Guides · Research ideas and questions

Hypothesis examples and how to write a testable hypothesis

A testable hypothesis predicts a direction or pattern, names what will be measured, and is stated before the data are seen, so the data could prove it wrong. The examples below contrast vague and testable marketing hypotheses, show how to set a competing hypothesis against your own, and explain why hypothesizing after the results are known turns a test into a description.

What makes a hypothesis testable

A hypothesis is a prediction that data can contradict. Nosek and colleagues separate prediction, where new data are used to confront the possibility that the prediction is wrong, from postdiction, where an explanation is built for data already seen. Only predictions give diagnostic evidence for an explanation. Postdictions are valuable for discovering ideas, but they cannot test them. The most common escape from a real test is vagueness: when a prediction is loose, many outcomes can be rationalized afterwards as support.^[1]

Earp and Trafimow explain the logic. Observing what a theory predicts does not prove the theory, because other causes could produce the same observation, whereas failing to observe it counts against the theory. A hypothesis therefore has to rule some outcomes out. They add that predictions always rest on auxiliary assumptions as well as on the theory, such as the assumption that a measure captures the construct it stands for, so a careful hypothesis makes those assumptions explicit.^[2]

A hypothesis needs a reason behind it

Sutton and Staw's essay on what theory is not treats hypotheses as bridges between a logical argument and the data, not as theory in themselves. Earlier findings alone cannot motivate a hypothesis; the author has to explain why the relationship should occur. The same holds for moderators: drawing a moderating variable in a diagram does not say why, or in which direction, it changes the effect.^[3]

The weakness is common in marketing. A systematic review of 235 B2B branding articles found that many quantitative studies relied on hypothesis testing without robust conceptual frameworks.^[4]

Testable versus untestable hypothesis examples

The pairs below are written for this guide. Each weak version could be reconciled with almost any result; each testable version names a direction, a comparison and a measurable outcome, and rests on the principle in the cited source.^[1][3]

• Vague: Retargeting influences purchase behaviour. Testable: First-time visitors randomly assigned to see a retargeting ad convert at a higher rate within 14 days than visitors randomly assigned to a holdout group. The vague version survives any result, which is exactly the looseness that lets outcomes be rationalized after the fact.^[1]
• Proposition, not yet a hypothesis: Market orientation matters for business performance. Testable: The positive relationship between market orientation and business performance is stronger in industries with more intense competition. Kohli and Jaworski stated this moderator as a research proposition; it becomes a testable hypothesis once competition and performance have agreed measures.^[5]
• Opposite moderators, each with a reason: Kohli and Jaworski also proposed that market turbulence strengthens the link between market orientation and performance, because customer preferences shift, while technological turbulence weakens it, because in fast-changing technology industries many new products originate in R&D rather than with customers. Stating the reason makes each direction something the data can reject.^[5][3]
• Obvious: Customer satisfaction increases loyalty. Stronger: predict the condition under which the expected relationship weakens or reverses, and explain why. A journal editor lists satisfaction leading to loyalty among hypotheses whose outcome is known before any test is run.^[6][7]
• Inflated: Each of five service quality dimensions positively influences satisfaction. Splitting one construct into dimensions and writing the same prediction for each adds hypotheses without adding ideas. One hypothesis about the dimension expected to behave differently, with the reason, is stronger.^[6]

How to write a competing hypothesis

The strongest tests set two plausible predictions against each other, so that either result teaches something. Huber, Payne and Puto staged such a contest. Most choice models of the time assumed regularity, meaning that adding an option cannot raise the choice share of an existing one, and the similarity hypothesis held that a new option takes share mainly from options like it. They showed that adding an option dominated by one existing item but not by another raises the share of the item that dominates it, violating both assumptions.^[8]

Written as a competing pair for a subscription business, where a new Plus plan costs more than Premium and offers less, but offers more than Basic:^[8]

• H1a, similarity: adding Plus lowers Premium's share, because the new plan is closest to Premium and draws buyers from it.^[8]
• H1b, attraction: adding Plus raises Premium's share, because Premium now clearly dominates an available option. Huber and colleagues note the managerial stakes: a product line can become more profitable by adding an option that almost nobody chooses.^[8]
• Established regularities also supply precise predictions. When Wilson and colleagues asked whether leisure physical activity follows the negative binomial distribution, the buying regularity predicted a graded spread of weekly engagement rather than a split into active and inactive people. The data could have contradicted it; instead most people were non- or light-engagers, with engagement varying by degree.^[9]

