Correlation vs Causation: Understanding the Critical Difference

Correlation vs Causation: The Difference That Changes Everything

Ice cream sales go up. Shark attacks go up. Both happen in summer. Does eating ice cream attract sharks? Obviously not. But this is exactly how correlation vs causation confusion plays out in real life — sometimes with absurd results, and sometimes with devastating consequences. The ability to tell a genuine causal relationship apart from a statistical coincidence is one of the most valuable skills in science, medicine, policy, journalism, and academic research.

Two variables can move together perfectly without one causing the other. They might both be driven by a third factor. They might share a coincidental pattern in a particular dataset. Or one might actually cause the other — but only careful study design can tell you which is true. Correlation does not imply causation is not just a statistical slogan. It is a foundational principle of scientific reasoning that protects us from bad decisions built on misleading data.

This article covers the full picture. You will learn what correlation and causation actually mean, why the difference matters, how confounding variables and spurious correlations fool researchers, what tools researchers use to establish genuine causation, and the historical and contemporary examples that have shaped how we think about this problem. Whether you are writing a research paper or trying to read a news headline critically, understanding this distinction changes how you see almost everything.

What Is Correlation? A Clear Definition

Correlation is a statistical measure that describes the degree to which two variables move together. When one goes up and the other tends to go up too, that is a positive correlation. When one goes up and the other tends to go down, that is a negative correlation. When there is no consistent pattern between them, they are uncorrelated.

The most common measure of correlation is the Pearson correlation coefficient, represented as r. It ranges from -1 to +1. An r of +1 indicates a perfect positive linear relationship. An r of -1 indicates a perfect negative linear relationship. An r of 0 means no linear relationship. Correlation is purely descriptive. It quantifies a pattern in data. It says nothing about why that pattern exists. For a deeper look at the statistical foundation, see this guide on simple linear regression, which uses correlation as its backbone.

What Is Causation? A Clear Definition

Causation — or a causal relationship — means that one variable directly produces a change in another. The cause must precede the effect in time. There must be a plausible mechanism through which the cause produces the effect. And the relationship must hold up when other explanations are ruled out. Causation is not something you can read directly off a correlation coefficient. It requires evidence of mechanism, temporal order, and the elimination of alternative explanations.

What Is the Difference Between Correlation and Causation?

The difference is fundamental. Correlation describes a pattern. Causation describes a mechanism. A correlation tells you that two things happen together. Causation tells you that one thing makes the other happen. Every causal relationship produces a correlation — but not every correlation reflects a causal relationship. That is the asymmetry that makes this distinction so important.

Think of it this way: if you observe that people who carry lighters are more likely to develop lung cancer, that is a genuine correlation. But lighters do not cause lung cancer. The confounder is smoking — people who smoke both carry lighters and develop lung cancer at higher rates. Hypothesis testing in statistics helps researchers systematically evaluate whether an observed association is likely real — but even a statistically significant result only confirms the pattern, not the causal story behind it.

Types of Correlation and How They Are Measured

Before you can understand why correlation vs causation matters, you need to understand what kinds of correlation exist and how statisticians measure them. Not all correlations are the same. Some are strong. Some are weak. Some describe linear relationships. Others describe rank-order relationships. Each type has a specific measure and a specific interpretation.

Positive Correlation

In a positive correlation, both variables increase together. As one goes up, the other tends to go up as well. The classic academic example: study hours and exam scores tend to be positively correlated. The more a student studies, the higher their score tends to be. But even here, the relationship is not perfectly causal — intelligence, prior knowledge, study quality, and test anxiety are all confounders that affect exam scores independently of raw study time.

Positive correlations produce an upward-sloping trend on a scatter plot. The Pearson correlation coefficient for a positive relationship is greater than zero. The closer to +1, the stronger the relationship. Understanding how these coefficients are computed is foundational to any statistics course. The regression analysis framework builds directly on correlation to model these relationships predictively.

Negative Correlation

In a negative correlation, one variable increases while the other decreases. Absenteeism and grade point average tend to be negatively correlated — students who miss more class tend to earn lower GPAs. Again, confounders exist: students who miss class may do so because of health issues, personal crises, or disengagement from academics, any of which independently affect academic performance.

Zero or No Correlation

When two variables show no consistent pattern of movement relative to each other, they have a correlation near zero. Shoe size and SAT score, for example, are uncorrelated. Knowing someone’s shoe size tells you nothing about their test performance. A correlation of zero means the variables are statistically independent in their linear relationship — though non-linear relationships may still exist.

