A hypothesis is a specific, testable, falsifiable statement predicting the relationship between two or more variables. It is not a topic, not an open question, and not a hope about what the data will show — it is a claim precise enough that a defined result could prove it wrong. Because “hypothesis” gets used loosely for several related-but-distinct things in research writing, this page first sorts out what a hypothesis is and isn’t, then covers its main types, what makes one testable, how to operationalize the variables inside it, how it connects to statistical testing, and the integrity rules — chiefly HARKing and preregistration — that keep a hypothesis a genuine prediction rather than a story written after the fact. For a step-by-step walkthrough of drafting your own, see the practical template and worked examples further down this page.
Hypothesis vs. research question, aim, prediction, and theory
These five terms describe different stages of the same reasoning chain, and mixing them up is one of the most common issues reviewers flag in a study’s introduction.
| Term | What it is | Example |
|---|---|---|
| Theory | A broad, evidence-supported explanatory framework describing why a relationship should exist. A single study rarely tests an entire theory directly — it tests a hypothesis derived from it. | Cognitive load theory predicts that working memory has limited capacity, which shapes how instructional material should be structured. |
| Research question | The open question a study is designed to answer. Doesn’t commit to a predicted answer. | Does split-attention instructional design affect learning outcomes compared with integrated design? |
| Aim / objective | A statement of intent or purpose (“to examine,” “to determine,” “to evaluate”) — broader than a hypothesis and doesn’t require a predicted direction. | To evaluate the effect of instructional format on learning outcomes in undergraduate students. |
| Hypothesis | A specific, falsifiable, predicted answer to the research question, stated as a testable claim about variables and their relationship. | Undergraduate students taught with integrated instructional material will score higher on a post-test than students taught with split-attention material. |
| Prediction | The concrete, specific outcome that should be observed in the data if the hypothesis is true — often the exact form a preregistration commits to before data collection. | The integrated-format group’s mean post-test score will exceed the split-attention group’s mean by a statistically detectable margin at p < .05. |
A simple ordering: the theory explains why you’d expect a relationship to exist; the research question asks whether it does; the hypothesis states, specifically and falsifiably, what you predict the answer is; the prediction spells out exactly what that would look like in the data; the aim describes what the study sets out to do about all of that. If you’re deciding between “conceptual framework” and “theoretical framework” for your own study, see Conceptual Framework vs. Theoretical Framework. Hypotheses are typically derived from theory deductively — reasoning from a general principle to a specific testable prediction — which is one reason the hypothetico-deductive model dominates quantitative research; see Inductive vs. Deductive vs. Abductive Reasoning for how that contrasts with the inductive, theory-building reasoning common in qualitative and exploratory work.
Types of hypotheses
A single study’s hypothesis is usually classified along several of these dimensions at once — for example, a hypothesis can be simultaneously alternative, directional, simple, and causal.
Null vs. alternative
- Null hypothesis (H0): states there is no effect, no difference, or no relationship between the variables — the default position a statistical test is set up to try to reject. See Null Hypothesis Explained for a full treatment.
- Alternative hypothesis (H1 or Ha): states that an effect, difference, or relationship does exist — usually the researcher’s actual prediction and the reason the study is being run.
For the full side-by-side comparison of how the two are written, tested, and reported, see Null Hypothesis vs. Alternative Hypothesis.
Directional (one-tailed) vs. non-directional (two-tailed)
A directional alternative hypothesis predicts which way an effect runs (an increase or a decrease) and is only appropriate when there’s a genuine prior basis for expecting the effect to run one way; it changes how the test is set up, concentrating the rejection region on one side of the distribution. A non-directional hypothesis predicts that a relationship exists without committing to its direction — the more conservative default when prior evidence doesn’t clearly favor one direction.
