2027 Public Health PhD Research: Planning Missing-Not-at-Random Sensitivity Analyses
Co-Founder and Chief Data Scientist
A public health dissertation can reach a misleading conclusion when people with unobserved outcomes differ systematically from those who remain in the study. Missing-not-at-random (MNAR) sensitivity analysis tests how conclusions change under plausible versions of that difference.
For PhD researchers planning surveys, cohorts, trials, or evaluations, the crucial decision comes before analysis: Which outcome matters, why might it be missing, and how large a departure from standard assumptions would change the finding? This guide turns those questions into a defensible analysis plan and shows what training and supervision the work requires.
Key Things You Should Know
- Plan around a defined population, outcome, time point, and estimand; record which observations could be missing and why before choosing a model.
- Compare the three missingness mechanisms, MCAR, MAR, and MNAR, but treat MNAR scenarios as assumptions to examine, not effects the observed data can generally identify.
- Specify plausible parameter ranges and a tipping point in advance, then report both the primary estimate and how conclusions change across scenarios.
What Is Missing-Not-at-Random Sensitivity Analysis in Public Health Research?
Data are missing not at random when the probability that a value is missing still depends on that unobserved value after accounting for information included in the analysis.
In an illustrative cohort studying depressive symptoms, participants with more severe symptoms might be less likely to complete follow-up even after researchers account for their recorded baseline symptoms and demographics. Whether that remaining dependence exists cannot generally be established from observed records alone.
An MNAR sensitivity analysis repeats the substantive analysis under explicitly different assumptions about unobserved values or their probability of being missing. Its purpose is not to discover the "true" missing values. It is to show which conclusions remain credible across defensible assumptions and which depend heavily on an unverifiable one.
Researchers entering from other disciplines may need foundational probability and regression preparation before advanced missing-data work. A public health bachelor's affordability guide can inform an earlier education decision, but a degree title alone does not establish readiness to design an MNAR analysis.
Why Should Public Health PhD Researchers Plan MNAR Sensitivity Analyses?
Missingness can change a population-health estimate, an exposure-outcome association, or an intervention comparison. If a study of access to care disproportionately loses participants whose health worsens, analysis of observed outcomes may portray a more favorable picture than the full study population would support. The direction and size of bias depend on the outcome, missingness process, design, and estimand; they should not be assumed.
Planning early helps researchers collect useful predictors of missingness, protect follow-up resources, obtain appropriate data permissions, and distinguish a prespecified sensitivity analysis from one selected because its result looks favorable. It also gives advisors and community or agency partners a chance to assess whether the proposed scenarios reflect how participation actually works.
A PhD is appropriate when the goal includes producing original methodological or substantive research under doctoral supervision. Someone primarily seeking to apply established techniques in an analyst role might instead compare targeted methods training with low-cost public health master's degrees, rather than assume a doctorate is necessary.
How Do You Design an MNAR Sensitivity Analysis Before Collecting Data?
Start with the scientific question, not the software. For an illustrative longitudinal study, the target might be the difference in mean symptom scores at a specified follow-up between two exposure groups in the originally enrolled population. That target determines whose missing outcomes matter and whether missed visits, permanent dropout, and missing baseline covariates require different plans.
A prospective protocol should make the following decisions reviewable:
- Define the estimand, analysis population, outcome scale, comparison, follow-up time, and primary analysis.
- Map likely missingness by variable and visit; ask participants and study staff about feasible barriers to follow-up without assuming their explanations prove an MNAR mechanism.
- Identify baseline and time-varying predictors worth measuring, including prior outcomes, access barriers, and study-contact history, subject to consent and data governance.
- Prespecify an MAR-based reference analysis, MNAR model or models, parameter ranges, and a rule for interpreting a tipping point.
- Plan retention efforts, outcome ascertainment, sample-size or simulation work where feasible, and documentation of protocol changes.
If the dissertation depends on follow-up that cannot be funded or staffed, simplify the question before enrollment. Budget and pacing choices for public health students are relevant to that feasibility decision, but program scheduling must be checked separately from the study's recruitment and data-collection timetable.
