A research hypothesis is a testable prediction about the relationship between two or more variables. To write one, put the variables, the relationship and the population together in one claim that data can prove wrong. It converts your research question into a claim which your data can support or fail to support. Research committees and examiners read the hypotheses very carefully. If the hypothesis is badly worded, mostly the analysis also won’t answer the question.
In this guide we cover what makes a hypothesis testable. Then the difference between null and alternative hypotheses, and between directional and non-directional ones. You will see how to get hypotheses out of your literature, with worked examples from different fields. Last part is about studies which don’t need a hypothesis at all. Last reviewed September 2026.
What makes a hypothesis testable
A good hypothesis is one specific statement which is falsifiable, meaning data can show it is wrong. See the difference in these two:
- Not testable: “Social media has an impact on youth.” Which social media? What impact, and on which youth? Any result will fit this sentence.
- Testable: “Among undergraduate students in Madurai, daily time spent on Instagram is negatively associated with self-reported sleep quality.” Here the variables, the population and the direction are all given. A survey can show this is false.
Check these four points in every hypothesis:
- Variables are named and measurable. Till you decide how each variable will be measured, there is no way to test the hypothesis.
- The relationship is stated. Write the kind of relationship you expect. It may be two groups differing, or one variable affecting another, or two just moving together.
- The population or context is stated. For whom or for what is the claim made?
- It is one claim. “X affects Y and Z” is actually two hypotheses. Make it two, since the data may support one part and not the other.
Many scholars are mixing up the hypothesis and the objective. The objective tells what you are going to do (“to examine the effect of X on Y”). The hypothesis tells what you expect to find (“X increases Y”).
Null and alternative hypotheses
In statistical testing the hypotheses come as a pair. The null hypothesis (H₀) is the claim that there is no effect and no difference. The alternative hypothesis (H₁) says that some effect is there. Your test calculates how likely results at least as extreme as yours would be if the null was true. This probability is the p-value. Usually the significance level is 0.05. If your p-value comes lower than the level you fixed, the null is rejected and you go with the alternative.
| Type | What it says | Example |
|---|---|---|
| Null (H₀) | There is no effect, difference or relationship | There is no difference in mean systolic blood pressure between patients on drug A and drug B. |
| Alternative (H₁ or Hₐ) | There is an effect, difference or relationship | Mean systolic blood pressure differs between patients on drug A and drug B. |
| Directional (one-tailed) | The effect goes in one stated direction | Patients on drug A have lower mean systolic blood pressure than patients on drug B. |
| Non-directional (two-tailed) | There is an effect, but the direction is not given | Mean systolic blood pressure differs between the two groups. |
| Associative / correlational | The two variables move together | Job satisfaction is positively associated with organisational commitment among bank employees. |
| Causal | A change in one variable brings a change in the other | Adding 20% fly ash increases the 90-day compressive strength of M30 concrete. |
Two points examiners pick up
You never “accept” or “prove” a hypothesis. The only two outcomes are that you reject the null or you fail to reject it. Suppose you could not reject the null. Still the effect may be there, and maybe your sample was too small to catch it. So in your results write “the data did not support H₁” or “H₀ was not rejected”. Don’t write “H₀ was accepted” or “it is proved that”.
Statistical significance is not size. With a large sample, even a very small, unimportant difference can come out significant. So give the effect size and the confidence interval also, and the reader can decide if the result matters in practice.
Directional or non-directional?
Give a direction only if earlier studies or theory give you a clear reason to expect it. A directional hypothesis is tested one-tailed. With a one-tailed test it is easier to get significance in that direction. But after that you cannot claim an effect in the opposite direction. Fixing the direction after looking at the data is not acceptable. When you are in doubt, go for a non-directional hypothesis with a two-tailed test. Most examiners and reviewers won’t question that.
Deriving hypotheses from your literature
Every hypothesis has to follow from the literature or from some theory. In the thesis or paper, write the reasoning in one or two paragraphs first, and only then state the hypothesis. For a quantitative thesis, mostly each arrow in your conceptual framework becomes one hypothesis. The usual pattern is like this:
- State what theory or earlier studies say. “The Job Demands–Resources model predicts that job resources reduce burnout, and studies of nurses in several countries support this.”
- Explain why it should apply to your context, or why it may be different there. “Private hospitals in India offer fewer formal job resources, which may make supervisor support more important.”
- State the hypothesis. “H2: Supervisor support is negatively associated with burnout among nurses in private hospitals in Chennai.”
Is your management or social science thesis built on a conceptual model? In that case number the hypotheses as H1, H2, H3a, H3b and so on, and show them as paths in the model diagram. Each arrow gets its own hypothesis, and each hypothesis points to one arrow or one set of arrows only.
How many hypotheses?
Write as many as your objectives and your model need, and no more. For a PhD with five objectives, eight to fifteen hypotheses is possible. Some Indian departments commonly see a thesis with forty hypotheses. Most of them are demographic comparisons (“there is no significant difference in X by age, gender, income, education…”). Mostly these don’t answer the research question, and the thesis looks weaker. Your department may still want those age and gender tables. Fine, give them in a descriptive section. Hypotheses are for relationships your study is really about.
When you test many hypotheses, the chance that some come out significant just by accident goes up. Say you run 20 independent tests at the 0.05 level. You would expect around one false positive even when no real effect is there. So if you are running many related tests, think about a correction, and discuss about it with your statistician or guide.
