Seriously? The "Largest Gun Study Ever: More Guns, More Murder"

I have gotten a number of emails about a study in the American Journal of Public Health entitled: "The Relationship Between Gun Ownership and Firearm Homicide Rates in the United States, 1981–2010." 

I don't have time to do more than glance at this paper, but here are some very superficial initial thoughts.

1) Using state level data the study claims a positive relationship between the percentage of suicides committed with guns (they call this the gun ownership rate rather than what it actually is) and the firearm homicide rate.  The big problem with their measure of gun ownership is that it picks up a lot of demographic information that may itself be related to homicide and to crime.  

2) Do we care about total murders or murders involving guns?

3) "None of the existing panel studies examined data more recent than 1999."  Presumably this is what is causing some left wing outlets to claim "Largest Gun Study Ever" (at last glance the link to that article was retweeted 1,077 times).  The authors seem completely unaware of the third edition of More Guns, Less Crime that looked at data up through 2005 -- six years longer than they claim.  Of course, my research also started with 1977, not 1981 as they did.  Of course, I have also used county and city level data and have many more observations than they have.  My research has run regressions with up to 96 times more observations that the 1,000 that they point to in this paper.  While I account for hundreds of factors, these guys account for almost none (6 in their final reported model (23 unreported in bivariate estimates -- meaning just running one of these variables at a time in explaining firearm murder rates).   It would be nice if Mr. Zack Beauchamp was notified that these authors are apparently unaware of any of my research since "1988" [sic] (they couldn't even get the year right for my first edition of MGLC).

4) No explanation is offered for why they leave Washington, DC out of their regressions.  I can offer one: it weakens their results.

5) Only a very small percentage of the prison population are there for murder.  Possibly a percent or two in any given year.  Do changes in the share of the prison population for larceny or burglary really help explain a lot of the variation in murder rates?  A more direct measure would be the arrest rate for murder and/or the number of people in prison for murder and/or the death penalty execution rate.

6) "To develop a final, more parsimonious model, we first entered all variables found to be significant in bivariate analyses (we used a Wald test at a significance level of .10) into 1 model. We then deleted variables found not to be significant in the presence of the other variables, assessing the significance of each variable with a Wald test at a significance level of .05."  -- The problem here is that the resulting statistical significance levels don't mean what these authors seem to think that that do.  The levels of significance for a regression assume a random draw.  If you 23 specifications and then pick the variables that are significant, the variables that you are picking were picked in a biased manner.

7) Six variables is what they finally include in their "Final Model."  Leaving out variables that affect the murder rate will cause the other variables to act as a proxy for these left out variables.  This gets back to my point (1).

8) Even if all these issues were dealt with, they have completely ignored the issue of causation.  Is it increased crime that results in more guns or the reverse?

Breastfeeding and intelligence: the problem with the newest research

Do I believe the end result?  Probably.  But I don't believe that this paper comes close to proving its case.  Fox News has this discussion of a recent study in JAMA Pediatrics.

. . . Belfort and her colleagues gathered data from 1,312 mothers and children in the United States, tracking everything from the mother’s frequency of breastfeeding to other factors including the mother’s intelligence, the mother and father’s education levels, measures of the home environment, the mother’s employment and the type of childcare the baby received.
While the link between breastfeeding and cognition had been previously explored, many earlier studies did not control for these additional factors.
“Many previous studies have been criticized because any link you might observe between breastfeeding and childhood intelligence could be explained by those other factors,” Belfort told FoxNews.com.
Belfort then performed a series of tests measuring the children’s cognitive development after infancy.  At age 3, the children underwent the higher Peabody Picture Vocabulary Test, a measure of receptive language, or how well a child understands language.
“A child’s receptive language is highly correlated with general intelligence as measured by more typical IQ tests,” Belfort said. . . . 

Unfortunately, the discussion at Fox News is really is too uncritical.  Of course, the bigger blame should be placed on JAMA Pediatrics   As with almost all medical studies, they ignore the issue of endogeniety when they can't do randomized experiments.  If I were do the experiment, I would control for family specific effects (that is a dummy variable for each family to pick up the average intelligence of children born into that family) and, after accounting for birth order, spacing between children, and the other controls included in the study, see if intelligence of children within the family varied with the amount of breastfeeding.  The problem with the current purely cross-sectional study is that the measurement error in how education and other factors are measured is large compared the differences they claim to find in intelligence.