data-scientist

name	data-scientist
description	Data science methodology and method-selection routing for quantitative research. Covers EDA, data validation, descriptive analysis, causal inference (IV, DiD, RD, synthetic control), clustering/PCA/UMAP, supervised ML, geospatial analysis, network analysis, and visualization design. Contains the canonical method-to-library routing tree, routed by execution language — Python: statsmodels (OLS/GLM/time series), pyfixest (FE/DiD), linearmodels (RE/GMM/SUR), svy (complex surveys), scikit-learn (clustering/prediction ML), geopandas (spatial), igraph (network analysis); R: r-stats (OLS/GLM/time series), fixest (FE/DiD), plm (panel/RE/IV), survey-r (complex surveys), tidymodels (ML), sf-terra (spatial), igraph-r (network analysis). Load the routed tool-specific skill before giving tool-specific advice or writing code — library skills encode environment constraints and curated caveats absent from general knowledge.
metadata	{"audience":"any-agent","domain":"research-methodology"}

Data Scientist Skill

Rigorous data science methodology and mindset for quantitative research in Python or R. Covers EDA, data validation, transformation verification, documentation standards, visualization design, descriptive analysis, statistical modeling, causal inference method selection (IV, DiD, RD, synthetic control), unsupervised analysis (clustering, PCA, UMAP), supervised ML methodology (prediction vs. inference, cross-validation, model interpretation, fairness), geospatial analysis, and network analysis (centrality, community detection, bipartite graphs). Provides methodology decisions and analytical approach guidance. Load the routed, language-appropriate tool-specific skill (Python: polars, statsmodels, plotnine, pyfixest, scikit-learn, geopandas, igraph; R: tidyverse, r-stats, ggplot2, fixest, tidymodels, sf-terra, igraph-r; etc.) before giving tool-specific advice or writing code — library skills encode environment-specific constraints and curated caveats that general knowledge lacks or gets wrong. Use for any data analysis, exploration, transformation, or modeling task — especially when choosing methods, checking assumptions, or structuring an analysis.

Establishes a rigorous, methodical approach to data science work. This skill is about how to think and work, not specific tools. The moment a specific tool enters the conversation — in advice and brainstorming as much as in code — load its specialized skill (Python: polars, plotnine, plotly, marimo; R: tidyverse, ggplot2, plotly-r, quarto; etc.): those skills know this environment's tooling in ways general knowledge does not.

Core Principles - NON-NEGOTIABLE

These five principles must guide ALL data science work. They are not optional.

Principle 1: Data Robustness First

ALWAYS check data before operating on it.

Before ANY analysis or transformation:

• Check shape, types, and memory usage
• Examine value distributions and ranges
• Identify and characterize missing values (count, percentage, pattern)
• Understand what uniquely identifies each row (granularity)
• Look for outliers and anomalies

Be VERBOSE about what you're checking and what you find. Never assume data is clean.

Python:

# ALWAYS start with this pattern
print(f"Shape: {df.shape}")
print(f"Columns: {df.columns.to_list()}")
print(f"Types:\n{df.dtypes}")
print(f"Null counts:\n{df.null_count()}")
print(f"Sample:\n{df.sample(5)}")

R:

# ALWAYS start with this pattern
cat("Shape:", nrow(df), "x", ncol(df), "\n")
cat("Columns:", paste(names(df), collapse = ", "), "\n")
str(df)
cat("Null counts:\n")
print(colSums(is.na(df)))
cat("Sample:\n")
print(df[sample(nrow(df), 5), ])

This principle applies only when you are conducting actual data work. Do NOT conduct net new analyses or data inspections when tasked with compiling past work (e.g., analytic notebook creation), or synthesizing prior analyses into a report (e.g., final report writing).

Principle 2: Documentation First

ALWAYS understand or create data documentation.

Before analysis:

• Seek data dictionaries, schemas, or documentation
• Understand where data comes from (provenance)
• Learn collection methods and their implications
• Identify known quality issues or caveats
• Clarify what each column means in business context

If documentation doesn't exist, CREATE IT as you learn about the data.

Principle 3: Verify Every Operation

NEVER assume a transformation worked correctly.

For EVERY data operation:

• Check row counts before and after
• Examine random samples of affected rows
• Validate that expected changes occurred
• Confirm no unintended side effects
• Document what you checked and what you found

Python:

# Before transformation
print(f"Before: {len(df)} rows, columns: {df.columns.to_list()}")
sample_before = df.filter(pl.col("id").is_in([1, 42, 100]))

# After transformation
print(f"After: {len(result)} rows, columns: {result.columns.to_list()}")
sample_after = result.filter(pl.col("id").is_in([1, 42, 100]))
print(f"Sample comparison:\nBefore:\n{sample_before}\nAfter:\n{sample_after}")

R:

# Before transformation
cat("Before:", nrow(df), "rows, columns:", paste(names(df), collapse = ", "), "\n")
sample_before <- df |> dplyr::filter(id %in% c(1, 42, 100))

# After transformation
cat("After:", nrow(result), "rows, columns:", paste(names(result), collapse = ", "), "\n")
sample_after <- result |> dplyr::filter(id %in% c(1, 42, 100))
cat("Sample comparison:\nBefore:\n")
print(sample_before)
cat("After:\n")
print(sample_after)

Principle 4: Thorough Code Documentation (ENFORCED)

Write extensive comments explaining your reasoning. This is MANDATORY, not optional.

