A lot of people pushed back against my post on preschool, so it looks like we need to discuss this in more depth.
A quick refresher: good randomized controlled trials have shown that preschools do not improve test scores in a lasting way. Sometimes test scores go up a little bit, but these effects disappear after a year or two of regular schooling. However, early RCTs of intensive “wrap-around” preschools like the Perry Preschool Program and the Abecedarians found that graduates of those programs went on to have markedly better adult outcomes, including higher school graduation rates, more college attendance, less crime, and better jobs. But these studies were done in the 60s, before people invented being responsible, and had kind of haphazard randomization and followup. They were also small sample sizes, and from programs that were more intense than any of the scaled-up versions that replaced them. Modern scaled-up preschools like Head Start would love to be able to claim their mantle and boast similar results. But the only good RCT of Head Start, the HSIS study, is still in its first few years. It’s confirmed that Head Start test score gains fade out. But it hasn’t been long enough to study whether there are later effects on life outcomes. We can expect those results in ten years or so. For now, all we have is speculation based on a few quasi-experiments.
Deming 2009 is my favorite of these. He looks at the National Longitudinal Survey of Youth, a big nationwide survey that gets used for a lot of social science research, and picks out children who went to Head Start. These children are mostly disadvantaged because Head Start is aimed at the poor, so it would be unfair to compare them to the average child. He’s also too smart to just “control for income”, because he knows that’s not good enough. Instead, he finds children who went to Head Start but who have siblings who didn’t, and uses the sibling as a matched control for the Head Starter.
This ensures the controls will come from the same socioeconomic stratum, but he acknowledges it raises problems of its own. Why would a parent send one child to Head Start but not another? It might be that one child is very stupid and so the parents think they need the extra help preschool can provide; if this were true, it would mean Head Starters are systematically dumber than controls, and would underestimate the effect of Head Start. Or it might be that one child is very smart and the so the parents want to give them education so they can develop their full potential; if this were true, it would mean Head Starters are systematically smarter than controls, and would inflate the effect of Head Start. Or it might be that parents love one of their children more and put more effort into supporting them; if this meant these children got other advantages, it would again inflate the effect of Head Start. Or it might mean that parents send the child they love more to a fancy private preschool, and the child they love less gets stuck in Head Start, ie the government program for the disadvantaged. Or it might be that parents start out poor, send their child to Head Start, and then get richer and send their next child to a fancy private preschool, while that child also benefits from their new wealth in other ways. There are a lot of possible problems here.
Deming tries very hard to prove none of these are true. He compares Head Starters and their control siblings on thirty different pre-study variables, including family income during their preschool years, standardized test scores, various measures of health, number of hours mother works during their preschool years, breastfedness, etc. Of these thirty variables, he finds a significant difference on only one: birth weight. Head Starters were less likely to have very low birth weight than their control siblings. This is a moderately big deal, since birth weight is a strong predictor of general child health and later life success. But:
Given the emerging literature on the connection between birth weight and later outcomes, this is a serious threat to the validity of the [study]. There are a few reasons to believe that the birth weight differences are not a serious source of bias, however. First, it appears that the difference is caused by a disproportionate number of low-birth-weight children, rather than by a uniform rightward shift in the distribution of birth weight for Head Start children. For example, there are no significant differences in birth weight once low-birth-weight children (who represent less than 10 percent of the sample) are excluded.
Second, there is an important interaction between birth order and birth weight in this sample. Most of the difference in mean birth weight comes from children who are born third, fourth, or later. Later-birth-order children who subsequently enroll in Head Start are much less likely to be low birth weight than their older siblings who did not enroll in preschool. When I restrict the analysis to sibling pairs only, birth weight differences are much smaller and no longer significant, and the main results are unaffected. Finally, I estimate all the models in Section V with low-birth-weight children excluded, and, again, the main results are unchanged.
Still, to get a sense of the magnitude of any possible positive bias, I back out a correction using the long-run effect of birth weight on outcomes estimated by Black, Devereux, and Salvanes (2007). Specifically, they find that 10 percent higher birth weight leads to an increase in the probability of high school graduation of 0.9 percentage points for twins and 0.4 percentage points for siblings. If that reduced form relationship holds here, a simple correction suggests that the effect of Head Start on high school graduation (and by extension, other outcomes) could be biased upward by between 0.2 and 0.4 percentage points, or about 2–5 percent of the total effect.
