Pregnancy-Associated Homicide and Denominators, Denominators, Denominators, Denominators
Multiple papers purporting to show that pregnant and postpartum women are at elevated risk of homicide...do not.
Last month the New York Times ran a story called The Tragic Link Between Pregnancy and Homicide:
It's a startling statistic: In the United States, a woman's risk of being killed grows when she becomes pregnant and after giving birth — by about 20 percent, on average. This increase is entirely driven by the murders of America's youngest mothers. When women under age 25 get pregnant, their odds of death by homicide more than double.
This claim is attributed to the research of Maeve Wallace, an assistant professor of Public Health at the University of Arizona. Let’s take a look at her papers on the topic:
Pregnancy-associated homicide and suicide in 37 US states with enhanced pregnancy surveillance, American Journal of Obstetrics & Gynecology, 2016
Homicide During Pregnancy and the Postpartum Period in Louisiana, 2016-2017, JAMA Pediatrics, 2020
Homicide During Pregnancy and the Postpartum Period in the United States, 2018–2019, Obstetrics & Gynecology, 2021
Trends in Pregnancy-Associated Homicide, United States, 2020, American Journal of Public Health, 2022
For a fun challenge, stop here and go have a look at Trends in Pregnancy-Associated Homicide, United States, 2020 to see if you can spot the error that several licensed and bonded peer reviewers missed while reviewing these four papers before I break it down for you. I’ll wait.
…
Back already? How’d you do?
Okay, let’s dig in.
In all of these papers, pregnancy-associated homicide is defined as homicide in which the victim is either pregnant or up to one year postpartum, which I believe is standard in the field.
This was a retrospective analysis of the 2020 mortality file released by the National Center for Health Statistics (NCHS), which includes all death records issued in the United States. These data were restricted to female-sex-assigned-at-birth decedents of reproductive age (10–44 years), and cases of pregnancy-associated homicide were those with a manner of death indicating homicide or an International Classification of Diseases, 10th Revision (ICD-10; Geneva, Switzerland: World Health Organization; 1992) code for assault as underlying cause of death (X85–Y09), in addition to a pregnancy checkbox value indicating that the decedent was pregnant or within 1 year of the end of pregnancy at the time of her death.
The pregnancy-associated homicide rate was then estimated as follows:
I obtained data on counts of live births by year (including by maternal age and race/ethnicity) from the NCHS natality files and used them to estimate annual pregnancy-associated homicide rates (deaths per 100 000 live births).
The homicide rate among women who were not pregnant was estimated the usual way:
I computed homicide rates (deaths per 100 000 population) of the nonpregnant, nonpostpartum population of reproductive age by taking the count of females aged 10 to 44 years (data obtained from the US Census’ American Community Survey) minus counts of live births in each year.
Here’s where things go wrong:
There were 189 pregnancy-associated homicides identifiable in the 2020 mortality file. The 2020 pregnancy-associated homicide rate was 5.23 deaths per 100 000 live births, up from 3.30 in 2018 (n = 125) and 3.95 in 2019 (n = 148), the 2 previous years of available data.
…
In 2020, there were 3.87 homicides of nonpregnant, nonpostpartum women of reproductive age per 100 000 population. Risk of homicide victimization was 35% higher among pregnant and postpartum women compared with nonpregnant, nonpostpartum women of reproductive age (homicide rate ratio = 1.35; 95% confidence interval = 1.17, 1.57).
Anyone? Anyone? Ballmer?
Close, Mr. Ballmer, but not quite. The answer is denominators.
Homicide rates for a subset of the population are calculated as the number of homicides per 100k people in the population at risk, i.e. the subset of the population whose homicide rate is being calculated. “At risk” does not, in this sense, denote elevated risk, only that any of these individuals who are murdered will be counted in the numerator of the homicide rate.
To calculate the homicide rate for women of reproductive ages who are neither pregnant nor up to one year postpartum (abbreviated as PPP from here on), the authors divide the number of homicides in this age group with non-PPP victims by the population at risk, i.e. the population of women in this age group minus the number of live births. For reasons which we will see shortly, this slightly overestimates the population at risk, but it’s reasonably close.
To calculate the pregnancy-associated homicide risk for a given year, the authors divide the number of homicides deaths among PPP women in that year by the number of live births in that year. This is not a proper homicide homicide rate, though—the population at risk is at least 77% greater than the number of live births, and perhaps as high as double!
Why? To keep things simple, suppose that one woman gets pregnant each day, and every pregnancy is successfully carried to term. This results in 365 live births per year. On any given day, how many women are pregnant or up to one year postpartum? 280 (the average length of a pregnancy) plus the 365 women who gave birth on each of the last 365 days. With 365 live births per year, at any given time there are 645 PPP women.
That is, assuming a constant birth rate, the population at risk is at least 1.77 times the number of live births. Due to miscarriages, stillbirths, and abortions, the population at risk is actually even larger than this, perhaps up to twice the number of live births.
