Datasets for Evaluation Practice

The course follows five real running cases, and every worked example in this book uses real data. Two are Texas panels we built; two are landmark public-use evaluation datasets; the fifth is a landmark small-N nonprofit program whose real published outcomes drive the exercises. All analysis is done in Microsoft Excel (enable the Data Analysis ToolPak).

Get the data and code — all public. The two Texas panels (Cases A and B), with codebooks and sources, live in the public companion repository github.com/scottlangford2/pa3311-textbook/data. This book’s own source and the NSW files are in github.com/scottlangford2/pa5312-textbook. Cases C and D also link to their original public archives (NBER, ICPSR) below.

Case A — Texas City Finance Panel (2013–2024)

One row per city per year — 1,180 Texas cities: sales-tax allocation, taxable sales, business outlets, local rate. Used for the economic-development sales tax case (Type A/B EDC adoption) — descriptive analysis, regression, difference-in-differences, cost-benefit. (Real; shared with the PA 3311 course. To run the DiD, each city’s tax-adoption year is added from the Texas Comptroller’s local sales-tax-rate history.)

Case B — Texas County Political Panel (2000–2024)

One row per county per presidential election — 254 counties × 7 elections: turnout, two-party vote share, income, poverty, education, race, metro status. Used for the countywide vote-center case — group comparisons, regression, difference-in-differences, interrupted time series. Download (public): TX_County_Political_Panel_2000_2024.csv · codebook · vote-center adoption. (Real. The panel now includes vote_center_adopt_year and a staggered, time-varying vote_center treatment indicator — the year each county joined the Texas Secretary of State’s Countywide Polling Place Program, chained from the SoS §43.007(j) biennial reports — so the DiD and interrupted-time-series designs run on real adoption timing. 100 of 254 counties had adopted as of 2024.)

Case C — National Supported Work (NSW) Demonstration

The classic randomized job-training experiment (mid-1970s): disadvantaged workers randomly assigned to subsidized employment or control, with earnings before and after. We use it for the workforce-training case — the experiment, the famous selection-bias contrast, regression/matching, and cost-benefit.

  • Download (in this repo): NSW_experimental.xlsx (445 rows — the randomized sample) · CSV · NSW_observational_cps.csv (program group + a non-experimental CPS comparison, for the selection-bias exercise) · codebook. Source: the Dehejia–Wahba public extract — users.nber.org/~rdehejia/data.
  • Key real numbers: experimental effect +$1,794 (treated $6,349 vs. control $4,555); the naive observational comparison gives −$8,498 (LaLonde 1986; Dehejia & Wahba 1999).

Case D — Moving to Opportunity (MTO)

A real HUD housing-voucher lottery (1994–): public-housing families randomly assigned to a low-poverty-restricted voucher + counseling, a standard voucher, or control. We use it for the housing-assistance / RCT case — random assignment, intention-to-treat vs. treatment-on-the-treated, and cost-effectiveness.

  • Data (free public-use): ICPSR Study 34563 (general-public access; ~3,300 adults) — icpsr.umich.edu. Export to CSV for Excel.
  • Key real findings: children who moved young earned ~$3,477 (31%) more as adults vs. a $11,270 control mean (Chetty, Hendren & Katz 2016); adult mental-health gains (Kling, Liebman & Katz 2007).

Case E — Perry Preschool (small nonprofit program)

A landmark small-N nonprofit program evaluation. From 1962 to 1967 the HighScope Educational Research Foundation — a nonprofit — ran a preschool in Ypsilanti, Michigan, for 123 low-income children at high risk of school failure, randomly assigning them to a program group (n = 58) or a no-program comparison group (n = 65) and following them for decades. We use it for the small-organization / grant-funded case — needs assessment, process monitoring on a shoestring, small-sample impact inference, and the most famous cost-benefit result in the field. Its lesson is that rigor does not require a big budget or a big sample: with 123 children and mostly yes/no outcomes, two-proportion tests and a transparent cost-benefit calculation carried the day.

  • Data (real, published aggregates): Perry_Preschool_outcomes.csv · codebook. These are the real published outcome percentages (with the whole-child counts they imply for n = 58 and n = 65), taken from the official age-40 report. Participant-level records are restricted, so the file is aggregate, not microdata — the Excel exercises compute directly from the published cell values.
  • Key real numbers: high-school graduation 65% vs. 45%; earned $20,000+ at age 40 60% vs. 40%; arrested five or more times 36% vs. 55% (Schweinhart et al. 2005). Cost-benefit: about $16 returned per $1 invested to society ($12.90 to the public), most of it from reduced crime; an independent reanalysis estimates a 7–10% annual social rate of return (Heckman et al. 2010).
  • Want real downloadable microdata? Use the Carolina Abecedarian Project — ICPSR / Child and Family Data Archive Study 4091 (icpsr.umich.edu; free account, public-use files, n ≈ 111), a comparable small-N early-childhood experiment with a citable benefit-cost ratio of about 7.3 (García et al. 2020). It is the “download and clean a real archived file yourself” companion to Perry.

Bring-Your-Own Program Data

For the final project you may evaluate a program of your choosing using one of the datasets above or a public agency’s data; we discuss sourcing and feasibility at the proposal stage.

A note on causal claims. Cases A and B are observational panels; Cases C, D, and E are randomized experiments (E on a very small sample). A central skill of this course is judging how credibly a given design and dataset support a causal claim — and stating the limits honestly. The NSW case makes the stakes vivid: the same data give +$1,794 with the experiment and −$8,498 without it. Case E makes a different point — a randomized design with only 123 children still produced credible, policy-moving evidence.