Four lessons | Four labs | One shared investigation
Code, data, and judgment.
A practical introduction to R, Quarto, and coding agents for international development policy.
Math Camp begins with questions, not commands. You will examine development data, make a claim precise, ask an agent for help, and decide what evidence is strong enough to trust.
Everyone receives a working dataset and code scaffold. Students new to programming can follow the main path. Experienced programmers can update the data, test alternate specifications, and challenge the agent.
Each vertical segment is a residual. OLS selects the intercept and slope that make Σe² as small as possible.
WHY THIS CAMP
A policy question becomes a computational investigation.
Math Camp is the opening chapter of API 209, not a compressed version of the fall statistics course. Its purpose is to give everyone a shared way to enter a data project: begin with a question, learn what the records can represent, make a small analytical move in R, and then decide whether an AI coding agent has actually helped. Students who have never programmed and students who already work comfortably with data will use the same evidence, but they will not need to take the same-sized steps.
The four lessons follow one investigation from first contact to handoff. We begin by reading the structure and documentation of a country-year dataset. We then build a traceable analysis table, use plots and a simple model to describe a relationship, and finally test whether another person can understand and rerun the work. Each lab turns the lesson into practice. Working groups are temporary spaces for comparison and debugging; there is no group submission and no grade.
R does the computation. RStudio provides the workspace in which we read, run, and revise code. Quarto can hold code, prose, figures, and slides in one reproducible document. Codex is the baseline coding agent, while Claude Code is a compatible alternative. These tools are connected, but they are not interchangeable. An agent can inspect several files and propose edits, yet it cannot decide whether the policy question is well framed or whether a numerical pattern deserves a causal interpretation. Those remain analytical judgments.
By the end of camp, each student may keep a small personal artifact: an R script, a short Quarto notebook, or another compact record of the investigation. The artifact is optional and ungraded. Its value is practical—it gives you code, questions, and verification habits that can travel into the semester.
AUGUST 2026
Twelve hours across three weeks.
The schedule alternates between guided lessons and coding labs. A lesson introduces one stage of the investigation; its lab gives you time to try that stage with classmates, compare decisions, and ask for help. The work accumulates, so each meeting begins where the previous one ended.
- 01 Meet the evidence + Lab 1Orient to R and coding agents, inspect the shared file, and build an evidence inventory.
- 02 Build the datasetSelect variables, transform rows, document decisions, and prepare an update request.
- 03 Lab 2 · Update sprintAsk an agent to extend a bounded part of the data, then verify provenance and coverage.
- 04 Compare patterns + Lab 3Use plots and a simple model to describe a relationship without exceeding the evidence.
- 05 Audit the handoffCheck the analysis, communicate uncertainty, and prepare the project for another reader.
- 06 Lab 4 · Mini hackathonRun a final audit, repair one weakness, and leave with a small personal artifact.
READING PATH
Four lessons, one investigation.
The lessons are chapters rather than isolated demonstrations. Each one answers a question that the next lesson needs: What is in the file? How should we construct an analysis table? What pattern can we describe? Can another person inspect and rerun what we did?
- 01Meet the evidence.Learn what R, RStudio, Quarto, Codex, and Claude Code do; then inspect rows, variables, types, keys, coverage, and missingness.
- 02Build the dataset.Choose variables, filter and group observations, join documentation, preserve provenance, and supervise a small data update.
- 03Compare patterns.Read distributions, make a purposeful plot, fit a simple regression, examine residuals, and state the boundary of an associational claim.
- 04Audit the handoff.Ask an agent to review the project, reproduce the important checks, improve the communication, and leave a transparent record.
POLICY QUESTIONS
One file. Four ways into development.
A shared dataset does not require a single substantive interest. The four questions below let students choose a development problem while the class keeps the same unit of observation, documentation, and core R workflow. That common structure makes it possible to compare reasoning across groups without turning the lab into four unrelated courses.
These are deliberately descriptive and associational exercises. Their purpose is to practice turning a broad concern into variables, comparisons, checks, and honest claim language. A visible pattern can motivate a better question; it does not by itself establish a policy effect.
- 01HEALTH · INCOME AND SURVIVALHow is national income associated with under-five mortality, and does the pattern differ by income group?
Use GDP per capita, under-five mortality, and income classifications to examine scale, transformation, and between-group comparisons.
- 02GENDER · SCHOOLING AND FERTILITYWhere is girls’ secondary enrollment associated with lower adolescent fertility, and where does the relationship depart from the overall pattern?
Use female secondary enrollment and adolescent fertility to think about missing coverage, regional context, and observations that do not follow the overall pattern.
- 03ACCESS · ELECTRICITY AND CONNECTIONDoes electricity access travel with internet access, and which countries closed the combined access gap fastest?
Use electricity access, internet use, year, and region to distinguish levels from changes and to compare trajectories over time.
- 04CLIMATE · ENERGY AND GROWTHCan renewable electricity expand while carbon intensity falls and income continues to grow?
Use renewable electricity, carbon intensity, and GDP per capita growth to practice multi-variable reasoning without collapsing a complex transition into one score.
THE WORKING PROTOCOL
The agent proposes. Evidence decides.
The labs repeat one short sequence so that using an AI agent becomes a visible analytical practice rather than a hidden shortcut. The order matters. Students first form an expectation, then make a small attempt, then ask for bounded assistance, and finally test the result against the data and documentation.
- 01FRAMEState the question before opening the agent.
Name the population, unit of observation, variables, comparison, time period, and the pattern you expect. This gives you something concrete to evaluate when code or advice arrives.
- 02HUMAN PASSRead or modify a small piece of R.
Run the smallest working example you can understand. Predict what should change when an input changes, and keep the output that records what actually happened.
- 03AGENT PASSRequest one bounded, reviewable task.
Give Codex or Claude Code a goal, relevant project context, constraints, and a completion check. Ask it to separate facts found in files from inferences it makes.
- 04VERIFYMake the result earn your trust.
Inspect rows, keys, types, ranges, missingness, figures, model output, and claim language. Classify important agent statements as confirmed, contradicted, or unresolved.
REFERENCE
Keep the working guides nearby.
The lessons explain the investigation in sequence. The references answer questions that recur across lessons: how to repair the setup, where a variable came from, what an AI term means, or how to find the editable source behind a rendered page. Use them as field guides rather than as additional assignments.