Correlation vs. Causation — Distinguish correlation from causation; identify confounders - College Political Science

SECTION 1

Historical Context & Motivation

The distinction between correlation and causation stands as one of the most consequential ideas in the history of empirical inquiry. Long before modern statistics formalized the concept, philosophers and early scientists grappled with the problem of attributing cause to observed regularities. In political science, the stakes are especially high: policy prescriptions, institutional reforms, and democratic accountability all hinge on whether analysts can correctly identify causal relationships rather than merely documenting statistical associations. A legislator who confuses a correlation between, say, police spending and crime rates with a causal relationship may advocate for policies that waste resources or even exacerbate the very problem they intend to solve.

1748 • Hume's Problem of Induction

David Hume argued in An Enquiry Concerning Human Understanding that we can never directly observe causation—only the constant conjunction of events. This philosophical challenge remains the foundation of causal inference debates.

1900 • Pearson's Correlation Coefficient

Karl Pearson formalized the product-moment correlation coefficient (r), giving researchers a precise numerical measure of the linear association between two variables. Pearson himself cautioned against interpreting correlations as evidence of causation.

1935 • Fisher and the Randomized Experiment

Ronald Fisher published The Design of Experiments, establishing random assignment as the gold standard for isolating causal effects and eliminating confounding variables.

1986 • Holland's Rubin Causal Model Framework

Paul Holland's influential article 'Statistics and Causal Inference' formalized the potential outcomes framework (often credited to Donald Rubin), giving social scientists a rigorous vocabulary for expressing causal claims in terms of counterfactual outcomes.

2021 • Nobel Prize for Natural Experiments

Joshua Angrist and Guido Imbens shared the Nobel Memorial Prize in Economics for their methodological contributions to causal inference using natural experiments—demonstrating the centrality of the correlation-causation distinction in modern social science.

The central question this lesson addresses is deceptively simple: when two political phenomena move together—rising inequality and declining voter turnout, or media exposure and political polarization—how do we determine whether one actually produces the other, rather than both being driven by a hidden third factor? Answering this question requires both statistical literacy and careful research design, skills that are indispensable for any serious student of political science.

SECTION 2

Core Principles & Definitions

Before moving to applications, it is essential to establish precise definitions. In casual conversation, people often say one thing "causes" another when they merely observe that the two occur together. In political science research methods, we insist on sharper terminology. A correlation exists when two variables systematically co-vary—as one changes, the other tends to change in a predictable direction. Causation requires something far more demanding: a demonstration that a change in one variable directly produces a change in the other, holding all else constant. The gap between these two concepts is where most analytical errors in public policy and political commentary arise.

Correlation

A statistical relationship in which two variables move together (positively or negatively). Correlation does not imply directionality—it simply quantifies the strength and direction of a linear association between X and Y.

Causation

A relationship in which a change in one variable (the independent variable) directly produces a change in another (the dependent variable). Establishing causation requires ruling out alternative explanations, particularly confounders and reverse causality.

Confounding Variable

A third variable that is associated with both the independent and dependent variables, creating a spurious correlation. Confounders make it appear that X causes Y when the real driver is Z, which influences both.

Spurious Correlation

A statistical association between two variables that arises not because one causes the other, but because both are influenced by a confounding variable—or by pure coincidence in a finite sample.

Reverse Causality

A situation in which the presumed direction of causation is backwards. For example, rather than democracy causing economic development, economic development may cause democratization—or both may reinforce each other simultaneously.

✦

KEY TAKEAWAY

Think of correlation like noticing that people carrying umbrellas tend to be in places where the sidewalk is wet. It would be absurd to conclude that umbrellas cause wet sidewalks—the real cause, rain, drives both. In political science, confounders play the role of rain: they are hidden common causes that make two variables appear linked even when neither directly influences the other. Every time you encounter a statistical association in a research article or policy report, your first question should be: what unmeasured factor could be driving both variables?

SECTION 3

Visual Explanation — How Confounders Create Spurious Associations

This diagram illustrates the three primary causal structures that can produce an observed correlation between variables X and Y. Panel A shows genuine causation where X directly produces Y. Panel B shows confounding, where a hidden third variable Z drives both X and Y, creating a spurious association (dashed red line). Panel C shows reverse causality, where the causal arrow actually runs from Y to X. The bottom row applies all three structures to a single political science example: the observed positive correlation between police presence and crime rates.