Having set up his experimental and control group, Deming does the study and determines how well the Head Starters do compared to their controls. The test scores show some confusing patterns that differ by subgroup. Black children (the majority of this sample; Head Start is aimed at disadvantaged people in general and sometimes at blacks in particular) show the classic pattern of slightly higher test scores in kindergarten and first grade, fading out after a few years. White children never see any test score increases at all. Some subgroups, including boys and children of high-IQ mothers, see test score increases that don’t seem to fade out. But these differences in significance are not themselves significant and it might just be chance. Plausibly the results for blacks, who are the majority of the sample, are the real results, and everything else is noise added on. This is what non-subgroup analysis of the whole sample shows, and it’s how the study seems to treat it.
The nontest results are more impressive. Head Starters are about 8% more likely to graduate high school than controls. This pattern is significant for blacks, boys, and children of low-IQ mothers, but not for whites, girls, and children of high-IQ mothers. Since the former three categories are the sorts of people at high risk of dropping out of high school, this is probably just floor effects. Head Starters are also less likely to be diagnosed with a learning disability (remember, learning disability diagnosis is terrible and tends to just randomly hit underperforming students), and marginally less likely to repeat grades. The subgroup results tend to show higher significance levels for groups at risk of having bad outcomes, and lower significance levels for the rest, just as you would predict. There is no effect on crime. For some reason he does not analyze income, even though his dataset should be able to do that.
He combines all of this into an artificial index of “young adult outcomes” and finds that Head Start adds 0.23 SD. You may notice this is less than the 0.3 SD effect size of antidepressants that everyone wants to dismiss as meaningless, but in the social sciences apparently this is pretty good. Deming optimistically sums this up as “closing one-third of the gap between children with median and bottom-quartile family income”, as “75% of the black-white gap”, and as “80% of the benefits of [Perry Preschool] at 60% of the cost”.
Finally, he does some robustness checks to make sure this is not too dependent on any particular factor of his analysis. I won’t go into these in detail, but you can find them on page 127 of the manuscript, and it’s encouraging that he tries this, given that I’m used to reading papers by social psychologists who treat robustness checks the way vampires treat garlic.
Deming’s paper very similar to Garces Thomas & Currie (2002), which does the same methodology on a different dataset. GTC is earlier and more famous and probably the paper you’ll hear about if you read other discussions of this topic; I’m focusing on Deming because I think his analyses are more careful and he explains what he’s doing a lot better. Reading between the lines, GTC do not find any significant effects for the sample as a whole. In subgroup analyses, they find Head Start makes whites more likely to graduate high school and attend college, and blacks less likely to be involved in crime. One can almost sort of attribute this to floor effects; blacks many times more likely to have contact with the criminal justice system, and there are more blacks than whites in the sample, so maybe it makes sense that this is only significant for them. On the other hand, when I look at the results, there was almost as strong a positive effect for whites (ie Head Start whites committed more crimes, to the same degree Head Start blacks committed fewer crimes) – but there were fewer whites so it didn’t quite reach significance. And the high school results don’t make a lot of sense however you parse them. GTC use the words “statistically significant” a few times, so you know they’re thinking about it. But they don’t ever give significance levels for individual results and one gets the feeling they’re not very impressive. Their pattern of results isn’t really that similar to Deming’s either – remember, Deming found that all races were more likely to benefit from high school, and no race had less crime. GTC also don’t do nearly as much work to show that there aren’t differences between siblings. Deming is billed as confirming or replicating GTC, but this only seems true in the sense that both of them say nice things about Head Start. Their patterns of results are pretty different, and GTC’s are kind of implausible.
And for that matter, ten years earlier two of these authors, Currie and Thomas, did a similar study. They also use the National Longitudinal Survey of Youth, meaning I’m not really clear how their analysis differs from Deming’s (maybe it’s much earlier and so there’s less data?) They first use an “adjust for confounders” model and it doesn’t work very well. Then they try a comparing-siblings model and find that Head Starters are generally older than their no-preschool siblings, and also generally born to poorer mothers (these are probably just the same result; mothers get less poor as they get older). They also tend to do better on a standardized test, though the study is very unclear about when they’re giving this test so I can’t tell if they’re saying that group assignment is nonrandom or that the intervention increased test scores. They find Head Start does not increase income, maybe inconsistently increases test scores among whites but not blacks, decreases grade repetition for whites but not blacks, and improves health among blacks but not whites. They also look into Head Start’s effect on mothers, since part of the wrap-around program involves parent training. All they find is mild effects on white IQ scores, plus “a positive and implausibly large effect of Head Start on the probability that a white mother was a teen at the first birth” which they say is probably sampling error. Like the later study, this study does not give p-values and I am too lazy to calculate them from the things they do give, but it doesn’t seem like they’re likely to be very good.
