Missing Data Handling

The methods a study uses to handle outcomes it never observed (dropouts, lost-to-follow-up, missed measurements). The threat to validity depends on the mechanism: MCAR (missing completely at random — benign), MAR (missing at random, explainable by the observed data), and MNAR (missing not at random — related to the unobserved value itself; the dangerous one). Sound approaches: multiple imputation and sensitivity analyses around the assumption; complete-case analysis can bias, and LOCF (last-observation-carried-forward) is outdated and biased.

A clean randomised trial springs leaks: each arm loses some outcomes, and here the arms lose different numbers (60 vs 90) — differential dropout, which on its own can break the balance randomisation created. The fork at the bottom is the decision: throw the gaps away (or paper over them with LOCF), or impute them properly and stress-test the assumption.

First ask why the data are missing. If truly MCAR, complete-case analysis is unbiased (just less precise). If MAR, multiple imputation using the observed predictors recovers an unbiased estimate. If MNAR — the sickest patients vanish because they were doing badly — no method fully fixes it, so you bracket the result with sensitivity analyses across plausible assumptions.

Missing outcomes are one of the commonest threats to a trial’s validity. Differential, outcome-related dropout produces attrition bias that can manufacture or erase an effect. How a paper handles its missing data — and whether the conclusion survives a sensitivity analysis — is a fast read on how much to trust it.

• MCAR · MAR · MNAR — MNAR is the dangerous one
• Prefer multiple imputation + sensitivity analysis
• LOCF and single-value carry-forward are outdated and biased

The US National Research Council’s expert panel report (The Prevention and Treatment of Missing Data in Clinical Trials, 2010; summarised in NEJM 2012) set the modern standard: LOCF and other single-imputation methods “should not be used” as the primary approach because they understate uncertainty and assume the dropout would have stayed unchanged. It urges prevention of missing data first, then methods like multiple imputation with pre-specified sensitivity analyses. No statistical method can rescue data that are MNAR — the only true fix is to minimise missingness by design and chase up follow-up.

• Blithely assuming MCAR — the strongest and least likely mechanism.
• Accepting LOCF as if it were conservative (it is biased, not safe).
• Ignoring differential dropout between the arms, which signals outcome-related (MNAR) loss.