Contents
On this page
- What is a statistical hypothesis?
- The hypothesis as a statement about a model parameter
- The null hypothesis as a computational device rather than a belief
- Conclusions of the American Statistical Association (ASA) statement
- A worked example of a commonly overlooked null hypothesis
- One null hypothesis, two p-values
- Common misinterpretations in draft manuscripts
- Translating a research aim into a testable null hypothesis
- Limitations of hypothesis testing
- Key points
- Application to your study
- Frequently asked questions
Academy · Research Data Analytics & AI · Statistical hypothesis
What Is a Statistical Hypothesis?
A statistical hypothesis is a statement about a parameter in a model, not a statement about the world. Both the usefulness of hypothesis testing and its capacity to mislead follow from this distinction. The resulting p-value is a property of the data under an assumed model; it is not, and has never been, the probability that the hypothesis is true.
Most researchers can state that the null hypothesis is the hypothesis of no effect, yet many then write results sections that treat a p-value as though it measured the probability that the finding is real. The discrepancy between these two positions is where reviewers identify problems, and it cannot be resolved by more careful wording alone. It requires an understanding of what the hypothesis is a statement about, what the p-value is computed from, and which of the two the conclusion concerns. This article sets out that distinction and the specific errors it prevents.
What is a statistical hypothesis?
A statistical hypothesis is a precise, testable statement about a population parameter, such as a mean, a proportion or a difference between groups, that can be evaluated with sample data. Research questions are turned into a pair of competing hypotheses, a null and an alternative, and a statistical test then measures how compatible the observed data are with the null.
Hypotheses matter because they fix, before the data are analysed, exactly what is being tested and in which direction. That protects the analysis from being shaped by the results, determines the sample size needed, and decides which test is appropriate.
Types of statistical hypothesis
| Type | What it states | Example |
|---|---|---|
| Null hypothesis (H0) | No effect, no difference or no association in the population | The new drug lowers blood pressure by the same amount as the standard drug |
| Alternative hypothesis (H1) | The effect, difference or association the study is looking for | The new drug lowers blood pressure by a different amount from the standard drug |
| Two-tailed hypothesis | An effect in either direction counts against the null | The mean differs from 120 mmHg (higher or lower) |
| One-tailed hypothesis | Only an effect in one pre-specified direction counts | The mean is lower than 120 mmHg |
Illustrative examples written for this guide. A one-tailed test must be justified before the data are seen; choosing the direction afterwards is a form of selective analysis.
The rest of this article looks more closely at what the hypothesis is a statement about, what a p-value does and does not measure, and the misinterpretations that most often appear in draft manuscripts.
Learning objectives
- What a statistical hypothesis is a statement about, and why that is a model parameter rather than a clinical claim
- Why the null is a computational device rather than a belief, and the role of the alternative hypothesis
- The six principles in the ASA statement on p-values, and the definition of a p-value it settles on
- Which of the 25 misinterpretations catalogued by Greenland and colleagues are most likely to occur in a draft manuscript
- How to translate a vague research aim into a testable null hypothesis on which a statistician can act
- What a hypothesis test cannot establish, including the two conclusions that authors most often seek from it
Common misconceptions
- “The p-value is the probability that the null hypothesis is true.” It is the probability, under an assumed model in which the null holds, of a summary of the data as extreme as the one observed. The hypothesis is conditioned on, not assessed.
- “P greater than 0.05 means there is no effect.” It means the data are not far from what the model predicts under the null. Greenland and colleagues, in the European Journal of Epidemiology, list treating a non-significant result as evidence of no effect among 25 documented misinterpretations.
- “A smaller p-value means a larger or more important effect.” P-values depend on sample size and on the model as much as on effect magnitude. Importance is not a statistical quantity at all.
- “The null hypothesis is what we believe before the study.” An effect of exactly zero is rarely considered plausible. The null is chosen because it pins the model down tightly enough to compute from.
- “One test, one p-value, one answer.” Different valid tests of the same null give different p-values on the same data, because each assumes a different model.