The diagram above is the single most important visual in this lesson. Notice that all three causal structures—true causation, confounding, and reverse causality—produce the exact same statistical correlation between X and Y. This is precisely why correlation alone can never establish causation. To distinguish among these structures, researchers must draw on theory, research design (especially random assignment or quasi-experimental methods), and careful reasoning about potential confounders. The police-crime example in the bottom panel is a classic case: an uncritical reading of the data might conclude that hiring more police officers increases crime, when in reality high-crime areas simply demand larger police forces (reverse causality), or urbanization drives both crime rates and police hiring simultaneously (confounding).

SECTION 4

The Statistical and Logical Framework

While political science is not exclusively a quantitative discipline, a basic understanding of the statistical machinery behind correlation and causal inference is essential. The Pearson correlation coefficient (r) provides the standard measure of linear association, while causal inference relies on a distinct logical framework rooted in counterfactual reasoning.

PEARSON CORRELATION COEFFICIENT

r = Σ[(Xᵢ − X̄)(Yᵢ − Ȳ)] / √[Σ(Xᵢ − X̄)² × Σ(Yᵢ − Ȳ)²]

Where Xᵢ and Yᵢ are individual observations, X̄ and Ȳ are the sample means, and Σ denotes summation over all observations. The coefficient r ranges from −1 (perfect negative linear relationship) to +1 (perfect positive linear relationship), with 0 indicating no linear association.

It is critical to recognize that r measures only linear association. Two variables can have a strong nonlinear relationship yet produce an r near zero. Moreover, even a very high r value (say, r = 0.95) tells us nothing about whether X causes Y, Y causes X, or some confounder Z drives both. The correlation coefficient is descriptive, not causal.

CAUSAL EFFECT (POTENTIAL OUTCOMES FRAMEWORK)

τᵢ = Yᵢ(1) − Yᵢ(0)

Where τᵢ is the individual causal effect for unit i, Yᵢ(1) is the outcome when unit i receives the treatment, and Yᵢ(0) is the outcome when unit i does not receive the treatment. The fundamental problem of causal inference is that we can never observe both Yᵢ(1) and Yᵢ(0) for the same unit—this is the counterfactual we must estimate.

AVERAGE TREATMENT EFFECT (ATE)

ATE = E[Yᵢ(1)] − E[Yᵢ(0)]

The ATE averages the individual causal effects across the population. Random assignment ensures that the treatment and control groups are comparable in expectation, allowing the simple difference in group means to serve as an unbiased estimate of the ATE.

Mill's Criteria for Causal Inference

Beyond the statistical framework, political scientists often invoke three logical requirements for a causal claim, derived from John Stuart Mill's methods of experimental inquiry. First, there must be covariation: X and Y must be correlated. Second, there must be temporal precedence: the cause X must precede the effect Y in time. Third, there must be non-spuriousness: the association must not disappear when confounding variables are controlled for. Correlation satisfies only the first criterion; establishing causation requires all three.

SECTION 5

Identifying and Classifying Confounders

Of the three threats to causal inference—confounding, reverse causality, and measurement error—confounding is the most pervasive in observational political science research. A confounding variable must satisfy two conditions simultaneously: it must be correlated with the independent variable (the presumed cause) and it must independently affect the dependent variable (the presumed effect). If a variable satisfies only one of these conditions, it is not a true confounder. Careful researchers construct directed acyclic graphs (DAGs) to map out potential confounders before conducting any statistical analysis, a practice that forces theoretical clarity about the causal mechanisms at work.

This framework applies the confounder identification process to a classic political science question: whether democracy causes economic growth. Education level and colonial history satisfy both criteria for valid confounders (correlated with both the independent and dependent variables), while latitude remains debated. The checklist at the bottom provides a reusable template for any research question.

Potential Confounder	Relationship with Democracy (X)	Relationship with GDP Growth (Y)	Verdict
Education Level	Higher education is associated with greater demand for democratic governance	Human capital drives productivity and economic growth	Valid confounder
Colonial History	British colonies more likely to develop democratic institutions	Colonial institutions shaped property rights and long-term growth trajectories	Valid confounder
Geographic Latitude	Weak correlation; mechanism debated	Correlates with disease burden, which affects growth	Debated
Oil Wealth	Resource curse literature links oil to authoritarianism	Oil revenue provides growth but may distort economic structure	Valid confounder

Common confounders in the democracy–economic growth debate

⚠

Mediators vs. Confounders

A common mistake is confusing a mediator with a confounder. A mediator lies on the causal pathway from X to Y (e.g., democracy → property rights → economic growth). If you statistically control for a mediator, you remove part of the causal effect you are trying to estimate, which is called over-controlling. Only control for variables that are not on the causal pathway between X and Y.