Finally, Deming’s work was also replicated and extended by a team from the Brookings Institute. I think what they’re doing is taking the National Longitudinal Survey of Youth – the same dataset Deming and one of the GTC papers used – and updating it after a few more years of data. Like Deming, they find that “a wide variety” of confounders do not differ between Head Starters and their unpreschooled siblings. Because they’re with the Brookings Institute, their results are presented in a much prettier way than anyone else’s:
The Brookings replication (marked THP here) finds sizes somewhat larger than GTC, but somewhat smaller than Perry Preschool. It looks like they find a positive and significant effect on high school graduation for Hispanics, but not blacks or whites, which is a different weird racial pattern than all the previous weird racial patterns. Since their sample was disproportionately black and Hispanic, and the blacks almost reached significance, the whole sample is significant. They find increases of about 6% on high school graduation rates, compared to Deming’s claimed 8%, but on this chart it’s hard to see how Deming said his 8% was 80% as good as Perry Preschool. There are broadly similar effects on some other things like college attendance, self esteem, and “positive parenting”. They conclude:
These results are very similar to those by Deming (2009), who calculated high school graduation rates on the more limited cohorts that were available when he conducted his work.
These four studies – Deming, GTC, CT, and Brookings – all try to do basically the same thing, though with different datasets. Their results all sound the same at the broad level – “improved outcomes like high school graduation for some racial groups” – but on the more detailed level they can’t really agree which outcomes improve and which racial groups they improve for. I’m not sure how embarrassing this should be for them. All of their results seem to be kind of on the border of significance, and occasionally going below that border and occasionally above it, which helps explain the contradictions while also being kind of embarrassing in and of itself (Deming’s paper is the exception, with several results significant at the 0.01 level). Most of them do find things generally going the right direction and generally sane-looking findings. Overall I feel like Deming looks pretty good, the Brookings replication is too underspecified for me to have strong opinions on, and the various GTC papers neither add nor subtract much from this.
I’m treating Ludwig and Miller separately because it’s a different – and more interesting – design.
In 1965, the government started an initiative to create Head Start programs in the 300 poorest counties in the US. There was no similar attempt to help counties #301 and above, so there’s a natural discontinuity at county #300. This is the classic sort of case where you can do a regression discontinuity experiment, so Ludwig and Miller decided to look into it and see if there was some big jump in child outcomes as you moved from the 301st-poorest-county to the 300th.
They started by looking into health outcomes, and found a dramatic jump. Head Start appears to improve the outcomes of certain easily-preventable childhood diseases 33-50%. For example, kids from counties with Head Start programs had much less anemia. Part of the Head Start program is screening for anemia and supplementing children with iron, which treats many anemias. So this is very unsurprising. Remember that the three hundred poorest counties in 1965 were basically all majority-black counties in the Deep South and much worse along every axis than you would probably expect – we are talking near-Third-World levels of poverty here. If you deploy health screening and intervention into near-Third-World levels of poverty, then the rates of easily preventable diseases should go down. Ludwig and Miller find they do. This is encouraging, but not really surprising, and maybe not super-relevant to the rest of what we’re talking about here.
But they also find a “positive discontinuity” in high school completion of about 5%. Kids in the 300th-and-below-poorest counties were about 5% more likely than kids in the 301st-and-above-poorest to finish high school. This corresponds to an average of staying in school six months longer. This discontinuity did not exist before Head Start was set up, and it does not exist among children who were the wrong age to participate in Head Start at the time it was set up. It comes into existence just when Head Start is set up, among the children who were in Head Start. This is a pretty great finding.
Unfortunately, it looks like this. The authors freely admit this is just at the limit of what they can detect at p < 0.05 in their data. They double check with another data source, which shows the same trend but is only significant at p < 0.1. "Our evidence for positive Head Start impacts on educational attainment is more suggestive, and limited by the fact that neither of the data sources available to us is quite ideal." This study has the strongest design, and it does find an effect, but the effect is basically squinting at a graph and saying "it kind of looks like that line might be a little higher than the other one". They do some statistics, but they are all the statistical equivalent of squinting at the graph and saying "it kind of looks like that line might be a little higher than the other one", and about as convincing. For a more complete critical look, see this post from the subreddit.
