2027 Writing a Missing-Data Plan Before Opening a Dataset for MPH and MS Students
Audited and Verified:
Our Editorial and Data Verification Process
Research.com uses government datasets, official education statistics and primary institutional information to support its career and degree guides. Our review process checks factual accuracy, occupational definitions, source dates and the requirements associated with the credentials discussed.
How This Article Was Fact-Checked
Our editorial team verifies salary estimates, employment projections, education requirements and program information against original data providers and official organizations. Salary figures are checked against the relevant occupational definitions and reporting periods. Information about licensure, accreditation and required in-person training is reviewed separately.
Fact-checked and verified on:
Co-Founder and Chief Data Scientist
Before opening a dataset, an MPH or MS student can decide what missing information would threaten the study - and what to do if it appears. That decision is harder to make impartially after seeing which records survive an analysis. A short, written missing-data plan helps students distinguish unanswered questions from records that were never eligible for a measure, choose defensible methods, and document changes.
This guide shows how to draft the plan, test its assumptions once access is approved, and use it in coursework or supervised research.
Key Things You Should Know
- Write the plan before inspecting record-level observations: define the study population, essential variables, expected reasons for missingness, and planned analyses.
- Separate structural non-applicability from missing values, and distinguish the three commonly discussed mechanisms - MCAR, MAR, and MNAR - without claiming that a dataset can prove which mechanism holds.
- Prespecify a primary analysis and at least one sensitivity analysis; obtain required access and review approvals before examining restricted data.
What is a missing-data plan in public health research?
A missing-data plan is a dated, revisable record of how a study will define, investigate, handle, and report incomplete information. It is written from the research question, study protocol, questionnaire, codebook, and data-collection process - not from an early look at which analytical choice produces the most favorable result.
For example, a hypothetical evaluation might estimate the association between participation in a community health program and a follow-up health measure among eligible enrollees. The plan would specify who belongs in that population, when the measure should have been collected, and how an absent follow-up value could affect the estimate. It would not label a participant who was never scheduled for follow-up as having declined to answer.
An MPH project may emphasize an actionable program estimate and clear reporting to a practice partner; a research-oriented MS thesis may require a more detailed model and sensitivity analysis. Neither degree label determines the method on its own. Students comparing public health graduate degrees and salary pathways should examine the actual methods courses and supervision available for this work.
Why should MPH and MS students create a missing-data plan before opening a dataset?
Missingness can change both who an analysis represents and what it estimates. If people facing transportation barriers are less likely to complete a follow-up visit, analyzing only completed visits may describe a selected group rather than all eligible participants. A larger sample of complete records does not, by itself, remove that bias.
Writing first makes decisions easier to audit. A student can distinguish a planned exclusion from one suggested by an unexpected result, explain departures to an advisor, and give a community partner an honest account of uncertainty. The plan also exposes feasibility problems early: a required outcome may not be collected at every site, or an intended adjustment variable may not exist in the approved extract.
This preparation need not depend on an expensive degree pathway. When weighing choosing a budget-friendly online MPH, students interested in applied analysis should ask whether they will receive feedback on real data-management decisions, not just whether a course lists statistical software.
What should a missing-data plan include for a public health dataset?
Keep the plan short enough to use, but specific enough that another analyst could understand its decisions. A practical draft can follow this sequence:
- Define the question and target: State the eligible population, outcome, exposure or intervention, comparison, time points, and quantity to be estimated. Specify whether the goal is description, association, or a causal estimate.
- Map the data: List required variables, their collection sources and dates, skip patterns, linkage steps, and any known site or survey-design differences. Identify which values are essential to eligibility or the analysis.
- Anticipate gaps: Record plausible reasons for nonresponse, missed visits, incomplete records, and unsuccessful linkage without assuming that any one reason has occurred.
- Set assessment rules: State how missingness will be summarized by variable, time point, and relevant groups; how impossible values and coded refusals will be distinguished from blanks; and which denominators will be reported.
- Choose analysis and reporting: Specify a primary method, assumptions, feasible sensitivity analyses, model checks, and how exclusions and uncertainty will appear in the final report.
- Record governance: Name the approved data source, access limits, supervisor or partner responsibilities, version date, and process for documenting amendments.
Draft these items from approved study materials and advisor input. If methods prerequisites are a barrier to doing the work responsibly, explore pathways into public health master's prerequisite completion rather than assuming admission or access to a dataset supplies the necessary statistical preparation.
