How to start medical research using public databases

A concrete task with RigorMed

1. Develop a research direction into a reviewable data-source comparison

   These are suggested task inputs and outputs to request, not a new execution, a customer outcome or a guarantee of completion.

   Materials to provide: A specific question, target population, required variables, and candidate official data access points and usage conditions.

   Requested outputs / delivery goal: A candidate-source comparison table, a materials checklist and a narrowed research question.

2. An example request

   Compare only verifiable official data sources. List design, population, variable coverage, access and licence conditions, and questions each cannot answer. Identify which documentation to open first; do not guess unknown fields.

3. What still needs review

   A public access point does not mean unrestricted use or existing product integration. List controlled applications and permission requirements separately; do not bypass them.

First identify the type of resource

'Medical public databases' is often used loosely, although access methods and research purposes differ substantially.

The first question is not 'Which database makes publication easy?' but 'Which source can lawfully and accurately answer my question?' More data cannot fix mismatched access conditions, designs, or variable coverage.

Seven checks from a research question to an executable plan

1. Specify the research question first

   At minimum, specify the population, main exposure or predictors, primary outcome, time window, and design. For a question such as whether a measure is associated with an outcome in adults, state where the population comes from, how the measure and outcome are defined, and whether the study concerns cross-sectional associations, cohort outcomes, or prediction.

   Do not discover a variable first and reverse-engineer an apparently usable topic.

2. Check access conditions and permitted uses

   Classify resources as public-use, available after registration, controlled-access, or unavailable. Retain official access points, versions, licenses or data-use agreements, application requirements, citation requirements, and redistribution restrictions.

   NHANES supplies public-use files and detailed documentation. MIMIC-IV currently requires credentialing, prescribed training, and a signed data-use agreement. SEER research data likewise require a request and confirmation of the relevant agreement. Descriptions such as 'public' or 'free' do not replace these checks.

3. Check whether the design fits

   A variable's presence does not mean it can answer every question. Cross-sectional surveys describe prevalence or contemporaneous associations but usually cannot establish temporal order from one observation. Hospital records concern particular care-seeking populations and do not automatically represent the general population.

   Specify the sampling frame, who may be absent, when observation starts and ends, whether exposure precedes outcome, and the population to which conclusions will be generalized.

4. Map variables individually

   Do not record column names alone. Bind each core variable to its official definition, file, cycle or version, measurement method, units, coding, eligible population, skip pattern, meaning of missing values, and permitted derivations.

   Across cycles, check changes in wording, laboratory equipment, detection limits, units, and codes. CDC/NCHS warns that ignoring questionnaire skips can produce incorrect denominators. Where laboratory methods change, check for official bridging studies or calibration equations. Matching names alone do not establish comparability.

5. Assess sample, event, and missingness constraints

   Total sample size is not the only criterion. Binary outcomes require both event and non-event counts; regression requires the number of prespecified parameters, not merely variables. Estimate complete-case exclusions. Complex surveys also require attention to weights, strata, clusters, and design degrees of freedom.

   Do not substitute a single rule of thumb such as 'ten events per variable' for a full assessment. Determine model complexity, outcome frequency, overfitting control, missing-data handling, validation, and target estimands before analysis.

6. Check publications and study registries together

   Use literature databases such as PubMed to check the closest studies, then registries such as ClinicalTrials.gov for relevant registered work that may remain unpublished. Retain strategies, dates, records, and exclusion reasons. Compare populations, exposures, outcomes, timing, designs, and methods, not just titles.

   'No identical topic found' does not establish novelty. It may reflect a narrow strategy, different terminology, limited coverage, or unpublished studies.

7. Freeze the plan before statistics

   Fix eligibility criteria, main exposures and outcomes, covariate rationale, the main model, missing-data strategy, sensitivity analyses, software environment, and reporting standards before formal analysis. Record reasons for later changes and distinguish prespecified from exploratory analyses.

   For observational studies, consult STROBE during planning and writing. With routinely collected health data, assess applicability of extensions such as RECORD. Reporting checklists do not replace sound designs but can expose methodological and disclosure gaps early.

Six common errors

1. Downloading data and obtaining significant results before specifying the question and covariate rationale;
2. Reading variable names without checking exact questions, methods, units, skips, and eligibility;
3. Pooling cycles without examining changes in codes, measurement methods, and weights;
4. Ignoring complex sampling and treating observations as independent and identically distributed;
5. Using total sample size alone without checking events, parameter budgets, missingness, and validation samples;
6. Calling a missed search result a 'first discovery' or interpreting an observational association as causation.

Checklist before starting

Before formal analysis, make sure you can answer these questions:

• Does the question specify population, exposure or predictors, outcome, timing, and design?
• Are data obtained through official access points, with versions and acquisition dates recorded?
• Have conditions, agreements, citation requirements, and redistribution restrictions been checked?
• Do collection purposes, sampling frames, and target populations support the proposed generalization?
• Do core variables have official definitions, methods, units, codes, and eligible populations?
• Are variables genuinely comparable across cycles or sources?
• Do samples, events, non-events, missingness, and prespecified parameters support the model?
• Have the closest publications and relevant registered studies been systematically checked?
• Were main analyses, sensitivity analyses, and reporting standards specified before examining results?
• Have responsible parties confirmed authorization, ethics determinations or exemptions, registration, and other governance requirements?

If any item is unanswered, the next step is usually to obtain source, variable, or planning evidence, not to start statistics.

Frequently asked questions