The hypothesis as a statement about a model parameter
A statistical hypothesis is a claim about the value of a parameter in a specified probability model. “The mean difference in HbA1c between arms is zero” is a statistical hypothesis. “The drug does not work” is not: it is a clinical claim, and it maps onto a statistical hypothesis only once the outcome, the population, the timing of measurement and the model have been specified.
The distinction is substantive. Every number a test produces is computed within the model, so the test can indicate only the compatibility of the data with that model and that parameter value. Greenland and colleagues, in the European Journal of Epidemiology, make this the organising point of their guide: a test assesses the entire model, not only the parameter of interest, so a small p-value may indicate that the distributional assumption is wrong rather than that the effect is real 9.
Two practical consequences follow. First, a hypothesis that cannot be written as a statement about a named parameter is not yet testable, and no amount of analysis will make it so. Second, reporting a test commits the author to the model, which is why the model belongs in the methods section alongside the hypothesis, not in a supplement.
The parts of a hypothesis test, and what each is not
| Element | What it is | What it is not |
|---|---|---|
| Statistical hypothesis | A statement about the value of a parameter in a specified model | A clinical or scientific claim. The claim has to be translated into a parameter first |
| Null hypothesis | The test hypothesis in which the targeted effect size is zero, or another fixed value chosen as the reference for the test | Anyone’s prior belief. It is a device that specifies the model tightly enough to compute from |
| Alternative hypothesis | The set of parameter values designated as being of interest, fixed before the data | Whatever the data turn out to suggest. Choosing it afterwards invalidates the p-value |
| Test statistic | A summary of the data whose distribution is known under the null | A measure of effect size. Q, t and z all grow with sample size at fixed effect |
| P-value | The probability, under a specified model, that the summary would be equal to or more extreme than observed | The probability that the null is true, that the result is due to chance, or that the finding will replicate |
| Significance level | A pre-chosen error rate for a decision procedure | A threshold of truth, or a boundary at which an effect begins to exist |
| Confidence interval | A range produced by a procedure with a stated long-run coverage property | A range containing the true value with that probability in a particular study |
Definitions follow Wasserstein and Lazar (The American Statistician, 2016) and Greenland et al (European Journal of Epidemiology, 2016). The null hypothesis definition — the test hypothesis in which the targeted effect size is zero — is Greenland and colleagues’.
The null hypothesis as a computational device rather than a belief
Greenland and colleagues define the null hypothesis as the test hypothesis in which the targeted effect size is zero 9. The word “targeted” is essential: the null is whichever value has been chosen as the reference, and zero is the conventional choice rather than a necessary one. It is possible to test a null of no difference, a null of a specified non-zero difference, or a null that an effect lies inside or outside an interval.
A treatment effect of exactly zero, to infinite precision, is rarely considered plausible. The null is useful because it specifies the model completely enough to derive the distribution of a test statistic; that is its sole function. Once it is understood as a computational anchor rather than a belief, several confusions are resolved. “Accepting the null” ceases to be meaningful, because the null was never being weighed as a candidate truth. “The study proved there was no difference” is untenable for the same reason.
The alternative hypothesis is the complement selected as being of interest, and it is where the substantive design decisions lie: whether the interest is in a difference in either direction, in one direction only, or in a region of equivalence 3. Those choices define different tests and different claims, and they are set out on our types of statistical hypothesis page. The essential point here is that the alternative must be fixed before the data are available, because it determines what the p-value means.
Conclusions of the American Statistical Association (ASA) statement
Wasserstein and Lazar, writing in The American Statistician for the American Statistical Association, set out six principles on p-values and define a p-value as the probability, under a specified statistical model, that a statistical summary of the data would be equal to or more extreme than its observed value 10. Each clause of that definition is significant. Under a specified statistical model: if the model changes, the number changes. A statistical summary of the data: the p-value concerns the data, not the hypothesis. Equal to or more extreme: it includes results that were not observed, which is why it depends on what the analyst would have done with other data.
