03 · What You Need to Know
What Can Change and What Each Change May Mean
Your Sample Size Can Change
A preregistration may specify a target sample size or stopping rule, but actual recruitment does not always cooperate.
You may fail to reach the intended sample because participants are unavailable, a collaborating site withdraws, funding ends, the recruitment window closes, or an unforeseen event interrupts data collection. In other cases, you may discover before examining the primary result that additional observations are feasible and scientifically worthwhile.
A change in sample size is not inherently improper. Its interpretation depends substantially on the reason and timing.
Stopping at 180 participants instead of a preregistered 200 because recruitment has become impossible is different from repeatedly examining a p-value and stopping as soon as it crosses a desired threshold. Likewise, increasing the sample according to a formal prespecified sequential design differs from extending recruitment because the initial result was almost statistically significant.
Report the sample size actually obtained, explain consequential differences from the registered target or stopping rule, and avoid obscuring whether interim knowledge of the outcome influenced the decision.
Recruitment and Sampling Procedures Can Change
The route by which observations enter the study may also need revision.
A planned recruitment channel may produce too few participants. Access to a population may be withdrawn. A sampling frame may turn out to contain errors. Researchers may need to recruit from an additional site or extend the recruitment period.
These changes can affect more than logistics. A different recruitment source may alter the composition of the sample and therefore the population to which the findings reasonably apply.
When the sampling process changes, consider both transparency and substantive consequences. Ask whether the revised sample differs in ways relevant to the research question, not merely whether the final number of observations is adequate.
Study Procedures Can Change
Procedures sometimes need to be modified because the original implementation is impractical, unsafe, unavailable, or demonstrably ineffective.
A laboratory task may contain a programming error. An intervention session may need to move online. A planned piece of equipment may become unavailable. Instructions may prove ambiguous during initial implementation.
The methodological consequences depend on what changed and when. A correction made before any participant completes the affected procedure may have different implications from changing the procedure halfway through data collection.
If participants experience materially different versions of the study, consider whether version or timing needs to be accounted for in analysis or interpretation.
Measures and Outcome Operationalization Can Change
Changes to measurement deserve particular attention because they can alter what the study is actually testing.
A researcher may discover that an instrument was administered incorrectly, a scale contains an error, a measure cannot be scored as expected, or a planned data source is unavailable. A modification or replacement may be necessary.
Changing a measure before observing the relevant outcomes can still affect comparability with the original plan. Changing the primary outcome after seeing which measured variable produces the strongest result raises a more serious concern because the observed evidence may have influenced which outcome received primary status.
Watch Out
Changing the label does not change the chronology. If an outcome was designated as secondary in the preregistration and promoted to primary after its favorable result was observed, report that change explicitly rather than presenting it as though it had always been the primary outcome.
Eligibility and Exclusion Criteria Can Change
Researchers may discover cases that the original exclusion rules did not anticipate: duplicate records, technical failures, impossible responses, contamination, protocol violations, or unusual observations that make the original rule difficult to apply.
You can revise or add an exclusion criterion when justified. But exclusions can directly change the analyzed sample and sometimes the result, so their timing matters.
If you decide on an exclusion after seeing that removing particular observations changes the primary finding, the resulting analysis has a different status from an exclusion rule established without knowledge of its effect.
When reasonable, sensitivity analyses showing the result with and without a consequential post-registration exclusion may help readers understand its influence.
Data-Cleaning and Processing Decisions Can Change
Raw data rarely arrive ready for analysis. Researchers may need to address coding errors, missing values, duplicate observations, transformations, scale construction, aggregation, or implausible values.
A preregistration may not anticipate every anomaly. When an unforeseen problem appears, researchers need not ignore it merely because no rule was registered.
Document consequential decisions, especially when alternative processing choices could produce meaningfully different analytical datasets. If the change occurred after inspecting the data, make that chronology clear.
The Statistical Analysis Can Change
Analysis plans are among the most consequential parts of many preregistrations, but they are not immune to revision.
A model may fail to converge. An assumption may be violated. A preregistered test may not actually correspond to the data structure. A coding mistake in the plan may become apparent. A methodological development may reveal a better approach.
Research on preregistration explicitly recognizes such circumstances as potential reasons for deviation. The appropriate response is not to conduct an analysis known to be defective solely because it was written down earlier.
Instead, distinguish the planned analysis from the revised one and explain why the latter was used. Depending on the circumstances and publication venue, it may also be useful or required to report the originally planned analysis.
