A research hypothesis states the answer you expect to a substantive research question. A statistical hypothesis translates a particular part of that prediction into a statement about a population parameter or distribution. The two are not interchangeable. Testing a hypothesis in a bachelor's or master's thesis also requires a defined measurement, a sample, an analysis plan and clear limits on the conclusion.
The example below takes a prediction about students through to a null hypothesis, an alternative hypothesis and a plan for the analysis. It also explains when you can reject the null, why a non-significant result does not establish the absence of an association, and what to do when your questionnaire does not measure what your methodology chapter says it does.
Research hypothesis vs statistical hypothesis: what are you testing?
“Students who organise their studying more effectively put off assignments less often” is a substantive prediction. Its concepts still need definitions. It does not yet specify how to measure study organisation, what counts as putting off an assignment, or how to assess an association. Adding the label H1 does not turn the sentence into a complete testing plan.
A statistical hypothesis refers to a specified model. For example, you might test whether the correlation between the scores on two named scales equals zero in a defined population. That test assesses a particular type of association between measured scores. It does not automatically settle an entire theory of study organisation or every possible relationship between students' behaviours.
If you are still defining your aim and research questions, start with our guide “Thesis research methodology: how to formulate the aim, research questions and hypotheses?”. Here, we focus on the next step: giving a hypothesis a form that can be assessed and deciding what would count as an answer.
One example, from research question to measurement
Study organisation and academic procrastination
Question: among master's students at a specified university, is better study organisation associated with lower academic procrastination?
Research hypothesis: better study organisation is associated with lower academic procrastination. “Associated with” describes a relationship. It does not assume that study organisation causes the difference.
Measurement plan: both characteristics are measured using suitable instruments described in the thesis. Higher scores on the first scale should indicate better organisation; higher scores on the second should indicate more procrastination. Before analysis, check the scoring instructions, reverse-scored items and permitted missing responses.
A single questionnaire item asking “Do you organise your studying well?” would not automatically be equivalent to a measure of the whole construct. You would need to justify what that response actually captures. We discuss the step from concept to indicator separately in our guide “Operationalising variables in a thesis: indicators, measurements and examples”.
Also decide who your conclusions are intended to describe. Responses from friends in one seminar do not become representative of all students in Poland because software produces a p-value. Recruitment, participation among those invited and the limits of the sampling approach must be explained independently of the test result.
How should you write the null and alternative hypotheses?
Define your notation at the start of the chapter. H1 can mean the first research hypothesis, but it is also commonly used for the statistical alternative. To avoid confusion, you could label substantive hypotheses RH1 and RH2, and use H0 and Ha for statistical hypotheses, explaining the convention you have chosen.
A two-sided correlation hypothesis
H0: the population correlation coefficient equals zero. Ha: the population correlation coefficient differs from zero.
For Pearson's correlation, this concerns the parameter rho and a linear association. For a rank-based analysis, specify the corresponding coefficient and interpretation. Do not insert the name of any available test into the same sentence without checking what that test actually assesses.
A two-sided alternative includes both positive and negative associations. A statistically significant positive association may therefore allow you to reject H0, while failing to support a prediction of a negative association. The statistical decision and the answer to the research hypothesis need two separate statements.
The null does not always mean “nothing is happening”. It can concern a specified mean, a particular difference or the fit of a distribution. In this example, zero linear correlation does not exclude every possible relationship, such as a curved one. The methods section must therefore identify the property being tested precisely.
A directional prediction and a one-sided test are different decisions
“Better organisation is associated with less procrastination” is a directional prediction. This does not automatically require a one-sided test. Choosing a one-sided or two-sided procedure should follow from the scientific rationale, study design and analysis rules established before you inspect the results.
For a one-sided test of a negative correlation, Ha specifies a value below zero, while H0 includes zero and positive values. The familiar equality-only notation for H0 then represents the boundary used in calculating the test. Explain the direction and procedure in your thesis rather than leaving an ambiguous shorthand.
Do not switch to a one-sided test because the two-sided result is close to 0.05. That changes the rule after seeing the data. An important, unplanned direction of association can still be discussed, but it must be clearly distinguished from the test specified in advance.
An association and a group difference answer different questions
In our example, both scores are numerical variables and the question concerns how they vary together. If you instead want to compare students who are employed with those who are not, you are changing the research question. You now ask about a difference in procrastination between defined groups. Employment and group membership also need clear definitions.
For a comparison of means, H0 might state that the difference between population means is zero. This is not the same as saying that the distributions are identical in every respect. Nor does choosing a rank-based test automatically justify describing its result as a difference in medians. Interpretation depends on the hypothesis tested and any additional assumptions.
