Who you live around surely matters. Poor kids in poor neighborhoods have worse outcomes than poor kids in rich neighborhoods. Children have educational outcomes which are correlated with those of their neighborhood, and those of their classmates. Sorting out whether this is causal, though, is difficult. People select into environments in ways that are not easily visible to the outside investigator. Charles Manski calls this the “reflection problem”, after a man in a mirror. An outside viewer is unable to distinguish between a person in the mirror reflecting the movement of a man, or causing it, unless they have outside information. In addition, changes in outcome could be due to having access to higher quality institutions, rather than having different people around you. This has a big impact on our policy choices — if it is simply institutional quality, then changing people’s peer groups is entirely unnecessary.
Nevertheless, numerous experiments and quasi-experiments suggest that peers do have a causal effect on outcomes, though the effects are generally small. The direction of the effects suggest that poor people pull each other down. Concentrated poverty creates misery greater than the sum of its parts, and we can perhaps improve outcomes by scattering the children of the poor amongst richer people. I do not wish to oversell these conclusions, however. There is no panacea for intergenerational poverty, as the primary for its persistence are the shared traits and cultures which parents pass down to their children. Improvement is possible, but difficult.
The cleanest sources of evidence on neighborhood effects are the “Moving to Opportunity” experiments. 4,600 families were divided into three groups, two of which received vouchers to pay for housing. For one of the groups receiving vouchers, it could only be used for low-income areas within the first year, while the other group could move anywhere and move immediately. (The third group received no vouchers, so as to be the controls). Numerous studies have been written using this data set. For example, Chetty, Hendren, and Katz (2016) look at the long run effects of moving to a better neighborhood on adolescents. Moving while young, while temporarily disruptive, has beneficial effects on income, educational attainment and eventual college attendance. For those below the age of 13, receiving the offer of a housing voucher increases income in their mid-20s by $1,624 per year. Accounting for not everyone taking the voucher, the treatment on the treated effect was $3,477, a 31% increase in income. These benefits fade out as someone gets older, suggesting that one is permanently shaped by the environment of their youth. They do not find a critical age, before which moves should occur. Rather, the improvement in income is linearly related to the number of years they spent in better neighborhoods. These effects of neighborhoods are very local — Chetty, Friedman, Hendren, Jones, and Porter (2018) find that the characteristics of census tracts one mile away have no predictive power for outcomes. Ludwig et al (2015), however, working with the same dataset, finds no economic effects whatsoever. Another study, from 2006, corroborates Ludwig, finding no difference in educational attainment (which is a very strong predictor of income). Instead, Ludwig et al. find positive impacts on well-being. The key choice here seems to be combining children of all age groups, rather than accounting for length of exposure. Whether we should allow that is a tough decision. P-values only measure the probability of something occurring due to chance if we have one look at one value. Once we have multiple comparisons, the odds of finding something which is only due to chance goes up. We face a trade-off between inclusiveness of results across many dimensions, and not believing false things. I do think the Chetty study is broadly correct in doing this, but only because the basic finding replicated elsewhere, as we’ll see in the next section.
Quasi-experimental evidence uses very large datasets of moves, combined with plausibly exogenous reasons to move to find the causal impact. Chetty and Hendren (2018) discover the same effects of Chetty, Hendren, and Katz — that the length of exposure matters. This is key to their identification strategy. If several people in a family move, they are going to have similar environments to each other, with the key difference that younger children are going to spend more time in the better environment. They find that children moving from one area to another converge to the new area’s income at a rate of 4% per year. In other words, a child moving from an area at the 20th percentile of income to the 80th percentile of income at age 9, and residing there until 23, will close 56% of the income gap. This is connected only to the duration of exposure — moving for a year, and then moving back, has the same effect at age 10 as it does at age 15. The results from Moving to Opportunity, then, is due only to older children having less time to absorb the norms and culture of the new area.
