Episode 18: Maya Rossin-Slater
Maya Rossin-Slater
Maya Rossin-Slater is an Assistant Professor of Health Research and Policy at Stanford University’s School of Medicine.
Date: December 10, 2019
Bonus segment on Professor Rossin-Slater’s career path and life as a researcher.
A transcript of this episode is available here.
Episode Details:
In this episode, we discuss Professor Rossin-Slater's work on the effects of violent assault during pregnancy on birth outcomes and infant health:
"Violence While in Utero: The Impact of Assaults During Pregnancy on Birth Outcomes" by Janet Currie, Michael Mueller-Smith, and Maya Rossin-Slater.
OTHER RESEARCH WE DISCUSS IN THIS EPISODE:
"Poverty, violence, and health the impact of domestic violence during pregnancy on newborn health" by Anna Aizer.
"Birth outcomes for Arabic-named women in California before and after September 11" by Diane S. Lauderdale.
"Stress and birth weight: evidence from terrorist attacks" by Adriana Camacho.
"Armed conflict and birth weight: Evidence from the al-Aqsa Intifada" by Hani Mansour and Daniel I. Rees.
"Prenatal exposure to violence and birth weight in Mexico: Selectivity, exposure, and behavioral responses" Florencia Torche and Andrés Villarreal.
"The hidden costs of war: Exposure to armed conflict and birth outcomes" by Florencia Torche and Uri Shwed.
"Scaring or scarring? Labour market effects of criminal victimisation" by Anna Bindler and Nadine Ketel.
Transcript of this episode:
Jennifer [00:00:00] Hello, Probable Causation listeners. Before we get started today, I want to encourage you to support the show on Patreon. For just five or ten dollars per month, you get access to exclusive bonus content, such as interviews with book authors hosted by David Eil and bonus segments with the scholars I interview on the show, talking about their background and life as a researcher. Plus, you'll know that your contributions help keep the show running, something for which the entire team is grateful. To subscribe go to Patreon.com/probablecausation. There's also a link on our website. Thank you in advance for your support. Now on to today's show.
Jennifer [00:00:43] Hello and welcome to Probable Causation, a show about law, economics, and crime. I'm your host, Jennifer Doleac of Texas A&M University, where I'm an Economics Professor and the Director of the Justice Tech Lab.
Jennifer [00:00:53] My guest this week is Maya Rossin-Slater. Maya is an assistant professor of health research and policy at Stanford University School of Medicine. Maya welcome to the show.
Maya [00:01:03] Thank you so much for having me on the podcast.
Jennifer [00:01:06] We are going to talk today about your work on the effects of violent assault, particularly intimate partner violence or domestic violence during pregnancy on birth outcomes and infant health. Before we dive into all of that, though, could you tell us about your research expertize and how you became interested in this topic?
Maya [00:01:24] Sure. So I'm an economist with research in areas of health, public and labor economics. I'm particularly interested in understanding the choices and constraints faced by families with children and how various public policies and other factors can influence their well-being, especially within disadvantaged populations in the United States as well as in other developed countries. A lot of my research has focused on identifying the impacts of adverse shocks as well as public policies during that in utero and early childhood periods. I think this work is important in part for understanding the persistence of inequality in America and identifying effective policies for diminishing the persistence and potentially improving the life chances of poor children. By now, there's a wealth of evidence documenting the lasting impacts of the early life environment on individuals outcomes throughout life and even into the next generation.
Maya [00:02:21] So if poor children happen to be born in environments where they're disproportionately exposed to adverse shocks in early life, and if those shocks themselves translate into worse outcomes for them in adulthood, then we can see how this can be a powerful mechanism for transmitting inequality from the fetal period to birth throughout childhood into adulthood and even into the next generation. So this suggests that public policies that can break the cycle can be potentially highly effective and important. So I became interested in the topic of domestic violence specifically and its impact on the early life period, in part just because of the alarming prevalence of this type of violence both around the world, but also here in the United States. And because just like many other adverse shocks that I've studied, poor and minority women are much more likely to experience this type of violence than their more advantaged counterparts. So while there's been a lot of work documenting the early life impacts of various factors from outside of the family environment, so things like air pollution or even say crime or violence in the local community or the neighborhood, we know much less about the influence of intra family dynamics. This is where I think this paper fits in. I'm really honored to have fantastic co-authors Janet Curry, sort of one of the leading experts and kind of one of the founders of the research on early childhood health and wellbeing within economics.
Maya [00:03:50] And then Mike Mueller Smith, who is an expert in the field of economics and crime. So we've teamed up for this project.
Jennifer [00:03:58] Yes, so your paper is titled "Violence While in Utero: The Impact of Assaults During Pregnancy on Birth Outcomes." As you mentioned, it's co-authored with Janet Curry and Mike Mueller-Smith. So give us some more background on intimate partner violence, which I'll probably refer to as IPV, a little bit less of a mouthful there. And violence during pregnancy in general. How prevalent are these types of assault?