The statement was unusual in being issued because of how routinely the quantity was being misused rather than because of a new technical result. Its principles cover what p-values can indicate, what they cannot — including the probability that the hypothesis is true or that the data were produced by chance alone — and the hazards of basing conclusions on whether a p-value crosses a threshold. It also makes the point that a p-value, or statistical significance, does not measure the size or the importance of an effect 10.
The constructive reading is that a p-value is one input among several, best reported next to an effect estimate and an interval, with the model stated. That is the form most journals now request, and it is what our biostatistics reviewers look for first.
Sources: Wasserstein RL, Lazar NA. The ASA’s Statement on p-Values: Context, Process, and Purpose. The American Statistician. 2016;70(2):129-133. Greenland S, Senn SJ, Rothman KJ, Carlin JB, Poole C, Goodman SN, Altman DG. Statistical tests, P values, confidence intervals, and power: a guide to misinterpretations. European Journal of Epidemiology. 2016;31:337-350.
A worked example of a commonly overlooked null hypothesis
Meta-analysis contains a null hypothesis that most readers never notice they are testing, and it is a clean illustration of the logic. The homogeneity null states that every included study estimates the same underlying effect, with observed differences due to sampling error alone 25. Cochran, in Biometrics, gave the statistic that tests it.
DerSimonian and Laird, in Controlled Clinical Trials, report it for Baum’s review of antibiotic prophylaxis in colon surgery: the within-study homogeneity statistic Q was 40.4 on 25 degrees of freedom 9,24. The structure of that sentence is the structure of every hypothesis test. There is a null (all studies share one true effect), a test statistic computed from the data (Q), and a reference distribution determined by the design rather than by the result (25 degrees of freedom, one fewer than the number of studies). Under the null, Q is expected to be about equal to its degrees of freedom, so 40.4 against 25 is more dispersion than the null predicts 25.
The test does not indicate how much the studies differ, or whether that difference matters for the decision at hand. For this reason Higgins and Thompson, in Statistics in Medicine, proposed three statistics — H, I-squared and R — to describe heterogeneity independently of the effect metric and the number of studies, illustrated on five datasets 19. The test answers a yes-or-no question about compatibility with a model; the descriptive statistics answer the substantive question. This gap between a test and a quantity recurs throughout statistical practice.
One null hypothesis, two p-values
If a p-value measured the credibility of a hypothesis, two valid tests of the same hypothesis on the same data would have to agree. They do not, and the clearest published demonstrations come from tests of funnel plot asymmetry. Sterne and colleagues, in the BMJ, report for the intravenous magnesium meta-analysis that the Harbord test gave P less than 0.001 while the Peters test gave P equal to 0.002 for the same null of no asymmetry on the same set of trials 13.
Both numbers are correct. They differ because the tests model the relationship between effect size and precision differently, and the p-value is computed from the model 13. Neither is the probability that asymmetry exists. The practical lesson is that choosing a test is a modelling decision with consequences, and reporting “asymmetry was significant” without naming the test conceals the information a reader most needs. The same applies to any outcome for which several tests are available: the test should be named and its choice justified.
This is also the underlying reason that pre-specification matters. Where several tests of the same null are defensible, selecting after seeing the results turns a legitimate choice into a biased one. Our statistical analysis plan page covers how to fix these choices in advance.
Common misinterpretations in draft manuscripts
Greenland and colleagues enumerate 25 common misinterpretations of P values, confidence intervals and power 9,24. All 25 merit reading; four of them account for most of the issues that reviewers raise.
Treating a non-significant result as evidence of no effect. A wide interval that includes zero is compatible with no effect and also with a large effect 9. The correct report is the estimate and the interval, with a statement of what the data cannot rule out.
Reading the p-value as a probability about the hypothesis. It is computed under the assumption that the null holds, so it cannot be a measure of whether that assumption holds.
Treating two studies as conflicting because one crossed a threshold and the other did not. Two studies with similar estimates can land on opposite sides of a conventional cut-off 6,9. Estimates and intervals should be compared, not labels.