Spurious Correlation

A spurious correlation is an apparent statistical relationship between two variables that share no genuine causal link. Both may be driven by a common third variable, or the association may be a coincidental artifact of the particular dataset. The website Spurious Correlations by Tyler Vigen has documented hundreds of striking examples — including the near-perfect correlation between U.S. per capita cheese consumption and the number of people who died by becoming tangled in their bed sheets. The correlation coefficient in this case is extremely high. The causal relationship is nonexistent.

Pearson vs Spearman: Measuring Correlation

The Pearson correlation coefficient (r) measures the strength and direction of the linear relationship between two continuous variables. It assumes that both variables are normally distributed and that the relationship between them is linear. The Spearman rank correlation coefficient (ρ) is a non-parametric alternative. It measures the monotonic relationship between two variables and is used when data are ordinal, non-normal, or when the relationship may be non-linear. Both measure association. Neither measures causation. For a full treatment of how these fit into broader statistical modeling, the guide on regression model assumptions is essential reading.

Why Does Correlation Not Imply Causation?

This is the question at the heart of correlation vs causation. Understanding why a statistical relationship — even a strong and consistent one — does not prove causation is not just a textbook exercise. It is the lens through which every research finding should be evaluated. There are four main reasons why correlation fails to establish causation, and each one has real-world consequences.

Reason 1: Confounding Variables

A confounding variable (also called a confounder or lurking variable) is a third variable that influences both of the variables you are studying, creating an apparent relationship between them that does not reflect a direct causal link. Confounders are the single most common explanation for misleading correlations in observational research.

The classic public health example: countries with higher rates of television ownership tend to have lower rates of certain diseases. Does television prevent disease? No. Wealth is the confounder. Wealthier countries have more televisions and better healthcare. The correlation between television ownership and health outcomes is driven entirely by national wealth — neither variable causes the other. Causal inference methods have been developed specifically to detect and account for confounders in observational data.

Reason 2: Reverse Causation

Reverse causation occurs when researchers assume the causal direction is A → B, but the true direction is actually B → A. A famous economic example: countries with strong stock markets tend to have higher GDP. But does GDP drive stock market performance, or does stock market performance drive GDP? The relationship is real. The causal direction is contested and probably bidirectional. In medicine, researchers once observed that people with lower cholesterol were more likely to have cancer. Initial interpretations suggested high cholesterol might protect against cancer. The actual explanation was reverse causation — people developing cancer were losing weight and consequently had lower cholesterol. The disease caused the low cholesterol, not the other way around.

Reason 3: Coincidence and Chance

Some correlations exist purely by chance in a given dataset. The Nicholas Cage problem — a near-perfect correlation between Nicolas Cage film releases and pool drowning deaths in the United States across a particular decade — is perhaps the internet’s most cited example of a purely coincidental correlation. Both variables fluctuated over the same time period. There is no mechanism, no confounder, no causal story. It is noise.

With large enough datasets and enough variables, spurious correlations are mathematically inevitable. A study testing 20 hypotheses at a p-value threshold of 0.05 would expect to find one false positive purely by chance. Understanding confidence intervals and the logic of statistical significance helps contextualize when an observed association is likely to be real versus likely to be noise.

Reason 4: Selection Bias

Selection bias occurs when the sample being studied is not representative of the population of interest, creating apparent correlations that do not hold in the broader population. The classic historical example is the 1936 Literary Digest poll that predicted Alf Landon would defeat Franklin Roosevelt in a landslide — based on over two million responses. The problem: the survey was sent to telephone and car owners, who were disproportionately wealthy and Republican. The sample was heavily biased, the correlation between sample opinion and actual vote was misleading, and Roosevelt won by a historic margin.

Famous Real-World Examples of Correlation vs Causation

The correlation vs causation distinction is most powerfully illustrated through concrete examples. The following cases come from medicine, economics, social science, and public health — and each one shows both what can go wrong when the distinction is ignored and what good causal reasoning looks like when it is applied correctly.

Smoking and Lung Cancer: How Causation Was Eventually Established

In the mid-20th century, the link between smoking and lung cancer was observed as a correlation in epidemiological studies. Tobacco industry representatives — and some scientists — argued that correlation alone could not prove causation. Maybe anxious people both smoked and developed cancer. Maybe genetics caused both the desire to smoke and the propensity for cancer. Maybe the correlation was confounded by urbanization or air pollution.

