Missing Data and Attrition Bias in Long-Term Ageing Studies
Key Takeaways
- Missing follow-up data can reduce precision and bias estimates when the probability of being observed is related to health, function, cognition, or the outcome under study. [1] [3]
- Attrition is often selective in ageing cohorts: participants who leave tend, on average, to be older, frailer, or cognitively less able than those who remain. [4] [5]
- The labels MCAR, MAR, and MNAR describe assumptions about why data are missing; they cannot usually be established from the observed dataset alone. [1] [2]
- Multiple imputation, likelihood-based models, weighting, and joint models can reduce bias under stated assumptions, but no method reconstructs unobserved outcomes without assumptions. [2] [6] [8]
Long-term ageing studies depend on repeated observations, yet the people who are hardest to reassess may also be those experiencing the greatest decline. Illness, cognitive impairment, disability, relocation, withdrawal, and death can all remove observations from later waves. When this loss is patterned rather than random, the remaining sample can present an overly healthy picture of ageing trajectories. [3] [4] [5]
Who This Is Useful For
This page is useful for readers interpreting cohort studies that follow cognition, frailty, biomarkers, disability, or health over many years. It is especially relevant when later analyses include substantially fewer participants than the baseline sample, or when a paper reports only complete cases without explaining who was lost and why. [3] [6]
Missing Data and Attrition Are Related but Not Identical
A value may be missing at one visit because an assessment was skipped, a device failed, or a participant was temporarily unavailable. Attrition is the loss of a participant from subsequent follow-up, although some people later return. Both create incomplete records, but attrition can also change who remains in the study population over time. [2] [3]
Attrition bias arises when continued participation is associated with variables relevant to the analysis. For example, if people with faster cognitive decline are less likely to complete later testing, an analysis restricted to returners may underestimate average decline. A 2024 individual-participant meta-analysis found that people who left longitudinal ageing studies had lower prior cognitive performance across all assessed domains than those who remained. [5]
Common Missingness Patterns
| Pattern or Event | Example | Potential Consequence |
|---|---|---|
| Intermittent missingness | A participant misses one assessment but returns at a later wave | Some within-person information remains, but the missed value may still be informative. [2] [6] |
| Monotone dropout | A participant withdraws and supplies no later measurements | The observed cohort becomes progressively selected if dropout relates to health or outcome. [4] [5] |
| Item non-response | A blood sample or physical test is absent while the rest of the visit is completed | Analyses of that measure may use a narrower and systematically different subset. [3] [6] |
| Death | Cognition or physical function is not observed after a participant dies | The outcome may be undefined rather than merely unrecorded, changing the estimand. [8] [9] |
These patterns are analytically distinct. In particular, methods that combine dropout and death without distinction can implicitly describe a hypothetical cohort in which nobody dies, whereas an analysis among those alive at each time point answers a different question. [8] [9]
MCAR, MAR, and MNAR
Missing completely at random (MCAR) means the probability that a value is missing does not depend on the relevant observed or unobserved data. Under missing at random (MAR), missingness may depend on observed information but, after conditioning on that information, not on the unseen value itself. Missing not at random (MNAR) means the probability of missingness still depends on unobserved information after the observed variables have been considered. [1] [2]
In an ageing cohort, missed grip-strength testing might be compatible with MAR if recorded mobility, prior grip strength, age, and health status adequately explain who does not complete the test. It may be MNAR if people skip testing because their current unmeasured strength is especially poor, even after those recorded predictors are considered. Whether the recorded variables are sufficient is an assumption, not something that the incomplete data can definitively reveal. [1] [6]
Why Ageing Studies Are Especially Vulnerable
Repeated waves create more opportunities for non-response, and older participants are more likely to experience illness, functional limitation, cognitive impairment, and death during follow-up. A systematic review of longitudinal studies of older people identified increasing age and cognitive impairment as consistent independent predictors of dropout, with frailty and poor health also associated with loss. [4]
The resulting bias is not always toward smaller effects. It depends on how participation relates jointly to the exposure and outcome. In one cognitive-ageing analysis, weighting for selective attrition made the estimated difference in cognitive decline between current and never smokers larger, showing how attrition had masked part of an adverse association. [7]
Why the Percentage Missing Is Not Enough
A low proportion of missing data can matter if the missing observations are concentrated among participants with particular outcome trajectories. Conversely, a larger proportion may cause less bias when missingness is well explained by recorded information and the analysis appropriately uses that information. Bias therefore depends on the missingness mechanism, the analysis target, and the relationships among variables, not on a universal percentage threshold. [1] [2]
Missing data also reduce information. Even when an estimate remains unbiased, fewer observed measurements generally widen uncertainty. Complete-case analysis can compound this loss by discarding participants who have useful observations on other variables or waves. [2] [10]
How Common Analytic Approaches Differ
| Approach | What It Does | Central Limitation |
|---|---|---|
| Complete-case analysis | Analyzes only records with all variables required for a model | Can lose precision and be biased unless restrictive conditions hold. [2] |
| Multiple imputation | Creates several plausible completed datasets and combines their estimates | Depends on the imputation model and usually an MAR assumption for the primary analysis. [2] [10] |
| Likelihood-based longitudinal models | Use the observed repeated measurements directly under a specified model | Validity depends on model form and assumptions about missingness. [2] [8] |