SECTION 6

Worked Example — Does Social Media Use Cause Political Polarization?

Consider a scenario drawn from contemporary political science research. A study of 2,000 American adults finds a positive correlation (r = 0.42) between daily hours spent on social media and a composite measure of political polarization (on a 1–10 scale). A newspaper headline declares: "Social Media Is Making America More Polarized." Let us work through the analytical steps necessary to evaluate this claim.

Evaluating the Social Media–Polarization Correlation

1
Step 1 — Confirm the Correlation

The study reports r = 0.42, which indicates a moderate positive correlation: as social media use increases, polarization scores tend to increase as well. This satisfies Mill's first criterion—covariation. However, we note that r² = 0.18, meaning social media use accounts for only about 18% of the variance in polarization scores.

Covariation confirmed: r = 0.42, r² = 0.18
2
Step 2 — Assess Temporal Precedence

The study uses cross-sectional data—all variables were measured at a single point in time. This means we cannot determine whether social media use preceded increased polarization or whether already-polarized individuals seek out more social media content that reinforces their views. Without longitudinal data or an experimental design, temporal precedence is unestablished.

Temporal precedence: NOT established (cross-sectional data)
3
Step 3 — Identify Potential Confounders

We brainstorm variables that could plausibly correlate with both social media use and political polarization. Age is one candidate: younger adults use social media more and may also exhibit different polarization patterns. Political interest is another: politically interested individuals may both consume more political media online and hold stronger ideological positions. Education, partisanship strength, and geographic location (urban vs. rural) are additional potential confounders.

Confounders identified: age, political interest, education, partisanship, geography
4
Step 4 — Consider Reverse Causality

It is entirely plausible that the causal arrow runs in the opposite direction: people who are already politically polarized may gravitate toward social media platforms where they can engage with like-minded communities and consume ideologically congruent content. This selection effect would produce the same positive correlation observed in the data without social media being the cause.

Reverse causality: plausible (polarized people self-select into social media)
5
Step 5 — Evaluate the Causal Claim

Given that temporal precedence is unestablished, multiple confounders remain uncontrolled, and reverse causality is plausible, the newspaper headline's causal claim is not supported by this study design. A more accurate statement would be: "Social media use is positively associated with political polarization, but the causal direction and role of confounders require further investigation using experimental or quasi-experimental methods." To strengthen causal inference, researchers could conduct a randomized controlled trial in which participants are randomly assigned to reduce social media use, as Allcott et al. (2020) did in their influential study.

Conclusion: Correlation established, causation NOT demonstrated

SECTION 7

Research Designs for Causal Inference — Strengths and Limitations

Political scientists employ a range of research designs to move beyond correlation and toward credible causal inference. Each design involves trade-offs between internal validity (the confidence that X actually causes Y within the study) and external validity (the extent to which findings generalize to other populations and contexts). Understanding these trade-offs is essential for critically evaluating empirical research in the discipline.

Research Design	How It Addresses Confounding	Strengths	Limitations
Randomized Controlled Trial (RCT)	Random assignment ensures treatment and control groups are equivalent in expectation, eliminating all confounders (observed and unobserved)	Highest internal validity; gold standard for causal inference	Often infeasible or unethical in political contexts (cannot randomly assign regime types); limited external validity
Natural Experiment	Exploits as-if random variation created by an external event (lottery, policy threshold, weather) to approximate random assignment	Leverages real-world variation; higher ecological validity than lab experiments	Requires convincing argument that the variation is truly "as-if" random; often narrow in scope
Regression with Controls	Statistically holds measured confounders constant by including them as covariates in a regression model	Widely applicable; can use existing datasets; relatively straightforward	Cannot control for unobserved confounders; results sensitive to model specification; omitted variable bias
Difference-in-Differences (DiD)	Compares changes over time between a treatment group and a control group, removing time-invariant confounders	Controls for all stable unobserved differences between groups	Requires parallel trends assumption; vulnerable to time-varying confounders
Instrumental Variables (IV)	Uses a variable (instrument) that affects X but has no direct effect on Y, isolating the exogenous variation in X	Can address both confounding and reverse causality simultaneously	Valid instruments are difficult to find; exclusion restriction is untestable; estimates local average treatment effect (LATE), not ATE

Research designs for causal inference in political science

✦

KEY TAKEAWAY

Think of causal inference as detective work. A correlation is like finding the suspect at the crime scene—it is a clue, but not proof. An RCT is like having security camera footage of the crime itself. A natural experiment is like finding an eyewitness whose testimony is credible because they had no stake in the outcome. Regression with controls is like cross-examining the suspect while accounting for known alibis—useful, but always vulnerable to the alibi you did not think to check (omitted variable bias). The strength of your causal claim depends entirely on the quality of your research design.