Method-driven change
The original analysis is revised because it is erroneous, infeasible, or inappropriate for the data-generating structure.
Result-driven change
The analysis is revised because another analytical choice produces a more desirable, striking, or statistically favorable result.
Both are deviations, but they raise quite different interpretive concerns.
Covariates, Transformations, and Model Specifications Can Change
Sometimes the headline analysis remains the same while apparently smaller choices change around it. Researchers may add or remove covariates, transform an outcome, change the random-effects structure, alter a correction for multiple comparisons, or modify how a predictor is coded.
These decisions can materially affect results. They should not be dismissed as technical details merely because the research question remains unchanged.
If a consequential specification differs from what was originally preregistered, report the difference and its rationale.
You Can Add New Analyses
Preregistration does not prohibit researchers from asking questions that become interesting only after the study begins.
You may notice an unexpected subgroup pattern. A secondary variable may suggest a new relationship. A reviewer may request an informative robustness analysis. A new theoretical explanation may emerge while interpreting the data.
Conducting additional analyses can be scientifically valuable. The key distinction is whether they were specified before the relevant results were known.
An analysis conceived after seeing a pattern should normally be identified as exploratory, post hoc, or otherwise clearly distinguished from the preregistered confirmatory analysis. This preserves the distinction between exploration and confirmation without treating either as inherently inferior.
You Can Develop New Hypotheses
Data can surprise you. That is not a defect in the scientific process.
An unexpected result may suggest a mechanism you had not considered. You can formulate a new hypothesis and investigate whether the observed data are consistent with it.
What you cannot legitimately do is retroactively make that hypothesis part of the original prediction. A hypothesis generated from the data and a hypothesis specified independently before observing those data do not provide identical tests of the idea.
A useful strategy is to report the new hypothesis as generated by the current evidence and, where feasible, test it prospectively using new data.
Primary and Secondary Analyses Can Change, but Relabeling Needs Transparency
Researchers sometimes discover that a prespecified secondary analysis is more informative than the primary analysis. There may be a sound substantive reason to emphasize it.
You may discuss that result prominently. What matters is preserving the fact that it was not originally designated as primary.
Similarly, a planned primary analysis that becomes impossible should not simply vanish from the narrative. Explain what prevented it from being conducted and what analysis replaced it.
Preregistration is most useful when it preserves this research history rather than forcing the final paper to mimic a plan that reality made impossible.
The Research Question Itself Can Evolve
A study can sometimes reveal that the original question was poorly formulated or that another question is more scientifically interesting.
You are free to pursue the new question. Preregistration does not grant ownership of your future thoughts to an earlier research plan.
But changing the question may fundamentally change the evidential status of the study. If the new question was generated after examining the same data used to answer it, the resulting analysis may be better understood as exploratory or hypothesis-generating rather than as a prospective test.
The appropriate response is transparent framing, not suppression of the new idea.
Some Changes Matter Much More Than Others
There is no sensible rule that treats every discrepancy between a preregistration and final paper as equally important.
| Type of change |
Possible concern |
What readers need to know |
| Recruitment procedure |
Different sample composition or generalizability |
What changed, why, and whether the resulting sample differs meaningfully |
| Sample size or stopping rule |
Outcome-dependent stopping or altered precision |
Why recruitment changed and whether results were inspected before the decision |
| Primary outcome |
Selective outcome reporting |
Original outcome designation and reason for changing it |
| Exclusion rule |
Result-sensitive sample selection |
When the criterion changed and its effect on the analyzed sample |
| Statistical model |
Additional analytical flexibility |
Why the planned model was replaced and when the problem was identified |
| Additional analysis |
Post hoc pattern selection |
Whether the analysis was planned or arose after examining the data |
| New hypothesis |
Presenting postdiction as prediction |
That the hypothesis was generated after the relevant evidence was observed |
The more directly a change affects the study's central claims, and the more opportunity the observed results had to influence the decision, the more important transparent reporting becomes.
Changes Should Be Documented, Not Hidden in the Final Paper
A deviation is useful information about the research process. It should not require a reader to place the preregistration and manuscript side by side and hunt for discrepancies.
Willroth and Atherton define a preregistration deviation broadly as a discrepancy between what researchers said they would do in the preregistration and what they actually did in the final article. Their work emphasizes that preregistration can remain useful despite deviations when those differences are documented transparently.
How to report those differences systematically deserves separate attention because a study may accumulate several changes across design, sampling, measurement, and analysis. Researchers should therefore plan to document preregistration deviations clearly rather than reconstructing them only when a reviewer notices a discrepancy.
