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: "Adding More Control Variables" May Be the Most Dangerous Move in Your Paper

In methodological discussions of empirical research, one question is asked with remarkable frequency yet answered systematically with remarkable rarity—

"How should I choose my control variables?"

Many people's default approach is: throw in every variable that might theoretically affect Y. The reasoning is "controlling for more is better than controlling for less—at least reviewers won't accuse me of omitted variables."

This intuition is wrong in some cases, and dangerously so. Not every variable that affects Y should be controlled—some variables, once controlled, not only fail to reduce bias but actually introduce new bias, and can even turn your otherwise clean causal identification into a complete mess.

The choice of control variables depends on your research objective. Research objectives can be broadly divided into two categories—this article will clarify the issue along these two paths.

Core message: If your goal is to "predict Y" or "explore the determinants of Y," control variables can be relatively broad—but you must still be wary of reverse causality and overfitting. If your goal is to "identify the causal effect of X on Y," the choice of control variables must strictly follow the principle of "blocking confounding paths without blocking causal paths." Including the wrong control variable can be more fatal than omitting one.

II. Two Maps: Two Research Objectives, Two Control Logics

Before discussing how to choose specifically, let us first distinguish between the two most common research scenarios.

Scenario A: Studying the determinants of Y (no single core explanatory variable).

Typical paper titles: "Analysis of Factors Influencing Subjective Well-Being of Chinese Residents," "A Study of Determinants of Firm Innovation Performance."

Your goal is to describe as comprehensively as possible which factors are associated with Y and which factors have greater "explanatory power." You do not have a "core coefficient" that needs to be prioritized for identification and protection—all X's are roughly equal in status. This type of research is closer in nature to description or prediction.

Scenario B: Studying the causal effect of X on Y (with a clearly defined core explanatory variable).

Typical paper titles: "Estimating Returns to Education—Based on Instrumental Variables," "The Effect of Minimum Wage on Employment—Evidence from a Natural Experiment."

Your core goal is singular: to unbiasedly estimate the coefficient β₁ of the core variable X. The entire purpose of control variables is to serve this goal—they are invited in to "take the bullets" (blocking confounding paths), not to "steal the show" (they must not interfere with the core causal chain). This type of research is causal identification.

The logic of choosing control variables differs completely across the two scenarios. We address each in turn below.

3.1 What Can You Include?

In description/prediction-oriented research, your goal is to maximize explanatory or predictive power for Y. Therefore:

Any variable that theoretically affects Y can, in principle, be included. Age, gender, education, region, policy environment—as long as theory or existing literature links it to Y, it can enter the model.
Your core focus is not on any single coefficient but on the explanatory power of the overall pattern. Hence there is no need to distinguish between "core variables" and "control variables"—all variables receive "equal treatment" in the model (even though their theoretical importance may differ).

3.2 But There Are Three Things You Cannot Do

First, you cannot include variables that are "affected by Y" (reverse causality).

You study "determinants of firm performance" and control for "whether the firm is listed." But firms with better performance are more likely to go public—listing is not a cause of performance but a consequence. Put it in the regression, and the interpretation of the coefficient becomes extremely difficult: does listing improve performance, or do high-performing firms find it easier to list? These two cannot be disentangled within a single regression coefficient.

Including a reverse-causality variable contaminates the coefficients of all variables correlated with it in the model—because the model is using a variable shaped by Y to "explain" Y.

Second, you cannot include tautological variables.

You study "factors affecting GDP" and include "consumption," "investment," "government expenditure," and "net exports" as independent variables. The sum of these four variables exactly equals GDP (by the expenditure approach)—you are running a regression on an identity. The R² will be astonishingly high, but this is not "explanatory power"; it is accounting. Doing this is equivalent to saying "GDP is GDP because consumption plus investment plus government expenditure plus net exports happens to equal it"—this carries zero information.

Third, multicollinearity does not cause bias, but in small samples it makes coefficient estimates unstable.

If two of your independent variables are highly correlated (e.g., "per capita GDP" and "per capita income"), their coefficient standard errors will inflate—you cannot distinguish "which one is doing the work." In prediction tasks, this may be acceptable (as long as overall predictive power is unaffected), but it becomes problematic when interpreting "the magnitude of each factor's effect."

IV. Scenario B: Studying the Causal Effect of X on Y—"Block Confounders, Do Not Block Causes"

This is the core of the article. For causal identification research, the choice of control variables follows a strict set of principles.

4.1 Causal Graph (DAG) Thinking—Think in Diagrams

Before explaining in words, let us introduce a thinking tool—the Directed Acyclic Graph (DAG). It depicts causal relationships among variables (arrows = direction of causation).

