Syllabus

Version 1 — August 2026. The most current version of this syllabus always lives at this page; substantive changes will be announced in class and on Canvas.

Course

AEM 6850 — Empirical Methods for Applied Economists

Fall 2026. Tuesdays and Thursdays, 8:40–9:55 am. First meeting: August 25. Last meeting: December 3. Location: Malott Hall, Room 406.

Instructor. Prof. Ariel Ortiz-Bobea. Office: 450B Warren Hall. Email: ao332@cornell.edu. Office hours: set in week 1, once I’ve seen the class’s schedule constraints — times posted on Canvas.

Credits. 3. Prerequisites: none. No prior programming, no econometrics. You need a laptop you can bring to every class.

What this course is

Every empirical project begins with data someone had to find, fetch, clean, check, and shape. Methods courses assume that work is already done. This course teaches it — across every kind of raw material an economist now meets:

messy tables · text · documents (PDFs & scans) · images · audio · maps & rasters

Two things make the 2026 version of this course different from what you may expect:

  1. AI is ambient. Modern AI tools write competent first-draft code and can read text, documents, images, and audio at scale. We use them openly, from early in the semester, the way you will use them in your research career.
  2. Verification is the graded skill. Precisely because models write the first draft, the human contribution shifts to knowing whether the result is right: predicting what code should produce, checking data against the world, and documenting the evidence. “It runs” is not “it’s right.” That posture — not any one package — is the course.

The semester runs in two parts. Part I (weeks 1–6): foundations — R, version control, AI tools, data wrangling, verification strategy, reproducible projects, graphics. Part II (weeks 7–15): applications — one unconventional data family per session, each walked through the same pipeline: what economic questions it answers → what’s out there → acquire → process → verify.

Homework is a single running investigation of a real, recent event — this year, the January 2025 Los Angeles wildfires — examined through a new data family each week: air-quality readings, damage inspections, parcel and demographic records, official PDFs, news text, press-conference audio, satellite imagery, wind fields. By the end you will have touched every family and built one coherent piece of empirical work, including its climate-justice dimension: who was in the path, and who bears the damage.

What you will be able to do

By December, you can:

  1. Acquire data from files, APIs, web pages, and licensed platforms — politely, reproducibly, with credentials handled correctly.
  2. Process any of the six data families into analysis-ready tables in R.
  3. Verify systematically: match every failure family (vanishing rows, silent type conversion, impossible values, unreconciled totals) to the technique that catches it, and validate model-constructed variables with gold sets and error analysis.
  4. Visualize for diagnosis and for publication.
  5. Reproduce: any of your projects re-runs end-to-end on a fresh machine from its repository.
  6. Delegate to AI tools honestly and safely — scope a task, predict the output, check the diff, disclose the use.

How the course works

The website is the course. Every lecture is a manual page you can read before class and revisit any time, paired with slides that mirror it section-for-section and a runnable R script for following along live. If you put in the time, everything is learnable from the site. Materials stay up permanently, archived by year.

A typical week.

  • Tuesday & Thursday, 8:40: a short Poll Everywhere quiz on the previous session opens most classes (from week 2). Then lecture + live coding + in-class exercises.
  • Thursday after class: homework posts. Due the following Thursday, 8:40 am, committed to your private course repository on GitHub.

Submission from day one. In the first class you’ll get a private repository in the course’s GitHub organization — no git knowledge needed (the browser handles it). We switch to proper version control in week 3, when you’ll discover you’ve been using it all along.

Schedule

# Date Topic
Part I — Foundations
1 Tue Aug 25 Overview
2 Thu Aug 27 R essentials I — objects, vectors, types, data frames
3 Tue Sep 1 R essentials II — matrices, lists, apply, and real files
4 Thu Sep 3 Basic plots I — graphics as instruments
5 Tue Sep 8 Basic plots II — layouts, shading, colour scales, axes, maps
6 Thu Sep 10 How to code — organizing a project, functions, checks, and debugging
7 Tue Sep 15 Projects & git — organization, paths, version control
8 Thu Sep 17 AI I — chatbots & assistants for research work
9 Tue Sep 22 AI II — AI in the coding loop
10 Thu Sep 24 Verification — strategies for data you can never fully know
11 Tue Sep 29 Reproducible projects — renv, Quarto, replication packages, the fresh-clone test
12 Thu Oct 1 Communicating with graphics — ggplot2 to publication grade
Part II — Applications
13 Tue Oct 6 The unconventional-data landscape
14 Thu Oct 8 Channels I — APIs: data on tap
Oct 13 fall break — no class
15 Thu Oct 15 Channels II — web scraping
16 Tue Oct 20 Text I — strings & regular expressions
17 Thu Oct 22 Text II — LLMs: classification & extraction
18 Tue Oct 27 Documents — PDFs & scans
19 Thu Oct 29 Images — photographs as data
20 Tue Nov 3 Satellite I — data types and local processing
21 Thu Nov 5 Satellite II — Earth Engine: move the computation, not the data
22 Tue Nov 10 Audio — speech as text, plus metadata that means something
23 Thu Nov 12 Time — dates, time zones, and the world clock
24 Tue Nov 17 Spatial I — vector data, projections, maps
25 Thu Nov 19 Spatial II — rasters & fast zonal statistics
26 Tue Nov 24 Spatial III — constructed exposures: wind, distances, paths
Nov 26 Thanksgiving — no class
27 Tue Dec 1 AI in research — ethics, disclosure, provenance
28 Thu Dec 3 Moving forward — what carries, and what the tools cannot do