First identify what kind of absence a code represents. A questionnaire item skipped because it does not apply, a refused answer, a laboratory result not yet returned, and a value lost during record linkage should not automatically share one missing-value code. Check the instrument and data dictionary before recoding anything.
The table distinguishes the statistical mechanisms often used to reason about genuinely missing values. These are assumptions about the process that produced the gaps, not diagnoses a missingness plot can establish.
| Mechanism | Meaning | Public health illustration |
| Missing completely at random (MCAR) | The probability a value is missing does not depend on observed or unobserved values relevant to the analysis. | A randomly selected set of otherwise comparable records loses a measure because of a processing failure. |
| Missing at random (MAR) | After accounting for observed information, missingness does not additionally depend on the missing value itself. | Follow-up completion differs by recorded clinic and age, but not by the unobserved outcome after those factors are accounted for. |
| Missing not at random (MNAR) | Missingness still depends on unobserved information after accounting for what was observed. | People with an unrecorded worsening outcome are less likely to report that outcome, even after adjustment for available measures. |
All three illustrations are hypothetical. Observed patterns can challenge an MCAR assumption, but they generally cannot prove MAR rather than MNAR. Students who need more practice with these distinctions may find a focused methods course more direct than pursuing lower-cost routes through online public health study solely to learn one technique.
Create a dated expectations log beside the draft plan. For each important variable, note when it should exist, plausible reasons it might not, the document or collaborator supporting that expectation, and what remains unknown. This preserves a record of reasoning without pretending to know the eventual pattern.
Before viewing observations, the following checks can make that log more useful:
- Read questionnaires, collection schedules, codebooks, and approved protocols for skip logic, optional questions, changing definitions, and known interruptions.
- Ask the data steward or project supervisor which missing-value codes, linkage rules, and collection sites apply; record the answer and its date.
- Write conditional statements such as "If follow-up is less complete at certain sites, compare site-specific denominators," rather than inventing missingness rates.
- Save the pre-access version. After authorized inspection, append observed findings and justified changes instead of silently replacing the original plan.
For students building a research portfolio after uneven undergraduate grades, this documented reasoning can demonstrate preparation, although it does not replace any program's admissions criteria. Guidance on MPH applications after uneven undergraduate grades is a separate decision from whether a particular analysis plan is sound.
The right method depends on the estimate, why values may be absent, the information available to explain missingness, and the data structure. No technique reconstructs unobserved facts with certainty. The comparison below describes common choices and their principal limits; students should select and check a method with an appropriately trained supervisor.
| Approach | Potential role | Important limitation |
| Complete-case analysis | Analyze records with all variables required for a specified model. | Can change the study population or bias an estimate when completeness relates to the outcome or other relevant factors. |
| Multiple imputation | Generate several plausible completed datasets, analyze each, and combine estimates while reflecting imputation uncertainty. | Depends on a defensible imputation model and assumptions about missingness; it does not validate MAR. |
| Inverse-probability weighting | Give greater weight to observed participants who resemble people less likely to have observed outcomes. | Requires a credible observation model and adequate representation of relevant groups; extreme weights can make estimates unstable. |
| Likelihood-based longitudinal models | Use available repeated measurements under the model's assumptions without discarding every partially observed participant. | Results still depend on model specification and assumptions about the unobserved measurements. |
A sensitivity analysis asks how conclusions change if an important assumption fails - for example, if unobserved follow-up outcomes are systematically worse than an MAR-based method predicts. Report the assumption and the range examined; do not present a single imputed dataset as recovered truth. Survey weights, clustering, and repeated measurements may require additional design-specific handling.
A targeted course or supervised project may be a better next step than another degree if the immediate gap is implementing one method. Compare that option with lower-cost routes through online public health study based on the depth of instruction and feedback actually needed.
Plan from documentation before accessing records, and check the conditions attached to the actual dataset. Public-use files, restricted extracts, identifiable records, and protected health information can have different access, storage, disclosure, and reporting rules. A data-use agreement may limit record linkage, sharing code or outputs, small-cell reporting, or use beyond the approved project. An advisor's permission to discuss a project is not necessarily authorization to open its data.
Missingness analysis can itself reveal sensitive information when results are broken down by small communities, clinics, or rare conditions. Specify the minimum variables and subgroup summaries needed, approved computing location, who may see record-level data, and how outputs will be reviewed. Ask the responsible institution whether human-subjects review or another determination is required; do not assume a course assignment is exempt, or that every secondary analysis needs the same review.