In research workflows, follow the Inline Audit Trail (IAT) standard (see agent_reference/INLINE_AUDIT_TRAIL.md). The IAT standard is enforced during QA review — scripts with sparse documentation receive WARNING findings.

Every code block should explain:

• WHAT you're trying to accomplish (the goal) → IAT Type 2: Intent Comment
• WHY you chose this approach (the reasoning) → IAT Type 3: Reasoning Comment
• WHAT assumptions you're making (the dependencies) → IAT Type 4: Assumption Comment

For tests and validations, explain:

• What behavior you're checking
• What would indicate success vs. failure
• Why this check matters

Principle 5: Focus on Research Questions

Balance rigor with usefulness.

Always consider:

• What question are we actually answering?
• What level of rigor does this decision require?
• Are there multiple valid approaches with different tradeoffs?
• Should I check with the user before proceeding?

CHECK IN with users when:

• Multiple valid methodologies exist
• Tradeoffs between precision and practicality arise
• Findings are surprising or counterintuitive
• Scope might need adjustment

Language Routing

This skill routes to language-specific library skills based on the execution language set in CLAUDE.md § User Preferences and propagated in the agent's prompt. When no language is specified, default to Python.

Method	Python Skill	R Skill
Data manipulation	polars	tidyverse
Static visualization	plotnine	ggplot2
Interactive visualization	plotly	plotly-r
Fixed effects / DiD	pyfixest	fixest
OLS / GLM / time series	statsmodels	r-stats
Panel / RE / IV / system	linearmodels	plm
Complex survey statistics	svy	survey-r
ML / clustering / PCA	scikit-learn	tidymodels
Geospatial	geopandas	sf-terra
Network analysis	igraph	igraph-r
Table formatting	great-tables	gt
Notebook	marimo	quarto

All decision trees below use Python skill names as the primary label, with the R counterpart noted inline as (Python)/(R) pairs. When the execution language is R, substitute the R skill from this table. Time-series estimation routes to statsmodels in Python and r-stats in R (see its references/time-series.md).

Related Skills - When to Load

Core Workflow Skills (Load Together):

• polars (R: tidyverse) - Required for DataFrame operations; data-scientist provides methodology, polars/tidyverse provides syntax
• marimo (R: quarto) - Required for creating validated notebooks; data-scientist defines validation patterns, marimo/quarto provides implementation

For Data Analysis Workflows:

In the research pipeline, data-scientist methodology is applied within the file-first execution pattern:

• Write script files FIRST (to scripts/stage{N}_{type}/) as .py (Python) or .R (R)
• Execute via Bash with automatic output capture wrapper script
• Validation results get automatically embedded in scripts as comments
• Marimo (Python) or Quarto (R) notebook assembles validated scripts for interactive review

Closely read agent_reference/SCRIPT_EXECUTION_REFERENCE.md for the mandatory file-first execution protocol covering complete code file writing, output capture, and file versioning rules.

Load for Specific Needs:

