01 · The Question
If the Data Give You a New Hypothesis, Are You Allowed to Report It?
You run a study expecting one result and discover something else. The unexpected pattern is intriguing. It fits a plausible theory, perhaps even better than your original prediction, and immediately suggests a new hypothesis.
Can that hypothesis appear in the paper?
Yes. Research would be strangely unproductive if investigators were forbidden from learning anything they had not predicted. The integrity problem begins when the manuscript erases the chronology and presents a hypothesis generated from the observed results as though it had been specified before those results were known.
03 · What You Need to Know
The Problem With HARKing Is the Rewritten Chronology
What HARKing Actually Means
Psychologist Norbert Kerr introduced the term HARKing, short for Hypothesizing After the Results are Known, in a 1998 article examining the practice.
Kerr defined HARKing specifically as presenting a post hoc hypothesis, one based on or informed by the observed results, in the research report as though it had been an a priori hypothesis.
That definition contains an important distinction.
Simply thinking of a hypothesis after seeing data is not enough. The reporting practice matters. A researcher can develop a new hypothesis after examining results and openly say so. That is hypothesis generation, not necessarily HARKing.
A Priori and Post Hoc Describe When the Hypothesis Arose
A priori hypothesis
A hypothesis formulated independently of the particular results being used to evaluate it.
Post hoc hypothesis
A hypothesis developed or materially informed after the researcher has seen the relevant results.
These terms concern chronology and information, not intellectual quality.
A post hoc hypothesis can be theoretically sophisticated. An a priori hypothesis can be poorly reasoned. The distinction matters because the evidential relationship between the hypothesis and data differs.
Why the Same Data Cannot Be Treated as Completely Independent Discovery and Confirmation
Suppose you measure 20 variables without predicting a particular association. One surprisingly strong relationship appears. After seeing it, you develop a theory that predicts exactly that relationship.
The theory may be excellent. But the observed relationship already influenced which hypothesis you chose.
If you then present that same relationship as a successful test of a prediction that supposedly existed before the analysis, the reader receives a misleading picture. The apparent match between prediction and evidence is partly guaranteed by the process that generated the prediction.
This does not render the finding worthless. It changes what the finding can establish.
HARKing Can Make Theory Look More Predictive Than It Was
Imagine a paper with a beautifully structured introduction. It reviews previous theory, develops a precise hypothesis, and then reports exactly the predicted result.
A reader reasonably infers that the theory generated the prediction before the outcome was known.
Now imagine that the researchers actually began with a different hypothesis, failed to support it, noticed an unexpected pattern, and then reconstructed the introduction around that pattern.
The final manuscript creates an illusion of prospective theoretical success. It makes the theory appear to have predicted evidence that actually helped create the theory.
This is one reason HARKing can distort evaluations of theories. If unsuccessful predictions disappear while successful post hoc explanations are routinely rewritten as predictions, published theories can appear more predictively successful than the underlying research process warrants.
HARKing Can Hide the Number of Possible Explanations
Unexpected results often permit several plausible explanations.
Researchers may consider mechanisms A, B, C, and D before deciding that B provides the most convincing narrative. If only B appears in the paper as the original hypothesis, readers do not see that the explanation was selected after observing the result.
The same logic resembles other forms of result-dependent selection. Andrade describes HARKing alongside practices such as cherry-picking and p-hacking because each can alter the relationship between the full research process and the polished account readers receive.
Not Every Post Hoc Explanation Needs to Be Banished to a Footnote
An unexpected result may become the most scientifically interesting finding in the study. You can discuss it seriously.
For example:
“We did not predict this interaction. One possible explanation is that prior experience changes how participants respond to the intervention. This interpretation should be treated as hypothesis-generating and tested prospectively.”
That statement does not weaken the scholarship. It tells readers exactly where theory generation occurred.
The researcher can then explain why the proposed mechanism is plausible, connect it to existing literature, consider alternatives, and design future work around it.
HARKing Is Not the Same as Exploratory Research
Exploratory research intentionally looks for patterns, relationships, mechanisms, or hypotheses that were not necessarily specified beforehand.
That is a legitimate scientific function.
HARKing occurs when the exploratory origin is concealed and the final report makes the hypothesis look prospective. In other words, the problem is not that researchers explored. It is that readers are told a confirmatory story about an exploratory process.
