Contents
On this page
- Four dimensions of classification
- One-sided and two-sided hypotheses and the consequences of post hoc selection
- Simple, composite, point and interval hypotheses
- Superiority, non-inferiority and equivalence hypotheses
- Omnibus and global null hypotheses, with two worked examples
- Selecting the hypothesis type according to the decision required
- Stating the hypothesis so that its type is explicit
- Errors associated with each hypothesis type
- Key points
- Application to your study
- Frequently asked questions
Academy · Research Data Analytics & AI · Types of statistical hypothesis
Types of Statistical Hypothesis: Classification and Appropriate Use
Statistical hypotheses differ along four dimensions: the number of parameter values the null hypothesis fixes, whether the alternative hypothesis is directional, whether the claim concerns difference or similarity, and whether one or several parameters are tested simultaneously. Each combination defines a different null hypothesis, supports a different conclusion and is subject to different errors. Selection of an inappropriate hypothesis type is the most common reason a correctly executed analysis cannot support the conclusion stated about it.
Equivalence and superiority are not the same claim tested at different thresholds; they have different null hypotheses, and a study designed for one cannot be reanalysed as the other. Nor is a one-sided test simply a more powerful two-sided test; Bland and Bland demonstrated the consequences for the error rate when authors treat it as such. This article sets out a classification of statistical hypotheses and the conclusions that each type of design can and cannot support.
Learning objectives
- The four dimensions along which statistical hypotheses differ, and how to locate a research question on them
- When a one-sided test is appropriate, and the specific increase in error rate that Bland and Bland quantified when it is not
- What simple and composite hypotheses are, and why almost every hypothesis stated in practice is composite
- How superiority, non-inferiority and equivalence differ in the location of the null hypothesis, and why the margin is central to the design
- What omnibus and global null hypotheses can and cannot indicate, with the homogeneity and funnel-asymmetry null hypotheses as worked examples
- The misinterpretations associated specifically with directional and interval-based hypotheses
Common misconceptions
- “A one-sided test is simply the more powerful option.” Bland and Bland, in the BMJ, showed that conducting one-sided tests in both directions and reporting the smaller P value increases the probability of a spurious significant result from approximately 5% to approximately 10%.
- “A non-significant superiority test shows the treatments are equivalent.” Equivalence has its own null hypothesis, located outside a pre-specified margin. Failure to reject a null hypothesis of no difference does not constitute evidence for it.
- “Non-inferiority and equivalence are the same thing.” Non-inferiority is one-sided; equivalence requires both bounds. A non-inferiority design cannot support a claim of equivalence.
- “The margin is a statistical detail.” The margin defines the hypothesis. Changing the margin changes the claim, which is why it must be justified clinically and fixed in advance.
- “A significant omnibus test identifies which group differs.” It indicates that the data are incompatible with a model in which no groups differ. Which group differs is a separate question, with its own error rate.
Four dimensions of classification
Textbook lists of hypothesis types combine distinct categories, which makes the topic appear more complex than it is. There are, in fact, four independent questions, and every stated hypothesis answers all four, whether or not this is recognised.
How many parameter values does the null fix? A simple hypothesis specifies the distribution completely, with a single value for every parameter. A composite hypothesis specifies a set of values. In practice almost every tested hypothesis is composite, because even “the mean difference is zero” leaves the variance unspecified.
Does the alternative have a direction? A two-sided alternative admits departures in either direction; a one-sided alternative admits departures in only one. This is a design decision with consequences for the error rate, not a reporting preference 20.
Is the claim about difference or about similarity? Superiority asks whether an effect exists; equivalence and non-inferiority ask whether it is small enough to be unimportant 12,17. These designs place the null hypothesis in different locations, which is the most important structural distinction in the taxonomy.
One parameter, or many at once? A focused hypothesis concerns one parameter. An omnibus or global hypothesis makes an assertion about a whole set: that all group means are equal, that all studies share one true effect, or that there is no relationship between effect size and precision. Omnibus tests answer a coarser question than most authors intend.
What a hypothesis is, as distinct from which type it belongs to, is covered on our what is a statistical hypothesis page.
One-sided and two-sided hypotheses and the consequences of post hoc selection
A two-sided test treats departures in either direction as evidence against the null hypothesis. A one-sided test considers one direction only and disregards the other, which is legitimate when a result in the opposite direction would genuinely not change the decision; that condition is rarer than is often claimed 20. If an unexpected harm would be consequential, the situation is not one-sided.