What task are you performing?
├─ Data visualization (any kind)
│   └─ Stage 8.2 — FIRST read visualization reference files below:
│       ├─ ./references/visualization-design.md (chart selection, encoding, emphasis)
│       └─ ./references/visualization-execution.md (color, labeling, accessibility, export)
│       THEN load the tool-specific skill:
│       ├─ Static plots → Load `plotnine` skill (Python) or `ggplot2` skill (R)
│       └─ Interactive plots → Load `plotly` skill (Python) or `plotly-r` skill (R)
├─ Descriptive analysis (subgroups, distributions, decompositions, trends)
│   └─ Stage 8.1 — FIRST read ./references/descriptive-analysis.md
│       THEN load the `polars` skill (Python) or `tidyverse` skill (R) (some methods
│       may also need `statsmodels`/`r-stats` for weighted SEs/formal tests or
│       `pyfixest`/`fixest` for descriptive FE regressions)
├─ Statistical modeling (regression, robustness checks)
│   └─ Stage 8.1 — FIRST read ./references/statistical-modeling.md
│       THEN load library skill:
│       ├─ Standard regression (OLS, logistic, GLM) → Load `statsmodels` skill (Python) or `r-stats` skill (R)
│       │   (Note: for OLS with clustered SEs, prefer `pyfixest`/`fixest` — native cluster support)
│       ├─ Fixed effects, IV with FE, or DiD → Load `pyfixest` skill (Python) or `fixest` skill (R)
│       ├─ Random effects, between, first difference, Fama-MacBeth → Load `linearmodels` skill (Python) or `plm` skill (R)
│       ├─ IV without FE (LIML, GMM) → Load `linearmodels` skill (Python) or `plm` skill (R)
│       ├─ System estimation (SUR, 3SLS) → Load `linearmodels` skill (Python) or `plm` skill (R)
│       ├─ Time series modeling (ARIMA/SARIMAX, VAR, forecasting, stationarity
│       │   tests, exponential smoothing) → Load `statsmodels` skill (Python) or
│       │   `r-stats` skill (R, see its references/time-series.md)
│       │   (to *describe* trends or seasonality without formal modeling, read
│       │   ./references/descriptive-analysis.md "Trend Analysis" section instead)
│       └─ Spatial regression (spatial lag, spatial error, GWR) → Load `geopandas` skill (Python) or `sf-terra` skill (R)
│           (Python: PySAL/spreg via geopandas; R: spdep/spatialreg via sf-terra; also read geospatial refs)
├─ Supervised ML (prediction, classification, risk scoring)
│   └─ FIRST read ./references/supervised-ml.md (when to use ML, how to validate, interpret, report)
│       THEN load `scikit-learn` skill (Python) or `tidymodels` skill (R) (algorithms, syntax, evaluation)
│       ├─ Model interpretation (SHAP, feature importance)
│       │   → Read supervised-ml.md "Interpreting ML Models" + scikit-learn/tidymodels interpretation refs
│       └─ Fairness assessment
│           → Read supervised-ml.md "Fairness" + scikit-learn/tidymodels fairness refs
├─ Unsupervised analysis (clustering, dimensionality reduction, pattern discovery)
│   └─ Stage 8.1 — FIRST read ./references/exploratory-unsupervised.md
│       THEN load `scikit-learn` skill (Python) or `tidymodels` skill (R)
│       ├─ Clustering → clustering.md, evaluation-unsupervised.md (scikit-learn refs; R: tidymodels unsupervised.md)
│       ├─ Dimensionality reduction → decomposition.md, manifold.md (scikit-learn refs; R: tidymodels unsupervised.md)
│       └─ Index construction via PCA → also read ./references/descriptive-analysis.md
├─ Causal / quasi-experimental analysis
│   └─ FIRST read ./references/causal-inference.md
│       THEN load appropriate library skill:
│       Python: pyfixest for DiD/IV/FE, linearmodels for panel RE/IV-GMM
│       R: fixest for DiD/IV/FE, plm for panel RE
│       For RD implementation (rdrobust) → also read ./references/causal-rd.md
│       For matching/IPW/AIPW implementation → also read ./references/causal-matching.md
│       For Heckman selection correction → also read ./references/causal-selection.md
│       For synthetic control implementation → also read ./references/causal-synth.md
│       For causal ML (DML, CATE, meta-learners, causal forests) → also read ./references/causal-ml.md
│       For mediation analysis (mechanisms, NDE/NIE) → also read ./references/causal-mediation.md
│       (reference files show Python implementations; in R, map to fixest/plm/r-stats
│       per the Language Routing table and the R library skills' own references)
├─ Complex survey data analysis (NHANES, ACS PUMS, CPS, ECLS-K, MEPS, etc.)
│   └─ FIRST read ./references/survey-analysis.md (methodology, pitfalls, weight selection)
│       THEN load `svy` skill (Python) or `survey-r` skill (R)
│       ├─ Survey-weighted descriptive statistics → svy/survey-r estimation refs
│       ├─ Survey-weighted regression (OLS, logistic, Poisson) → svy/survey-r regression refs
│       ├─ Survey design setup / replicate weights → svy/survey-r design-weights refs
│       └─ Advanced models not in svy (ordinal, survival, IV) → rpy2 + R survey package
│           (Python: see svy skill "rpy2 Bridge" section; R: use `survey-r` skill directly — rpy2 bridge not needed)
├─ Creating formatted tables (data summaries, regression output)
│   └─ Python: Load `great-tables` skill (grammar-of-tables display tables,
│       HTML/LaTeX export). No modelsummary equivalent — for regression tables
│       use library-specific output (e.g., pyfixest etable(), statsmodels summary())
│       R: Load `gt` skill (gt for data tables, modelsummary for regression
│       tables, kableExtra for simple Quarto tables)
├─ Communicating to non-technical audiences
│   └─ Load `science-communication` skill
├─ Geospatial / spatial analysis (any kind)
│   └─ FIRST read methodology reference files:
│       ├─ ./references/geospatial-analysis.md (spatial thinking, methods, interpretation)
│       └─ ./references/geospatial-operations.md (joins, weights, interpolation, operations)
│       THEN load `geopandas` skill (Python) or `sf-terra` skill (R)
├─ Network / graph analysis (relationships, centrality, community detection,
│   bipartite/two-mode data, paths/components, ego networks, network visualization)
│   └─ FIRST read ./references/network-analysis.md (when a network frame fits,
│       node/edge/directedness/weight conceptualization, centrality selection,
│       community detection + seed discipline, bipartite projection, disconnected-graph
│       and weights-as-distances guardrails, reproducibility requirements)
│       THEN load `igraph` skill (Python) or `igraph-r` skill (R)
│       (ERGM / statistical network models are not currently covered — see the
│       reference's "Out of Current Scope" note; escalate to orchestrator)
└─ Not currently covered by DAAF skills:
    ├─ Bayesian modeling (PyMC, bambi / brms) → escalate to orchestrator
    ├─ Survival / time-to-event analysis → escalate to orchestrator
    └─ Deep learning (PyTorch, TensorFlow / torch for R) → escalate to orchestrator

The THEN-load steps apply to advisory and brainstorming turns as much as implementation. Recommending a method, reviewing a plan, or talking through an approach that names a tool needs the routed library skill loaded just as much as writing code does. The library skills encode environment-specific constraints (which estimators and export backends are actually installed and working here) and curated caveats that general knowledge lacks or gets wrong — for example, when a familiar tool is statistically inappropriate for the data at hand. Naming a tool in advice without loading its skill risks recommending an approach this environment cannot run, or one the skill's curated caveats explicitly warn against. The norm extends one hop further: once the routed library skill is loaded, its own reference-file routing carries the same advisory-inclusive expectation — answer from its routed reference files, not just its SKILL.md overview.