Austin Bradford Hill, a British epidemiologist at the London School of Hygiene and Tropical Medicine, developed his nine criteria for causal inference in 1965 specifically to address this challenge. His criteria — strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, and analogy — provided a framework for concluding that the smoking-lung cancer relationship was causal even without a randomized controlled trial. His work, along with that of Richard Doll and A.B. Hill in the British Doctors Study, remains one of the most consequential applications of causal reasoning in the history of public health. The study is available through the original BMJ publication.

Ice Cream and Drowning: The Classic Confounder Example

Ice cream sales are positively correlated with drowning deaths. Both peak in summer. The confounder is warm weather: people swim more and buy more ice cream when it is hot. Neither causes the other. This example appears in virtually every introductory statistics course in the United States and the United Kingdom precisely because it illustrates the confounder mechanism so clearly. The lesson: before concluding that A causes B, always ask what third variable might be driving both.

Shoe Size and Reading Ability in Children

Young children with larger shoe sizes tend to have better reading ability. Does shoe size cause reading improvement? No. Age is the confounder. Older children have larger feet and have had more years of reading instruction. Once you control for age, the correlation between shoe size and reading disappears entirely. This is a textbook illustration of how controlling for confounders changes the interpretation of an apparent relationship.

The Firefighter Paradox: More Firefighters, More Damage

Across cities, the number of firefighters dispatched to a fire is positively correlated with the amount of damage caused by that fire. Cities that send more firefighters to fires have more fire damage — on average. Does this mean firefighters cause damage? No. The severity of the fire determines both how many firefighters are sent and how much damage results. Severity is the confounder. This example is particularly instructive for public policy: naive correlation-based reasoning would lead to the absurd conclusion that reducing fire department response would reduce fire damage.

Hormone Replacement Therapy: A Cautionary Tale in Medicine

For decades, observational studies showed that women who used hormone replacement therapy (HRT) had lower rates of heart disease than those who did not. Many physicians concluded that HRT protected the heart. When the Women’s Health Initiative — a large randomized controlled trial conducted by the National Institutes of Health — tested HRT directly, it found the opposite: HRT was associated with increased risk of heart disease in certain populations. The observational data had been confounded by the “healthy user effect” — women who chose HRT tended to be healthier, more health-conscious, and from higher socioeconomic backgrounds than those who did not. These factors independently reduced their heart disease risk, creating a spurious protective correlation. This remains one of the most cited examples of why data type and study design determine what conclusions are valid.

How Do Researchers Establish Causation?

If correlation is not enough, what does it take to establish causation? This is one of the central questions in research design. The short answer: it takes careful study design, the right methods, and — in many cases — convergent evidence from multiple approaches. Here are the primary methods researchers use to move from association to causation.

Confounding Variables, Selection Bias, and Common Research Pitfalls

The most consequential threats to valid causal inference in research are confounding variables, selection bias, and measurement error. Each one can create or distort correlations in ways that make a non-causal relationship look causal — or hide a real causal relationship under noise. Understanding these threats is essential for reading research critically and for designing studies that produce trustworthy conclusions.

What Is a Confounding Variable?

A confounding variable is any variable that is causally related to both the independent variable and the dependent variable you are studying. Because it influences both, it creates a statistical relationship between them that does not reflect any direct causal connection. Confounders are the primary reason that observational studies cannot establish causation without additional methods. Researchers control for confounders by including them in regression models, by matching, by stratification, or by using the design-based methods described above.

The Healthy User Effect

The healthy user effect is one of the most pervasive confounders in medical research. People who comply with medical recommendations — taking vitamins, exercising, following prescribed treatments — tend to be healthier overall and engage in fewer risky behaviors than those who do not comply. This means that observational studies of any health behavior frequently produce inflated estimates of benefit. The HRT-heart disease example discussed earlier is the canonical case. Recognizing the healthy user effect is essential when reading any observational health research.

Berkson’s Paradox

Berkson’s paradox (named after Joseph Berkson of the Mayo Clinic) is a form of selection bias that occurs specifically in hospital-based studies. If two diseases are independently distributed in the general population but both cause hospitalization, then a hospital study will find a spurious negative correlation between them — simply because patients with one disease are less likely to also have the other (they are hospitalized for the first one). This statistical artifact has caused researchers to draw incorrect conclusions about the relationships between diseases when studying hospital samples rather than population samples.