There is one other slightly similar regression discontinuity study, Carneiro and Ginja, which regresses a sample of people on Head Start availability and tries to prove that people who went to Head Start because they were just within the availability cutoff do better than people who missed out on Head Start because they were just outside it. This sounds clever and should be pretty credible. They find a bunch of interesting effects like that Head Starters are less likely to be obese, and less likely to be depressed. They find that non-blacks (but not blacks) are less likely to be involved in crime (which, remember, is the opposite finding as the last paper about Head Start and crime and race). But they don’t find any effect on likelihood to graduate high school or be involved in college. Also, they bury this result and everyone cites this paper as “Look, they’ve replicated that Head Start works!”
A few scattered other studies to put these in context:
In 1980, Chicago created “Child Parent Centers”, a preschool program aimed at the disadvantaged much like all of these others we’ve been talking about. They did a study, which for some reason published its results in a medical journal, and which doesn’t really seem to be trying in the same way as the others. For example, it really doesn’t say much about the control group except that it was “matched”. Taking advantage of their unusually large sample size and excellent follow-up, they find that their program made children stay in school the same six months longer as many of the other studies find, had a strong effect on college completion (8% vs. 14% of kids), showed dose-dependent effects, and “was robust”. They are bad enough at showing their work that I am forced to trust them and the Journal of the American Medical Association, a prestigious journal that I can only hope would not have published random crap.
Havnes and Mogstad analyze a free universal child-care program in Norway, which was rolled out in different places at different times. They find that “exposure to child care raised the chances of completing high school and attending college, in orders of magnitude similar to the black-white race gaps in the US”. I am getting just cynical enough to predict that if Norway had black people, they would have a completely different pattern of benefits and losses from this program, but the Norwegians were able to avoid a subgroup analysis by being a nearly-monoethnic country. This is in contrast to Quebec, where a similar childcare program seems to have caused worse long-term outcomes. Going deeper into these results supports (though weakly and informally) a model where, when daycare is higher-quality than parental care, child outcomes improve; when daycare is lower-quality than parental care, child outcomes decline. So a reform that creates very good daycare, and mostly attracts children whose parents would not be able to care for them very well, will be helpful. Reforms that create low-quality daycare and draw from households that are already doing well will be harmful. See the discussion here.
Then there’s Chetty’s work on kindergarten, which I talk about here. He finds good kindergarten teachers do not consistently affect test scores, but do consistently affect adult earnings, similar to fade-out arguments around preschool. This study is randomized and strong. Its applicability to the current discussion is questionable, since kindergarten is not preschool, having a good teacher is not going to preschool at all, and the studies we’re looking at mostly haven’t found results about adult earnings. At best this suggests that schooling can have surprisingly large and fading-out-then-in-again effects on later life outcomes.
And finally, there’s a meta-analysis of 22 studies of early childhood education showing an effect size of 0.24 SD in favor of graduating high school, p less than 0.001. Maybe I should have started with that one. Maybe it’s crazy of me to save this for the end. Maybe this should count for about five times as much as everything I’ve mentioned so far. I’m putting it down here both to inflict upon you the annoyance I felt when discovering this towards the end of researching this topic, and so that you have a good idea of what kind of studies are going into this meta-analysis.
What do we make of this?
I am concerned that all of the studies in Parts I and II have been summed up as “Head Start works!”, and therefore as replicating each other, since the last study found “Head Start works!” and so did the newest one. In fact, they all find Head Start having small effects for some specific subgroup on some specific outcome, and it’s usually a different subgroup and outcome for each. So although GCT and Deming are usually considered replications of each other, they actually disprove each other’s results. One of GCT’s two big findings is that Head Start decreases crime among black children. But Deming finds that Head Start had no effect on crime among black children. The only thing the two of them agree on is that Head Start seems to improve high school graduation among whites. But Carneiro and Ginja, which is generally thought of as replicating the earlier two, finds Head Start has no effect on high school graduation among whites.
There’s an innocent explanation here, which is that everyone was very close to the significance threshold, so these are just picking up noise. This might make more sense graphically:
It’s easy to see here that both studies found basically the same thing, minus a little noise, but that Study 1 has to report its results as “significant for blacks but not whites” and Study 2 has to report the opposite. Is this what’s going on?