Reading a confidence interval as containing the true value with 95% probability. The coverage property belongs to the procedure across hypothetical repetitions, not to the single interval obtained 9.
None of these is corrected by hedged language. They are corrected by reporting an effect estimate with its interval, naming the model, and writing the conclusion about the estimate rather than about the test.
Translating a research aim into a testable null hypothesis
The most common statistical problem in a submitted manuscript is not a wrong test; it is that no hypothesis was ever stated precisely enough to be tested. “We investigated whether X helps” is an aim. It names no outcome, no population, no timepoint, no comparator and no parameter, which means any of several analyses could be presented as testing it ; this flexibility is precisely what pre-specification is intended to remove.
An aim, and the hypothesis it has to become
An aim, not a hypothesis
We investigated whether structured discharge planning helps reduce readmissions in elderly patients, and whether patient satisfaction improves. Statistical significance was assessed at the 5% level.
Which outcome? “Readmissions” could be all-cause or condition-specific, at any window, counted per patient or per event.
Which parameter? A risk difference, a risk ratio and an odds ratio give different numbers and different nulls.
Which population and comparator? Neither is defined, so the hypothesis cannot be stated as a parameter.
How many primary hypotheses? Two outcomes are named without saying which is confirmatory, leaving the error rate undefined.
What model? Unstated, although the p-value is computed within the model, so this is not a minor detail.
Testable, and pre-specifiable
The primary null hypothesis was that the 30-day all-cause readmission risk does not differ between structured discharge planning and usual care in patients aged 65 and over discharged from an acute medical ward, against a two-sided alternative, with the risk difference as the parameter of interest. The test was a two-sided Wald test from a log-binomial model adjusted for the pre-specified stratification factors of centre and admission category, with the estimate reported as a risk difference with a 95% confidence interval; the two-sided significance level was 5% for this single primary comparison. Patient satisfaction was a secondary outcome and its hypothesis was not tested confirmatorily; its estimate and interval are reported without a p-value.
The outcome is named, timed and defined as all-cause, so only one quantity can be computed.
The parameter is named, which is what makes the null a statement with a value rather than a direction.
Population and comparator are specified, so the hypothesis is about a definable contrast.
One primary comparison is declared and the secondary outcome is explicitly not confirmatory, which keeps the error rate meaningful.
The model and the adjustment set are stated, so a reader can see what the p-value is conditional on.
The second version is not more cautious; it is narrower. It relinquishes the freedom to present any of a dozen analyses as the test of the stated aim, the very freedom that renders the first version unusable. It also adds, alongside the test, an estimate and an interval, reported for the secondary outcome without a p-value at all. That is the form towards which the ASA statement’s principles direct authors: the test is one input, and the estimate is the finding.
Limitations of hypothesis testing
Three requests arise repeatedly, and none can be answered by a test, however it is conducted.
It cannot establish that an effect is absent. A test of a null of zero can fail to reject; that is not evidence for the null 9,20. If the absence of a meaningful effect is the actual research question, a hypothesis built around a margin is required; this is the equivalence and non-inferiority family covered on our types of statistical hypothesis page 1,15.
It cannot indicate whether a result matters. The ASA statement is explicit that a p-value does not measure the size or the importance of an effect. Importance is a clinical, economic or policy judgement made on the estimate and its interval, and it rests with the investigator rather than with the analysis.
It cannot repair a design. A test inherits every assumption of the model and every limitation of the data. Where allocation was not concealed, where the outcome was not pre-specified, or where the sample could not have detected a plausible effect, the p-value is a well-computed number about a study that cannot answer the question. This is why hypothesis specification belongs in the protocol stage — see clinical study protocol — and why a statistician consulted at analysis is being asked to fix something that is no longer fixable.
