Understanding the Perry Preschool Longitudinal Follow-Up
The Perry Preschool Project ran from 1962 to 1967 in Ypsilanti, Michigan. It targeted 5 African American children who scored in the bottom quarter on intelligence testing and came from low-income households. They were placed into two groups. One group received high-quality preschool instruction using the High Scope approach. The control group did not receive any intervention. Both groups were followed through age 40 and beyond. The results got widely cited in policy debates around early childhood funding. The study found several measurable differences between the two groups by the time participants reached age 40. Those in the intervention group earned higher monthly incomes on average. They were more likely to hold a job continuously. They graduated from high school at a higher rate. Arrest records showed fewer felony convictions among participants in the preschool group. A subsequent cost-benefit analysis by the economist Heckman estimated a return of roughly $7 to $12 for every dollar invested in the program. The High Scope method itself relied on what the researchers called "plan-do-review." Children planned their activities before starting them. They carried out those plans. Then they reflected on what happened afterward. Teachers acted as facilitators rather than instructors. They used open-ended questioning and scaffolding techniques during child-led play. The classrooms had low student-to-teacher ratios. Every session followed a predictable daily structure. The material was concrete and hands-on.
How to Access the Original Data and Analysis
The main data archives are held at the High Scope Press in Ypsilanti. The federally funded data repository ICPSR also hosts cleaned datasets from the Perry project. You can request access through the University of Michigan's survey research center. The original documentation includes participant rosters, testing scores, attendance logs, and follow-up survey instruments from waves conducted at ages 14, 15, 19, 27, and 40. There is also a separate dataset that tracks earnings and tax records through administrative matching. If you need the full technical report, look for the publications by Schweinhart, Barnes, and Weikart. The most cited summary is the age-40 follow-up report from 2005. The dataset files are typically available in SPSS and SAS formats. Some of the coding books are quite thin because the original researchers made assumptions about how to code missing values that modern analysts might question. I ran into a problem a few years ago when I tried to merge the economic outcomes file with the psychological well-being wave. The participant ID scheme changed between the age-27 and age-40 datasets. One file uses a zero-padded numeric key while the other uses a mixed alphanumeric code. I ended up writing a small Python script to clean the IDs before merging. Without fixing that mismatch first, the merge dropped about 12 percent of the sample because of duplicate entries created by the inconsistent formatting. I kept a log of every transformation so anyone else working with the file could replicate the process.
Common Misreadings and Where the Study Falls Short
People often treat the Perry results as proof that any preschool program works. That reading is wrong. The intervention was expensive and intensive. It ran for two years. Each child attended five days a week with two teachers present. Home visits were conducted once every two weeks. The teacher-student ratio was roughly one to six. Most modern public preschool programs do not match that level of staffing or frequency. The study does not tell you whether a standard state-funded pre-K program will produce the same outcomes. Another misreading involves causation. The sample was small. There were only 123 original participants. Randomization was used, but the groups were not perfectly balanced at baseline. A few covariates differed slightly between the treatment and control groups. Some later analyses adjusted for those differences and found the effect sizes shrink somewhat. The core findings still hold directionally, but the magnitude of the estimated returns drops when you apply modern causal inference adjustments. That is worth knowing if you are building a model around this data. The economic follow-up data is also incomplete. Income records were derived from administrative sources, which means self-employment income, cash payments, and work in the informal economy are undercounted. The study captured formal earnings well, but it missed a meaningful slice of how these individuals actually made money. If you are using the dataset for labor market analysis, you should treat the income variable as a floor estimate, not a ceiling.
Get the Full Details
Working With the Dataset in Practice
The variables are dated. Some of the survey instruments used scales that are obsolete now. The vocabulary in the open-ended follow-up questions reflects 1960s and 1970s conventions. When you code qualitative responses, you will encounter terminology that is either outdated or loaded. I stopped trying to recode everything through a modern framework and instead documented the original wording before applying any translation. That way, readers can see what the interviewers actually asked rather than a sanitized version that smooths over historical context. The attrition rate is not trivial. A portion of the original cohort could not be located at later follow-up points. The researchers did their best to track participants across decades, but some data is missing for specific waves. When analyzing outcomes, you should check whether the missingness correlates with treatment status. In the Perry case, attrition was roughly balanced between groups, which reduces the risk of selection bias, but it still matters for statistical power. Dropping cases listwise will hurt your precision without adding credibility. One thing that trips up people new to this dataset is the timing of the follow-up intervals. The waves are not evenly spaced. You have age 4, age 11, age 14, age 15, age 19, age 27, and age 40. There is no adolescence-to-early-adulthood bridge between 19 and 27. If you are running growth models or event history analysis, that gap will show up. You cannot interpolate cleanly across eight years without making assumptions you should state explicitly. I usually suggest using piecewise linear splines anchored at the known time points rather than forcing a smooth curve through a wide empty stretch.
Bottom Line
The Perry Preschool data remains one of the most important longitudinal datasets in education research. It is useful for studying long-term returns to early intervention, for testing life-course models, and for evaluating policy arguments. It is not useful as evidence that every early childhood program will deliver similar results. The intervention was unusually resourced. The sample was narrow. The follow-up protocols reflected the methods and limitations of their era. Use the data carefully. Document your cleaning steps. Report the caveats alongside the findings.