I made a table. I am really really not confident in this table. On one level, I am fundamentally not confident that what I am doing is even possible, and that the numbers in these studies are comparable to one another or mean what it looks like they mean. On a second level, I’m not sure I recorded this information correctly or put the right numbers in the right places. Still, here is the table; red means the result is significant:
This confirms my suspicions. Every study found something different, and it isn’t even close. For example, Carneiro & Ginja finds a strong effect of lowering white crime, but GCT finds that Head Start nonsignificantly increases white crime rates. Meanwhile, GCT find a strong and significant effect lowering black crime, but Carneiro and Ginja find an effect of basically zero.
The strongest case for the studies being in accord is for black high school graduation rates. Both Deming and Ludwig+Miller find an effect. Carneiro and Ginja don’t find an effect, but their effect size is similar to those of the other studies, and they might just have more stringent criteria since they are adjusting for multiple comparisons and testing many things. But they should have the more stringent criteria, and by trying to special-plead against this, I am just reversing the absolutely correct thing they did because I want to force positive results in the exact way that good statistical practice is trying to prevent me from doing. So maybe I shouldn’t do that.
Here is the strongest case for accepting this body research anyway. It doesn’t quite look like publication bias. For one thing, Ludwig and Miller have a paper where they say there’s probably no publication bias here because literally every dataset that can be used to test Head Start has been. For another, although I didn’t focus on gender or IQ on the chart above, most of the studies do find that it helps males and low-IQ people more with the sorts of problems men and low-IQ people usually face, which suggest it passes sanity checks. Most important, in a study whose results are entirely spurious, there should be an equal number of beneficial and harmful findings (ie they should find Head Start makes some subgroups worse on some outcomes). Since each of these studies investigates many things and usually finds many different significant results, it should be hard to publication bias all harmful findings out of existence. This sort of accords with the positive meta-analysis. Studies either show small positive results or are not signficant, and when you combine all of them into a meta-analysis, they become highly significant, look good, and make sense. And this would fit very well with the Norwegian study showing strong positive effects of childcare later in life. And Chetty’s study showing fade-out of kindergarten teachers followed by strong positive effects later in life. And of course the Perry Preschool and Abecedarian studies showing fade-out of tests scores followed by strong positive effects later in life. I even recently learned of a truly marvelous developmental explanation for why this might happen, which unfortunately this margin is too small to contain – expect a book review in the coming weeks.
The case against this research is that maybe the researchers cheated to have there be no harmful findings. Maybe the meta-analysis just shows that when a lot of researchers cheat a little, taking care to only commit minor undetectable sins, that adds up to a strong overall effect. This is harsh, but I was recently referred to this chart (h/t Mother Jones, which calls it “the chart of the decade” and “one of the greatest charts ever produced”):
This is the outcome of drug trials before and after the medical establishment started requiring preregistration (the vertical line) – in other words, before they made it harder to cheat. Before the vertical line, 60% of trials showed the drug in question was beneficial. After the vertical line, only 10% did. In other words, making it harder to cheat cuts the number of positive trials by a factor of six. It is not at all hard to cheat in the research of early childhood education; all the research in this post so far comes from the left side of the vertical line. We should be skeptical of all but the most ironclad research that comes from the left – and this is not the most ironclad research.
The Virtues of Rationality say:
One who wishes to believe says, “Does the evidence permit me to believe?” One who wishes to disbelieve asks, “Does the evidence force me to believe?” Beware lest you place huge burdens of proof only on propositions you dislike, and then defend yourself by saying: “But it is good to be skeptical.” If you attend only to favorable evidence, picking and choosing from your gathered data, then the more data you gather, the less you know. If you are selective about which arguments you inspect for flaws, or how hard you inspect for flaws, then every flaw you learn how to detect makes you that much stupider.
This is one of the many problems where the evidence permits me to disbelieve, but does not force me to do so. At this point I have only intuition and vague heuristics. My intuition tells me that in twenty years, when all the results are in, I expect early childhood programs to continue having small positive effects. My vague heuristics say the opposite, that I can’t trust research this irregular. So I don’t know.
I think I was right to register that my previous belief preschool definitely didn’t work was outdated and under challenge. I think I was probably premature to say I was wrong about preschool not working; I should have said I might be wrong. If I had to bet on it, I would say 60% odds preschool helps in ways kind of like the ones these studies suggest, 40% odds it’s useless.
I hope that further followup of the HSIS, an unusually good randomized controlled trial of Head Start, will shed more light on this after its participants reach high school age sometime in the 2020s.