Bland and Bland, in their BMJ Statistics Notes, set out the arithmetic of this misuse 20. If one-sided tests are run in both directions and the smaller P value is reported, the probability of a spurious significant result rises from about 5% to about 10% 10,20. This is not a subtle inflation but a doubling, and it occurs without being apparent, because the reported analysis resembles a single legitimate test.
Three rules follow, none of them onerous. Sidedness should be declared in the protocol or the statistical analysis plan, not in the paper. A one-sided choice should be justified by stating what action would be taken if the effect were in the opposite direction. Where there is uncertainty, a two-sided test should be used, because the penalty for a two-sided test when a one-sided test would have been defensible is a modest loss of power, whereas the penalty in the other direction is an invalid error rate. The statistical analysis plan page describes where these choices are recorded.
The error-rate cost of choosing sidedness after seeing the data. Source: Bland JM, Bland DG. Statistics Notes: One and two sided tests of significance. BMJ. 1994;309(6949):248.
Simple, composite, point and interval hypotheses
A simple hypothesis fixes every parameter of the distribution, so that exactly one distribution is compatible with it. “The data are normal with mean 0 and variance 1” is a simple hypothesis. Such hypotheses occur mainly in theory, because real hypotheses leave nuisance parameters unspecified.
A composite hypothesis specifies a set. “The mean difference is zero” is composite, because the variance is unspecified and any value of it is compatible with the null hypothesis. Every one-sided alternative is also composite, since it covers a whole half-line. This has two practical consequences. First, composite null hypotheses require the distribution of the test statistic to be the same, or approximately the same, across the whole set; this is where assumptions about variance and distributional form enter, and where they can fail. Second, power for a composite alternative is not a single number but varies across the alternative, which is why a power calculation always specifies the particular effect size at which it was computed.
A related distinction is that between point and interval hypotheses. A point null hypothesis fixes one value. An interval null hypothesis asserts that the parameter lies within a region or, in equivalence testing, outside one. Interval hypotheses provide the structural response to the objection that no effect is ever exactly zero, and they are what make claims of similarity testable at all.
Superiority, non-inferiority and equivalence hypotheses
Misstatement is most damaging in this area, because the three designs appear similar in a results table but support entirely different conclusions.
In a superiority test the null hypothesis is no difference and the alternative is that a difference exists. Rejection of the null hypothesis supports a claim that an effect is present. Failure to reject it supports no conclusion about similarity.
In a non-inferiority test the null hypothesis is that the new treatment is worse than the comparator by at least the margin, and the alternative is that it is not 12,15. This is intrinsically one-sided, because only the unfavourable direction is of interest. Rejection of the null hypothesis supports a claim that the new treatment is not meaningfully worse, not that it is as good, and certainly not that it is better 12.
In an equivalence test the null hypothesis is that the true difference lies outside the margin in either direction, and rejecting it requires demonstrating that the whole confidence interval falls inside the margin on both sides 17. For this reason equivalence is more difficult to demonstrate than non-inferiority, and a non-inferiority trial cannot be re-presented as demonstrating equivalence.
In all three designs the margin defines the study. It must be justified on clinical grounds, as the largest difference that would not change practice, and fixed before the data are available, because a margin chosen afterwards can be placed wherever the interval happens to end 1,2. Reviewers are aware of this, and an unjustified margin is one of the few statistical objections that cannot be resolved in a response letter.
Where the null sits, and what rejecting it licenses
| Type | Null hypothesis | Sided | What rejection supports | What it does not |
|---|---|---|---|---|
| Superiority | The difference is zero | Usually two-sided | An effect is present, in the direction estimated | Nothing about similarity if the null is not rejected |
| Non-inferiority | The new treatment is worse by at least the margin | One-sided by construction | The new treatment is not meaningfully worse | That it is as good, or better, or equivalent |
| Equivalence | The true difference lies outside the margin, either side | Both bounds required | The two are interchangeable within the stated margin | Anything, unless the whole interval lies inside the margin |
| Simple | Every parameter fixed at a single value | Either | That the data are incompatible with that one distribution | Which parameter is responsible |
| Composite | The parameter lies in a specified set | Either | Incompatibility with the whole set, under the model’s assumptions | A single power figure — power varies across the set |
| Omnibus / global | A statement about many parameters at once, e.g. all studies share one effect | Typically one-sided on a dispersion statistic | That the data are incompatible with the collective assertion | Which element departs, or by how much |
Structural characterisation of each hypothesis type. The homogeneity and funnel-asymmetry nulls used as worked omnibus cases are taken from DerSimonian and Laird (Controlled Clinical Trials, 1986) and Sterne et al (BMJ, 2011) respectively; the one-sided error-rate figures are from Bland and Bland (BMJ, 1994).