Key points
Key points
A statistical hypothesis is a statement about a parameter inside a model, and a p-value is a property of the data computed under that model assuming the null holds. It is not the probability that the hypothesis is true, it does not measure effect size or importance, and a non-significant result is not evidence of no effect. The hypothesis should be written so that it names an outcome, a population, a comparator, a timepoint, a parameter and a model, and should be fixed before the data are available. The estimate and its interval should then be reported as the finding, with the p-value as one input beside them.
Application to your study
Application to your study
- The primary hypothesis should be written as a sentence containing a parameter. A sentence that contains no named quantity states an aim rather than a hypothesis.
- The outcome should be named fully. The outcome, population, comparator, timepoint, and whether the outcome is all-cause or specific should be stated. Any ambiguity here constitutes analytical flexibility that reviewers are likely to query.
- The number of primary hypotheses should be declared. A single primary hypothesis is the easiest to defend; more than one requires a stated approach to the error rate.
- The model should be reported in the methods alongside the hypothesis. The p-value is conditional on the model, so a reader cannot interpret one without the other.
- The test should be named and its choice justified. Where several valid tests of the same null exist, they give different p-values, so the choice is a reportable decision.
- Every “no significant difference” sentence should be revised. It should be replaced with the estimate, the interval, and the effects that the data cannot rule out.
- The draft should be checked against the ASA principles. The two principles on thresholds and on effect size are particularly relevant, since they identify most overreach in results sections.
- Where the research question concerns the absence of an effect, the hypothesis should be revised. A failure to reject the null will not answer the question; a margin-based hypothesis will.
Frequently asked questions
It is a precise statement about the value of a parameter in a statistical model — for example, that the mean difference between two arms is zero, or that a risk ratio equals one. It is deliberately narrower than a research question. “Does the drug work?” is a research question; converting it into a statistical hypothesis requires naming the outcome, the population, the comparator, the timepoint, the parameter used to measure the effect, and the model under which it will be estimated. If these cannot be specified, the hypothesis is not yet testable, and a statistician cannot remedy the omission at the analysis stage.
Wasserstein and Lazar, in The American Statistician, define it as the probability, under a specified statistical model, that a statistical summary of the data would be equal to or more extreme than its observed value. Three consequences follow. It is computed assuming the null is true, so it cannot be the probability that the null is true. It is conditional on the whole model, so a small p-value may reflect a wrong distributional assumption rather than a real effect. And it includes results you did not observe, which is why it depends on the design. The ASA statement sets out six principles covering these points and the hazards of threshold-based conclusions.
No, and this is one of the 25 misinterpretations catalogued by Greenland and colleagues in the European Journal of Epidemiology. A non-significant result means the data are not far from what the model predicts under the null; a wide interval that includes zero is also compatible with a clinically important effect. The correct report is the estimate with its confidence interval, plus an explicit statement of what the data cannot rule out. If demonstrating the absence of a meaningful effect is the objective, a hypothesis built around a pre-specified margin is required rather than a null of exactly zero.
Not always, and treating it as fixed causes problems. Greenland and colleagues define the null as the test hypothesis in which the targeted effect size is zero, with “targeted” signalling that the reference value is chosen by the investigator. Zero is conventional because it is a natural reference point and because it specifies the model completely, which is all the null needs to do; its purpose is computational, not doctrinal. It is equally possible to test a null of a specified non-zero difference, which is what non-inferiority trials do, or a null that an effect falls outside an interval, which is what equivalence testing does.
Yes, and the published examples are instructive. Sterne and colleagues, in the BMJ, report that for the intravenous magnesium meta-analysis the Harbord test gave P below 0.001 and the Peters test gave P equal to 0.002 for the same null of no funnel plot asymmetry on the same trials. Both are valid; they differ because they model the data differently, and the p-value is computed inside the model. This is why naming the test is not a formality, and why choosing a test after seeing which one gives the smaller p-value converts a legitimate modelling decision into a biased one.