Visualization loading order matters: The reference files provide design principles (what chart to use, how to direct attention, how to handle color accessibly). The tool skills provide syntax (how to code it). Read the design guidance first so implementation choices are principled, not ad-hoc.

For Domain-Specific Analysis (e.g., CCD Education Data):

• Load relevant *-data-source-* skill first to understand domain-specific data caveats
• Then apply data-scientist methodology with that context

Prerequisite Knowledge: This skill assumes familiarity with:

• Python or R programming basics
• DataFrame concepts (rows, columns, filtering)
• Basic statistical concepts (mean, distribution, correlation)

Important: This skill provides the METHODOLOGY. The specialized skills provide TOOL KNOWLEDGE. Use both together — on advisory and brainstorming turns as much as when writing code. A methodology answer that names a tool without its skill loaded rests on general knowledge, which misses the environment constraints and curated caveats the tool skills encode.

Reference File Structure

File	Purpose	When to Read
eda-checklist.md	Detailed EDA procedures and validation checks	Starting analysis on new data
data-documentation.md	Understanding and creating data documentation	Working with unfamiliar data
transformation-validation.md	Validating data operations	Before/after any transformation
code-documentation.md	Writing thorough comments and docs	Writing any analysis code
research-questions.md	Framing questions, stakeholder communication	Scoping work, presenting findings
visualization-design.md	Chart selection, visual encoding, emphasis, integrity	Before creating any visualization
visualization-execution.md	Color palettes, accessibility, labeling, typography, export	When producing figures
descriptive-analysis.md	Summary statistics, subgroups, distributions, decompositions, weighting, inequality, correlation, missing data	Stage 8 analysis when the research contribution is descriptive
statistical-modeling.md	Model selection, assumption checking, robust inference, coefficient interpretation, robustness checks	Stage 8.1 analysis involving regression, modeling, or hypothesis testing. For formal time-series estimation (ARIMA/SARIMAX, VAR, forecasting), load the statsmodels skill directly
causal-inference.md	Causal identification, DAGs, RCTs, IV, RD, DiD, synthetic control, matching	Stage 8.1 analysis requiring causal claims
causal-rd.md	Regression discontinuity implementation: rdrobust API (sharp, fuzzy, kink), bandwidth selection, manipulation testing, covariate balance, visualization, diagnostics	Stage 8.1 analysis using regression discontinuity designs
causal-matching.md	Matching (NN, caliper, Mahalanobis, exact, CEM), IPW, doubly robust/AIPW implementation with sklearn + statsmodels + scipy + polars; balance diagnostics; inference

Validation Tracking

For multi-step transformations, track validation state with a simple dict (Python) or named list (R):

Python:

validation_log = {}

# After each transformation step:
validation_log["Filter to high schools"] = {
    "pre_rows": pre_rows,
    "post_rows": result.shape[0],
    "status": "PASSED" if result.shape[0] > 0 else "FAILED",
}

# Print summary at end:
for step, info in validation_log.items():
    print(f"  [{info['status']}] {step}: {info['pre_rows']:,} → {info['post_rows']:,}")

R:

validation_log <- list()

# After each transformation step:
validation_log[["Filter to high schools"]] <- list(
  pre_rows = pre_rows,
  post_rows = nrow(result),
  status = if (nrow(result) > 0) "PASSED" else "FAILED"
)

# Print summary at end:
for (step in names(validation_log)) {
  info <- validation_log[[step]]
  cat(sprintf("  [%s] %s: %s -> %s\n", info$status, step,
              format(info$pre_rows, big.mark = ","),
              format(info$post_rows, big.mark = ",")))
}

This is inline code, not a separate module. Never create a validation.py / validation.R or import a validation class.

Quick Decision Trees

"I'm starting a new analysis"

Starting new analysis?
├─ Do I have data documentation?
│   ├─ Yes → Read it thoroughly first
│   │         → ./references/data-documentation.md
│   └─ No → Create it as you explore
│           → ./references/data-documentation.md
├─ Have I profiled the data?
│   └─ No → Run full EDA checklist
│           → ./references/eda-checklist.md
├─ Do I understand the research question?
│   └─ Unclear → Clarify with stakeholder
│                → ./references/research-questions.md
└─ Ready to analyze → Document as you go
                      → ./references/code-documentation.md

"I have unfamiliar data"

Unfamiliar data?
├─ Step 1: Basic inspection
│   └─ Shape, types, head/tail/sample
│      → ./references/eda-checklist.md
│      IF data contains geometry column, lat/lon, GEOID, or FIPS codes:
│      → also read ./references/geospatial-analysis.md
│      IF data has a temporal index (dates, repeated periods, panel structure):
│      → also read ./references/descriptive-analysis.md "Trend Analysis" section
│        (for forecasting or formal time-series modeling, load `statsmodels` skill)
├─ Step 2: Understand granularity
│   └─ What does each row represent?
│   └─ What columns uniquely identify a row?
├─ Step 3: Check data quality
│   └─ Missing values, duplicates, outliers
│      → ./references/eda-checklist.md
├─ Step 4: Seek documentation
│   └─ Data dictionary, schema, provenance
│      → ./references/data-documentation.md
└─ Step 5: Document findings
    └─ Create documentation if none exists

"I need to transform data"

