What does evolution have to do with the cost of police in Switzerland? Probably not much.

Lucerne, Switzerland. Photo by Jamie McHale.

ResearchBlogging.orgSo a little while ago, I was perusing the latest from PLoS ONE while doing some low-attention-requiring lab work Monday afternoon, and a title caught my eye: “A test of evolutionary policing theory with data from human societies.” Oh, hey. That looked interesting.

The paper’s author, Rolf Kümmerli, claims to have found evidence for a particular kind of evolutionary model of cooperation in recent economic data from Switzerland. The problem Kümmerli addresses is a classic one: from the perspective of natural selection, individuals (apparently) have little evolutionary incentive to cooperate, unless they’re relatives. And yet, we see cooperation in human societies.

One solution to this quandary has been group-level selection, which is a whole ‘nother kettle of worms. Another is that policing behavior could evolve to help keep groups of less-closely-related people cooperative. Of course, modern human societies are a long way past the days when most of us lived in villages that were also basically big extended families. Kümmerli proposes that we might use some sort of rule of behavioral thumb to (unconsciously) assess how likely it is we’re interacting with close relatives—which is less likely in bigger communities, and communities with larger immigrant populations.

Er, what?

Kümmerli compiled data from the Swiss national government, comparing crime rates and police expenditures to the population size percentage of foreign nationals living in every Swiss canton, or administrative region. Rather than just use the raw population size or percentage of foreigners in each canton, he constructed an index that combines the two. And he did, indeed, find that as this index of “dissimilarity” increases, so do crime rates and expenditures on police.

Kümmerli concludes that his data support the “evolutionary policing theory.” But what has he actually shown? Crime happens for lots of reasons, not necessarily because people somehow “know” to behave more cooperatively in small towns. Most glaringly, Kümmerli’s data set includes no data on poverty, which seem like an obvious alternative explanation for the pattern—bigger communities with more immigrants also often have more poor people, and poverty is certainly related to crime rates.

Swiss police. Photo by Kecko.

Fortunately, the data Kümmerli uses, and many more variables, are all freely available online through the Swiss Statistical Encyclopedia. So I took a couple hours to play around with the raw numbers. I did all my statistical work in good old R.

For each of the 26 cantons, I compiled the number of reported crimes in 2009, the number of citizens (in thousands) in 2009, the percentage of foreign residents in 2009, the percentage of unemployed residents in 2010, annual expenditures on police in 2008, and—just for the heck of it—the percentage of commuters using public transit in 2000. As in Kümmerli’s data set, each statistic is the most recent value available. I didn’t try to replicate Kümmerli’s “dissimilarity” index because it’s not clearly explained in the paper; but I did log-transform the crime rate, the number of citizens, police expenditures, the unemployment rate, and the transit use rate to make them better conform to a normal distribution.

Here’s what the simple linear relationships among all those variables look like. Apologies for the complicated graphic, but this is a complex data set.

Linear relationships (upper triangle) and correlation coefficients (lower triangle) among variables from the Swiss Statistical Encyclopedia. Grapic by jby.

In the upper triangle of this matrix, you can see scatter plots with linear regression lines estimated from the data. Regression lines are colored according to statistical significance, corrected for multiple testing: red lines are “very” significant, orange just significant; grey lines indicate relationships no stronger than expected by chance. The bottom triangle gives the raw correlation coefficient between the variables, on a scale where 1 means a perfect relationship and 0 means no relationship.

What you should notice first is that top row of scatterplots, which show that crime rates have strong linear relationships with every other variable in the dataset, from population size to mass transit use. But that makes a certain amount of sense—all these variables are interrelated. Larger communities tend to attract more immigrants and tend to have better public transit systems that support more use. Communities with more unemployed people might have higher mass transit use, since cars are expensive. So, lots of correlation—but is there any causation in there?

There are a number of ways to tackle that question. A relatively easy one is to use multiple regression and a “model comparison” approach. This essentially builds a statistical model in which multiple variables—population, foreign residents, unemployment, mass transit use—are used to predict a single variable, crime rates. The procedure then compares the model’s AIC score, an index of the model’s ability to predict crime rates from the other variables, to models with each of the individual variables removed. If removing a variable makes a “significant” reduction in AIC—which is typically understood to be a difference of at least 2 AIC points, then that variable contributes significantly to predicting crime rates.

A Swiss public transit police car. Photo by Kecko.

It turns out that all the variables I considered make a significant difference in a multiple linear regression model trying to predict Swiss crime rates. But they aren’t equally important. Removing unemployment from consideration made a difference of 4.9 AIC points, removing the percentage of foreigners made a difference of 5.9, and removing the percentage of people using mass transit made a difference of 13.6. But removing the number of citizens made a difference of 92.9 points—an order of magnitude bigger difference than the other variables.

So it looks like the strongest pattern in Kümmerli’s data is just the effect of larger communities—they have more crime.

This is not what we scientists call a “surprise.”

Moreover, it’s not particularly informative for the purpose of the question Kümmerli sets out to answer—we don’t really know how population size actually relates to humans’ tendency to be less “cooperative,” or to need police to make them cooperative. Larger population does seem to be related to more crime, but it’s also related to more mass transit use—and mass transit use strikes me as a pretty cooperative behavior.

Admittedly, that’s a pretty off-the-cuff assessment based on a couple hours of fiddling around with simple statistical analysis of an easy-to-access public data set. But I strongly suspect that you could say exactly the same thing of Kümmerli’s paper. ◼


Kümmerli, R. (2011). A test of evolutionary policing theory with data from human societies. PLoS ONE, 6 (9) DOI: 10.1371/journal.pone.0024350