Homework

Thirteen weekly assignments, each a self-contained piece of the semester’s investigation using that week’s tools. Every assignment states its goal precisely and includes reference checkpoints — published totals, documented counts, known facts — so you can know you’re right before submitting. Several assignments are match-the-target: the finished table, figure, or map ships with the assignment, and your job is to rebuild it exactly from raw data.

Mechanics.

  • Posted Thursday after class; due the following Thursday 8:40 am. HW13 spans Thanksgiving and gets two weeks.
  • Submitted by committing to your private course repository (browser upload for HW1–3; git from HW4).
  • Every assignment ships a starter, and every week’s starter includes a known-good checkpoint of the running project — falling behind one week never snowballs.
  • Late work: homework is not accepted late, and your lowest homework score is dropped. The weekly checkpoint means you can always rejoin cleanly. Documented emergencies: talk to me.

What’s graded (every week, same four dimensions):

Dimension Question it asks
Reproduces Does it re-run from a fresh copy on another machine?
Correct Do your outputs match the stated checkpoints?
Verified Does your work check itself — predictions, assertions, reconciliations, a diagnostic plot read in one sentence?
Craft Structure, clarity, commit history, complete disclosure

Weights shift across the semester: early assignments weight Correct most (HW1: 15 / 60 / 15 / 10); by the second half Verified dominates — as AI writes more of your code, the checking becomes your contribution. Each assignment states its weights.

Assignment starters and submissions live in private repositories in the course’s GitHub organization; sign-up and access instructions are posted on Canvas only.

Quizzes & attendance

Short Poll Everywhere quizzes at 8:40 in most classes from week 2, typically two per week: 2–4 questions on the previous session — prediction (“how many rows come out?”), spotting (“this runs clean; what’s wrong?”), judgment (“which check catches this?”). They test understanding you can’t look up, they’re auto-graded to Canvas, and they double as attendance.

  • Your lowest two quiz scores are dropped; this is the absence buffer.
  • I am flexible about absences when you are transparent with me. Tell me before class when life happens; the dropped scores and the website exist so that missing a class is recoverable.
  • Quizzes are answered in the room, on your own device, through your own Cornell account. Answering for someone else, or having someone answer for you, is academic misconduct — see below.

Grading

Component Weight
Homework (13, lowest dropped) 70%
Quizzes (lowest two dropped) 30%

Your cumulative checks library (checks.R) and validation work are graded where they live — inside each homework’s Verified dimension.

Letter grades follow the standard Cornell scale. This course is designed so that steady weekly work is rewarded; there are no exams and no final project.

AI policy

This course assumes you will use AI tools — that’s part of what it teaches. The rules:

  1. Permitted by default. Chatbots, coding assistants, and agents are allowed on all homework unless a specific exercise says otherwise. The early weeks are deliberately structured so working by hand teaches you the vocabulary you’ll later verify in.
  2. Disclosure is mandatory. Every submission includes a short disclosure block: which tools, for what, and what you checked. Disclosed AI use is never penalized. Undisclosed AI use is misrepresentation and is treated as an integrity violation.
  3. No-AI zones. In-class quizzes and designated in-class exercises (announced as such) are your own unassisted work.
  4. You own every number. “The model did it” is not a defense. If it’s in your submission, you are asserting you checked it.
  5. Protect data and people. Never paste licensed data, other students’ work, or personal information into external tools. Licensed datasets (e.g., parcel records) stay inside the private course infrastructure.

Academic integrity

Cornell’s Code of Academic Integrity applies. In this course specifically: collaboration on ideas is welcome; submitted work must be your own (with AI use disclosed); quiz impersonation and credential sharing are treated as serious violations; and honest disclosure of a shortcut you took will always be treated more generously than its discovery. I am flexible with people and inflexible with deception.

Materials & costs

Everything is free.

  • Software: R and RStudio (installed in class 1), git and a GitHub account, a Cornell Poll Everywhere account. AI tools at free tiers suffice; course API keys are provided where needed.
  • No textbook. The course website is the text. Optional references, all free online: R for Data Science (2e), Healy’s Data Visualization, Happy Git with R.

Accessibility & support

Students with disabilities: I want this course to work for you. Please share your SDS accommodation letter early, and talk to me — the course’s multiple formats (readable manual, slides, scripts, recorded material where available) exist so there are several ways in. Wellness: if life goes sideways during the semester, tell me early; late-semester rescue is much harder than mid-semester adjustment. Cornell Health and cares.cornell.edu are there for bigger things.

Colophon

This site is built with Quarto, versioned on GitHub, archived by year, and produced with AI assistance — disclosed here, per the course’s own policy. The instructor wrote and verified every claim.