Do not split a scale score into “low” and “high” simply to fit a familiar test. The split discards information and its threshold needs justification. First identify the association or difference you want to estimate; then choose an analysis suited to the measurement and study design.
What belongs in a hypothesis-testing plan?
A list reading “mean, standard deviation, t-test, correlation” is not yet a plan. It does not show which result will answer each question or what will happen if the data do not support the intended procedure. For every main hypothesis, write a short plan covering the decisions below.
- Variable and score: the instrument, scoring procedure, direction of the scale and unit of analysis. State whether one row represents a person, a visit or, for example, a school class.
- Study material: eligibility criteria, recruitment, planned sample size and its justification. There is no single respondent count suitable for every hypothesis.
- Data preparation: duplicate detection, missing-data rules, exclusions and handling of unusual observations. Do not remove responses because they weaken the result you expected.
- Main analysis: the parameter or comparison, procedure, assumptions, sidedness and significance level. Also specify a justified course of action if the assumptions are not met.
- Reporting: the estimated association or difference, uncertainty, p-value, number of observations analysed and limits of the conclusion. If there are multiple tests, explain how multiplicity will be handled.
Checking the distribution of one column is not enough. Independence of observations, how groups were formed, the shape of an association and properties of model residuals may also matter. The appropriate checks depend on the analysis. For a complex project, it is worth addressing these decisions before data collection through the service “Statistical analysis for bachelor’s and master’s theses”.
When should hypotheses be set, and when can they change?
Record hypotheses intended for testing and your main analytical decisions before examining the results. Keep a dated version of the plan. The Center for Open Science presents preregistration as a way to distinguish planned analyses from ideas developed after looking at the data. Preregistration itself does not replace sound design or an assessment of measurement quality.
Changes may be necessary: an instrument becomes unavailable, recruitment takes a different course, or a pilot reveals a confusing question. Explain what changed, when and why. A correction before the main data collection is different from adjusting a hypothesis to match a pattern already noticed in a completed dataset.
Additional analyses can be useful when they answer a meaningful question. Do not present them as planned from the outset. For example, an association found in only one of several examined subgroups requires more cautious interpretation and attention to the number of comparisons performed.
What does a p-value tell you?
The significance level alpha, for example 0.05, is set before testing. In an appropriately constructed procedure, it controls the risk of a Type I error: rejecting a true H0. It does not mean there is a five per cent probability that this particular conclusion is wrong. With multiple comparisons, also specify how error risk will be controlled across the relevant family of tests.
A Type II error occurs when a false H0 is not rejected. Its probability depends on factors including the true effect, sample size, variability and procedure. Statistical power is the probability of rejecting H0 under a specified alternative. Sample-size planning therefore requires assumptions about the effect you want to detect. More responses alone cannot repair an unsuitable measure or biased sampling.
A p-value assesses the compatibility of the data with a specified null hypothesis under the chosen model and its assumptions. It is not the probability that your research hypothesis is true. The ASA also emphasises that statistical significance does not establish an effect's size or practical importance. A conclusion needs more than a comparison between one number and a threshold.
Three ways an analysis might turn out
Negative association, H0 rejected: the direction agrees with the prediction. Still assess the magnitude, confidence interval and credibility of measurement and sampling. Do not add a causal conclusion.
Positive association, H0 rejected in a two-sided test: the direction is opposite to the prediction. You cannot describe the research hypothesis as supported merely because the result is statistically significant.
H0 not rejected: the analysis did not provide sufficient grounds for rejecting it. Assess the range of values compatible with the estimate. A wide confidence interval may leave both small and substantively meaningful associations plausible.
If your aim is to demonstrate that a difference is practically negligible, an ordinary p-value above 0.05 is insufficient. You need a design that can assess equivalence or estimate the effect with suitable precision, together with a previously justified threshold of practical importance. Do not substitute a standard test of no difference for that objective.
Three wording changes that alter the meaning
- “Study organisation affects procrastination” should become a prediction about an association if you have a one-time survey. “Affects” requires a defensible causal inference, not simply a significant correlation.
- “The hypothesis is true because p < 0.05” should be split into the decision about H0 and an interpretation of the estimated effect. Report direction, magnitude, uncertainty and limitations.
- “There is no association because p > 0.05” should become a statement that this analysis did not provide sufficient grounds for rejecting H0. The uncertainty of the estimate helps determine how far your conclusion can go.
What if your questionnaire cannot test the hypothesis?