| Inverse-probability weighting | Gives greater weight to observed participants who resemble those likely to be lost | Depends on correctly modeling observation or retention probabilities and measuring their predictors. [6] [7] |
| Joint or partly conditional models | Represent longitudinal outcomes together with dropout, survival, or the population still alive | Different models answer different scientific questions and introduce additional assumptions. [8] [9] |
Multiple imputation preserves uncertainty by analyzing several completed datasets rather than treating one filled-in value as known. Its performance depends on including useful predictors of both missingness and the incomplete variables, aligning the imputation model with the analysis, and representing longitudinal structure adequately. [2] [10]
Sensitivity Analysis and Transparency
Because assumptions about unobserved data are not fully testable, sensitivity analyses examine how results change under plausible departures from the primary assumptions. Substantive conclusions that persist across credible scenarios are less dependent on any one missing-data model; conclusions that change materially should be interpreted as assumption-sensitive. [6] [11]
Transparent reporting includes the number observed at each wave, reasons for loss where known, baseline differences between retained and lost participants, the variables used in imputation or weighting, and the assumptions and sensitivity analyses. A methodological survey found that reporting and handling were often incomplete in longitudinal studies of older adults, with complete-case analysis the most common approach among studies that excluded observations for missingness. [3]
Questions to Ask When Reading a Study
- How many participants contributed at each wave? A baseline sample size alone does not show the information available for later analyses. [3]
- Why were observations missing? Temporary non-response, withdrawal, illness, and death do not necessarily support the same analysis. [8] [9]
- Who was more likely to leave? Differences in age, cognition, frailty, health, or exposure status can signal selective attrition. [4] [5]
- What assumptions did the method require? Naming multiple imputation or weighting is not a substitute for explaining the model and predictors used. [2] [10]
- Did the conclusion survive alternative assumptions? Sensitivity analyses help show whether the main result depends on an unverifiable missingness scenario. [6] [11]
What This Does Not Mean
- It does not mean every study with attrition is invalid; the direction and size of bias depend on the data-generating process and analysis. [1] [2]
- It does not mean a high retention percentage guarantees unbiased results; selective loss can matter even when relatively few participants leave. [2]
- It does not mean multiple imputation automatically removes bias; its conclusions inherit assumptions from the imputation model. [2] [10]
- It does not mean death should always be imputed or always be excluded; the treatment of death depends on the outcome and the scientific estimand. [8] [9]
Practical Interpretation Examples
- If a cohort reports little cognitive decline among ten-year survivors: the result may accurately describe survivors while underrepresenting people who died or left after poorer cognitive performance. [5] [8]
- If complete cases look healthier at baseline than excluded participants: the final estimate describes a selected subset and may not transport directly to the original cohort. [3] [4]
- If imputed and complete-case estimates agree: agreement is informative, but both approaches can share bias when their assumptions fail in similar ways. [2] [6]
Related Reading
Summary
Missing data in long-term ageing studies are not merely empty cells. They can reflect health and survival processes that reshape the observed cohort. Sound interpretation therefore depends on who was observed, why others were not, what population and outcome the analysis targets, and how strongly the conclusion relies on assumptions about unseen data. [3] [5] [6] [8]
References
- Rubin, D. B. (1976). Inference and missing data. Biometrika. https://doi.org/10.1093/biomet/63.3.581
- Sterne, J. A. C., White, I. R., Carlin, J. B., et al. (2009). Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. BMJ. https://www.bmj.com/content/338/bmj.b2393
- Okpara, C., Edokwe, C., Ioannidis, G., et al. (2022). The reporting and handling of missing data in longitudinal studies of older adults is suboptimal: a methodological survey of geriatric journals. BMC Medical Research Methodology. https://pmc.ncbi.nlm.nih.gov/articles/PMC9040343/
- Chatfield, M. D., Brayne, C. E., and Matthews, F. E. (2005). A systematic literature review of attrition between waves in longitudinal studies in the elderly shows a consistent pattern of dropout between differing studies. Journal of Clinical Epidemiology. https://pubmed.ncbi.nlm.nih.gov/15649666/
- Hernandez, R., Jin, H., Lee, P.-J., et al. (2024). Attrition from longitudinal ageing studies and performance across domains of cognitive functioning: an individual participant data meta-analysis. BMJ Open. https://bmjopen.bmj.com/content/14/3/e079241
- Duchesneau, E. D., Shmuel, S., Faurot, K. R., et al. (2023). Missing data approaches in longitudinal studies of aging: a case example using the National Health and Aging Trends Study. PLOS ONE. https://pmc.ncbi.nlm.nih.gov/articles/PMC10249888/
- Weuve, J., Tchetgen Tchetgen, E. J., Glymour, M. M., et al. (2012). Accounting for bias due to selective attrition: the example of smoking and cognitive decline. Epidemiology. https://pmc.ncbi.nlm.nih.gov/articles/PMC3237815/
- Wen, L., Muniz Terrera, G., and Seaman, S. R. (2018). Methods for handling longitudinal outcome processes truncated by dropout and death. Biostatistics. https://doi.org/10.1093/biostatistics/kxx045
- Kurland, B. F., Johnson, L. L., Egleston, B. L., and Diehr, P. H. (2009). Longitudinal data with follow-up truncated by death: match the analysis method to research aims. Statistical Science. https://doi.org/10.1214/09-STS293
- White, I. R., Royston, P., and Wood, A. M. (2011). Multiple imputation using chained equations: issues and guidance for practice. Statistics in Medicine. https://onlinelibrary.wiley.com/doi/10.1002/sim.4067
- Little, R. J., D'Agostino, R., Cohen, M. L., et al. (2012). The prevention and treatment of missing data in clinical trials. New England Journal of Medicine. https://www.nejm.org/doi/10.1056/NEJMsr1203730
This content is provided for educational purposes only and does not constitute medical advice.