Omnibus and global null hypotheses, with two worked examples
Some hypotheses make an assertion about an entire collection of parameters simultaneously. They are useful, but they answer a coarser question than authors generally intend.
The homogeneity null in meta-analysis states that every included study estimates the same underlying effect 23. Cochran, in Biometrics, gave the statistic. DerSimonian and Laird, in Controlled Clinical Trials, report it for Baum’s review of antibiotic prophylaxis in colon surgery, in which the within-study homogeneity statistic Q was 40.4 on 25 degrees of freedom, indicating more dispersion than the null hypothesis predicts, since under the null hypothesis Q is expected to be close to its degrees of freedom 6,21. However, the test does not indicate how much the studies differ, which is why 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 16. The test provides a verdict on a model; the statistics provide the quantity. The fixed versus random effects page describes how the result should be used.
The funnel asymmetry null states that there is no relationship between estimated effect and precision 10,19. Sterne and colleagues, in the BMJ, make the limits explicit: tests of funnel plot asymmetry should generally not be used when a meta-analysis contains fewer than 10 studies, and the minimum may be substantially higher when heterogeneity is substantial 10,20. This is a statement about the informativeness of a hypothesis test, not about publication bias, and it illustrates a general principle: a global null hypothesis can be untestable in practice while remaining perfectly well defined.
Selecting the hypothesis type according to the decision required
The design should be derived from the conclusion that the study is intended to support. If the intended conclusion is “treatment A is better than B”, a superiority hypothesis is required, two-sided unless there is a stated reason why the opposite direction is irrelevant. If it is “A is not meaningfully worse than B, and has some other advantage”, non-inferiority with a clinically justified margin is required. If it is “A and B are interchangeable”, equivalence is required, and both bounds must be met. If it is “something differs among these five groups”, an omnibus test is sufficient; however, if the question of interest is which groups differ, the comparisons and the error rate should be planned in advance rather than treating the omnibus result as permission for further exploration.
Two further checks identify most problems. First, the hypothesis type required should be the type for which the sample size was computed. Non-inferiority and equivalence sample sizes depend on the margin, and a study powered for superiority is usually underpowered for equivalence 12,17. Second, the hypothesis type should not change between the protocol, the analysis and the abstract. A design that begins as superiority, fails, and is reported as showing no difference has changed hypothesis type without any change in its data 8,13.
Stating the hypothesis so that its type is explicit
A reviewer should be able to identify, from a single sentence of the methods, the position of the study on each of the four axes. The weak version below does not permit this, which is why the analysis plan that follows it could be almost anything.
The same study, with and without a stated hypothesis type
Type invisible
We investigated whether the once-daily regimen helps compared with the twice-daily regimen. The difference between groups was compared and significance assessed at the 5% level. As no significant difference was found, the two regimens can be considered comparable.
Which type? “Helps compared with” could be superiority, non-inferiority or equivalence, and the three have different nulls.
Sided? Unstated, so the error rate is unknown; the doubling described by Bland and Bland is precisely what this lack of transparency permits.
Where is the margin? Absent. Without one, no claim about similarity is testable at all.
The conclusion does not follow. “No significant difference, therefore comparable” is the non-significance-as-no-effect error applied to a design claim.
Type and margin declared
This was a non-inferiority comparison. The null hypothesis was that the 12-week response rate with the once-daily regimen is lower than with the twice-daily regimen by at least the non-inferiority margin of 7 percentage points on the risk difference scale, against the one-sided alternative that it is not; the test is one-sided by construction. The margin was set before recruitment at the largest reduction in response that a prescribing clinician in our advisory group judged acceptable in exchange for the adherence advantage of once-daily dosing, and was fixed in the statistical analysis plan. Non-inferiority is concluded only if the lower bound of the two-sided 95% confidence interval for the risk difference lies above the margin. We do not test equivalence, and a finding of non-inferiority will not be reported as showing the regimens equivalent.