Consider a simple causal structure:

       Ability (Z)
      /          \
     ↓            ↓
  Education (X) → Wage (Y)

Z (ability) affects both X (education) and Y (wage) → Z is a confounder.
X (education) affects Y (wage) → this is the causal path you want to identify.

Now, if your regression includes only X but not Z, the coefficient on X will capture both the "education→wage" path and the "education←ability→wage" path—the coefficient is biased. To obtain an unbiased causal effect, you need to control for Z in the regression.

Each chain in this graph corresponds to a decision of "control or not." Below we unpack the three most common structures one by one.

4.2 Structure One: Confounder—Must Control

Z affects both X and Y. For "education→wage," Z is "ability."

Conclusion: Z must be controlled. Without controlling Z, the coefficient on X is biased—the correlation between X and Y is contaminated by Z's confounding. After controlling Z, X and Y are compared at the same level of Z, and this "backdoor path" (X ← Z → Y) is blocked.

Principle for selecting confounders: Consider all variables that theoretically affect both X and Y and include them as controls. If you are unsure whether a variable affects "both," look for clues in the existing literature or run robustness checks (whether the direction of change in the core coefficient when controlling versus not controlling matches expectations).

4.3 Structure Two: Mediator—Do Not Control

    X → M → Y

X affects Y through M. In the example "education→occupational prestige→wage," occupational prestige M is a mediator.

Conclusion: If you want to estimate the total effect of X on Y—do not control M.

Controlling M is statistically equivalent to "blocking off" a real causal channel through which X affects Y. The coefficient on X you estimate will be only the "direct effect" (X→Y, the path bypassing M), not the total effect you want (direct effect + indirect effect through M).

Exception: If your research question happens to be "after controlling for occupational prestige, does education still have an additional direct effect on wage"—then you do need to control M. But in that case you clearly know you are estimating the direct effect and no longer claim "the total effect of education on wage." This is a distinction at the level of research objectives and must be clearly stated in the paper.

4.4 Structure Three: Collider—Absolutely Do Not Control

    X → C ← Y

C is a common outcome of both X and Y. X and Y are originally uncorrelated (or correlated only through other paths), but if you control C, you will instead create a spurious association between X and Y.

Example: X = talent, Y = effort, C = "admission to an elite university." Talent and effort may be uncorrelated in the population. However—among students admitted to elite universities, talent and effort exhibit negative correlation: admitted students with relatively low talent tend to be exceptionally hardworking; those with extremely high talent but less effort are also admitted. If you control for "whether admitted to an elite university" (i.e., look only at elite students), you will "manufacture" the statistical phenomenon that "talent and effort are negatively correlated," which does not hold in the population.

This is collider bias—after controlling C, the association between X and Y is distorted, and what you obtain is a spurious correlation contaminated by selection effects.

Common colliders in empirical work:

"Whether survived to the survey date"—X and Y may both affect survival probability. If you control for this in an analysis restricted to survivors, you introduce survivor selection bias.
"Whether completed schooling"—when studying the effect of education on income, if your sample contains only graduates, you have implicitly conditioned on the collider "whether graduated."
"Whether participated in the survey"—same logic.

Conclusion: Absolutely do not control. Colliders are the most dangerous trap in control variable selection—because they "look like important control variables" (they are related to both X and Y), but in reality controlling them makes your analysis worse.

4.5 Structure Four: Variables Affecting Only Y, Not X—Controllable but Not Necessary

    X → Y ← W

W affects only Y, not X. For example, when studying "education (X) on wage (Y)," W = "local weather" affects only wage (affecting output in outdoor work such as agriculture and construction) and is unrelated to a person's education level.

Conclusion: Controlling W does not reduce bias (because it is not a confounder), but it may improve precision. Because W explains part of the variation in Y, the residual variance σ² decreases → standard errors fall → t-values increase → real effects become easier to detect.

However, if the sample size is limited, adding unnecessary control variables consumes degrees of freedom, causing standard errors to increase slightly instead. In practice—when the sample is sufficiently large (n > 100, k < 10), these "harmless but beneficial" variables can be included; when the sample is tight, prioritize coverage of confounders rather than minimizing residual variance.

