Data Quality Checker · 6 min read

How to handle missing data in research without weakening your results

A practical guide to diagnosing why values are missing, choosing between deletion and imputation, and reporting the decision so reviewers accept it.

Diagnose the pattern before you fix anything

Missing values are not all the same. Values missing completely at random carry no information; values missing because of another variable in your data are missing at random; values missing because of the unmeasured value itself are missing not at random and are the dangerous case.

Run the Data Quality Checker first and read the per-column missingness rates. A column missing 3% of values in a scattered way is a different problem from a column missing 40% of values concentrated in one group of respondents.

Choose a method that matches the pattern

Listwise deletion is defensible only when missingness is small and completely at random; otherwise it biases your sample. Mean substitution is almost never defensible because it shrinks variance and distorts correlations.

For scattered missingness in continuous variables, multiple imputation is the standard. For categorical variables, treat 'missing' as an explicit category when the absence itself is meaningful, such as an unanswered income question.

Set a threshold and stick to it

Decide in advance what proportion of missingness makes a variable unusable — many supervisors accept 5% for deletion and question anything above 20% for imputation — and record the rule in your methodology before you look at results.

Choosing a threshold after seeing which option gives a significant result is the fastest route to an unreproducible study.

Report the decision in full

State how many cases were affected, in which variables, what method you used, and what happened to your key estimates under an alternative method. A one-paragraph sensitivity note pre-empts the most common reviewer objection.

Chat on WhatsApp