The type is named in the first clause, so every later choice can be read against it.
Sidedness follows from the type rather than being chosen, which removes the discretion.
The margin has a value, a scale, a clinical justification and a date before recruitment.
The decision rule is stated as an interval condition, not a p-value threshold.
The claim the design cannot support is explicitly ruled out in advance.
The second version is not merely more cautious wording around the same study; it is a different hypothesis, with a different sample size calculation. This is the purpose of the taxonomy: the hypothesis type must be chosen at the design stage, because by the time of analysis the available conclusions are already fixed. The decision rule is an interval condition rather than a p-value, which is the form in which claims of similarity must be stated.
Errors associated with each hypothesis type
Greenland and colleagues, in the European Journal of Epidemiology, enumerate 25 misinterpretations of P values, confidence intervals and power, and several are specific to particular hypothesis types rather than general 6,21.
With one-sided hypotheses, the characteristic error is reporting a one-sided p-value without saying so, which halves an apparently two-sided number and is indistinguishable from a legitimate result unless the sidedness is declared 20.
With interval and equivalence hypotheses, it is the interpretation of a non-significant superiority test as evidence of equivalence, that is, the misinterpretation of a non-significant test as evidence of no effect, presented as a design claim 6,17. The interval, not the p-value, is what supports an equivalence conclusion.
With composite hypotheses, it is the citation of a single power figure as though power were a property of the study rather than of a combination of study and effect size. Power varies across a composite alternative, so “the study had 80% power” is incomplete without the effect size at which it was computed 6.
With omnibus hypotheses, it is the treatment of a significant global test as identifying the source of the departure, followed by unplanned pairwise comparisons at the nominal level. The biostatistics page describes how these comparisons should be handled.
One gap in the literature should be acknowledged. The literature is clear that margins should be clinically justified and pre-specified, but it offers no general method for setting them, and conventions differ widely between fields and regulators 1,2,12. Authors choosing a margin should therefore expect to defend a judgement rather than to apply a rule.
Key points
Key points
Statistical hypotheses differ in the number of parameter values the null hypothesis fixes, whether the alternative is directional, whether the claim concerns difference or similarity, and whether one or several parameters are tested simultaneously. The most important structural distinction is the location of the null hypothesis: superiority places it at no difference, non-inferiority at the margin on one side, and equivalence outside the margin on both sides. Rejection of each supports a different conclusion, and a study designed for one cannot be reanalysed as another. The hypothesis type and sidedness should be declared before data are collected, any margin should be justified clinically, and a non-significant superiority test should not be interpreted as evidence of similarity.
Application to your study
Application to your study
- The intended conclusion should be formulated first, and the hypothesis type that supports it selected accordingly. Superiority, non-inferiority and equivalence support three different conclusions and require three different designs.
- Sidedness should be declared in the protocol, with a rationale. If a result in the opposite direction would change the decision, the test is two-sided, and a one-sided test is not appropriate.
- A claim of similarity requires a documented margin. The margin should be recorded with its scale, its clinical justification and the date on which it was fixed. Without a margin, a claim of similarity cannot be tested.
- The sample size should be confirmed as computed for the hypothesis type in use. A study powered for superiority is usually underpowered for equivalence.
- Power should be reported with the effect size at which it was computed. Power varies across a composite alternative hypothesis, so a percentage alone is incomplete.
- Comparisons following an omnibus test should be planned before the result is known. A significant global test does not identify which element departed from the null hypothesis, and unplanned pairwise tests at the nominal level inflate the error rate.
- The appropriateness of asymmetry tests should be established. Sterne and colleagues advise against their use with fewer than 10 studies, and recommend a higher threshold when heterogeneity is substantial.
- The protocol, analysis and abstract should be compared for changes in hypothesis type. A superiority study that did not demonstrate superiority and is reported as showing no difference has implicitly changed its hypothesis type.
