Empirical questions are answerable through systematic observation or experiment; hypotheses are specific, testable predictions. Strong hypotheses specify predicted relationships between variables, are grounded in theory or prior evidence, and can be rejected by data. The difference between directional (one-tailed) and nondirectional (two-tailed) hypotheses has implications for statistical testing and study design.
Take broad interests (e.g., 'Does social media affect mood?') and iteratively refine into testable hypotheses. Practice stating hypotheses in 'If-then' form. Review published hypotheses and evaluate their clarity and testability.
From the empirical cycle, you know that science advances through iteration: observations raise questions, questions generate hypotheses, hypotheses generate predictions, predictions are tested, and results feed back into theory. The current skill is about one specific step — forming good hypotheses — which turns out to require considerably more precision than everyday language suggests. The difference between a vague curiosity and a testable hypothesis is the difference between something interesting and something scientifically useful.
A research question and a hypothesis are not the same thing. A research question states what you want to know: "Does social media use affect adolescent mood?" A hypothesis is a specific, directional prediction about what you expect to find and why: "Adolescents who spend more than three hours per day on social media will report lower mood scores than those who spend less than one hour, because passive consumption produces upward social comparison." The hypothesis does three things the question does not: it specifies the direction of the expected relationship, names the variables precisely enough to be measured, and grounds the prediction in a mechanism (upward social comparison) derived from prior theory. A hypothesis that cannot say *why* it predicts what it predicts is not well-grounded — it's a guess dressed up as a hypothesis.
The key quality criterion is falsifiability: a hypothesis must be stated in a form where specific data could contradict it. "People sometimes behave irrationally" is not falsifiable, because you would never specify what counts as evidence against it — any single example of irrational behavior confirms it, and no amount of rational behavior disproves it. "Sleep-deprived participants will recall fewer words on a 20-item list than control participants" is falsifiable: you operationalize "sleep-deprived," measure "words recalled," and the data can come back showing controls performed worse, or no difference — either would contradict the hypothesis. Falsifiability is why operational definitions are so tightly linked to hypothesis quality. A hypothesis is only as falsifiable as its variables are precisely defined.
The directional versus nondirectional distinction has real consequences for statistical analysis. A directional (one-tailed) hypothesis commits to the sign of the expected effect: "sleep-deprived will do *worse*, not better." This allows you to concentrate statistical power in one tail of the distribution, making it easier to detect an effect in the predicted direction. A nondirectional (two-tailed) hypothesis predicts only a difference without specifying direction. The choice should be determined by the strength of your prior evidence, not convenience: directional hypotheses are only appropriate when theory and prior results strongly support a specific direction. If you predict a direction because it's the only direction that would be publishable, not because the evidence warrants it, that is a form of bias baked in before data collection — which is exactly why preregistration (the next topic) exists to prevent it.