tooluniverse-dataset-discovery

Dataset Discovery

When to Use

• User asks "find me data about X" or "where can I get data on Y"
• User wants to analyze a relationship between variables
• User needs specific study designs (longitudinal, cross-sectional, experimental)
• User asks about specific surveys or cohorts

Step 1: Understand What the Research Question Requires

Before searching, determine the minimum data requirements:

Study design needed:

• "Does X predict CHANGES in Y over time?" → longitudinal (same people measured repeatedly). Cross-sectional data CANNOT answer this — don't settle for it.
• "Is X associated with Y?" → cross-sectional is sufficient (one-time measurement)
• "Does intervention X cause outcome Y?" → experimental (clinical trial with controls)
• "What genes/proteins are involved in X?" → omics (sequencing, expression, proteomics)

Variables needed:

• List the specific exposure, outcome, and confounder variables
• For each variable, note the measurement type (continuous, categorical, biomarker vs self-report)
• Identify minimum confounders needed (age, sex are almost always required; domain-specific confounders depend on the question)

Population needed:

• Age range, geography, clinical status, sample size requirements
• Power analysis: to detect a small effect (r=0.1), you need ~800 subjects at 80% power

Step 2: Search Strategy

Search from broadest to most specific. Use find_tools to discover available dataset search tools — don't rely on memorized tool names.

Layer 1 — Cross-repository search (cast wide net): Search tools that index datasets across thousands of repositories. These find datasets you didn't know existed.

• Search by: research topic keywords, variable names, population descriptors
• Look for: DOI-registered datasets, repository listings, government data portals
• re3data_search_repositories/re3data_get_repository — a registry of 3,000+ research data REPOSITORIES themselves (not individual datasets), covering every discipline; use this one level up from a dataset search when the question is "what repository should I even be looking in for this kind of data" (verified live: searching returns real health-research repositories, e.g. the "EPIC study" cohort and "Pacific Islands Families Study"). Once you've identified the right repository from re3data, go search inside it with the domain-specific tools in Layer 2 below.

Layer 2 — Domain-specific repositories: Search repositories specialized for your data type.

• Health surveys: CDC, NHANES (US; search by variable name, not topic keywords); DHS Program (DHSProgram_search_indicators/_get_data) for the international/LMIC equivalent -- fertility, maternal/child mortality, nutrition, immunization, HIV indicators across national surveys, filterable by country and year
• Genomics (open): SRA, ENA, ArrayExpress, GEO, and DDBJ (DDBJ_search_entries/_get_entry/_get_cross_references) -- the third INSDC archive alongside SRA/ENA, plus GEA/MetaboBank entries; also GSA (GSA_get_accession) for China's national genomics archive
• Genomics (controlled-access, human studies): dbGaP (DbGaP_search_studies/_get_study) and EGA (EGA_get_study/_get_dataset/_get_study_datasets) -- both require a separate data-access request process; these tools return study/dataset metadata, not the underlying genotype data itself
• Proteomics: PRIDE, MassIVE
• Metabolomics: MetaboLights, Metabolomics Workbench
• Clinical: ClinicalTrials.gov (for trial data with results)
• Consortium research portals (AD Knowledge Portal, NF-OSI, PsychENCODE, HTAN, DREAM challenges and others): Synapse_search_entities finds public Synapse projects/folders/files/tables/datasets by keyword (node_type narrows it), Synapse_get_entity returns the curated annotations (assay, species, tissue, disease), and Synapse_list_children browses a project. Anonymous and read-only: restricted entities return a clear "not public" error and file contents are not downloaded -- data access still goes through synapse.org.

Layer 3 — Literature-based discovery: Many datasets aren't in any repository — they're described in paper methods sections.

• Search PubMed/EuropePMC for papers that analyzed the relationship you're interested in
• Read their methods: "We used data from [DATASET NAME]" tells you exactly what exists
• Check supplementary materials for deposited data (GEO/SRA accession numbers)
• This is often the MOST effective strategy for finding niche datasets

Step 3: Evaluate Dataset Fitness

For each candidate dataset, assess these dimensions:

Variables:

• Does it contain your SPECIFIC exposure and outcome variables?
• Are they measured the way you need? (biomarker vs self-report, continuous vs categorical)
• Are key confounders available? (missing confounders = biased analysis)

Design match:

• If you need longitudinal: does it follow the SAME individuals over time? How many waves? What's the follow-up interval?
• Beware: "repeated cross-sections" (different people each wave) are NOT longitudinal
• If you need experimental: is there a proper control group? Randomization?

Sample:

• Is the sample large enough for your analysis? (logistic regression needs ~10 events per predictor)
• Does the population match? (age range, geography, clinical characteristics)
• Are there subgroups you need? (stratified by sex, race, disease status)

Access:

• Publicly downloadable (best) vs registration required (days) vs collaboration agreement (months) vs restricted (may be impossible)
• Data format: CSV/TSV (easy), XPT/SAS (need conversion), proprietary database (may need special software)

Quality:

• Is it from a well-known study with published methods? (NHANES, HRS, UK Biobank = high quality)
• Has it been used in peer-reviewed publications? (indicates data is usable)
• What's the response rate / missingness pattern?

Step 4: Download and Analyze

Don't stop at finding datasets — download and analyze them. Write and run Python code via Bash. Never describe what you "would do" — execute it.

Data Loading Cookbook

Choose the loader that matches your data source. When unsure of the format, download a small sample first and inspect.

import

Merge, Clean, Analyze

Always run the code and report actual numbers (β, p-value, CI, N).

Step 5: Report Honestly

Structure the report as:

1. Best available dataset — name, what it contains, access method, key limitation
2. Analysis results — actual statistics (β, p-value, CI, N) from running the code
3. Alternative datasets — ranked by fitness, with tradeoffs
4. What CANNOT be answered — if no dataset matches the study design needed, say so clearly
5. Recommended next steps — apply for access to longitudinal data, replicate in other cohorts

Critical honesty rules:

• Never claim a dataset answers a temporal question if it's cross-sectional
• Distinguish "data exists but needs registration" from "data doesn't exist"
• Report actual computed statistics, not hypothetical analyses
• State the strongest analysis possible with available data, even if it's weaker than what was asked

LOOK UP, DON'T GUESS

Never assume a dataset exists — search for it. Never assume access is public — check. Never assume variables are measured the way you need — verify the codebook.