Simpson’s Paradox

Simpson’s paradox is perhaps the most counterintuitive finding in statistics. It occurs when a trend appears in several groups of data but disappears — or reverses — when these groups are combined. The University of California, Berkeley admissions case is the most famous example: aggregate data showed that men were admitted at a higher rate than women, suggesting sex discrimination. But when applicants were broken down by department, women were admitted at higher rates than men in most departments. The aggregate correlation reversed because women applied in disproportionately large numbers to the most competitive departments. Understanding Simpson’s paradox is crucial for anyone doing descriptive and inferential statistics without stratifying appropriately.

The paradox has practical implications in medical research and policy. A treatment may appear effective in aggregate data but harmful when individual patient groups are analyzed separately — or vice versa. Stanford’s Encyclopedia of Philosophy provides a rigorous treatment of Simpson’s paradox and its implications for causal reasoning.

Logical Fallacies Related to Correlation and Causation

The misinterpretation of correlational data is not just a statistical error — it is a logical one. Several named logical fallacies describe the specific ways that people incorrectly reason from correlation to causation. Understanding these fallacies makes you a better critical thinker and a more careful reader of research. They appear constantly in journalism, politics, and everyday conversation.

Post Hoc Ergo Propter Hoc

Post hoc ergo propter hoc is Latin for “after this, therefore because of this.” This fallacy concludes that because one event followed another, the first caused the second. A student eats a specific breakfast, then aces an exam, and concludes the breakfast caused the success. A government introduces a policy, the economy improves, and the government claims credit. Temporal sequence is a necessary condition for causation — but it is not sufficient. Many things happen after other things without being caused by them. The scientific method was specifically developed to move beyond post hoc reasoning by requiring controlled experiments and systematic evidence.

Cum Hoc Ergo Propter Hoc

Cum hoc ergo propter hoc means “with this, therefore because of this.” This is the classic correlation-causation fallacy: concluding that because two things occur together, one must cause the other. Countries with more television sets have lower birth rates — therefore television prevents births? No. Wealth is the confounder. Two things happening simultaneously or correlating statistically tells you nothing about which, if either, causes the other.

The Ecological Fallacy

The ecological fallacy occurs when correlations observed at the group or aggregate level are incorrectly applied to individual cases. A country may have high rates of both red wine consumption and low heart disease rates. But it does not follow that any individual who drinks red wine will have lower heart disease risk. The relationship observed at the population level may be driven entirely by confounders that operate at the group level and have nothing to do with the mechanism of red wine on individual physiology. Committing the ecological fallacy in academic assignments — particularly in sociology, public health, and economics — is a significant analytical error. The guide on critical thinking in assignments addresses how to avoid this class of mistake.

The Texas Sharpshooter Fallacy

The Texas sharpshooter fallacy involves identifying clusters of data after the fact and then arguing they represent meaningful patterns. The name comes from the image of a shooter who fires at a barn wall, then draws a target around the bullet holes and claims to be an excellent shot. In statistics, this describes data mining without hypothesis pre-registration: searching through large datasets for any pattern that achieves statistical significance, then reporting it as if it were a pre-planned finding. The Open Science Foundation and journals like Nature and Science have pushed for pre-registration of research hypotheses specifically to prevent Texas sharpshooter analysis from contaminating the literature.

How to Identify Correlation vs Causation in News, Media, and Academic Papers

The correlation vs causation problem does not live only in textbooks. It appears every week in news headlines, government reports, social media posts, and even peer-reviewed journals. Developing the ability to identify when a causal claim is built on correlational evidence is a critical literacy skill for students in every academic discipline — and for any citizen navigating a world saturated with data.

Red Flags in News Headlines

News headlines that make causal claims from observational studies are extremely common and frequently misleading. Key red flag phrases to watch for include: “X causes Y,” “X leads to Y,” “X is linked to Y” (presented as though the link were causal), “X increases the risk of Y” (when based solely on observational data), and “scientists discover that X causes Y” (when the cited study is cross-sectional or retrospective). None of these headline structures are automatically wrong. But they require scrutiny. Good research skills include going back to the primary study, reading the methods section, and determining whether the study design supports the causal language being used.