Obviously, one cannot simply run regressions and call it a day. To argue causality, one must defend it. They use four strategies. First, they control for the observable characteristics of the family, such as college education. A potential source of bias would be college educated people being more likely to migrate to places where there are greater returns to skill, and the actual convergence in income is driven by that college education. Their results are robust to this. Next, you need to check against changes in family environment over time. Suppose your mother lands a new job, which enables your family to move to a better area. It may be the new job which is responsible for the convergence in income, not the area. They cannot observe everything which changes, of course, but they can observe income and marital status. Their results are robust to this, too.
The next strategy, looking for a plausibly exogenous source of a move, is one in common with several other papers. Paper use the unexpected destruction of a neighborhood, such as the demolition of public housing, as an instrumental variable. Public housing, such as the Cabrini-Green in Chicago, were and are notorious for crime, squalor and concentrated poverty. For the demolition of the actual Cabrini-Green neighborhood, the effects on people, especially children, were positive. (Of course, another study disagrees). Sudden shocks, like Hurricane Katrina, led to more than one million people having to move. Imberman, Kugler, and Sacerdote (2012) find that being around higher-achieving students raises outcomes, and being around lower achieving students lowers outcomes, with no changes in average outcomes. Nakamura, Sigurdsson, and Steinsson (2021) use your house getting run over by lava flows as the exogenous instrument, and find positive impacts on income. (Although, this paper is mainly about mobility, not peer effects, and so shouldn’t be too heavily weighted in your mind). In Chetty and Hendren, they use both natural disasters and plant closures to check the accuracy of their results. Unlike the prior studies in the “shock” literature, they identify the change in ZIP codes first, and check to see if there is a corresponding shock — I don’t think this is particularly important. Their results are robust to this.
The last step is using some extremely clever tests across subgroups. I emphasize extremely clever because I was so impressed with some of these ideas. The first step is to consider how areas improve or worsen over time. The outcomes of people who move converge to those from the area in their own birth cohort, and are unaffected by those before or after, which is precisely in line with predictions. The next is to check the distribution of income. Not all places have the same standard deviation — the example that they give in the paper is that residents of San Francisco are more likely to have very high or very low incomes than those in Boston. The distribution of the people who move mimics the distribution of that area, suggesting very strongly that it is in fact the place which is causing it. How improbable would it be that unobserved sorting happens to precisely mimic the structure, as well as the mean! Finally, they consider places where there are differences in gender outcomes. In some places, boys do relatively better — in others, girls. If a family with a daughter and a son move to an area where daughters do better, then if the place has an impact, the daughter should do relatively better. If they move somewhere where sons do better, then the converse. This degree of sorting would be highly implausible otherwise. It remains possible that the families which did choose to move had traits which would later correlate with success. These traits, though, would have to have effects proportionate to the years spent in a neighborhood for each child, not affect income, be correlated with moving due to a natural disaster, and replicate the exact distribution of each place in multiple different ways.
There is a long line of literature on peer effects in education. I pray you forgive me taking a synoptic view of the literature; to every relevant study would be exhausting. These are necessarily more limited than neighborhood studies, because they are generally only varying conditions within a classroom. Other peers in the neighborhood, or often in the school, are not varied. The studies tend to find that having smarter peers around tends to improve outcomes, although one can improve outcomes even more by separating students by ability (tracking) and having teachers teach to the classes level. This is different from the sort of peer effects we’re concerned with, though — that there is an improvement in teaching efficiency is different from the effect of peers on learning. Some studies though, like Lavy, Paserman, and Schlosser (2008), do find that low achieving peers drag down the top. The percentage of students who had repeated a grade negatively affects test scores among Israeli students. Hoxby 2000, using administrative data in the 1990s from Texas, and employing a strategy which uses idiosyncratic variation in ability from year to year, finds that a 1 point increase in your peers reading scores increases a student’s reading scores by .15 to .4 points. Angrist and Lang (2004) study Metco, a Boston desegregation program, and find absolutely no effects, besides a small increase in the test scores of minority third graders. (So, probably nothing). These effects are often small, such that the benefits of tracking dominate. Duflo, Dupas, and Kremer (2011), using an RCT in Kenya, show that while having higher achieving peers makes students achieve more, separating classes by ability makes teachers far more effective. Hoxby’s estimates are on the high-end of the literature, Angrist on the low end. There are at least a dozen more studies on similar lines, with estimates scattered between the two. There has been no satisfactory meta-analysis, unfortunately, likely due to the heterogeneity of designs and things being estimated. Meta-analyses are ideal when aggregating randomized controlled trials which have near identical designs, but are econometrically unsatisfactory when we aren’t sure we’re actually estimating the same thing at all. Most do find positive results, however.