Transforming data?
├─ Before transformation:
│   ├─ Document current state (shape, sample)
│   ├─ Identify what SHOULD change
│   └─ Identify what should NOT change
│      → ./references/transformation-validation.md
├─ After transformation:
│   ├─ Verify shape changes are expected
│   ├─ Check random sample of results
│   ├─ Validate invariants (sums, counts)
│   └─ Document what you verified
│      → ./references/transformation-validation.md
└─ For joins specifically:
    ├─ Check for unintended row duplication
    ├─ Check for unintended data loss
    └─ Validate join keys match expectations

"I need to communicate findings"

Communicating findings?
├─ Have I documented limitations?
│   └─ No → List caveats and assumptions
│           → ./references/research-questions.md
├─ Am I making causal claims?
│   └─ Yes → Ensure justified; prefer correlational language
│            → ./references/research-questions.md
├─ Is uncertainty quantified?
│   └─ No → Add confidence intervals or ranges
└─ Have I checked in with stakeholder?
    └─ No → Validate findings align with expectations

"I need to create a visualization"

What kind of visualization task?
├─ Choosing what chart to make
│   └─ → ./references/visualization-design.md
├─ Directing attention / emphasis strategy
│   └─ → ./references/visualization-design.md
├─ Selecting colors or ensuring accessibility
│   └─ → ./references/visualization-execution.md
├─ Labeling, titling, or annotating
│   └─ → ./references/visualization-execution.md
├─ Making it publication-ready (export, DPI, fonts)
│   └─ → ./references/visualization-execution.md
├─ Ensuring project consistency (theme, palette)
│   └─ → ./references/visualization-execution.md
├─ Mapping geographic data (choropleth, dot density, proportional symbols)
│   └─ → ./references/geospatial-analysis.md (map design, classification)
│       THEN → ./references/visualization-execution.md (color, accessibility)
│       THEN load `geopandas` skill (Python) or `sf-terra` skill (R)
└─ Tool-specific syntax (geoms, traces, themes)
    ├─ Static plots → Load `plotnine` skill (Python) or `ggplot2` skill (R)
    └─ Interactive plots → Load `plotly` skill (Python) or `plotly-r` skill (R)

"I need to analyze patterns in this data"

Am I trying to describe or explain?
├─ Describe (characterize what exists)
│   └─ → ./references/descriptive-analysis.md
│       ├─ Subgroup comparisons → Stratification section
│       ├─ Distribution shape → Distributional analysis section
│       ├─ Trends over time (characterize, smooth, decompose) → Trend analysis section
│       ├─ Gaps between groups → Decomposition methods section
│       ├─ Composite measure → Index construction section
│       ├─ Inequality → Inequality measurement section
│       └─ Weighted estimates → Weighted analysis section
├─ Model (regression, hypothesis testing)
│   └─ → ./references/statistical-modeling.md
│       ├─ Choose model by outcome type → Model selection framework
│       ├─ Check assumptions → Assumption checking protocol
│       ├─ Choose standard errors → SE type decision table
│       ├─ Interpret coefficients → Interpretation guide
│       └─ Forecast or formally model temporal dynamics (ARIMA/SARIMAX, VAR,
│           stationarity tests) → Load `statsmodels` skill (see its
│           "I need to analyze time series" decision tree)
└─ Explain causally
    └─ → ./references/causal-inference.md
        └─ Select method per Method Selection Guide

"I need to establish a causal relationship"

What variation identifies the causal effect?
├─ Random assignment → RCT analysis
│   → ./references/causal-inference.md
├─ I can control for all confounders → Regression / matching
│   → ./references/causal-inference.md (methodology)
│   → ./references/causal-matching.md (implementation: PSM, IPW, AIPW, balance)
├─ There's a valid instrument → IV / 2SLS
│   → ./references/causal-inference.md
├─ There's a score cutoff → Regression discontinuity
│   → ./references/causal-inference.md (methodology)
│   → ./references/causal-rd.md (implementation: rdrobust, bandwidth, diagnostics)
├─ Policy changed for some groups → Difference-in-differences
│   → ./references/causal-inference.md
├─ Few treated units, long pre-period → Synthetic control
│   → ./references/causal-inference.md (methodology)
│   → ./references/causal-synth.md (implementation, packages, inference)
├─ Outcome only observed for a selected subset → Heckman selection correction
│   → ./references/causal-selection.md (implementation: manual Probit+OLS, FIML, bootstrap)
├─ I want to understand the mechanism (T→M→Y) → Mediation analysis
│   → ./references/causal-mediation.md (implementation: statsmodels Mediation, bootstrap, sensitivity)
├─ Many confounders / want ML nuisance estimation → DML
│   → ./references/causal-ml.md (manual DML, EconML, DoubleML)
├─ Want to explore treatment effect heterogeneity → CATE / causal forests
│   → ./references/causal-ml.md (meta-learners, causal forests, diagnostics)
└─ Not sure → Start with a DAG
    → ./references/causal-inference.md

"I need to discover structure or groupings in data"