Simple vs. complex
A simple hypothesis predicts a relationship between exactly one independent variable and one dependent variable. A complex hypothesis predicts relationships involving more than one independent or dependent variable at once — for example, that two predictors jointly affect an outcome, or that one predictor affects two separate outcomes. Complex hypotheses are harder to test and interpret cleanly, because a mixed result doesn’t clearly confirm or refute the whole claim; where possible, split a complex hypothesis into separate, individually testable simple hypotheses.
Associative vs. causal
An associative hypothesis predicts that two variables are related — that they co-occur or vary together — without claiming that one causes the other. A causal hypothesis predicts that changes in the independent variable directly produce changes in the dependent variable, which is a much stronger claim requiring a design capable of supporting it (random assignment, manipulation of the independent variable, and control of confounds). See Correlation vs. Causation and Experimental vs. Quasi-Experimental Design for what it actually takes to justify a causal hypothesis rather than an associative one.
Statistical hypothesis vs. research hypothesis
The research hypothesis is the substantive, plain-language prediction (“students taught with method A will score higher than students taught with method B”). The statistical hypothesis is its formal restatement in terms of population parameters — the form a significance test actually operates on, e.g. H0: μA = μB versus H1: μA ≠ μB. Every research hypothesis intended for quantitative testing needs a corresponding statistical hypothesis before analysis begins; see T-Test and What Is a P Value? for how that formal statement gets tested in practice.
What makes a hypothesis testable and falsifiable
The philosopher of science Karl Popper argued that what separates a scientific claim from a non-scientific one is not that it can be verified, but that it can, in principle, be shown to be false. A hypothesis that no conceivable observation could ever contradict is not testable, no matter how reasonable it sounds. This falsifiability criterion is the first filter any hypothesis needs to pass.
In practice, falsifiability comes down to two requirements:
- Every variable must be operationalized. “Improves wellbeing” is not testable; “increases scores on a validated wellbeing index” is. See the next section for how operationalization works.
- The predicted relationship must be specific enough to fail. A hypothesis that predicts “some kind of effect” can never be wrong, because almost any result confirms it. A testable hypothesis commits to a direction, a comparison, or a threshold that a specific pattern of data would contradict.
A useful check: write down, before collecting any data, what result would prove the hypothesis wrong. If you cannot describe that result, the hypothesis is not yet in testable form. In the Popperian sense, hypotheses are never proven true outright — they can only be falsified, or survive repeated attempts at falsification.
Operationalizing the variables in a hypothesis
A construct is an abstract idea — stress, engagement, quality of life, learning. A hypothesis can’t be tested until every construct in it has been operationalized: defined in terms of exactly how it will be measured or manipulated. “Stress” becomes “salivary cortisol level” or “score on a named, validated stress inventory”; “engagement” becomes “minutes of active participation logged in a learning platform.” See Operationalizing Variables for the general definition and worked examples.
Operationalization is also where a hypothesis’s independent variable (the one manipulated or used as the predictor) and dependent variable (the outcome being measured) get pinned down concretely — see Types of Variables for how independent, dependent, and confounding variables relate to each other. A hypothesis that names its variables only in the abstract (“exercise,” “performance”) but never states how each will actually be measured looks testable on paper but isn’t testable in practice.
Worked examples across disciplines
Each pair below shows a weak, untestable version and a stronger, operationalized revision. These are constructed to illustrate the structure — not drawn from or attributed to any specific real study.
Psychology
Weak: “Mindfulness meditation is good for stress.”
Stronger (directional, associative): “Among working adults, completing a daily 10-minute guided mindfulness meditation practice for four weeks will be associated with lower self-reported stress (Perceived Stress Scale) than a waitlist control group.”
Biomedical / clinical
Weak: “This intervention helps patients with diabetes.”
Stronger (directional, causal, RCT-appropriate): “Among adults with type 2 diabetes on stable metformin therapy, adding a structured weekly dietitian-led coaching program will produce a greater reduction in HbA1c at 12 weeks than metformin therapy alone.”
Education
Weak: “Group work improves learning.”
