As I mentioned in my previous post, the recent Brennan Center report on the effect of incarceration on crime identified fourteen possible factors that could explain crime trends, but included only eight in their regressions. So I wanted to think a bit about how omitted variable bias might throw off their findings. Last post I focused on just one, the failure to control for trends in crack use, and suggested that its exclusion likely leads to report to understate the crime-reducing impact of incarceration.
In my next few posts (spreading these out over several as a concession to “wonky” + “long” = “unreadable”), I want to consider the remaining five variables that didn’t make the cut. I feel like four of them don’t really raise any concerns, but one—the adoption of CompStat—does.
First, the four that don’t matter so much. These are trends in inflation, consumer confidence, lead exposure, and abortion. In this post I’ll consider the first two, and I’ll look at lead and abortion in the next.
For inflation and consumer confidence, my guess is that both their direct effects on crime and their correlations with incarceration rates are weak. According to the Brennan report, the evidence linking inflation to crime comes only from national-level studies (since inflation is not gathered at the state or local level), which link it primarily to property crime. It is easy to see why one might think there is a connection between inflation and property crime, as the figure below suggests.
But this is the trouble with national-level data. While inflation and property crime track each other closely, especially during the 1960s and 1970s, the correlation is likely spurious. A lot was changing during that time, which was a period of great social and economic upheaval. Surely these huge forces were driving both inflation and crime. Especially with national-level data, which just gives you one time series to work with, it is easy to get overlapping patterns that are random or spurious. (See this for more spurious-correlation awesomeness.)
As for consumer confidence, it too is measured only at the national level (from a survey of only 500 households), and thus faces a greater risk of spurious correlation. Furthermore, the paper proposing this connection that the report cites uses only a handful of explanatory variables and thus likely suffers from OVB itself. It’s almost certain that its estimate of consumer confidence is picking up a lot of other stuff.
But there is an even deeper reason to be wary of linking consumer confidence overall to crime rates. As David Weisburd and others have shown, crime is intensely geographically concentrated, not just within a state, not just within a city, not just within neighborhoods, but within blocks of those neighborhoods: New York City is more violent than Westchester, Brooklyn is more violent than Staten Island, East New York (a high-crime Brooklyn neighborhood) is more violent than Park Slope (Brooklyn’s hatchery), and there are persistently “good blocks” and “bad blocks” in East New York.
Given this concentration, it is unlikely that national-level surveys of just a few households are really going to capture the nature of confidence where crime is most densely located. Perhaps not rigorous evidence, but I do remember an episode of the Chris Rock Show from 2000 in which Rock goes to the South Bronx to ask people there whether they are feeling the benefits of the dot-com economic boom; the answers are predictable (and the clip sadly not on YouTube, as far as I can tell). I would expect the true effect of confidence on crime to thus be slight.
Or, put more carefully, the consumers whose confidence we measure are systematically not the consumers who are either committing or experiencing crime. Not only would these consumers be unlikely to show up in a 500-person survey just by chance alone, but the very nature of their more-difficult lives suggests that they will be systematically under-sampled.
Finally, even if you think there is a strong relationship between either of these and crime, I’m hard pressed to see much of a connection between them and incarceration rates. No previous study has ever thought to look at them (at least as of the time I wrote this review), and it strikes me that any effect they do appear to have is more likely due to the underlying economic shifts driving inflation and confidence (like trends in personal income, overall state economic output, unemployment rates, and maybe inequality), all of which are actually easier to measure at the state level anyway.
So dropping inflation and consumer confidence these variables thus shouldn’t bias the report’s estimate of incarceration at all. Which is good.
Posted by John Pfaff on February 20, 2015 at 10:12 AM