Missing completely at random (MCAR) means missingness does not depend on observed or unobserved study values. Missing at random (MAR) means that, conditional on appropriately included observed information, missingness does not depend on the still-unobserved value. MNAR allows that remaining dependence. These labels describe assumptions about a specified missingness process, not permanent properties of a dataset.
MAR is often a useful reference, especially when the study collects strong predictors of both outcome and response. But an observed comparison of responders with nonresponders cannot establish MAR: the nonresponders' missing outcomes remain unknown.
MCAR-based complete-case analysis is not automatically a safe benchmark, and neither imputation nor a significant missingness-model coefficient proves MNAR.
Compare assumptions relevant to the decision: an MAR reference, unfavorable and favorable departures that stakeholders consider plausible, and an extreme stress test clearly labeled as such.
Researchers changing fields through graduate study should check their quantitative preparation rather than assume conditional public health graduate entry includes the advanced coursework needed to evaluate these assumptions.
Pattern-mixture models describe outcome distributions within response patterns, then specify how outcomes missing within a pattern differ from outcomes that were observed. A common sensitivity device adds a prespecified delta shift to imputed outcomes.
If higher scores mean worse health, a positive delta could represent worse outcomes among participants whose follow-up score is missing. The shift's scale, timing, and affected group must be explicit.
Selection models jointly describe the outcome and the probability it is observed. A sensitivity parameter can express how an unobserved outcome changes the odds of response, conditional on measured variables. This may align well with a proposed dropout mechanism, but its parameter is not usually identified by observed data without additional assumptions or information.
Neither family is universally superior. Choose the formulation that stakeholders can interpret, that fits the data structure, and that the research team can check and implement reproducibly.
Prior grades alone do not determine that capability; prospective students reviewing MPH applications after uneven undergraduate grades should distinguish admissions options from the later methods preparation a research-intensive project demands.
Multiple imputation under MAR can provide a reference analysis; a delta-adjusted imputation analysis then changes imputed values according to stated MNAR scenarios. Likelihood-based pattern-mixture or selection models and Bayesian models can examine more complex longitudinal structures.
A tipping-point analysis asks at what parameter value the substantive decision changes. Simulation can assess whether a proposed design and analysis are workable under specified, hypothetical data-generating processes.
Inverse-probability weighting ordinarily relies on response being explainable by measured information, plus adequate probabilities of observation. Weighting by itself does not repair dependence on an unmeasured outcome. Likewise, a sophisticated model does not compensate for an inappropriate estimand, poor outcome measurement, or inadequate positivity.
Choose methods according to the missing-data pattern, outcome type, clustering or survey design, available predictors, and the team's ability to diagnose model behavior. Students considering budget planning for online MPH study should confirm whether advanced longitudinal analysis, missing-data electives, computing support, and supervised research are actually accessible; an MPH methods course is not automatically specialist training.
A sensitivity parameter should have a clear interpretation on the outcome or response scale. In a delta-adjusted analysis, specify whether the shift applies to every missing outcome, only permanent dropouts, or particular visits or groups.
In a selection model, explain what a chosen change in response odds represents. Avoid presenting either parameter as an estimate learned solely from the incomplete study data.
Build a defensible range using relevant external studies, earlier waves or linked outcomes when authorized, substantive knowledge from clinicians or community partners, and documented elicitation from experts.
Each source has limits: observed earlier outcomes may not predict later missing outcomes well, and expert judgments are judgments rather than measurements. Include a wider stress-test range when the defensible range could miss an important reversal.
For a doctoral project, ask whether an advisor or collaborator can scrutinize parameter interpretation before committing to a complex model. University scholarship options for MPH applicants may matter to someone building prerequisites first, but financing earlier study does not substitute for that methodological supervision.
Report the target population and estimand, missingness by variable and time point, reasons for missingness when known, the primary MAR-based analysis, and each MNAR scenario's assumptions.
Show estimates and uncertainty across the prespecified parameter range, along with the tipping point if one exists. State whether that tipping point falls inside or outside the range judged substantively plausible.
Separate three conclusions: what the observed data show, what each assumption adds, and what remains unknowable without further information. Do not describe a stable result across a narrow range as proof that missingness cannot matter.
Document deviations from the protocol, software and reproducible code, model checks, and any restrictions on sharing data. Obtain required institutional review, data-use authorization, and community or partner permission before accessing restricted records or releasing outputs.