Frequently asked questions
It is more useful to consider four axes than a single list. The first is how many parameter values the null hypothesis fixes: a simple hypothesis fixes all of them, a composite hypothesis specifies a set, and nearly every hypothesis tested in practice is composite. The second is whether the alternative has a direction, which gives one-sided and two-sided tests. The third is whether the claim concerns difference or similarity, which separates superiority from non-inferiority and equivalence. The fourth is whether one parameter or many are tested simultaneously, which distinguishes focused from omnibus or global hypotheses. Every stated hypothesis answers all four of these questions, whether or not this was intended.
Only when a result in the opposite direction would genuinely not change the decision, which is rarer than is often claimed. If an unexpected harm or an unexpected benefit would be consequential, the situation is two-sided. The choice must also be declared in advance. Bland and Bland, in the BMJ, showed the consequence when it is not: running one-sided tests in both directions and reporting the smaller P value raises the probability of a spurious significant result from about 5% to about 10%. Because the cost of defaulting to a two-sided test is a modest loss of power, whereas the cost of the alternative is an invalid error rate, a two-sided test is the appropriate default when there is uncertainty.
The difference lies in the location of the null hypothesis and in the number of bounds that must be cleared. In non-inferiority the null hypothesis is that the new treatment is worse than the comparator by at least the margin; the test is one-sided by construction because only the unfavourable direction is of interest, and rejection supports the claim that the new treatment is not meaningfully worse. In equivalence the null hypothesis is that the true difference lies outside the margin on either side, so the whole confidence interval must be shown to fall within the margin at both ends. Equivalence is therefore the more demanding claim, and a non-inferiority trial cannot be re-presented as having demonstrated it.
No. That inference is among the misinterpretations Greenland and colleagues catalogue in the European Journal of Epidemiology, and it is the single most common statistical objection raised about comparative studies. A non-significant superiority test indicates that the data are not far from what a null hypothesis of no difference predicts; a wide interval that spans zero is also compatible with a difference large enough to change practice. An equivalence claim requires a hypothesis with a pre-specified margin and a demonstration that the confidence interval lies inside it. If no margin was set in advance, that claim is not available, and the appropriate report is the estimate with its interval.
A simple hypothesis specifies the distribution completely, with every parameter at a single value, so that exactly one distribution is compatible with it. A composite hypothesis specifies a set of values. Nearly all practical hypotheses are composite: “the mean difference is zero” leaves the variance unspecified, and every one-sided alternative covers a whole half-line. The consequences are practical rather than theoretical. The reference distribution of the test statistic must hold across the whole set, which is where variance and distributional assumptions become relevant, and power varies across a composite alternative, so a power figure is meaningless without the effect size at which it was computed.
An omnibus or global hypothesis makes an assertion about a whole collection of parameters simultaneously: that all group means are equal, that all studies in a meta-analysis share one true effect, or that there is no relationship between effect size and precision. Rejecting it indicates only that the data are incompatible with that collective assertion; it does not indicate which element departs, or by how much. The homogeneity test illustrates this. DerSimonian and Laird report Q of 40.4 on 25 degrees of freedom in Baum’s colon surgery review, which indicates more dispersion than the null hypothesis predicts, but Higgins and Thompson’s H, I-squared and R statistics were required to describe its extent.
There is no general method, and this is a genuine gap rather than a question the literature has resolved. The agreed principles are that the margin must represent the largest difference that would not change practice, must be justified on clinical grounds rather than chosen for statistical convenience, and must be fixed before the data are seen, because a margin set afterwards can be placed wherever the confidence interval happens to end. Conventions differ substantially between therapeutic areas and between regulators, so authors should expect to defend a documented judgement, recorded with its date and its reasoning, rather than to apply a rule.
The null hypothesis is perfectly well defined (no relationship between estimated effect and precision), but the test is uninformative when there are few studies. Sterne and colleagues, in the BMJ, advise that tests for funnel plot asymmetry should generally not be used when a meta-analysis contains fewer than 10 studies, and note that the minimum may be substantially higher when heterogeneity is substantial. This illustrates a general principle concerning hypothesis types: a global hypothesis can be well specified and yet untestable in practice. The appropriate report is that the number of studies did not permit a meaningful test, not a non-significant result presented as reassurance.