If restricted-data training or secure computing affects the time and cost of study, include those requirements when considering public health graduate tuition after institutional funding. Funding comparisons should not assume that a school will provide access to a desired partner dataset.
In epidemiology and biostatistics courses, the plan connects a research question to denominators, bias, and model assumptions. In an MPH applied project, it can help a practice partner understand which program indicators are interpretable. In an MS thesis, it may become part of a fuller analysis protocol reviewed by a research advisor. Each assignment still has its own scope, supervision, and approval requirements.
A feasible learning sequence is to draft the plan from a protocol, review it with an instructor or site supervisor, inspect approved data, then submit an annotated revision explaining what changed. The assessable skill is not naming a sophisticated technique; it is making a justified choice, checking whether its assumptions are plausible, and communicating limitations.
Application convenience is a separate consideration. Students comparing public health master's programs that omit reference letters should still ask to see course descriptions and examples of supervised data work; an easier application does not establish methods depth.
Look for evidence that students practice decisions before analysis, not merely clean a supplied spreadsheet afterward. Ask whether courses cover codebooks, reproducible records of data transformations, missingness mechanisms, privacy, sensitivity analysis, and interpretation for public health audiences. Then ask who provides feedback on actual project plans and whether restricted or partner data can be used under approved conditions.
Compare a syllabus and project rubric with your starting skills. A beginner may need foundational statistics and data-management instruction; a student preparing an MS thesis may need advanced modeling and an available methods advisor. Confirm course sequencing, software access, project supervision, and any required on-site work directly with the program. A listed elective is not a guaranteed seat or a formal concentration.
Accreditation is worth checking for the precise school, unit, program, and degree, but it does not by itself establish the depth of missing-data training. Use accreditation-focused online MPH comparisons as a starting point, then verify current scope and curriculum with the institution.
What careers use missing-data and public health data-management skills?
Public health data analysts, research coordinators, evaluation specialists, epidemiologists, and biostatistics staff all encounter incomplete information, but their responsibilities differ. An entry-level coordinator may track collection gaps and maintain a codebook under supervision. An analyst may implement approved cleaning and reporting rules. Designing complex sensitivity analyses or signing off on consequential estimates generally calls for deeper methods preparation and experience.
For an early-career applicant, a strong work sample is a deidentified or wholly simulated plan that states the target population, distinguishes types of missingness, and explains the limits of an analysis. Never put restricted records or an employer's confidential procedures into a public portfolio. Job titles and degree expectations vary by employer; an MPH or MS alone does not confer authority over data governance or establish readiness for an advanced research role.
When comparing positions, treat average earnings after an MPH as broad context, not a prediction for a particular data role. Compare the actual job's duties, required methods, supervision, and pay range. A supervised project or focused statistical training may be a more direct investment than another credential.
Other Things You Should Know About Public Health
No. A dated document is enough for the initial plan. Use approved software later to implement and reproduce the analyses your course or project requires.
No. A plan can be an internal course or project document. Preregistration records specified decisions in a designated registry before the relevant analysis; whether it is appropriate depends on the study and supervisor's requirements.
Yes, if you clearly label it as simulated and do not imply that it represents authorized fieldwork, real participants, or findings from a partner organization.
Keep the original dated draft, check the dictionary against your assumptions, and record any corrections in a new version before proceeding with analysis.
- When Is a Complete-Case Approach to Missing Data Valid? The Importance of Effect-Measure Modification - PMC https://pmc.ncbi.nlm.nih.gov/articles/PMC7705610/
- Types of Missing Data - NCBI - NIH https://www.ncbi.nlm.nih.gov/books/NBK493614/
- Dealing with Missing Data | va.gov https://www.herc.research.va.gov/include/page.asp
- Document Your Process - Data Management and Sharing https://laneguides.stanford.edu/DataManagement/Documenting
- Getting Started with Multiple Imputation for Longitudinal Data | virginia.edu https://library.virginia.edu/data/articles/getting-started-multiple-imputation-longitudinal-data
- Accounting for missing data in statistical analyses: multiple imputation is not always the answer https://research-information.bris.ac.uk/en/publications/accounting-for-missing-data-in-statistical-analyses-multiple-impu/
- Missing Data and Multiple Imputation | Columbia University Mailman School of Public Health https://www.publichealth.columbia.edu/research/population-health-methods/missing-data-and-multiple-imputation