SECTION 8

Connection to Advanced Causal Inference Theory

The foundational concepts of correlation, causation, and confounding connect directly to more advanced frameworks that form the frontier of causal inference in political science and the broader social sciences. Understanding where these basic concepts fit within the larger methodological landscape prepares you for upper-division and graduate-level research design courses.

Foundational Concept	Advanced Extension	Key Insight
Confounding variable	Directed Acyclic Graphs (DAGs)	Judea Pearl's graphical framework provides formal rules (d-separation) for determining which variables to control for and which to leave unadjusted
Potential outcomes (τᵢ)	Heterogeneous Treatment Effects	The ATE may mask important variation: democracy may boost growth in some countries but not others. CATE (Conditional ATE) and machine learning methods allow exploration of treatment effect heterogeneity
Reverse causality	Simultaneous Equations / Granger Causality	When causation runs in both directions simultaneously (democracy ↔ growth), structural equation models and time-series methods attempt to disentangle reciprocal causation
Spurious correlation	Sensitivity Analysis / Bounds	Methods like Oster (2019) and Cinelli & Hazlett (2020) formally quantify how strong an unobserved confounder would need to be to fully explain away an estimated causal effect

From foundational concepts to advanced causal inference methods

The rapid methodological advances of the past two decades—often called the credibility revolution in the social sciences—have fundamentally raised the bar for causal claims. Whereas political scientists in the 1980s might have reported a regression coefficient and interpreted it causally, contemporary standards demand explicit identification strategies that address confounding, reverse causality, and selection bias. The correlation-causation distinction is not merely an introductory lesson; it is the conceptual foundation upon which the entire modern edifice of causal inference is built.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A researcher finds a strong positive correlation (r = 0.72) between a country's military spending as a percentage of GDP and its number of international conflicts over the past decade. Does this correlation demonstrate that higher military spending causes more international conflicts? Explain your reasoning using the three criteria for causal inference.

Show answer

No, the correlation alone does not demonstrate causation. While the first criterion (covariation) is satisfied by the positive r = 0.72, the other two criteria remain unmet. Temporal precedence is ambiguous: countries may increase military spending in response to existing conflicts (reverse causality), meaning conflicts precede spending increases. Non-spuriousness is also unestablished: a confounding variable such as a country's geographic location in a conflict-prone region could independently drive both higher military spending and more conflicts. Additionally, regime type could confound the relationship, as authoritarian governments may simultaneously invest heavily in the military and pursue aggressive foreign policies. To establish causation, one would need a research design—such as an instrumental variable or a natural experiment—that isolates the exogenous component of military spending variation.

Question 2: BASIC CALCULATION

A study reports that the correlation between voter turnout and median household income across 50 U.S. states is r = 0.58. Calculate r² and interpret it in the context of this research question. What proportion of the variation in voter turnout remains unexplained by income alone?

Show answer

The coefficient of determination is r² = (0.58)² = 0.3364, or approximately 33.6%. This means that about 33.6% of the variation in voter turnout across states can be statistically accounted for by differences in median household income. Conversely, approximately 66.4% of the variation in voter turnout remains unexplained by income alone. This substantial unexplained variation suggests that many other factors—such as state electoral laws, demographic composition, political competitiveness, and civic culture—also influence turnout. Critically, even the 33.6% explained variance does not imply that income causes higher turnout; the association could be driven by confounders like education levels, which are correlated with both income and political participation.

Question 3: INTERMEDIATE

A political scientist hypothesizes that proportional representation (PR) electoral systems cause higher levels of women's representation in national legislatures. She runs a cross-national regression and finds that PR systems are associated with 12 percentage points more women in parliament compared to majoritarian systems (p < 0.01). Identify at least three plausible confounding variables and explain how each could bias the estimated effect. For each confounder, specify the likely direction of bias.

