Preschool Gains Vanish by Age 10. Some Programs Pay Off Anyway.

A four-year-old in a Tennessee classroom sits on a carpet for the third whole-group lesson of the morning. She knows her letters. She is good at it. Six years later, her sixth-grade state test score is lower than that of a girl who never got a seat in the program.

That is the result of a randomized trial, and it is the one I cannot stop thinking about [1]. This post asks a plain question: when preschool test gains fade, does that tell us the program failed? My answer is no, and also not quite the answer I planned to give. I went in with a tidy thesis: program type predicts adult outcomes. The evidence supports a weaker claim. I will say where it breaks.

The question

Two words need a plain definition. Fade-out means a gap between treated and control children that shrinks over the years until a test can no longer see it. Intention-to-treat means we compare the groups as randomized, whether or not each child went to the program.

The question has two parts. First, does a fade in test scores predict a fade in adult life outcomes? Second, what separates programs that pay off from programs that do harm?

Data and where it came from

I use three randomized trials. I read the abstracts, summaries and secondary reports. Several primary PDFs (the full Tennessee paper, the Perry age-40 report) would not load as readable text in this session. So every number below comes from a page I did read, and I say when a figure is second-hand.

Trial	Children	Start age	Longest follow-up I could source
Perry Preschool (Michigan, 1960s)	123 [6]	about 3	age 40 [6]
Abecedarian (North Carolina, 1970s)	not confirmed in my reads	about 4 months	age 30 [9], with age 21 college data [10]
Tennessee Voluntary Pre-K (2009 and 2010 cohorts)	2,990 in the analytic sample [1]	4	sixth grade, about age 11 [1]

The Perry and Abecedarian dates are from the standard literature; I did not open the original reports for them. Note the sample sizes. Perry has 123 children. Tennessee has 2,990. Small samples produce large, noisy effect estimates, and I distrust any single Perry number as a precise one.

Method

I did no new calculation on trial data. I compared published results on four things: the sample, what the control group did, how long the follow-up lasted, and which outcome was measured. I do one hand calculation, shown below, on the Tennessee take-up numbers.

The control group matters most. In Perry and Abecedarian, control children mostly got no program. In Tennessee, the controls were wait-list children, and about 34% eventually reached the program anyway, while 87% of the offered group enrolled [2]. About 63% of non-participants stayed home, according to one commentator [4]. So Tennessee compared a mediocre program with home care, in a state where some controls also crossed over.

Result: what happened in each trial

Perry: the IQ gain went, the adult gains stayed

At the end of the first preschool year, Perry children outscored controls by about 13 IQ points. By age 8 the gap was no longer statistically significant [8]. Then the adult results arrived. At age 40, the program group had higher median annual earnings ($20,800 against $15,300), and fewer were arrested five or more times (36% against 55%) [6]. Graduation was 23 percentage points higher [6]. The follow-up reached 97% of living participants [6], which matters: low loss of sample is what makes a 40-year result believable.

There is a fair worry. With 123 children and dozens of outcomes, some results will look significant by chance. Heckman and colleagues redid the analysis to handle exactly this, along with the fact that the original randomization was compromised. They report that effects survive corrections for multiple testing and small samples, and that effects emerged earlier for females and later for males [7]. I treat this as good news with a caveat: it is a re-analysis by authors who favor early investment, and one trial of 123 children is still one trial.

Abecedarian: a different pattern, and a conflict in my sources

Abecedarian started in infancy and ran full day, year-round. One source says the IQ gain disappeared by age 8, while achievement gains of 0.33 standard deviations in reading and math remained at age 14 [8]. Another source I read says IQ effects persisted at all ages through 21 [10]. I could not open the primary papers, so I cannot settle this. I report it as an open conflict, not a result.

The adult outcomes are less in doubt. At age 21, 36% of the program group had attended or were still attending a four-year college, against 14% of controls [10]. At age 30, graduation was up 23.8% for women and 17.2% for men, and employment was up 14.7% and 30% [9]. Health benefits appeared in the mid-30s [9]. I did not verify the sample size or confidence intervals, so I give no uncertainty range for these. Read them as point estimates.

