Data Finder

Find and assess datasets for your research question. Two Explorer agents search in parallel across data source categories; an Explorer-Critic then stress-tests each candidate against the research design.

Input: $ARGUMENTS — a topic, or from spec to read the research question from quality_reports/.

Step 1: Read Research Context

1. Find the most recent quality_reports/project_spec_*.md or quality_reports/specs/*.md — extract:
   • Research question
   • Empirical strategy (DiD, RDD, IV, etc.)
   • Treatment variable (what varies)
   • Outcome variable (what we measure)
   • Controls needed
   • Time period of interest
   • Geography (national, state, county, individual)
   • Unit of observation (individual, household, firm, establishment)
2. Read references/domain-profile.md if it exists — extract the Common Datasets section (domain-specific datasets to check first).
3. If no research spec exists, extract the variables and strategy from $ARGUMENTS directly. If the request is vague, ask: "What are the treatment and outcome variables, and what empirical strategy did you have in mind?"

Step 2: Dispatch Two Explorer Agents in Parallel

Split the source categories between two Explorer agents to parallelize the search.

Explorer A — Institutional Data:

Task prompt: "You are an Explorer agent. Research question: [question].
Empirical strategy: [strategy].
Variables needed — Treatment: [X], Outcome: [Y], Controls: [list],
Time period: [period], Geography: [geo], Unit: [unit].
Domain datasets (check first): [list from domain-profile if available].

Your source categories to search:
1. Public microdata (CPS, ACS, NHIS, MEPS, SIPP, QWI)
2. Administrative data (Medicare/Medicaid, IRS, SSA, vital statistics, court records)
3. Survey panels (PSID, HRS, Add Health, NLSY97/79, BHPS/UKHLS)

For each dataset found, produce the full Explorer report format.
Follow the Explorer agent instructions."

Explorer B — Broader and Alternative Sources:

Task prompt: "You are an Explorer agent. Research question: [question].
Empirical strategy: [strategy].
Variables needed — Treatment: [X], Outcome: [Y], Controls: [list],
Time period: [period], Geography: [geo], Unit: [unit].
Domain datasets (check first): [list from domain-profile if available].

Your source categories to search:
1. International data (World Bank, OECD, Eurostat, IMF, IPUMS International)
2. Novel/alternative (satellite, web scraping, proprietary, RCT registries)
3. Any field-specific datasets not covered by Explorer A

For each dataset found, produce the full Explorer report format.
Follow the Explorer agent instructions."

Step 3: Dispatch Explorer-Critic

After both Explorer agents complete, dispatch the Explorer-Critic with the full combined dataset list.

Task prompt: "You are an Explorer-Critic agent. Research question: [question].
Empirical strategy: [strategy].
Variables needed — Treatment: [X], Outcome: [Y], Controls: [list],
Time period: [period], Geography: [geo], Unit: [unit].

Here is the combined dataset list from the Explorer agents:
[paste all Explorer findings]

Apply the 5-point critique to each dataset:
1. Measurement validity
2. Sample selection
3. External validity
4. Identification compatibility
5. Known issues

Produce adjusted feasibility grades and deal-breaker flags.
Follow the Explorer-Critic agent instructions."

Step 4: Produce Ranked Output

After the Explorer-Critic completes, compile the final ranked report:

1. Sort datasets by adjusted feasibility grade (A first, then B, then C, then D).
2. Within each grade, sort by identification compatibility score (highest first).
3. Separate out deal-breaker datasets into the rejection table.

Step 5: Save Report

Save to quality_reports/data_exploration_[sanitized_topic].md:

# Data Exploration: [Topic]

**Date:** [YYYY-MM-DD]
**Research question:** [one sentence]
**Empirical strategy:** [method]
**Variables sought:** Treatment = [X], Outcome = [Y], Controls = [list]


## Top Candidates (Grade A–B)

### 1. [Dataset Name] — Grade: A/B

**Provider:** [Name] | **Access:** [Public/Restricted/etc.] | **URL:** [link]

**Coverage:** [time period] | [geography] | [unit of observation] | N ≈ [size]

**Key Variables:**
- Treatment proxy: [variable]
- Outcome: [variable]
- Controls available: [list]

**Explorer-Critic Assessment:**
- Measurement validity: [1-2 sentences]
- Sample selection: [1-2 sentences]
- External validity: [1-2 sentences]
- Identification compatibility: [focused on the proposed strategy]
- Known issues: [specific documented problems]

**Bottom line:** [1-2 sentences — viable and under what conditions]


[Repeat for all A and B grade datasets]


## Accessible With Effort (Grade C)

[Brief summaries — name, access path, main limitation, why C not B]


## Rejection Table

| Dataset | Reason for Rejection | Deal-breaker? |
|---------|---------------------|---------------|
| [Name] | [Explorer-Critic finding] | YES/NO |


## Recommended Path Forward

1. **Best dataset:** [Name] — [one sentence why]
2. **Fallback if [best] unavailable:** [Name] — [why it's second choice]
3. **Access path for [best]:** [download URL, application URL, IRB requirements, restricted-data steps]

### Ingest Recipe for [best]

A copy-pasteable load-and-clean block for the recommended dataset. Tailor to whether the source is a flat file, an API/portal, or a non-tabular source (PDF, scraped HTML).

**R:**
```r
# Download (or login + download — note in a comment if manual)
library(tidyverse)
df <- readr::read_csv("data/raw/[file].csv")  # or haven::read_dta, arrow::read_parquet

df_clean <- df %>%
  rename(...) %>%
  mutate(...) %>%
  filter(...)

arrow::write_parquet(df_clean, "data/processed/[name].parquet")

Python:

import pandas as pd
df = pd.read_csv("data/raw/[file].csv")  # or pd.read_stata, pd.read_parquet
df_clean = (
    df.rename(columns={...})
      .assign(...)
      .query("...")
)
df_clean.to_parquet("data/processed/[name].parquet")

If the source is a PDF, scraped HTML page, or government portal API, replace the load step with the appropriate extraction recipe — tabulizer/pdftools/rvest/tidycensus/fredr in R, or pdfplumber/camelot/pandas.read_html/census/fredapi in Python. The /data-analysis skill's Phase 0.5 documents the full set; cross-reference it if extraction is non-trivial.

Known traps for [best]

• [E.g., FIPS codes read as integers — drop leading zeros]
• [E.g., Survey weights required — srvyr::as_survey_design()]
• [E.g., Top-coding on income — note threshold]

Next Steps

• /data-analysis [dataset] — begin analysis with the recommended dataset (Phase 0.5 handles non-tabular sources; Phase 3.5 runs design-specific identification diagnostics)
• /lit-review [topic] — check if papers in the literature use these datasets (helps validate choice)

## Important

- **Identification compatibility is the most important criterion.** A perfectly accessible dataset that can't support the proposed empirical strategy is useless. The Explorer-Critic's grade on this dimension should drive the recommendation.
- **Access level affects timeline.** An FSRDC dataset may take 1-2 years to access. A public download can start today. Make this tradeoff explicit.
- **Don't reject C-grade datasets outright.** A FSRDC dataset with perfect identification fit may be the right choice for a dissertation. Present the access path clearly.