Show answer

Three plausible confounders include: (1) Cultural attitudes toward gender equality—Nordic countries tend to have both PR systems and progressive gender norms, so the positive effect of PR on women's representation may be partially attributed to cultural factors rather than the electoral system itself, biasing the estimate upward. (2) Level of economic development—wealthier nations are more likely to adopt PR systems and to have higher women's representation due to greater educational attainment and labor force participation among women, again biasing the estimate upward. (3) Left-party strength—countries with strong social democratic parties are more likely to adopt PR systems and to implement gender quotas, meaning the observed association may partly reflect party ideology rather than electoral rules per se (upward bias). To address these confounders, the researcher could add control variables, use a difference-in-differences design comparing countries before and after switching to PR, or seek an instrumental variable that predicts electoral system adoption but does not independently affect women's representation.

Question 4: APPLIED

In 2020, Allcott, Braghieri, Enikopolov, and Gentzkow published a study in which they paid Facebook users to deactivate their accounts for four weeks before the U.S. midterm elections. They found that deactivation reduced political polarization by 0.16 standard deviations. Explain why this study design provides stronger evidence for a causal claim than a cross-sectional survey finding the same correlation. What is the key design feature that addresses confounding, and what limitations remain?

Show answer

This study provides stronger causal evidence because it uses random assignment to create treatment (deactivation) and control (continued use) groups. Random assignment is the key design feature: it ensures that, in expectation, the two groups are identical on all observed and unobserved characteristics—eliminating confounding. This means differences in polarization between groups can be attributed to Facebook deactivation rather than to pre-existing differences in personality, political interest, education, or any other confounder. The study also establishes temporal precedence, since deactivation occurs before the outcome is measured. However, limitations remain. First, compliance may be imperfect—some participants in the treatment group may have reactivated their accounts. Second, external validity is limited: participants who agreed to deactivate for payment may differ from the general population. Third, the study captures only short-term effects; long-term consequences of social media use may differ. Finally, the study measures the effect of complete deactivation, which may not generalize to marginal reductions in usage.

Question 5: CRITICAL THINKING

Consider the following claim: "Since we can never run a true randomized experiment on most political phenomena (we cannot randomly assign countries to be democracies or dictatorships), causal inference in political science is fundamentally impossible." Construct a rigorous counterargument to this claim, drawing on at least two quasi-experimental methods discussed in this lesson. For each method, provide a concrete political science example and explain how it approximates the logic of random assignment without actually randomizing.

Show answer

The claim overstates the limitation. While true randomization of political phenomena like regime type is indeed impossible, quasi-experimental methods allow researchers to approximate the logic of random assignment by exploiting naturally occurring variation. First, instrumental variables (IV) can address the democracy-growth question: Acemoglu, Johnson, and Robinson (2001) used historical settler mortality rates as an instrument for institutional quality, arguing that European colonizers established different institutions in places where they could settle permanently versus places where high mortality prevented settlement. Because settler mortality was determined by geographic and epidemiological factors unrelated to modern economic growth except through its effect on institutions, it approximates exogenous variation in institutional quality—mimicking random assignment of institutions. Second, regression discontinuity designs (RDD) exploit arbitrary thresholds: Lee (2008) studied the incumbency advantage in U.S. House elections by comparing candidates who barely won to those who barely lost, arguing that near the 50% vote threshold, the outcome is effectively random. This quasi-random assignment allows credible causal estimates of how holding office affects future electoral performance. Both methods achieve causal identification not through randomization but through carefully argued sources of exogenous variation, demonstrating that causal inference—while more difficult in observational settings—is far from impossible in political science.

SUMMARY

Lesson Summary

Summary

This lesson established the critical distinction between correlation (a statistical association between two variables) and causation (a relationship in which one variable directly produces a change in another). We traced the intellectual history of this distinction from Hume's skepticism about induction through Pearson's correlation coefficient, Fisher's experimental designs, and the modern potential outcomes framework. Establishing causation requires satisfying three criteria: covariation, temporal precedence, and non-spuriousness—of which correlation satisfies only the first.

The most common threat to causal claims in political science is confounding—a third variable that correlates with both the independent and dependent variables, producing a spurious correlation. Reverse causality is equally dangerous: the causal arrow may run opposite to the researcher's assumption. To move beyond correlation, political scientists employ a toolkit of research designs—including randomized controlled trials, natural experiments, difference-in-differences, and instrumental variables—each providing a different strategy for isolating causal effects. The ability to distinguish correlation from causation and to identify confounders is not merely an academic exercise; it is the skill that separates rigorous political analysis from uninformed speculation.