Examples of research hypotheses from several fields
Each row shows a research question turned into a testable hypothesis, and the test that normally checks it. Which test is right for you depends on your data and your design. Our statistical test guide has more on which test goes with which type of data.
| Field | Research question | Hypothesis (H₁) | Likely test |
|---|---|---|---|
| Management | Does work–life balance affect nurses’ intention to leave? | Work–life balance has a negative effect on turnover intention among nurses in private hospitals in Chennai. | Regression or SEM |
| Education | Does peer tutoring improve mathematics scores? | Class IX students taught with peer tutoring score higher on a standardised mathematics test than students taught conventionally. | Independent t-test or ANCOVA with pre-test |
| Agriculture | Does drip irrigation change sugarcane yield? | Mean yield per hectare is higher under drip irrigation than under furrow irrigation. | t-test or ANOVA across plots |
| Public health | Is screen time linked to sleep in adolescents? | Daily screen time is negatively correlated with sleep duration among adolescents aged 13–17. | Pearson or Spearman correlation |
| Commerce | Does gender relate to choice of investment avenue? | Choice of preferred investment avenue is associated with gender among salaried investors. | Chi-square test of independence |
| Engineering | Does the proposed algorithm beat the baseline? | The proposed model achieves higher F1-score than the baseline on the test set across 10 runs. | Paired t-test or Wilcoxon signed-rank |
Each hypothesis above gives a population, a measurable variable and a relationship. From the hypothesis and the data type you get the test, and from the test you get the sample size. For power analysis you must know your hypothesis and your test before data collection starts. Power analysis is explained with more detail in our guide to sample size calculation.
When you do not need a hypothesis
Some studies don’t test hypotheses at all, and forcing one into such a study makes the work weaker.
- Descriptive studies estimate how common something is, for example “the level of awareness of crop insurance among farmers”. For these you need research questions and no hypotheses.
- Qualitative studies deal with meanings and processes. A scholar doing interviews or case work mostly starts from research questions. The theory may come out from the data, instead of the data testing a theory.
- Exploratory studies are done in new areas. There may not be enough earlier researches to justify a prediction.
- Design and development work in engineering usually gives performance objectives (“to achieve accuracy above the current baseline”) and not formal hypotheses. Still, comparisons with the baseline can be written as hypotheses and tested.
Sometimes the committee ask for hypotheses even in a study like these. Then explain your design and propose research questions in place of hypotheses. Most committees accept a clear reason. For where the synopsis comes into this, read our PhD synopsis guide.
Hypothesis checklist and common mistakes
Go through this list for each of your hypotheses:
- Every variable is named, and you can measure each one with the data you will really collect.
- The population or the context is written in the hypothesis itself.
- The prediction is such that your data could prove it wrong.
- It makes only one claim, so if you have joined two claims with “and”, split them.
- It comes out of your literature review or a theory. A hunch is not enough.
- If you have given a direction, earlier studies support that direction.
- You already know which statistical test you will use on it, and you know this before data collection.
Common mistakes
- Writing the hypothesis only after the analysis, so that it matches whatever came significant. The name for this is HARKing (hypothesising after the results are known), and examiners don’t let it go easily.
- Giving only null hypotheses as a list, with no reasoning behind them. In many Indian theses this is the convention. If the department insists, follow that format. Just add the alternative hypothesis and its rationale next to each null.
- Writing hypotheses on variables which the questionnaire is not measuring.
- Reporting “H₀ accepted” or “hypothesis proved” in the results chapter.
- Taking “no significant difference” to mean that there is surely no difference.
The hypotheses are only one part of the methods chapter, and our guide on writing the research methodology chapter covers the rest of it. Some scholars want a methodologist to look at their model and hypotheses before they collect data. Our research methodology consulting does that review.
Want a specialist to help with this in your own work? Research Proposal and Synopsis: a free first consultation, then a written scope and fixed quote.
Sources
FAQ
Questions scholars ask
Should I state the null or the alternative hypothesis in my thesis?
Conventions differ. Many journals and international examiners want the alternative hypothesis, written as a prediction. Many Indian departments want the null. Most readers are fine if you give both. Giving H₁ with its rationale and making the null clear in the results chapter is also accepted. To be safe, see what the recent theses from your own department have done and follow that.
Can a hypothesis be wrong?
Yes, and an unsupported hypothesis is still a finding, so no need to worry. Write the result as it came out. Then give the likely reasons and compare with earlier studies. Examiners mostly think better of a thesis which is honest about it.
What significance level should I use?
In most fields the usual convention is 0.05. Some fields use 0.01. Fix the level before you start the analysis and mention it in the methodology chapter. Also put the exact p-value for each test in the results, because “p < 0.05” alone hides it.
Is a hypothesis the same as a research question?
No. The research question asks something (“Does X affect Y?”). The hypothesis answers it in advance with a prediction, and the data then test that prediction (“X increases Y”). In quantitative studies you usually find both. Qualitative studies usually have only the questions.
How do I write hypotheses for a structural equation model?
For each path in your model, write one hypothesis giving the two constructs and the direction. Mediation and moderation hypotheses you have to state separately. For example, “Organisational commitment mediates the relationship between job satisfaction and turnover intention.” How to test and report them is shown in our mediation and moderation guide.