How to Read a Research Paper for Causal Claims

When reading a research paper, the methods section determines everything about what conclusions are valid. Ask these questions: Is this an observational study, a quasi-experimental design, or a randomized controlled trial? If observational, what confounders did the authors control for — and what confounders might they have missed? Does the study establish temporal order (did the cause precede the effect)? Is the association statistically significant, and what is the effect size? Do the authors themselves describe the relationship as correlational or causal?

The STROBE statement (Strengthening the Reporting of Observational Studies in Epidemiology) provides reporting guidelines for observational studies. The CONSORT statement does the same for randomized trials. Studies that follow these guidelines are more transparent about their design limitations. The ability to evaluate research methodology is now explicitly tested in standardized assessments including the MCAT, GRE, and LSAT in the United States, as well as in A-Level and university-level science and social science courses in the UK.

Checking Claims on Social Media

Social media has created a perfect environment for correlation-causation confusion to spread. A graph with two lines moving together, a dramatic headline, a compelling story — none of these constitute evidence of causation. Before sharing or accepting a claim about what causes what, three questions are essential. First: what type of study generated this data? Second: what alternative explanations exist for the observed pattern? Third: have any independent studies replicated this finding? For academic work, using citation tools and checking primary sources directly is the most reliable defense against misinformation built on correlational data.

Correlation vs Causation in Statistics Courses, Assignments, and Exams

For students in college and university, correlation vs causation is not just a conceptual topic — it is a tested skill. It appears in statistics assignments, research methods courses, critical thinking assessments, psychology practicals, economics problem sets, and sociology research papers. Knowing the concepts is necessary. Being able to apply them correctly to specific scenarios is what determines your grade.

How This Concept Appears in Statistics Assignments

In introductory statistics courses at universities like Harvard, Stanford, Oxford, and University College London, correlation vs causation questions typically take several forms. You may be asked to compute a correlation coefficient and then interpret whether the association is causal. You may be given a research scenario and asked to identify confounders. You may be presented with a study design and asked whether the conclusions the authors draw are warranted. Or you may be asked to propose a study design that could establish causation where an existing study only shows correlation.

Each type of question tests something different. Computation tests procedural knowledge. Interpretation tests conceptual understanding. Study design evaluation tests methodological literacy. Getting a strong grade requires fluency across all three. The statistics assignment help resources at Ivy League provide support for all three levels — from formula application to study design critique.

The Difference Between Qualitative and Quantitative Evidence of Causation

Students in social science, public health, and interdisciplinary programs often encounter this distinction in the context of qualitative vs quantitative data. Quantitative studies produce correlations that can be measured precisely. Qualitative studies — interviews, ethnographies, case studies — generate narrative evidence of mechanism and process. Neither alone is sufficient for strong causal inference. The strongest causal evidence typically combines quantitative associations with qualitative evidence of mechanism — showing both that A and B move together and why A would plausibly produce B.

Correlation in Regression Analysis Assignments

A large proportion of statistics assignments involve regression analysis, which is built on correlation but interpreted in a causal-adjacent way. Multiple linear regression allows researchers to estimate the association between a predictor and an outcome while holding other predictors constant — which approximates controlling for confounders. But controlling for measured confounders in a regression model does not eliminate all confounding. Unmeasured confounders remain, and the causal interpretation of regression coefficients requires additional assumptions that are rarely fully satisfied in observational data. The regression model assumptions guide covers these issues in detail, including what happens to causal interpretation when key assumptions are violated.

Writing About Correlation vs Causation in Research Papers

Academic writing norms require precision about what your data can and cannot support. If you are reporting results from an observational study, you must use appropriately hedged language: “associated with,” “correlated with,” “was positively related to,” rather than “caused,” “produced,” or “led to.” Most university-level research methods courses and style guides — including APA 7th edition — explicitly address this requirement. Using causal language when your design only supports correlational conclusions is considered a significant methodological error in academic writing. The guide on conducting research for academic essays addresses this distinction in the context of academic writing standards.

Advanced Concepts: Directed Acyclic Graphs, Mendelian Randomization, and Causal AI

For students and researchers working at an advanced level, the correlation vs causation distinction is formalized in a rich set of mathematical and statistical frameworks that go far beyond correlation coefficients. These tools allow researchers to represent causal assumptions explicitly, test them, and use them to draw causal inferences from observational data under specific conditions.

