4

The Register reports, somewhat skeptically, that an american consultant has produced an analysis claiming that climate change will lead to an increase in crime.

The paper, published in the Journal of Environmental Economics and Management (original article paywalled) claims that:

Between 2010 and 2099, climate change will cause an additional 22,000 murders, 180,000 cases of rape, 1.2 million aggravated assaults, 2.3 million simple assaults, 260,000 robberies, 1.3 million burglaries, 2.2 million cases of larceny, and 580,000 cases of vehicle theft in the United States.

Is this analysis remotely credible, or is this just another example of someone leaping on the Climate Change bandwagon?

matt_black
  • 56,186
  • 16
  • 175
  • 373
  • Someone with access to the pay-walled article is going to have to go through the methods since that is what will likely what will be needed in a good answer. – rjzii Feb 25 '14 at 22:27
  • You've presented a false dichotomy: "credible analysis" or "somebody leaping on the bandwagon"... what about a third option of them just being wrong. –  Feb 26 '14 at 00:42
  • Piracy is, however, [down](http://www.venganza.org/about/open-letter/). – Oddthinking Feb 26 '14 at 01:51
  • @Articuno you are right there are three possibilities, but I'd group "good but wrong" in the credible analysis bucket. I suppose I'm being opinionated and pejorative, but it looks like too much very poor quality science is gaining attention *because it is associated with climate change*. I think it is important to highlight this before it undermines science itself. – matt_black Feb 26 '14 at 10:14
  • As others have said, correlation is not causation, but if you pick two variables that both have an trend over some period then the two variables would be correlated whether the trends have the same cause or not. The best guide to whether the correlation is meaningful is to ask if there is a plausible explanation for the correlation that also explains the strength of the observed effect. In this case, I suspect both rising crime and temperature are affected by economic developments, rather than temperature being the cause of crime. –  Feb 26 '14 at 11:07

1 Answers1

12

Correlation doesn't imply causation.

The author of that paper, no matter how convoluted he is trying it to appear, basically just assumes that there are no other parameters in crime rate per capita than temperature and precipitation. Also the model is linear and lacks any proper validation. The assumption that relation between temperature changes within a year and crime rates can be interpolated and imposed on general climate change pattern is completely baseless.

I document a striking relationship between monthly weather patterns and crime rates. Across a variety of offences, higher temperatures cause more crime.

For it to be proven, one must show that the relation between temperature and crime is stronger than between time of the year and crime (as well as region and crime - yes, south regions of US are more criminal and warmer, but there is no proof that one is related to another), that is the difference is outside the standard error (at least). The uncertainty calculation is omitted from paper and it is not clear if the resulting numbers are artefacts or statistically significant results. Any sort of cross-validation is also absent.

The methodology of data selection looks like cherry-picking as well (removing outliers is quite a controversial topic in statistics, and this probably is not the worst example of it, but if someone removes data from sample usually it means it doesn't support the theory).

I take several final steps to clean the crime data. First, I drop all county-by-year records in which U.S. Census estimates indicate that the county had a population of fewer than 1,000 persons. Second, I drop all county-by-year records in which zero crimes were reported in all months, or in which weather data are missing for at least one month. Third, I eliminate outliers (almost all of which appear to be reporting errors) by dropping county-by-year observations in which the crime rate in any month is greater than twice the value of the 99th percentile crime rate for the entire sample. Finally, to minimize problems with heteroskedasticity in the data, I drop counties in which the mean crime rate is above the 99th percentile or below the 1st percentile for the entire sample.

Moreover, crime rate plots by year (presented by the author) clearly show that the crime rate is going down since 90's while climate is slowly warming up. This is in direct contradiction with the conclusions of the paper.

enter image description here

All in all, while the general idea of the claim under investigation seems plausible, the methodology of this paper is far from drawing any robust conclusions. It lacks a few important steps which should be included in building any statistical forecasting model (fundamental model justification, variable selection, dealing with collinearity issues, cross-validation, uncertainty handling, etc.).

sashkello
  • 3,762
  • 3
  • 28
  • 37
  • Perhaps you could quote parts of the paper that suffer from the issues you outline. e.g. you say "Correlation doesn't imply causation", but does the paper get that wrong? – Oddthinking Feb 25 '14 at 23:41
  • 3
    @Oddthinking There is no cross-validation present in the paper at all, so there is nothing to quote really... – sashkello Feb 26 '14 at 00:06
  • 2
    something else the paper doesn't take into account is whether in warmer regions there are more criminal laws on the book, i.e. whether it's easier to commit a crime there. And given the calling for making disagreement with their agenda a crime by many in the AGW industry, in a warming world (as they claim ours is) if their agenda were implemented a lot of people would be criminals, namely everyone who doesn't agree with them (which is a lot of people). – jwenting Feb 26 '14 at 07:51