Research ethics and integrity · Lesson 2 of 3

What you owe everybody who reads it

Decide the analysis before you see the result. If you decide after, say so — exploratory is a respectable label.

9 min read · Free — no account needed to read

The severe categories are simple to state. Fabrication is inventing data. Falsification is altering it, including manipulating images or quietly dropping inconvenient cases. Plagiarism is presenting someone else's words, ideas or data as your own — including self-plagiarism, which is republishing your own work as new.

The more common failures are subtler and are usually committed by people who believe they are being reasonable.

p-hacking is trying analyses until something clears the threshold: excluding outliers after seeing the result, adding covariates, splitting subgroups, stopping collection when it looks good. Each step is individually defensible, and together they make significance nearly guaranteed. The tell is deciding the analysis after seeing the data.

HARKing — hypothesising after results are known — is presenting a finding you discovered as one you predicted. Exploratory findings are valuable and must be labelled as exploratory, because a prediction that came after the fact has not been tested by anything.

Selective reporting is running many outcomes and reporting the ones that worked. Cherry-picking is the same move applied to a literature.

The protections are mostly structural rather than moral. Write your analysis plan before you see the data. Report everything you measured, including what did not work. Label exploratory work as exploratory — it costs nothing and protects everything. Keep your data and code so a claim can be checked. Pre-registration formalises all of this and is increasingly expected.

Authorship belongs to people who made a substantive intellectual contribution: gift authorship and omitting a real contributor are both misconduct. Citation means crediting the source of an idea, citing what you actually read rather than what the citing paper said, and not padding.

See it happen

Your main analysis gives p = 0.08. Removing two unusual cases gives p = 0.03.

The read that arrives on its own

The two cases are clearly atypical, and excluding outliers is standard practice — report the clean result.

The same situation, with this idea

You notice the order of events. You are excluding them *because* of what it does to the result, which is the definition of the problem. Two legitimate routes exist: apply an exclusion rule decided before seeing the data, or report both analyses and say plainly that the result depends on two cases. The second is completely honest and often more informative than a significant result would have been.

After this, you can

notice when a defensible decision is being made because of what it does to the outcome.

The takeaway

Decide the analysis before you see the result. If you decide after, say so — exploratory is a respectable label.

Easy to take the wrong way

That these are rare acts by dishonest people. They are ordinary decisions made under pressure, which is why the defences are procedural.

Try this

Write your analysis plan for your current project now, before the data exists, and date it.