Directed Acyclic Graphs (DAGs)

A directed acyclic graph (DAG) is a visual representation of a causal model. Each variable is represented as a node. Causal relationships are represented as directed arrows from cause to effect. The “acyclic” requirement means there are no feedback loops — a variable cannot cause itself through any path in the graph. DAGs allow researchers to explicitly state their assumptions about confounders, mediators, and colliders, and to determine which variables should (and should not) be controlled for in a statistical analysis.

The concept of a collider is particularly important. A collider is a variable that is caused by two other variables in your graph. Conditioning on a collider — controlling for it in a regression — can actually create a spurious association between variables that are otherwise unrelated. This is the mechanism behind Berkson’s paradox. DAG methodology, developed primarily by Judea Pearl at UCLA, is now taught in advanced epidemiology and biostatistics programs at institutions including Johns Hopkins Bloomberg School of Public Health and the London School of Hygiene and Tropical Medicine. Pearl’s work on the do-calculus and causal inference represents the most complete mathematical treatment of causation currently available.

Mendelian Randomization

Mendelian randomization is a method that uses genetic variants as instrumental variables to study the causal effects of modifiable risk factors on health outcomes. Because genetic variants are assigned at conception — randomly, from the perspective of the individual’s environment — they act like natural randomization. If a genetic variant is associated with higher levels of a specific biomarker (say, LDL cholesterol), and people with that variant have higher rates of heart disease, researchers can infer that LDL cholesterol has a causal effect on heart disease — using the genetic variant as the instrument to avoid confounding by lifestyle, socioeconomic status, or other factors.

Mendelian randomization has been used to establish or challenge causal claims about a wide range of biomarkers and exposures in chronic disease research. The UK Biobank, 23andMe, and large genome-wide association studies have made this method increasingly powerful and widely applied in the 21st century.

Causal Machine Learning and AI

One of the most active research frontiers in statistics and data science is the integration of causal inference methods with machine learning. Traditional machine learning systems are extremely good at detecting correlations in large datasets — this is essentially what they do. But a model trained on correlational data will fail when the data distribution changes, when interventions are applied, or when predictions are used to make decisions that themselves change the system. Researchers at institutions including Carnegie Mellon University, MIT, and ETH Zurich are developing methods for causal representation learning — machine learning systems that do not just predict outcomes but model the causal structure of the data-generating process.

The distinction between a predictive model (based on correlation) and a causal model (based on mechanism) is one of the most important distinctions in modern data science. A predictive model tells you what will happen given current patterns. A causal model tells you what will happen if you intervene and change something. Only causal models are appropriate for policy decisions.

Related Statistical Concepts Every Student Should Know

Mastering correlation vs causation is easier when you understand the broader landscape of related statistical and methodological concepts. The following terms appear consistently across statistics, research methods, and data science curricula at universities in the United States and the United Kingdom.

LSI and NLP Keywords: A Student’s Reference

When researching this topic in academic databases or writing essays on correlation vs causation, the following related terms and phrases are frequently used in the literature and in course materials. Familiarity with these terms will help you navigate research papers, textbooks, and course assessments:

• Statistical association — the general term for any relationship between variables detected through data
• Causal inference — the field dedicated to drawing causal conclusions from data
• Confounding variable / lurking variable — a third variable that creates a spurious correlation
• Spurious correlation — an association with no genuine causal basis
• Observational study — a study in which researchers do not assign treatments; can only establish correlation
• Randomized controlled trial (RCT) — the gold standard for establishing causation
• Counterfactual reasoning — asking what would have happened in the absence of the treatment
• Bradford Hill criteria — nine criteria for evaluating whether an epidemiological association is likely causal
• Simpson’s paradox — a trend that reverses when data are disaggregated
• Berkson’s paradox — selection bias in hospital-based samples
• Post hoc fallacy — concluding that sequence implies causation
• Ecological fallacy — applying group-level associations to individuals
• Effect modification / interaction — when a third variable changes the direction or strength of a causal effect
• Internal validity — the degree to which a study correctly identifies a causal relationship within its sample
• External validity — the degree to which a study’s findings generalize to other populations and settings
• Replication crisis — the widespread failure of published scientific findings to hold up in independent replications
• P-hacking / data dredging — selectively reporting analyses that yield significant results
• Pre-registration — specifying hypotheses and methods before data collection to prevent post hoc analysis

Understanding these terms positions you to read the literature review section of any academic paper with genuine critical insight — not just passive comprehension.

Frequently Asked Questions: Correlation vs Causation