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
- 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
- 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
- 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
- Bland JM, Bland DG. Statistics Notes: One and two sided tests of significance. BMJ. 1994;309(6949):248. doi:10.1136/bmj.309.6949.248 read 9 October 2026. doi.org/10.1136/bmj.309.6949.248
- 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
- Fleiss JL. Measuring nominal scale agreement among many raters. Psychological Bulletin. 1971;76:378-382. read 9 October 2026. trialsjournal.biomedcentral.com
- 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). Types of Statistical Hypothesis: Classification and Appropriate Use. Pubrica. https://pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/
Pubrica Academy. “Types of Statistical Hypothesis: Classification and Appropriate Use.” Pubrica, 9 Oct. 2026, pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/.
Pubrica Academy (2026) Types of Statistical Hypothesis: Classification and Appropriate Use. Pubrica. Available at: https://pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/ (Accessed: 15 October 2026).
Pubrica Academy. “Types of Statistical Hypothesis: Classification and Appropriate Use.” Pubrica. October 9, 2026. https://pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/
Pubrica Academy, ‘Types of Statistical Hypothesis: Classification and Appropriate Use’ (Pubrica, 9 October 2026) <https://pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/> accessed 15 October 2026.
Pubrica Academy. Types of Statistical Hypothesis: Classification and Appropriate Use. [Internet]. Pubrica; 2026 Oct 9 [cited 2026 Oct 15]. Available from: https://pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/
Pubrica Academy, “Types of Statistical Hypothesis: Classification and Appropriate Use,” Pubrica, Oct. 9, 2026. [Online]. Available: https://pubrica.com/academy/concepts-definitions/types-of-statistical-hypothesis/
@misc{types_of_statistical_hypothesis_2026,
author = {{Pubrica Academy}},
title = {Types of Statistical Hypothesis: Classification and Appropriate Use},
howpublished = {Pubrica},
year = {2026},
month = {oct},
url = {https://pubrica.com/academy/concepts-definitions/types-of-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.
Hypothesis type declaration for a protocol or analysis plan
One declaration should be completed for each confirmatory hypothesis before any outcome data are seen. The four sections correspond to the four axes; an axis that cannot be specified remains open, and a reviewer is likely to identify it.
HYPOTHESIS TYPE DECLARATION
Study: ______________________ Version / date: ______________
Hypothesis number: ____ of ____ confirmatory hypotheses
THE SENTENCE WE WANT TO BE ABLE TO WRITE
______________________________________________________
(if this sentence is about similarity, sections 3 and 4
below are mandatory, not optional)
AXIS 1 - DIFFERENCE OR SIMILARITY
[ ] Superiority H0: difference = 0
[ ] Non-inferiority H0: new is worse by at least the margin
[ ] Equivalence H0: difference lies outside the margin, either side
[ ] Other (state): ___________________________________
H0 written out in full: ______________________________
H1 written out in full: ______________________________
AXIS 2 - DIRECTION
[ ] Two-sided
[ ] One-sided. What we would do if the effect went the other
way, and why that makes the other direction irrelevant:
___________________________________________________
Declared in: [ ] protocol [ ] SAP Date fixed: _______
AXIS 3 - MARGIN (similarity claims only)
Margin value: ____ Scale (risk difference / ratio / etc.): ____
Clinical justification, and who provided it: ___________
Date fixed, relative to first recruitment: _____________
Decision rule, stated as an interval condition:
____________________________________________________
AXIS 4 - SCOPE
[ ] Single parameter
[ ] Omnibus / global over: ____________________________
If rejected, the pre-planned follow-up comparisons:
__________________________________________________
Multiplicity approach for those comparisons: _______
PARAMETER AND MODEL
Parameter: ___________________________________________
Model, with nuisance parameters noted: _______________
Named test: __________________________________________
Significance level: ____
POWER
Power: ____% computed at an effect size of: __________
Type the sample size was computed FOR: _______________
[ ] this matches Axis 1 above
DRIFT CHECK (complete at write-up)
Type in protocol: ______ in analysis: ______ in abstract: ______
[ ] all three agree
[ ] no claim of similarity appears anywhere unless Axis 3 was completed 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 determine the hypothesis type, sidedness and margin at the protocol stage, when these choices can still be made, and compute the sample size for the design that will be reported rather than the design that is easiest to power. For completed analyses, the team checks for the changes in hypothesis type that reviewers commonly identify, such as a superiority study reported as demonstrating comparability, a one-sided p-value presented as two-sided, or an equivalence claim without a supporting margin, and indicates which claims the design can support.