Identify the reason first. Is a variable missing, has a scale been scored incorrectly, is one group barely represented, or are responses incomplete? Incorrect scoring can be repaired using the instrument's instructions. You cannot reconstruct an answer to a question that was never asked.
Distinguish three situations: the hypothesis was assessed and received support, it was assessed without sufficient support, or it cannot be assessed reliably. Do not describe the third situation as a refutation. Discuss with your supervisor whether additional data collection, a narrower objective or an explicit limitation is appropriate. Preserve an accurate account of how the research unfolded.
Not every thesis requires statistical hypotheses. A qualitative analysis of experiences may use research questions, and describing response distributions does not always require a test. A Polish ZPSB thesis guide discusses how the need for hypotheses depends on the type of research. Check your own programme's requirements separately; another institution's guide is not your university's rulebook.
Check alignment before submitting the chapter
- Each research hypothesis answers an identifiable question rather than simply echoing the thesis title.
- The hypothesis, instrument, dataset and analysis plan refer to the same variables.
- You distinguish a numbered research hypothesis from the symbol for a statistical alternative.
- The plan states what will be tested and which results will accompany the p-value.
- Conclusions stay within the limits of the design, measurement and participant selection.
- Changes and additional analyses are transparent, and lack of support is not concealed.
Sources and further reading
The materials below support the distinctions and interpretation principles discussed here. Sources checked on 28 October 2026.
- Stefanowski, J. Metodologiczne aspekty prowadzenia badań naukowych. Poznan University of Technology, lecture materials, shortened version 2019. Polish PDF on research problems, hypotheses and procedures.
- Jankowska, A. (2022). Poradnik dyplomanta dla studentów studiów I stopnia na kierunku Pedagogika. ZPSB. Polish academic guide, research-design section.
- Statistics Poland. Testowanie hipotezy statystycznej, terms used in official statistics. Definition of statistical hypothesis testing.
- Wasserstein, R. L., Lazar, N. A. (2016). The ASA's Statement on p-Values: Context, Process, and Purpose. The American Statistician, 70(2), 129-133. DOI: 10.1080/00031305.2016.1154108. Official PDF introducing the ASA's six principles.
- Center for Open Science. Preregistration. Planning and distinguishing confirmatory from exploratory analyses.
Questions about hypotheses and analysis
Does every thesis or survey require hypotheses?
No. Whether hypotheses are appropriate depends on the objective and nature of the research. Testing relationships requires a different approach from research focused on description or exploring experiences. I adapt the approach to the project and the requirements of the relevant degree programme.
If you formulate hypotheses, plan the data and analysis that will allow you to actually test them. Do not add them solely to fill out the chapter scheme.
What if the results do not support the hypothesis?
I describe this in line with the analysis. A lack of support for the hypothesis is a research finding, not a reason to change the respondents' answers.
It also does not automatically mean the aim was not achieved: the aim may have been to check a specific relationship. I explain what follows from the data and what limitations must be considered. A lack of support for the hypothesis does not by itself prove that the relationship definitely does not exist.
What to do if the hypothesis could not be tested?
Write what prevented its evaluation: lack of necessary measurement, incomplete data, or other limitations. Distinguish the inability to evaluate from a result that does not support the hypothesis. Agree on the next steps with the supervisor, without adding answers or changing the data.
Is a p-value < 0.05 sufficient to confirm a hypothesis?
No. With a pre-set significance level of 0.05, a p-value < 0.05 allows rejecting the null hypothesis within the adopted model and its assumptions. It does not mean that the research hypothesis is true with 95% probability.
Describe the direction and size of the effect, the confidence interval, the sample size, and the limitations of measurement and participant selection. Also consider multiple testing. A significant result alone does not prove causality or the practical importance of the difference.
Do you help with the methodology of original research?
Yes. I offer support with the methodology of original research, including defining the research objective and subject, research questions, hypotheses, variables, indicators, method, technique, and research instrument.
The most important thing is that all elements of the methodology are consistent with one another. The research instrument, such as a questionnaire or interview guide, should follow from the research questions and enable you to collect the data needed for the subsequent analysis of results.
Connect your hypotheses to data you can actually collect
Has your supervisor queried your hypotheses, or are you unsure how to turn them into an analysis? Send me your thesis topic, research questions, instrument and programme requirements. I will check how they fit together and help define the methodology or statistical support you need. If you have collected the data, include an anonymised file and a description of the variables. We will agree the approach, timeframe and price before work begins.
Discuss your hypotheses and analysis plan with Wiola
