A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
A confounding variable is any factor other than the IV that systematically differs between conditions and could explain the DV results. Internal validity refers to the degree to which a study justifies a causal inference — high internal validity means confounds have been ruled out. Common threats include selection bias (non-equivalent groups), history effects (external events during the study), maturation (natural change over time), regression to the mean, and demand characteristics. Experimental control and random assignment are the primary defenses against confounding.
Read descriptions of quasi-experiments (no random assignment) and identify which specific threats to internal validity apply and why they cannot be ruled out.
From experimental research design, you know that the purpose of an experiment is to isolate causation — to show that changes in the independent variable (IV) caused changes in the dependent variable (DV). A confounding variable is anything that threatens this isolation. Precisely: a confounder is a variable that (1) systematically differs between conditions, and (2) could plausibly explain the DV results independently of the IV. Both conditions must be met. A variable that varies randomly across participants is not a confound — it is noise. A variable that varies systematically and provides an alternative causal story is the problem.
Consider a study testing whether a new tutoring program improves test scores. Students who sign up for the program are compared against those who do not. Even if program students score higher, you cannot attribute this to the program, because students who volunteer for tutoring are likely more motivated to begin with. Selection bias is the confound: the groups differed in motivation before the intervention started, and motivation independently predicts test performance. The difference in test scores is explained (at least partly) by pre-existing group differences, not by the program. This is what makes it a confound rather than just a coincidental correlation — it is a systematic alternative explanation.
Random assignment is the primary tool for defeating confounds because it distributes all pre-existing differences — known and unknown — roughly equally across conditions by chance. When you randomly assign participants to the tutoring program or the control group, the two groups should be equivalent in motivation, ability, family support, and every other variable you could think of (plus the ones you have not thought of). The IV manipulation becomes the only systematic difference. This is why experimental design with random assignment earns the label "high internal validity" — confounds have been neutralized by design, not just measured and controlled statistically.
The major threats to internal validity each represent a specific failure mode. History effects occur when an external event happens during the study and affects one group more than the other. Maturation is a problem in longitudinal designs: participants naturally change over time, and this change could be mistaken for a treatment effect. Regression to the mean occurs when participants are selected because they scored at an extreme, and subsequent scores naturally drift toward average regardless of any intervention. Demand characteristics occur when participants infer the study's hypothesis and modify their behavior accordingly. Each threat is worth naming specifically because each has a specific remedy — and recognizing the threat is the first step toward designing around it.