Other studies have used random assignment of roommates, starting with Sacerdote 2001, who used Dartmouth undergrads. He found that roommates had a small, positive effect on GPA — those paired with a smarter roommate had a higher GPA. Choice of fraternity or sorority also positively covaries with that of your roommate, and unlike GPA, the decisions of your whole dorm also affect frat choice. Another study, Carrell, Fullerton, and West, (2009), used the random assignment of students in the Air Force Academy to peer groups of 30, and found very large effects — a 100 point increase in average group SAT score raised GPA by .45 points (out of 4!). These effects are biggest in freshman year, and fade in later years, which makes sense, as the restrictions on interaction outside of the group are only for the first year.
We must be cautious in interpreting these studies, however. Everyone in the Sacerdote study, for example, has already passed the bar of making it into Dartmouth. Much as how your performance on the SAT soundly predicts how well you will do in college, but is very bad at predicting how well someone will do in the particular college they go to, peer effects might not generalize to other cases. This is why the Moving to Opportunity studies are particularly valuable in assessing the effects of environment on the most important areas of interest. I shall end the discussion of studies pertaining to education here; those wishing to read in more depth are advised to consult Sacerdote’s 2011 chapter on peer effects in education. I also want to mention Andres Barrios-Fernandez’s superb summary article on “Peer Effects in Education”, which was invaluable as a guide to the literature.
There are several conclusions I think we can take from this. First, while outcomes can be driven by environment, we will have a very hard time changing people’s environment for the better. Remember that in heritability, environment is a residual — nothing but an error term. It is everything which cannot be explained by genes, and while inclusive of the following ideas, does not refer purely to parenting, culture, peers, natural environment, or anything else. It is all of it, and many other things besides. There may be no combination of policies which will remove the correlation of social status and income from generation to generation — or at the very least, none that we would regard as conscionable. There is a tradeoff between freedom and preventing poverty. In the studies where poor or lower-achieving people have less freedom to choose who they associate with, they do better than in those where the intervention is not so drastic, and they can slip into bad habits. I believe that this is due to the people around them, and not due to institutional quality.
Second, the case for immigration is better. Poor countries are full of poor people. It is reasonable to think that they tear each other down, just as poor people do in America. Many people worry that if we allow immigrants to the United States, they will make this country like the country they left, but I don’t think that’s so. People converge to the income of the neighborhood they moved to. We must combine support for immigration with a harsher view towards the worst among us. Disruptive students drag the whole class down, and a very small fraction of people are responsible for most crime. We should not be hamstrung by moral equivalency — we must not let some people make life worse for others.
Thirdly, since we find ourselves driven by concerns of fairness and inequality, we may need to focus on welfare as a palliative. Perhaps some poverty is due to poverty traps, where subsidies can get people out of the hole. Whatever there is, it isn’t much. We should expect to care for people for life, not once in childhood and then never again. The primary reason people are poor is who they are. They cannot help this, and I have no doubt that everyone would choose to be born smarter, but it cannot be much helped after someone is born. Just as we believe that someone with Down syndrome or severe cerebral palsy should be supported by the state, so too should the state support those who are unable to cut it in society.
Fourthly, you should be a bit more skeptical of folk wisdom. Things can be obvious, and then not hold up to empirical investigation. An important role for economics should be to examine the bases on which governments and people evaluate policy, and checking whether they’re actually right. The literature on peer effects has been well-done and comprehensive, and shows pretty clearly that while there are likely effects, they are small. Do not place your hopes on critical ages and thresholds — there are no easy ways out.