Stronger (directional, associative): “Among 9th-grade algebra students, participation in weekly structured peer-tutoring sessions will be associated with higher end-of-unit test scores than no peer tutoring.”
Social science / economics
Weak: “Minimum wage affects employment.”
Stronger (non-directional, given genuinely mixed prior evidence): “Among small retail employers in [region], a minimum-wage increase of [amount] will be associated with a change in part-time staffing levels within six months of implementation.”
How a hypothesis connects to statistical testing
A quantitative hypothesis isn’t just a sentence — it’s the thing a statistical test is built to evaluate, and misreading what the test result actually means is one of the most persistent errors in reporting.
- What the null actually asserts: that, at the population level, there is no difference, no effect, or no relationship between the variables — expressed as an equality between population parameters (e.g., two population means are equal).
- What rejecting the null tells you — and doesn’t: rejecting H0 means the observed data would be unlikely if H0 were true, at the significance threshold chosen. It does not prove the alternative hypothesis correct, does not give the probability the alternative is true, and does not by itself tell you whether the effect is large enough to matter practically — see Statistical Significance vs. Clinical Significance for why a statistically significant result and a meaningful one are separate questions.
- What failing to reject the null tells you — and doesn’t: the data didn’t provide enough evidence to rule out “no effect” at the chosen threshold. It is not evidence that the null is true — absence of evidence is not evidence of absence — and it is frequently a symptom of an underpowered study rather than a genuinely null result; see Power Analysis and Sample Size Calculation.
- Why “accepting the null” is the wrong phrase: the correct term is “fail to reject” the null, not “accept” it. A non-significant result is also consistent with a true, real, but small effect the study simply lacked the power to detect — “accepting” the null implies a certainty the test never provides.
- Type I and Type II error: a Type I error rejects a true null hypothesis (a false positive); a Type II error fails to reject a false null hypothesis (a false negative). Both are always possible, and the choice of significance threshold and sample size trades one risk against the other — see Type I and Type II Errors.
For the mechanics of running the test itself, see T-Test and What Is a P Value?.
The integrity dimension: hypotheses must come before the data
A confirmatory statistical test only means what it claims to mean if the hypothesis it’s testing was specified before the data were seen. Formulating or revising a hypothesis after looking at the results, then presenting it as though it had been predicted in advance, is a distinct and well-documented problem called HARKing (Hypothesising After Results are Known) — see HARKing and, for how it differs from the related practice of manipulating analyses to reach significance, P-hacking vs. HARKing.
Two mechanisms exist specifically to enforce the distinction between confirmatory work (testing a hypothesis stated in advance) and exploratory work (finding patterns after the fact, which is legitimate but must be labeled as such):
- Preregistration — publicly time-stamping a study’s hypotheses, methods, and analysis plan before data collection begins, so that anyone can later check what was predicted in advance versus what emerged from the data. See Preregistration of a Study Protocol.
- Registered reports — a publishing format where a journal peer-reviews and grants in-principle acceptance to a study’s introduction, hypotheses, and methods (Stage 1) before the results exist, removing the incentive to revise the hypothesis to fit whatever the data turn out to show. See Registered Reports.
Neither mechanism bans exploratory analysis — both simply require that exploratory findings be reported as exploratory, not repackaged as if they had been predicted from the start.
When no hypothesis is appropriate
Not every study needs a hypothesis, and the absence of one is not a weakness. Exploratory, descriptive, and much qualitative research proceeds from a research question without a predicted answer stated in advance — the goal is to characterize, describe, or generate understanding of a phenomenon that isn’t yet understood well enough to support a specific, falsifiable prediction. Forcing a hypothesis onto genuinely exploratory work can actively bias data collection and analysis toward confirming a preconception rather than discovering what’s actually there. See Descriptive Research for where this applies, and How to Write a Research Question for how to frame a strong question when no hypothesis is called for.