This distinction is central to reporting exploratory findings responsibly.
Post Hoc Does Not Mean False
A hypothesis does not become incorrect merely because it was generated after seeing data.
Many important scientific ideas arise from unexpected observations. Serendipity would have a rather difficult publication record under a rule forbidding post hoc thinking.
The relevant question is how much evidential weight the current dataset should carry as a test of a hypothesis that the dataset itself helped generate.
A useful next step is often to test the hypothesis with new observations collected under conditions in which the prediction is specified before the relevant outcome is known.
Preregistration Can Establish the Timeline More Clearly
One practical difficulty is that readers usually cannot tell from a finished manuscript when a hypothesis was formulated.
A prospectively timestamped protocol, preregistration, or analysis plan can document hypotheses before the relevant results are observed. Researchers can then add exploratory hypotheses later without pretending those additions were part of the original plan.
Preregistration does not make a hypothesis theoretically sound, nor does it prevent researchers from exploring. Its value here is partly temporal: it provides evidence about which questions preceded the results.
HARKing and P-Hacking Are Different
The terms are sometimes grouped together, but they describe different parts of the research process.
| Practice |
What is result-dependent? |
Illustrative example |
| HARKing |
The reported hypothesis or its apparent timing |
An unexpected relationship is rewritten as an original prediction |
| P-hacking |
Data collection or analytical choices |
Several models are tried until one produces statistical significance |
| Selective reporting |
Which findings readers receive |
Unfavorable relevant results are omitted while favorable findings are emphasized |
The practices can occur together. A researcher might p-hack a significant result, construct a hypothesis explaining it, and selectively report the successful analysis. But one does not logically require the others.
Whether Every Form of HARKing Is Equally Harmful Is a More Nuanced Question
Kerr's original treatment did not reduce the issue to a simplistic claim that every conceivable form of HARKing has identical costs. He explicitly described the balance of possible costs and benefits as complex and called for research and debate about the practice.
That nuance should not be confused with permission to disguise post hoc hypotheses casually. The cleanest practical solution is usually available: tell readers when the hypothesis arose.
Watch Out
Changing the tense of the sentence does not solve HARKing. Writing “we hypothesized” rather than “we predicted” still implies a chronology if the manuscript presents the hypothesis as motivating the analysis. The important issue is whether readers can tell that the hypothesis was developed after the relevant result was known.
04 · A Practical Example
The Same Unexpected Finding Can Be HARKed or Reported Transparently
Hypothetical Example
An Unexpected Effect of Prior Experience
A researcher predicts that a new learning tool will improve assessment scores for all students. The overall effect is inconclusive. During exploratory analysis, however, the researcher discovers a strong interaction: students with substantial prior experience appear to benefit more than beginners.
Version A: HARKing
Rewrite the rationale
The introduction is revised to argue that theory specifically predicts stronger effects among experienced students.
Rewrite the hypothesis
The manuscript states that the researchers hypothesized an interaction with prior experience.
Present confirmation
The observed interaction is described as support for the predicted hypothesis.
Version B: Transparent Post Hoc Hypothesis
Report the original prediction
The researchers state that the planned overall-effect hypothesis was not clearly supported.
Identify the discovery
They explain that an exploratory analysis revealed a possible interaction with prior experience.
Develop the hypothesis
They propose a theoretical mechanism that could explain why experienced students benefit more.
Set the next test
They describe the interaction as a hypothesis for prospective evaluation in independent data.
The underlying dataset is identical in both versions.
What changes is what readers are told about the intellectual chronology. Version A converts discovery into apparent prediction. Version B allows the unexpected finding to remain interesting without giving it an evidential history it never had.
07 · A Quick Checklist
Before Reporting a Post Hoc Hypothesis
Check the chronology of the hypothesis:
Did you formulate the hypothesis before or after seeing the relevant result?
Did the observed direction, subgroup, association, or effect influence the hypothesis you ultimately chose?
Does the manuscript make the timing of the hypothesis clear to readers?
Have you preserved the original hypotheses even when they were unsupported?
Are post hoc analyses distinguished from genuinely prespecified confirmatory analyses?
Does your interpretation acknowledge that the same data helped generate and evaluate the new hypothesis?
Have you avoided rewriting existing theory as though it necessarily predicted the unexpected result beforehand?
Would independent data provide a useful prospective test of the newly generated hypothesis?