A research question states what the investigator wishes to find out and may be loosely framed; a statistical hypothesis must be precise enough to compute from. The useful test is whether a parameter appears. “Whether structured discharge planning reduces readmissions” contains no parameter, no population, no timepoint and no comparator, so several different analyses could all claim to answer it. “The 30-day all-cause readmission risk difference between structured discharge planning and usual care in patients aged 65 and over is zero” can be tested in exactly one way once the model is named. Pre-specification is what removes the flexibility the loose version leaves open.
It is more informative for most purposes, because it reports the estimate together with the range of values compatible with the data, rather than compressing both into a single threshold comparison. But it is not immune to misreading: Greenland and colleagues include among their 25 misinterpretations the belief that a 95% interval contains the true value with 95% probability, when the coverage property belongs to the procedure across hypothetical repetitions rather than to any particular interval. The practical recommendation is to report the effect estimate, the interval, and the model, and to treat any p-value as one input beside them.
The hypothesis should be checked for six elements: the outcome, defined precisely enough that only one quantity could be computed; the population; the comparator; the timepoint; the parameter in which the effect is expressed; and the model under which it will be estimated. It should then be asked whether any element would change if the data were seen first; any element for which the answer is yes is likely to be questioned by reviewers. Fixing all six in a statistical analysis plan before unblinding is the only reliable way to ensure that the resulting p-value means what it will be claimed to mean.
References
- US Food and Drug Administration. Guidance for industry: non-inferiority clinical trials. US Department of Health and Human Services (guidance). 2010. read 9 October 2026. trialsjournal.biomedcentral.com
- European Medicines Agency, Committee for Medicinal Products for Human Use. Guideline on the choice of the non-inferiority margin. European Medicines Agency (guideline). 2005. read 9 October 2026. trialsjournal.biomedcentral.com
- ICH Steering Committee. Statistical principles for clinical trials (ICH E9). International Conference on Harmonisation (guideline). 1998. read 9 October 2026. trialsjournal.biomedcentral.com
- Mayo DG. Rejecting statistical significance tests: defanging the arguments. JSM Proceedings (American Statistical Association). 2020:236-256. read 9 October 2026. ww2.amstat.org
- Wasserstein R, Schirm A, Lazar N. Moving to a world beyond "p < 0.05". The American Statistician. 2019;73(S1):1-19. doi:10.1080/00031305.2019.1583913 read 9 October 2026. doi.org/10.1080/00031305.2019.1583913
- Amrhein V, Greenland S, McShane B. Retire statistical significance. Nature. 2019;567(7748):305-307. doi:10.1038/d41586-019-00857-9 read 9 October 2026. doi.org/10.1038/d41586-019-00857-9
- Ionides EL, Giessing A, Ritov Y, Page SE. Response to the ASA's statement on p-values: context, process, and purpose. The American Statistician. 2017;71(1):88-89. doi:10.1080/00031305.2016.1234977 read 9 October 2026. doi.org/10.1080/00031305.2016.1234977
- Ahlbom A. Significance testing: why does it prevail? European Journal of Epidemiology. 2017;32(1):1-2. doi:10.1007/s10654-016-0220-0 read 9 October 2026. doi.org/10.1007/s10654-016-0220-0
- Greenland S, Senn SJ, Rothman KJ, Carlin JB, Poole C, Goodman SN, Altman DG. Statistical tests, P values, confidence intervals, and power: a guide to misinterpretations. European Journal of Epidemiology. 2016;31:337-350. doi:10.1007/s10654-016-0149-3 read 9 October 2026. doi.org/10.1007/s10654-016-0149-3
- Wasserstein RL, Lazar NA. The ASA's Statement on p-Values: Context, Process, and Purpose. The American Statistician. 2016;70(2):129-133. doi:10.1080/00031305.2016.1154108 read 9 October 2026. doi.org/10.1080/00031305.2016.1154108