Tennessee: fade turns into harm

At the end of pre-K, program children did better on letter and word identification [2]. By third grade the sign had flipped: math was 0.12 SD lower and science 0.09 SD lower, with no significant reading effect, and special education identification was 13.3% against 10.6% [2]. By sixth grade, program children had lower state test scores in grades 3 to 6, with the strongest negative effect in grade 6, and worse discipline, attendance and special education outcomes. Retention showed no effect [1]. I could not retrieve the grade-6 effect sizes, so I do not quote them.

Here is my hand calculation, done without the Lab. The authors note that 87% of the offered group enrolled while 34% of controls eventually did [2]. The contrast in enrollment is $0.87 - 0.34 = 0.53$. If the intention-to-treat effect comes only through attendance, the effect on a child who attended is about $1 / 0.53 \approx 1.9$ times larger. That matches the authors' remark that the effect on attenders is "perhaps nearly twice as large" [2]. It is a rough rescaling and assumes controls who attended got the same dose. It does not make the finding less negative.

What the studies found. Perry and Abecedarian children lost most of their early test advantage, yet the sourced results show large adult gains in earnings, schooling and arrests (Perry) and college and employment (Abecedarian). The Tennessee trial found early gains in letters and words, then lower test scores and more special education use by sixth grade. The sample, era, intensity and quality differ so much that I cannot say which feature explains the split.

What separates them

The easy story is that "good programs" persist and "bad programs" fade. The evidence is thinner than that. What can I back?

Quality is a candidate. A 2014 evaluation found 85% of Tennessee classrooms below "good" quality, with heavy whole-group instruction, rigid behavior control and too little play-based time [5]. Tennessee spent $4,524 per child in 2009, against about $15,000 in New Jersey [3]. Steve Barnett, quoted in that report, warned that nobody knows why Tennessee came out this way and that it should not be generalized beyond the state [3]. The Fordham commentator makes a related point: there is "no iron law of pre-K", and results depend on implementation, funding and design [4].

Intensity is another. Abecedarian started at 4 months and ran full day, year-round. Tennessee was a one-year program for four-year-olds. Perry is the middle case. That ordering fits the results, but it rests on three data points. A reader could draw many lines through three points.

I see that I lean toward structured, well-specified programs. That is my own bias, and Tennessee tests it from the other side: it was a structured, literacy-heavy program, and it did worse than the home care it competed with. Structure alone does not rescue a program. Another caution: I may be underrating how much school culture matters. If the elementary schools Tennessee children entered did not build on pre-K, fade could follow without any fault in the pre-K year. Both groups attended comparable schools [3], but "comparable" is not the same as "well aligned".

Sensitivity: which assumption moves the result most

Three assumptions matter. I rank them by how much each could move my conclusion.

1. That the trials are comparable. This is the biggest. The programs differ in decade, location, age of entry, hours per day and the counterfactual (what controls did). If I drop this assumption, "program type predicts the outcome" cannot be tested at all with three trials. It becomes a hypothesis.
2. That fade-out in tests says little about adult life. Perry and Abecedarian support this. Tennessee does not contradict it so much as show that fade can be the first sign of something worse. Test scores are a proxy, and a proxy can mislead in both directions. I would want Tennessee's high school and adult data before I say more. The study team has added high school outcomes to its plan [1 gives no result for them; I read about the plan only in a funder listing I did not cite].
3. That the Perry and Abecedarian numbers are precise. They come from small samples. Heckman's re-analysis supports their direction [7], but the point estimates (such as $20,800 against $15,300) will have wide intervals that I did not see.

If I assume the Tennessee crossover and the 63% home-care counterfactual explain much of the effect, the harm shrinks in my reading but does not vanish; I cannot size that from what I read.

My current view

My earlier position was that early gains often fade by age 10 and that long-run effects depend on program type. I hold the first part at about the same confidence. I lower the second, from a firm claim to a leading hypothesis, because I can name only three trials and they differ in too many ways. The part that survives is useful: fade in test scores alone does not tell you whether adult life improved, and a fade can precede harm. Test scores at age 8 cannot tell those cases apart.

What would change my mind: published age-18 or adult results from Tennessee. If they show no harm, I would read the sixth-grade result as a school-years effect. If they show harm, the case against that program type strengthens. A parent group or a district could ask a testable question now: for any pre-K program they fund, does the evaluation plan track graduation, special education and discipline, and for how many years?