Hypothesizing after the results are known

Writing a hypothesis after seeing which result came out, and presenting it as if it had been predicted, is known as HARKing. Nosek and colleagues describe it as circular reasoning: the data generate the hypothesis and then serve as its test. Hindsight bias makes it easy to do without noticing, because an explanation built after the fact feels as if it had been expected all along.^[1]

It is not rare. In a survey of psychologists reported by Earp and Trafimow, 54% admitted claiming to have predicted an unexpected finding. Golder and colleagues surveyed marketing scholars and found that 28% develop hypotheses at later stages of research; with 140 responses and a 15.4% response rate, treat that figure as indicative rather than precise.^[2][10]

Exploring data is legitimate; disguising exploration as a test is not. Golder and colleagues argue for empirics-first research that starts from data, yet call reporting relationships found that way as prior hypotheses unacceptable, and recommend reporting both the original and the revised hypothesis when a hypothesis changes during review. An editor of an interactive marketing journal likewise criticizes hypotheses written retroactively after data-driven analysis.^[10][6]

Protect the test before you collect data

Flexible analysis can make a false hypothesis look supported. Simmons, Nelson and Simonsohn simulated 15,000 samples with no true effect and showed that combining four common choices (two correlated dependent variables, adding ten observations per cell, controlling for gender, and dropping one of three conditions) raised the rate of significant results at p < .05 to 60.7%.^[11]

Their requirements for authors include deciding the rule for stopping data collection in advance, collecting at least 20 observations per cell, listing every variable measured and reporting every condition run. Nosek and colleagues add preregistration, writing down the hypotheses and analysis plan before seeing outcomes. When the data already exist, one option is to split them: explore one part, and keep the other sealed as a holdout for testing what the exploration suggested.^[11][1]

The evidence behind this

• Research methods pitfalls: ratio variables and censored data
• Marketing knowledge: how it is made and when to trust it
• Choice architecture and decision biases: what the research shows

References

1. Brian A. Nosek, Charles R. Ebersole, Alexander C. DeHaven, David T. Mellor (2018). The Preregistration Revolution. Proceedings of the National Academy of Sciences.
2. Brian D. Earp, David Trafimow (2015). Replication, falsification, and the crisis of confidence in social psychology. Frontiers in Psychology.
3. Robert I. Sutton, Barry M. Staw (1995). What Theory is Not. Administrative Science Quarterly, 40(3), 371–384.
4. Christian Kowalkowski, Joona Keränen, Edwin J. Nijssen (2026). How B2B branding research measures up: A systematic review of key themes, theories, and methods. Industrial Marketing Management.
5. Ajay K. Kohli, Bernard J. Jaworski (1990). Market Orientation: The Construct, Research Propositions, and Managerial Implications. Journal of Marketing.
6. Cheng Lu Wang (2022). Editorial – The misassumptions about contributions. Journal of Research in Interactive Marketing, 16(1), 1–2.
7. Cheng Lu Wang (2025). Editorial: Demonstrating contributions through storytelling. Journal of Research in Interactive Marketing.
8. Joel Huber, John W. Payne, Christopher Puto (1982). Adding Asymmetrically Dominated Alternatives: Violations of Regularity and the Similarity Hypothesis. Journal of Consumer Research.
9. Amy L. Wilson, Byron Sharp, Cathy Nguyen, Svetlana Bogomolova (2017). Expanding Marketing Empirical Generalisations to Health Behaviours: Physical Activity Is Not So Different From Buying Behaviour, After-All. Australasian Marketing Journal.
10. Peter N. Golder, Marnik G. Dekimpe, Jake T. An, Harald J. van Heerde, Darren S. U. Kim, Joseph W. Alba (2022). Learning from Data: An Empirics-First Approach to Relevant Knowledge Generation. Journal of Marketing.
11. Joseph P. Simmons, Leif D. Nelson, Uri Simonsohn (2011). False-Positive Psychology. Psychological Science.

Bibliographic details only. Follow the DOI to read each study at its publisher.