- Schiller P, Burchardi N, Niestroj M, Kieser M. Quality of reporting of clinical non-inferiority and equivalence randomised trials: update and extension. Trials. 2012;13:214. doi:10.1186/1745-6215-13-214 read 9 October 2026. doi.org/10.1186/1745-6215-13-214
- Riley RD, Higgins JPT, Deeks JJ. Interpretation of random effects meta-analyses. BMJ. 2011;342:d549. doi:10.1136/bmj.d549 read 9 October 2026. doi.org/10.1136/bmj.d549
- Sterne JAC, Sutton AJ, Ioannidis JPA, Terrin N, Jones DR, Lau J, Carpenter J, Rucker G, Harbord RM, Schmid CH, Tetzlaff J, Deeks JJ, Peters J, Macaskill P, Schwarzer G, Duval S, Altman DG, Moher D, Higgins JPT. Recommendations for examining and interpreting funnel plot asymmetry in meta-analyses of randomised controlled trials. BMJ. 2011;343:d4002. doi:10.1136/bmj.d4002 read 9 October 2026. doi.org/10.1136/bmj.d4002
- Wangge G, Klungel OH, Roes KC, de Boer A, Hoes AW, Knol MJ. Room for improvement in conducting and reporting non-inferiority randomized controlled trials on drugs: a systematic review. PLoS One. 2010;5(10):e13550. read 9 October 2026. trialsjournal.biomedcentral.com
- Fleming TR. Current issues in non-inferiority trials. Statistics in Medicine. 2008;27:317-322. read 9 October 2026. trialsjournal.biomedcentral.com
- Le Henanff A, Giraudeau B, Baron G, Ravaud P. Quality of reporting of noninferiority and equivalence randomized trials. JAMA. 2006;295:1147-1151. read 9 October 2026. trialsjournal.biomedcentral.com
- Piaggio G, Elbourne DR, Altman DG, Pocock SJ, Evans SJ; CONSORT Group. Reporting of noninferiority and equivalence randomized trials: an extension of the CONSORT statement. JAMA. 2006;295:1152-1160. read 9 October 2026. trialsjournal.biomedcentral.com
- D'Agostino RB. Non-inferiority trials: advances in concepts and methodology. Statistics in Medicine. 2003;22:165-167. read 9 October 2026. trialsjournal.biomedcentral.com
- Higgins JPT, Thompson SG. Quantifying heterogeneity in a meta-analysis. Statistics in Medicine. 2002;21(11):1539-1558. doi:10.1002/sim.1186 read 9 October 2026. doi.org/10.1002/sim.1186
- Jones B, Jarvis P, Lewis JA, Ebbutt AF. Trials to assess equivalence: the importance of rigorous methods. BMJ. 1996;313:36-39. read 9 October 2026. trialsjournal.biomedcentral.com
- Arthur W, Bennett W, Huffcutt A. Choice of software and programs in meta-analysis research: does it make a difference? Educational and Psychological Measurement. 1994;54:776-787. read 9 October 2026. link.springer.com
- Eysenck HJ. Meta-analysis and its problems. BMJ. 1994;309:789-792. read 9 October 2026. link.springer.com
- Begg CB, Mazumdar M. Operating characteristics of a rank correlation test for publication bias. Biometrics. 1994;50(4):1088-1101. doi:10.2307/2533446 read 9 October 2026. doi.org/10.2307/2533446
- DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials. 1986;7(3):177-188. doi:10.1016/0197-2456(86)90046-2 read 9 October 2026. doi.org/10.1016/0197-2456(86)90046-2
- Cochran WG. The combination of estimates from different experiments. Biometrics. 1954;10(1):101-129. doi:10.2307/3001666 read 9 October 2026. doi.org/10.2307/3001666
- Veroniki AA, Jackson D, Bender R, Kuss O, Langan D, Higgins JPT, Knapp G, Salanti G. Methods to calculate uncertainty in the estimated overall effect size from a random-effects meta-analysis. doi:10.1002/jrsm.1319 read 9 October 2026. doi.org/10.1002/jrsm.1319
- Stang A, Deckert M, Poole C, Rothman KJ. Statistical inference in abstracts of major medical and epidemiology journals 1975-2014: a systematic review. European Journal of Epidemiology. doi:10.1007/s10654-016-0211-1 read 9 October 2026. doi.org/10.1007/s10654-016-0211-1
Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
Pubrica Academy. (2026, October 9). What Is a Statistical Hypothesis?. Pubrica. https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/
Pubrica Academy. “What Is a Statistical Hypothesis?.” Pubrica, 9 Oct. 2026, pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/.