What kind of structure am I looking for?
├─ Groups of similar observations (typology, classification)
│   ├─ Know number of groups → K-means or GMM (scikit-learn / tidymodels)
│   ├─ Don't know number → Try multiple k + validation (scikit-learn / tidymodels)
│   ├─ Arbitrary shapes / noise → DBSCAN or HDBSCAN (scikit-learn — Python only;
│   │   the R `dbscan` package is not installed and tidymodels does not cover it)
│   └─ Uncertain → Read exploratory-unsupervised.md "Algorithm Selection" table
├─ Reducing a large variable set
│   ├─ Linear reduction → PCA (scikit-learn / tidymodels)
│   └─ Visualizing structure → t-SNE or UMAP (scikit-learn / umap-learn; R: uwot)
│       └─ CAUTION: visualization only — not for analysis
│           → ./references/exploratory-unsupervised.md
├─ Using unsupervised results in subsequent regression
│   └─ Read exploratory-unsupervised.md "The Classify-Analyze Problem"
└─ Predicting outcomes with ML methods
    └─ Load `scikit-learn` skill (Python) or `tidymodels` skill (R) → classification/regression refs

"I'm working with geographic/spatial data"

Geospatial analysis task?
├─ Understanding spatial concepts, methods, or statistical theory
│   └─ (MAUP, CRS, autocorrelation, spatial regression)
│       → ./references/geospatial-analysis.md
├─ Choosing a spatial analysis method
│   └─ → ./references/geospatial-analysis.md (decision guide)
├─ Performing spatial joins, overlays, or weights construction
│   └─ → ./references/geospatial-operations.md
│       THEN load `geopandas` skill (Python) or `sf-terra` skill (R)
├─ Interpreting Moran's I, LISA, or spatial regression results
│   └─ → ./references/geospatial-operations.md (interpretation sections)
├─ Interpolation or areal interpolation
│   └─ → ./references/geospatial-operations.md
├─ Extracting raster values into polygons (zonal statistics)
│   └─ → ./references/geospatial-operations.md
├─ Debugging spatial operation failures or geometry errors
│   └─ → ./references/geospatial-operations.md (geometry validity)
├─ Making maps or choosing classification schemes
│   └─ → ./references/geospatial-analysis.md (map design)
│       THEN → ./references/visualization-execution.md (color, accessibility)
└─ Working with or advising on geopandas/PySAL/rasterio or sf/terra/spdep
    └─ Load `geopandas` skill (Python) or `sf-terra` skill (R)

"I need to analyze relationships or network structure"

Is a network frame appropriate for this question?
├─ Question is about attributes of individual units (how much, how many, what predicts Y)
│   └─ A network frame is NOT needed — use descriptive/regression/ML routing above.
│       Relational data can often be analyzed as ordinary tabular features.
├─ Question is about relationships, connectivity, or position within a web of ties
│   (who is central, who bridges groups, what clusters exist, how far apart are units)
│   └─ A network frame fits → ./references/network-analysis.md, then load
│       `igraph` skill (Python) or `igraph-r` skill (R)
│
├─ Which centrality answers my research question?
│   ├─ Activity / volume of direct ties → degree centrality
│   ├─ Brokerage / bridging between groups → betweenness
│   │   (NOTE: betweenness is NOT a resilience/robustness metric — do not read it that way)
│   ├─ Reach / average closeness to all others → closeness
│   │   (requires a connected component — see disconnected-graph guardrail below)
│   └─ Influence via being connected to well-connected units → eigenvector / PageRank
│       → ./references/network-analysis.md "Centrality Selection"
│
├─ I want to find groups / communities
│   └─ Community detection → prefer Leiden over Louvain (Louvain can yield
│       disconnected communities); ALWAYS set a seed (results are non-deterministic)
│       and record it; in R, cluster_leiden/cluster_louvain require an UNDIRECTED graph
│       → ./references/network-analysis.md "Community Detection"
│
├─ My data is two-mode / bipartite (e.g., people × events, authors × papers)
│   └─ Construct a bipartite graph, then project to one mode if needed — but
│       projection loses information and inflates ties; interpret projected
│       edges with care → ./references/network-analysis.md "Bipartite / Two-Mode Data"
│
├─ Before computing closeness/betweenness or trusting distances
│   ├─ Check connectivity/components first (these measures are ill-defined across
│   │   disconnected components) → network-analysis.md "Disconnected-Graph Guardrail"
│   └─ Confirm whether edge weights mean distance or strength — igraph treats a
│       `weight` attribute as DISTANCE and uses it automatically; pass weights=NA
│       to suppress → network-analysis.md "Weights-as-Distances Trap"
│
└─ I need statistical inference / generative network models (ERGM, SAOM, TERGM)
    └─ NOT currently covered by DAAF skills → escalate to orchestrator
        (noted as a future extension in network-analysis.md)

"I need to predict an outcome or classify observations"

Am I predicting or explaining?
├─ Explaining (estimating a causal or associational parameter)
│   └─ Use pyfixest/statsmodels (Python) or fixest/r-stats (R) — see statistical-modeling.md
├─ Predicting (minimizing prediction error on new data)
│   ├─ Read supervised-ml.md for methodology
│   ├─ Load `scikit-learn` skill (Python) or `tidymodels` skill (R)
│   ├─ Tabular data → Start with logistic regression baseline,
│   │   then try HistGradientBoosting/LightGBM (Python) or ranger/xgboost (R)
│   ├─ Text data → See supervised-ml.md "When Deep Learning Methods Are Appropriate"
│   ├─ Need to explain predictions → scikit-learn/tidymodels interpretation refs (SHAP)
│   └─ Need fairness assessment → scikit-learn/tidymodels fairness refs
└─ Not sure → Read supervised-ml.md "Prediction vs. Inference"

Essential Workflows

New Data Workflow

When you receive new data, ALWAYS follow this sequence:

1. Load and inspect (do not transform yet)

   Python:

   # Load data
   df = pl.read_csv("data.csv")  # or scan_csv for lazy
   
   # Immediate inspection
   print(f"Shape: {df.shape}")
   print(f"Columns: {df.columns}")
   print(f"Types:\n{df.dtypes}")
   print(f"First 5 rows:\n{df.head()}")
   print(f"Last 5 rows:\n{df.tail()}")
   print(f"Random sample:\n{df.sample(5)}")
   

   R:

   # Load data
   library(tidyverse)
   library(arrow)
   df <- read_csv("data.csv")  # or read_parquet() for parquet
   
   # Immediate inspection
   cat("Shape:", nrow(df), "x", ncol(df), "\n")
   cat("Columns:", paste(names(df), collapse = ", "), "\n")
   glimpse(df)
   cat("First 5 rows:\n")
   print(head(df, 5))
   cat("Last 5 rows:\n")
   print(tail(df, 5))
   cat("Random sample:\n")
   print(df[sample(nrow(df), 5), ])
   

2. Check data quality

   Python:

   # Missing values
   print(f"Null counts:\n{df.null_count()}")
   print(f"Null percentages:\n{df.null_count() / len(df) * 100}")
   
   # Duplicates
   print(f"Duplicate rows: {len(df) - len(df.unique())}")
   
   # Unique values per column
   for col in df.columns:
       print(f"{col}: {df[col].n_unique()} unique values")
   

   R:

   # Missing values
   cat("Null counts:\n")
   print(colSums(is.na(df)))
   cat("Null percentages:\n")
   print(round(colSums(is.na(df)) / nrow(df) * 100, 2))
   
   # Duplicates
   cat("Duplicate rows:", nrow(df) - nrow(distinct(df)), "\n")
   
   # Unique values per column
   for (col in names(df)) {
     cat(col, ":", n_distinct(df[[col]]), "unique values\n")
   }
   

3. Understand distributions

   Python:

   # Numerical columns
   print(df.describe())
   
   # Categorical columns - value counts
   for col in df.select(pl.col(pl.String)).columns:
       print(f"\n{col}:\n{df[col].value_counts().head(10)}")
   

   R:

   # Numerical columns
   summary(df |> select(where(is.numeric)))
   
   # Categorical columns - value counts
   df |>
     select(where(is.character)) |>
     names() |>
     purrr::walk(\(col) {
       cat("\n", col, ":\n", sep = "")
       print(head(count(df, .data[[col]], sort = TRUE), 10))
     })
   

4. Identify granularity

   Python:

   # What uniquely identifies a row?
   # Test candidate keys
   candidate_keys = ["id", "user_id", ["user_id", "date"]]
   for key in candidate_keys:
       cols = [key] if isinstance(key, str) else key
       unique_count = df.select(cols).n_unique()
       print(f"{cols}: {unique_count} unique vs {len(df)} rows")
   

Transformation Workflow

For ANY data transformation:

1. Document pre-state

   Python:

   # Record state before transformation
   pre_shape = df.shape
   pre_columns = df.columns.copy()
   pre_sample = df.sample(10, seed=42)  # Reproducible sample
   pre_sum = df.select(pl.col("amount").sum()).item()  # If applicable
   

   R:

   # Record state before transformation
   pre_shape <- c(nrow(df), ncol(df))
   pre_columns <- names(df)
   set.seed(42)
   pre_sample <- df[sample(nrow(df), 10), ]  # Reproducible sample
   pre_sum <- sum(df$amount, na.rm = TRUE)  # If applicable
   

2. Perform transformation with comments

   Python:

   # GOAL: Filter to active users and calculate total spend
   # REASONING: We only want users who logged in within 30 days
   # EXPECTED: Fewer rows, same columns, preserved sum for included rows
   result = (
       df
       .filter(pl.col("last_login") > cutoff_date)  # Remove inactive
       .group_by("user_id")
       .agg(pl.col("amount").sum().alias("total_spend"))
   )
   

   R:

   # GOAL: Filter to active users and calculate total spend
   # REASONING: We only want users who logged in within 30 days
   # EXPECTED: Fewer rows, same columns, preserved sum for included rows
   result <- df |>
     dplyr::filter(last_login > cutoff_date) |>  # Remove inactive
     summarise(total_spend = sum(amount, na.rm = TRUE), .by = user_id)
   

3. Validate post-state

   Python:

   # Verify transformation results
   post_shape = result.shape
   print(f"Shape: {pre_shape} -> {post_shape}")
   
   # Check sample of results
   sample_ids = pre_sample["user_id"].to_list()[:3]
   print(f"Sample before:\n{pre_sample.filter(pl.col('user_id').is_in(sample_ids))}")
   print(f"Sample after:\n{result.filter(pl.col('user_id').is_in(sample_ids))}")
   
   # Validate invariants where applicable
   # (e.g., sum should be preserved or explainably different)
   

   R:

   # Verify transformation results
   post_shape <- c(nrow(result), ncol(result))
   cat("Shape:", paste(pre_shape, collapse = "x"), "->", paste(post_shape, collapse = "x"), "\n")
   
   # Check sample of results
   sample_ids <- pre_sample$user_id[1:3]
   cat("Sample before:\n")
   print(pre_sample |> dplyr::filter(user_id %in% sample_ids))
   cat("Sample after:\n")
   print(result |> dplyr::filter(user_id %in% sample_ids))
   
   # Validate invariants where applicable
   # (e.g., sum should be preserved or explainably different)
   

4. Document what you verified

Analysis Workflow

From question to answer:

1. Clarify the question

   • What decision will this inform?
   • What would a "good" answer look like?
   • What level of rigor is required?