A dissertation timeline should reserve time for revising and checking these analyses, not just fitting the primary model. Students comparing inexpensive public health research doctorates should verify funding duration, computing access, dissertation supervision, and continuation rules rather than comparing advertised tuition alone.
No degree label establishes proficiency in MNAR methods. The strongest fit is a research-focused PhD environment where a prospective student can find advanced biostatistics or epidemiologic methods coursework, faculty actively working with incomplete longitudinal or survey data, access to appropriate computing, and supervision that connects methods to a feasible public health question.
A biostatistics PhD may fit someone developing new missing-data methods; an epidemiology or public health PhD may fit someone applying and evaluating methods within substantive research.
Before applying, examine current course schedules and faculty work, then ask who could supervise the analysis, whether a second methods mentor is available, what data access requires, and how dissertation funding and milestones work. Ask whether a proposed dataset actually contains the variables and follow-up needed.
Admission does not confer data-use rights. A public health doctoral degree comparison guide can help frame program questions, but verify the exact PhD rather than treating a practice-focused DrPH as interchangeable.
If the objective is competent use of established analyses rather than an original dissertation contribution, a methods-intensive MS, supervised research appointment, or targeted advanced course may be more direct. Compare those routes against the work you intend to produce, not degree prestige.
What Careers Use MNAR Sensitivity Analysis and Advanced Epidemiologic Methods?
Missing-data expertise is useful in academic epidemiology and biostatistics, health-services research, government and nonprofit evaluation, clinical research, and research teams working with surveys or longitudinal records.
Responsibilities may include designing follow-up, writing analysis plans, implementing sensitivity analyses, reviewing collaborators' assumptions, and explaining uncertainty to decision-makers. Job titles vary, and many analyst positions require sound application of established methods rather than a PhD or new methods research.
A PhD is most compelling for roles centered on independent research, methodological development, or leading a sustained research agenda; experience, publications, collaboration, and employer requirements still matter.
Someone comparing doctoral options can use guides on reviewing selectivity in public health doctoral study as one admissions consideration, but easier entry is not evidence of suitable faculty mentorship or a viable missing-data dissertation.
For the next step, write a one-page proposed analysis specifying the estimand, missingness process, MAR reference, MNAR parameter, plausible range, and decision-changing result. Take it to a potential supervisor or methods collaborator. Their ability to challenge and improve that plan is more informative than a program's general claim to teach advanced research.
Other Things You Should Know About Public Health Programs
Yes. Anonymity may affect willingness to answer, but respondents can still skip questions because of the unreported answer itself. Document item nonresponse and consider whether that mechanism could change the estimate.
They can reveal barriers to participation and inform plausible scenarios, provided interviewing is approved and appropriate. Interview accounts cannot establish the values of outcomes missing from the quantitative study.
It depends on the data, original permissions, proposed use, and institutional rules. Check with the responsible review office and data steward before accessing restricted information; do not infer an exemption merely because collection is complete.
Where permissions allow, share a redacted analysis plan, variable definitions, parameter specifications, software version, and code that runs on simulated data. Check disclosure and data-use rules before sharing even aggregate outputs.
- Handling missing data in clinical research | nih.gov https://pubmed.ncbi.nlm.nih.gov/36150546/
- Analytical approaches and estimands to take account of missing patient-reported data in longitudinal studies - PMC https://pmc.ncbi.nlm.nih.gov/articles/PMC6489631/
- Types of Missing Data - NCBI - NIH https://www.ncbi.nlm.nih.gov/books/NBK493614/
- Missing Data in Clinical Research: A Tutorial on Multiple Imputation - ScienceDirect https://www.sciencedirect.com/science/article/pii/S0828282X20311119
- Chapter: 5 Principles and Methods of Sensitivity Analyses | The National Academies Press https://www.nationalacademies.org/read/12955/chapter/7
- Fitting MNAR models in missingHE | r-project.org https://cran.r-project.org/web/packages/missingHE/vignettes/Fitting_MNAR_models_in_missingHE.html
- Pattern-Mixture Models for MNAR Sensitivity Analysis | CASRAI https://casrai.org/guides/pattern-mixture-models-mnar-sensitivity-analysis