V. Good Controls vs. Bad Controls—A Quick Reference Table

Variable Type	Causal Structure	Should You Control?	Why?
Confounder	X ← Z → Y	✅ Must control	Failure to control leads to omitted variable bias
Variable affecting only Y	X → Y ← W	⭕ Optional	Does not reduce bias, but may improve precision
Variable affecting only X	Z → X → Y	⭕ Optional	Does not affect bias; controlling does not improve precision either
Mediator	X → M → Y	❌ Do not control (if estimating the total effect)	Blocks the real causal path
Collider	X → C ← Y	❌ Absolutely do not control	Creates a spurious X-Y association
Variable affected by Y (Reverse Causality)	Y → V	❌ Cannot control	Coefficient interpretation becomes impossible
Instrumental variable	Z → X → Y (Z not directly related to Y)	❌ Cannot be used as a control variable	If the IV is used as a control, it is no longer an IV
Tautological variable	X + V ≡ Y (identity)	❌ Cannot control	Not explanatory power; it is accounting

VI. Practical Framework—Steps for Choosing Control Variables

If you are studying the causal effect of X on Y (Scenario B), follow these steps:

Step One: Draw the causal graph. Draw all variables you believe are related to X or Y on a DAG. Label the direction of each arrow between each variable and X and Y. This does not require "proving" causation—it merely externalizes your theoretical assumptions onto paper, allowing reviewers and readers to see what your assumptions are.

Step Two: Identify confounding paths. Find all variables that "affect both X and Y." These are the ones you must prioritize controlling.

Step Three: Identify mediator paths and collider structures. Find mediators between X and Y—if you are estimating the total effect, mark them as "do not control." Find common outcomes of X and Y—mark them as "absolutely do not control."

Step Four: Consider practical feasibility. Among the data you can obtain, which confounders are observable? If key confounders are unobservable, you may need to switch identification strategies (IV, DID, RDD, etc.) rather than relying solely on multiple regression.

Step Five: Run robustness checks. Report regression results without control variables (containing only the core X), as well as results with control variables added incrementally. If the core coefficient remains roughly stable across columns, this supports the claim that "confounders have limited influence"; if the coefficient changes dramatically as controls are added, confounding bias is severe.

Step Six: Discuss in the paper—not merely list. Do not fob readers off with "we controlled for age, gender, education, and region"—explain why these particular variables need to be controlled: "We control for education because education simultaneously affects political participation (X) and income level (Y)—in prior literature, education is considered one of the strongest confounders of political participation (citation)." This passage itself demonstrates causal reasoning rather than mechanical listing.

VII. Summary: A Formula to Help You Remember

$\text{Bias} = \text{Effect of the omitted confounder} \times \text{Correlation between the confounder and X}$

The core task of choosing control variables: make this "bias" as close to zero as possible—by including, one by one, variables Z for which $\text{Cov}(Z, X) \neq 0$ and $\beta_Z \neq 0$ .

But simultaneously ensure: do not treat mediators and colliders as confounders to be controlled—on the causal graph, confounders lie "upstream" of X and Y (pointing into X and Y), mediators lie "in between" X and Y (on the causal chain), and colliders lie "downstream" of X and Y (jointly pointed to by X and Y).

Not everything that affects Y belongs in the model. Once certain variables are included, your causal identification goes from "imperfect but defensible" to "completely unreliable."

One sentence to close:

"Choosing control variables is not a technical operation—it is the projection of your causal theory at the variable level. Every arrow you draw on the causal graph ultimately corresponds to a decision of whether a variable in the regression table is controlled or not. Draw the graph wrong, and the control variables will be chosen wrong. Choose the control variables wrong, and the coefficient is not the causal effect you intended."

VIII. Presentation Suggestions for Bilibili/WeChat Official Account

Bilibili video: The three causal structures (Confounder / Mediator / Collider) should be animated—demonstrating the change in association "before vs. after control" for each. For collider bias in particular: use the example of "negative correlation between talent and effort among elite university students" to show how conditioning (looking only at elite students) can manufacture out of thin air a relationship that does not exist in the population.
WeChat Official Account: The good controls vs. bad controls comparison table should be made into a two-column card design with "✅ Green Light / ❌ Red Light." Each of the three DAG structures should be accompanied by a hand-drawn illustration of a concrete example (ability→education→wage, etc.). The six-step practical framework should be made into a vertical flowchart.
Recommended titles:
- Main title: "How to Choose Control Variables in Empirical Research? — Not Everything That Affects Y Belongs in the Model"
- Alternative title: "Including the Wrong Control Variable Is More Fatal Than Omitting One"
Key quote:

"The choice of control variables is the projection of your causal theory at the variable level. Draw the causal graph correctly, and you will know which to control and which not to. Draw the graph wrong, and the meaning of the coefficient will not be the causal effect you imagine."