2. Assess data fitness

   • Does the data contain what we need?
   • What are the limitations?
   • Are there gaps or quality issues?

3. Choose methodology

   • What approaches are valid?
   • What are the tradeoffs?
   • CHECK WITH USER if multiple valid options

4. Execute with verification

   • Follow transformation workflow
   • Document each step thoroughly

5. Validate findings

   • Do results make sense?
   • Cross-check with known facts
   • Identify limitations and caveats

6. Communicate with appropriate uncertainty

Quick Checklists

Initial Data Inspection Checklist

• Loaded data successfully
• Checked shape (rows x columns)
• Reviewed column names
• Checked data types
• Examined head, tail, and random sample
• Counted missing values per column
• Checked for duplicate rows
• Identified unique value counts per column
• Generated summary statistics
• Identified granularity (what uniquely identifies a row)
• Documented findings

Pre-Transformation Checklist

• Documented current shape and columns
• Saved sample of data for comparison
• Recorded relevant aggregates (sums, counts)
• Stated what SHOULD change
• Stated what should NOT change
• Explained WHY this transformation is needed

Post-Transformation Checklist

• Verified shape change matches expectations
• Compared sample before/after
• Validated invariants are preserved
• Checked for unintended nulls
• Checked for unintended duplicates
• Documented what was verified and results

Documentation Checklist

• Data source documented
• Each column defined
• Missing value conventions explained
• Granularity/unit of observation stated
• Known quality issues noted
• Transformation history recorded

Marimo Integration

When working in marimo notebooks:

Cell Organization Pattern

# Cell 1: Imports and setup
import marimo as mo
import polars as pl

# Cell 2: Data loading (separate cell for reactivity)
df = pl.read_csv("data.csv")

# Cell 3: Data inspection (markdown + code)
mo.md("## Data Inspection")
# ... inspection code ...

# Cell 4: Data quality checks
mo.md("## Data Quality")
# ... quality checks ...

# Cell 5+: Analysis cells, each with:
# - Markdown explaining goal
# - Code with thorough comments
# - Validation of results

Using Reactivity for Validation

# Create interactive validators
validation_column = mo.ui.dropdown(
    options=df.columns,
    label="Select column to validate"
)

# Reactive validation display
mo.md(f"""
### Validation for `{validation_column.value}`
- Null count: {df[validation_column.value].null_count()}
- Unique values: {df[validation_column.value].n_unique()}
- Sample values: {df[validation_column.value].head(5).to_list()}
""")

Documentation Cells

Use markdown cells liberally:

• Before code: explain what you're about to do and why
• After code: summarize findings and implications
• At decision points: document choices made

Quarto Integration

When working in Quarto notebooks (R pipelines):

Chunk Organization Pattern

# Chunk 1: Library loading
library(tidyverse)
library(arrow)

# Chunk 2: Data loading (separate chunk)
df <- read_parquet("data.parquet")

# Chunk 3: Data inspection (with narrative)
#| echo: true
# ... inspection code ...

# Chunk 4: Data quality checks
# ... quality checks ...

# Chunk 5+: Analysis chunks, each with:
# - Narrative text explaining goal
# - Code with thorough IAT comments
# - Validation of results

Inline Validation

# Validate inline — no separate validation files
stopifnot(nrow(df) > 0)
cat("Shape:", nrow(df), "x", ncol(df), "\n")
cat("Columns:", paste(names(df), collapse = ", "), "\n")

Documentation Chunks

Use narrative text chunks liberally:

• Before code: explain what you are about to do and why
• After code: summarize findings and implications
• At decision points: document choices made

Topic Index

Topic	Reference File
Initial data inspection	./references/eda-checklist.md
Missing value analysis	./references/eda-checklist.md
Distribution analysis	./references/eda-checklist.md
Outlier detection	./references/eda-checklist.md
Uniqueness and cardinality	./references/eda-checklist.md
Correlation analysis	./references/eda-checklist.md
Data dictionaries	./references/data-documentation.md
Data provenance	./references/data-documentation.md
Working with undocumented data	./references/data-documentation.md
Questions to ask about data	./references/data-documentation.md
Before/after validation	./references/transformation-validation.md
Join validation	./references/transformation-validation.md
Aggregation validation	./references/transformation-validation.md
Schema validation (inline)	./references/transformation-validation.md
Common transformation errors	./references/transformation-validation.md
Comment philosophy	./references/code-documentation.md
Docstring patterns	./references/code-documentation.md
Notebook documentation	./references/code-documentation.md
Test documentation	./references/code-documentation.md
Research question formulation	./references/research-questions.md
Rigor vs. practicality	./references/research-questions.md

Citation Responsibility

When analytical methods from this skill's reference materials are used in DAAF analyses, the research-executor includes relevant citations in its structured output. Focus on primary citations — the papers and tools that directly enable the analytical results, not every background reference mentioned in the reference files.

Each citation must include a brief rationale explaining why it is included, so the researcher can make informed decisions about what to keep in the final report.

For the master citation index and inclusion thresholds, consult agent_reference/CITATION_REFERENCE.md.