and the coefficient changes from "one more year of schooling → 0.08 yuan/hour more" to "one more standard deviation of education → 0.25 standard deviations more in wages." On the right, the "key of taking ln" changes the ruler from "absolute length" to the elasticity language of "relative change"—"1% more education → 0.5% more wages." Act 3 "Why ln Wins": use three animations to show ln's triple benefits—elasticity interpretation (the native language of economics), right-skewed distribution pulled toward symmetry (histogram animation), and heteroskedasticity alleviation (residual plot animation). Closing image: a road sign pointing toward "taking ln" reading "99% of empirical papers go this way." Note: the property that t-values are unchanged can be demonstrated with a simple numerical example—display the t-values from the original and standardized regressions side by side, showing they are identical.
WeChat Official Account: The three-column comparison table of standardization vs. taking ln (operation, coefficient meaning, distributional impact, disciplinary tradition, cross-study comparability) should be made into a horizontal infographic as the centerpiece. The coefficient interpretation formula boxes for log-log/semi-log/log-linear models should each be paired with a numerical example. The four points under "What you lose after standardization" should be made into collapsible cards. The decision table (seven scenarios × three operations × rationale) should be made into a flowchart or matrix diagram. The conversion formula between standardized and original coefficients should be paired with a numerical calculation example.
Recommended titles:
- Main title: 《Can Data Be Standardized in Empirical Analysis? How Do You Interpret Coefficients After Standardization?》
- Alternative title: 《Standardization vs. Taking ln: Why Is ln Everywhere in Empirical Papers While Z-scores Are Rarely Seen?》
- New media title: 《"Standardize All the Data Before Running the Regression"—This Operation Looks Smart, But You May Be Losing the Most Important Thing》
Key quotes:

"Standardization brings all variables onto the same 'standard deviation' track—fair and comparable, but the cost is that you can no longer say plain-language things like 'one more year of schooling,' because your coefficient now speaks of 'one more standard deviation of schooling.'"

"Taking ln does not change the comparability of units—it does something smarter: it shifts your question from 'how many units change' to 'how many percent change.' Percentages are the native language of economics."

"Standardization changes the ruler; taking ln reshapes the distribution—the two solve problems at different levels. Standardization cannot give you normality, and taking ln cannot give you comparability across units. But in empirical work, what is most often needed is the former, and what is most often questioned is the latter—this is why ln wins."

"Never standardize dummy variables—a 'female' variable with mean 0.48 and standard deviation 0.5, once standardized, does not contain more information than a clean 0/1."