Common mistakes when writing a hypothesis
- Untestable or unfalsifiable claims. Predictions that no possible result could contradict — often because they invoke unmeasurable constructs, are stated as universal truths, or are so vague that any outcome “counts” as support.
- Vague or unoperationalized variables. Naming a construct (“stress,” “quality of life,” “engagement”) without specifying how it will actually be measured leaves the hypothesis untestable in practice, even if it sounds testable in principle.
- Circular reasoning. A hypothesis that is true by definition of its own terms — e.g., predicting that a “more effective” intervention will produce “better outcomes,” where effectiveness and better outcomes are defined identically — tests nothing.
- Restating the research question as a statement. Turning “Does X affect Y?” into “X affects Y” is not yet a hypothesis if it still doesn’t specify direction, magnitude, population, or measurement — it’s a question wearing a period instead of a question mark.
- Bundling multiple predictions into one complex hypothesis when a set of simple hypotheses would test more cleanly (see Simple vs. complex, above).
- Value-laden or unmeasurable terms. Words like “better,” “worse,” “improve,” or “harm” need an explicit, measured definition attached — otherwise the hypothesis imports a judgment call the data can’t actually adjudicate.
- Writing the hypothesis after seeing the results. See HARKing, above — a well-documented and distinct problem in its own right.
A practical template for writing a hypothesis
A workable quantitative hypothesis generally states: a specific population or sample, the independent variable and how it’s operationalized, the dependent variable and how it’s measured, and the predicted relationship between them (and, if directional, which way it runs).
A simple template: “Among [population], [operationalized independent variable] will be [associated with / lead to an increase or decrease in] [operationalized dependent variable].”
Before finalizing, check the hypothesis against the falsifiability test above: what specific result would prove it wrong? If that result is easy to state, the hypothesis is in testable shape. For how the hypothesis fits into the broader introduction of a manuscript, see Structuring the Introduction of a Research Paper.
Frequently asked questions
What’s the difference between a hypothesis and a research question?
A research question is the open question a study is designed to answer; a hypothesis is a specific, falsifiable, predicted answer to that question. Not every study needs a hypothesis — see When no hypothesis is appropriate, above.
What’s the difference between a null and an alternative hypothesis?
The null hypothesis (H0) states there is no effect or relationship; the alternative (H1/Ha) states that one exists. Statistical tests are built to reject or fail to reject H0 — they don’t test Ha directly. See Null Hypothesis vs. Alternative Hypothesis.
What’s the difference between an associative and a causal hypothesis?
An associative hypothesis predicts that two variables are related without claiming one causes the other; a causal hypothesis predicts that changes in one variable directly produce changes in another, which requires a design (typically an experiment with random assignment) capable of supporting that stronger claim.
What’s the difference between a simple and a complex hypothesis?
A simple hypothesis predicts a relationship between one independent and one dependent variable; a complex hypothesis involves more than one independent or dependent variable at once, and is generally harder to test and interpret cleanly.
Can a hypothesis be proven?
No, not in the sense of being established as permanently true. Statistical testing can reject a null hypothesis or fail to reject it, but it does not “prove” the alternative hypothesis correct. In the Popperian sense, hypotheses can only be falsified or survive attempts at falsification — they’re never verified outright.
Why is “accepting the null hypothesis” incorrect phrasing?
Because a non-significant result only means the data didn’t provide enough evidence to reject “no effect” at the chosen threshold — it doesn’t establish that no effect exists. The correct phrase is “fail to reject” the null, not “accept” it.
Does every study need a hypothesis?
No. Exploratory, descriptive, and much qualitative research proceeds from a research question without a predicted answer stated in advance. A hypothesis belongs in a study designed to test a specific predicted relationship, typically — though not exclusively — in quantitative and experimental work.
Where does a hypothesis go in a manuscript or protocol?
Typically at the end of the introduction, immediately after the research question and gap statement, and often restated in the methods section alongside the statistical analysis plan that will test it.