Pubrica Academy (2026) What Is a Statistical Hypothesis?. Pubrica. Available at: https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/ (Accessed: 15 October 2026).
Pubrica Academy. “What Is a Statistical Hypothesis?.” Pubrica. October 9, 2026. https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/
Pubrica Academy, ‘What Is a Statistical Hypothesis?’ (Pubrica, 9 October 2026) <https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/> accessed 15 October 2026.
Pubrica Academy. What Is a Statistical Hypothesis?. [Internet]. Pubrica; 2026 Oct 9 [cited 2026 Oct 15]. Available from: https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/
Pubrica Academy, “What Is a Statistical Hypothesis?,” Pubrica, Oct. 9, 2026. [Online]. Available: https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/
@misc{what_is_statistical_hypothesis_2026,
author = {{Pubrica Academy}},
title = {What Is a Statistical Hypothesis?},
howpublished = {Pubrica},
year = {2026},
month = {oct},
url = {https://pubrica.com/academy/concepts-definitions/what-is-statistical-hypothesis/},
note = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Primary hypothesis specification for use in a protocol
One of these should be completed for each confirmatory hypothesis, before any outcome data are seen. If a line cannot be completed, the hypothesis is not yet testable; that is a substantive finding, not a formatting problem.
PRIMARY HYPOTHESIS SPECIFICATION
Study: ______________________ Version / date: ______________
1. THE CLAIM IN WORDS
What we want to find out: ____________________________
2. THE HYPOTHESIS AS A PARAMETER STATEMENT
Population: __________________________________________
Intervention / exposure: _____________________________
Comparator: __________________________________________
Outcome, defined so only one quantity can be computed:
____________________________________________________
Timepoint: ___________________________________________
Parameter (risk difference / risk ratio / odds ratio /
mean difference / hazard ratio): _____________________
Null hypothesis (H0): ________________________________
Alternative (H1), and its direction: _________________
3. THE MODEL THE P-VALUE WILL BE CONDITIONAL ON
Model / distributional assumption: ___________________
Covariates, pre-specified: ___________________________
Clustering or repeated measures structure: ___________
Analysis population: _________________________________
Missing data approach: _______________________________
4. THE TEST
Named test: __________________________________________
Why this test rather than the alternatives available: _
____________________________________________________
Sided: [ ] two-sided [ ] one-sided, justified below
Significance level: ____
5. ERROR RATE ACROSS HYPOTHESES
Number of confirmatory hypotheses in this study: ____
If more than one, the multiplicity approach: _________
Outcomes explicitly NOT tested confirmatorily: _______
6. WHAT WILL BE REPORTED
[ ] effect estimate
[ ] confidence interval
[ ] model stated in the methods
[ ] p-value reported beside the estimate, not instead of it
[ ] for secondary outcomes: estimate and interval, no p-value
7. PRE-SPECIFICATION CHECK
Where is this document registered or filed? __________
Date fixed, relative to database lock: _______________
[ ] nothing in sections 2-4 can change after unblinding
8. INTERPRETATION RULES AGREED IN ADVANCE
If the null is not rejected, the sentence we will write:
____________________________________________________
(it must report the estimate, the interval, and what
the data cannot rule out - not "no effect")
The smallest effect we would regard as important, and
why: ________________________________________________ Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.
How Pubrica can help
Pubrica’s biostatisticians write hypothesis specifications at the protocol stage, where the choices are still open, and review results sections for the specific overreach that reviewers catch — non-significance reported as absence of effect, p-values standing in for effect sizes, tests named only as “appropriate”. Where a study has already been analysed, Pubrica states clearly which claims the data support and which require rewriting around the estimate.