You Have Missing Data. Now What?

Missing data is one of the most common — and most consequential — challenges in medical device clinical trials. Whether it's a patient who missed a follow-up visit, a device explanted before the study endpoint, or a participant who simply withdrew consent, missing data can quietly undermine the integrity of your results and, more importantly, your FDA submission.

Why Missing Data Is a Regulatory Problem

FDA reviewers scrutinize missing data carefully — and for good reason. When data are absent, the underlying assumption is always: why? Was the device failing? Was the patient experiencing adverse events? Did the site lose interest in the study?

The mechanism by which data go missing — what statisticians call the missing data mechanism — directly affects whether your analysis will hold up to regulatory scrutiny. FDA guidance and the ICH E9(R1) addendum both make clear that a credible primary analysis must be accompanied by a transparent missing data strategy, defined prospectively in the Statistical Analysis Plan.

The Three Missing Data Mechanisms

• Missing Completely at Random (MCAR): Missingness has nothing to do with the study or the patient's underlying status. This is the most favorable scenario — and the rarest.

• Missing at Random (MAR): Missingness is related to observed data, but not to the unobserved outcome itself. Many standard imputation approaches rely on this assumption.

• Missing Not at Random (MNAR): Missingness is related to the unobserved outcome itself — the most problematic scenario. MNAR requires more sophisticated handling and honest sensitivity analyses.

In practice, you can never fully verify which mechanism applies. That's why sensitivity analyses aren't optional — they're essential.

Common Approaches to Handling Missing Data

Complete Case Analysis — Only subjects with complete data are analyzed. Simple, but it discards information, reduces power, and can introduce bias if the missing subjects differ systematically from those included.

Single Imputation (e.g., Last Observation Carried Forward) — LOCF was once ubiquitous in device trials but has fallen out of favor with regulators. It can be either conservative or anti-conservative depending on the trajectory of the outcome, and it artificially inflates precision by treating imputed values as observed.

Multiple Imputation (MI) — The current gold standard for many applications. MI creates multiple plausible datasets under a specified model, analyzes each, and combines results using Rubin's Rules. Properly implemented under MAR, it preserves variability and produces valid standard errors.

Mixed Models for Repeated Measures (MMRM) — For continuous longitudinal outcomes, MMRM uses all available data without explicit imputation and produces valid estimates under MAR. It's increasingly the preferred primary analysis approach in device and drug trials alike.

Pattern Mixture Models and Selection Models — Used when MNAR is plausible, these approaches explicitly model the relationship between missingness and the outcome. They require assumptions that cannot be verified from the data, which is precisely why they must be paired with sensitivity analyses.

What FDA Expects to See

A defensible missing data strategy in your SAP should include a clear definition of what constitutes missing data for each endpoint, a stated assumption about the missing data mechanism, a pre-specified primary analysis approach that accounts for missing data, pre-specified sensitivity analyses testing robustness under alternative assumptions including plausible MNAR scenarios, and a tipping point analysis for binary endpoints.

Regulators don't expect perfection. They expect transparency, planning, and scientific rigor.

The Importance of Prevention

The best missing data strategy is the one you never need to use. Robust study design — adequate follow-up windows, clear discontinuation criteria, thorough site training, and a well-defined estimand — will always be more defensible than even the most sophisticated imputation approach applied after the fact.

Under the ICH E9(R1) framework, defining your estimand upfront forces you to think carefully about what should happen to intercurrent events before a single patient is enrolled. That clarity pays dividends when it comes time to handle the inevitable gaps in your data.

BSC's Approach

At Biomedical Statistical Consulting®, missing data strategy isn't an afterthought — it's built into protocol and SAP development from day one. We work with medical device sponsors to define estimands, pre-specify sensitivity analyses, and develop imputation strategies that will withstand FDA review.

Want to learn more? Contact BSC to speak with us.