Note that the rate of homicides in PPP women per 100k live births can be mechanically increased or decreased by altering the definition of postpartum so as to alter the ratio of live births to population at risk. When counting only women who are pregnant or up to three months postpartum, the population at risk will be roughly equal to the number of live births (a bit more when considering abortions, miscarriages, and stillbirths), but the population at risk grows by an order of magnitude if we extend the postpartum period by nine years.
Surely this was corrected for?
There were 189 pregnancy-associated homicides identifiable in the 2020 mortality file. The 2020 pregnancy-associated homicide rate was 5.23 deaths per 100 000 live births
According to the CDC, there were 3.61 million live births in 2020, or 36.1 * 100k. 189 / 36.1 is 5.24, close enough to the 5.23 above to make it clear that no such adjustment was made.
In theory, this methodology inflates the pregnancy-associated homicide rate by up to 100%; if we take the mortality data at face value and correct for this, we find that homicide risk is actually substantially reduced in PPP women.
That said, we probably shouldn’t take the data at face value. Likely there’s substantial underreporting of pregnancy and postpartum status in homicide victims; perhaps police routinely test female murder victims for pregnancy to aid the investigation, but late postpartum status may be more difficult to identify, and even if this is known, it may not be consistently reported, as discussed in the NYT story:
According to Dr. Wallace, from 2018 through 2022, at least 837 women were killed during pregnancy or within the following year, although because the pregnancy checkbox is new and not always consistently used, that number is most likely an undercount.
Furthermore, this may be a bit of a red flag for data reliability:
Among all incidents, 81% involved firearms, 55% occurred in the home, and 54% of victims were pregnant at the time of their death whereas the remaining victims were up to 1 year postpartum.
A woman who carries an infant to term spends about 43% of the PPP term pregnant, but 54% of the victims were pregnant. This is likely explained by some combination of:
Homicide risk being greater during pregnancy than during the postpartum year.
Abortions and miscarriages leading to many women spending time pregnant but not postpartum.
Underreporting of postpartum status in homicide victims.
In Pregnancy-associated homicide and suicide in 37 US states with enhanced pregnancy surveillance, Wallace et al make an attempt to correct for undercounting and find an RR of 1.841 for PPP women, but if we adjust this to use the correct denominator, the RR is no longer statistically significantly different from 1.0.
What about this, though?
When women under age 25 get pregnant, their odds of death by homicide more than double.
If homicide risk is more than doubled in younger PPP women, that would survive correction for the incorrect denominator, right?
Maybe, but probably not. This claim seems to be based on table 2 of Homicide During Pregnancy and the Postpartum Period in the United States, 2018–2019. The largest RR for homicide risk of PPP vs. non-PPP women in is seen in the age 10-19 bracket: 6.67 (4.77–9.33). However, this is misleading, for a few reasons:
Homicide rates tend to be greatly elevated in communities with high rates of young pregnancy and single motherhood, so even if PPP women under the age of 25 are at higher risk of homicide than non-PPP women of the same age nationwide, this does not necessarily imply that getting pregnant causally increases young women’s risk of homicide victimization.
The number of pregnancy-associated homicides in this age bracket is small, and the CI is wide.
In the 10-19 age range, both pregnancy and homicide victimization are heavily concentrated in the later teen years, while the age distribution for non-PPP women and girls in the 10-19 range is much more uniform. Because of this, the homicide rate for non-PPP women and girls age 10-19 is around 1.5, less than half of the 4.1 rate for non-PPP women 20-34. This inflates the RR for this age group relative to what we would see with finer-grained age controls.
A far more sophisticated analysis would be required to detect a true elevation in homicide risk causally attributable to pregnancy in this age group.
Overall, I remain agnostic on the question of whether PPP status increases the risk of homicide victimization for women generally or for specific demographic subsets of women. The answer to this question is highly sensitive to assumptions about the degree of underreporting of PPP status for homicide victims, and as discussed above, we don’t really have solid data on this.
But regardless of the degree of underreporting of PPP status in death certificates, the methodology here is clearly flawed, and creates a spurious finding of greatly elevated homicide risk in pregnant and postpartum women.
I don’t want to be too hard on Wallace and co-authors personally—anyone can make a mistake—but the fact that four papers with the same basic math error got through peer review in four different journals is disappointing; I would have hoped that at least one reviewer would have caught this at some point.
The assumption of 53% misclassification is based on a 2011 paper by Isabelle Horon and Diana Cheng, which takes an interesting approach of first identifying known pregnancy-associated deaths and then checking the death certificates to see whether the appropriate boxes had been checked. They find that only 45% of pregnancy-associated homicide death certificates had the appropriate box checked; I’m not sure why Wallace et al used 53% misclassification instead of 55%. In any case, this is based on a total of 42 homicides in Maryland between 2001 and 2008, so the estimate is imprecise and may or may not generalize to other states and later years.