Maya [00:04:20] Right, so intimate partner violence, or IPV, as you say, is a crime that involves physical, sexual or psychological harm by a current or former partner or spouse, and it's actually quite prevalent in our society. So estimates suggest that in the US, about 32 percent of women experience physical, intimate partner violence at some point during their lifetime. Another statistic is that IPV accounts for over one seventh of all of violent crime in the United States today. Pregnancy and the postpartum period elevate the risk of experiencing domestic violence for women. For instance, studies indicated to indicate that the prevalence rate of physical or sexual abuse among pregnant and postpartum women ranges from seven to twenty three percent, with more recent studies reporting higher rates. Since domestic violence is notoriously underreported, these numbers likely represent lower bounds on the true prevalence rate. One last disturbing statistic on this topic is the IPV related homicide is actually one of the leading causes of death during pregnancy for women.
Jennifer [00:05:26] So what mechanism should we have in mind for why IPV during pregnancy might affect the child's outcomes?
Maya [00:05:33] So, I think there's at least there's several mechanisms that are potentially relevant. So first, there's the direct physical channel. So blunt trauma to the woman's abdomen could result in all kinds of complications issues. It could result in placental abruption, for example, which could lead to early onset of labor. There could be other complications that are dangerous for both the woman and her unborn child, such as the rupture of the woman's uterus. Then there are more indirect channels. We know, for instance, from other work that stress during pregnancy can harm the fetus because the woman's body essentially releases too much cortisol that affects fetal development, experiencing violence and more generally, just being in a violent relationship is probably a fairly stressful experience, so that could be another important channel. Then the woman might also respond to the stress of violence by taking up unhealthy behaviors such as smoking or drinking alcohol during pregnancy, which in turn have their own independent negative effects on the baby's health. A last factor is that a controlling partner might restrict where the woman can go, which could affect her ability to receive various services, including prompt and regular prenatal care. This could, for instance, affect the likelihood of complications during pregnancy, get detected or get detected early enough and again could lead to harm to the unborn child.
Jennifer [00:06:59] So before this study, what had we known about the effects of IPV and violence during pregnancy?
Maya [00:07:06] So, I would say before the study, a lot of the literature on the topic had been in the field of public health. What these papers typically do is they take survey data on self reported IPV prevalence and information on infant health and show that women who experience IPV during pregnancy tend to have worse birth outcomes than women without IPV exposure. Basically these studies compare women who do and do not report experience active during pregnancy, and then they try to account for other differences between these two groups by including some sort of standard controls for things like education, race, age. But, you know, as we know, it is quite hard to interpret these estimates as causal because exposure to HPV during pregnancy is by no means random. It's nowhere near like a randomized control trial, which we would use to estimate causal effects. So, for instance, poor minority and otherwise disadvantaged women are much more likely to experience IPV than their more advantaged counterparts. Then these same groups are also more likely to experience poor infant health outcomes. So it's very hard to disentangle the causal effect of IPV from the influences of all of these other factors that are different between women who do and do not experience IPV during pregnancy, even in models that include controls for some basic observable characteristics. The issue here is that we really worry about unobservable characteristics that could also be different. So I should say that one important innovation on this literature is a paper by Anna Iser from 2011. So what she did is to use data on hospitalizations for assault during pregnancy, which are overwhelmingly due to intimate partner violence. She identified the effects of this type of hospitalization on birth outcomes using an instrumental variable strategy. So her basic strategy was to essentially exploit geographic and temporal variation in the enforcement of laws against domestic violence in the state of California. Then she used that as an instrument for her being hospitalized for assault during pregnancy. She found a fairly large negative effect of hospitalization on birth weight. So we essentially wanted to build on this work, but we use both different data as well as a different identification strategy. I should say there's also a literature on exposure to neighborhood violence or even more global events that involve violence like wars and terrorism on infant health. One major difference between that research and what we do in this paper is that those studies typically measure potential exposure rather than actual victimization. They usually focus on maternal stress during pregnancy as the key channel through which that potential exposure to violence can affect infant health.
Maya [00:09:55] So our paper instead focuses on identifying the direct consequence of violent crime on the victims themselves as well as their unborn children.
Jennifer [00:10:04] All right. So let's talk more about the empirical challenges to studying this issue. As you've alluded to, there are many.
Jennifer [00:10:10] So what did you see as the primary hurdles to measuring the causal effects of IPV during pregnancy that you all had to overcome in order to do the study?
Maya [00:10:19] Right. As we just discussed, I think the lack of random variation and IPV exposure presents a major identification challenge. We can't run a randomized experiment in the setting. I's really critical that we find a research design that is able to tell us something causal rather than just a correlation. Then on top of that, there's a data limitation. Broadly speaking, we have a lot more detailed data on individuals who are accused of committing crimes than we do on individuals who are victims of crimes. This makes a lot of sense, of course, because we're concerned about protecting the confidentiality of victims. But it also presents a challenge for research. There's some survey data sets that do ask people about victimization, but these likely suffer from at least some measurement error and potentially bias because people tend to underreport crimes and especially domestic violence. The types of people who self report experiencing IPV are likely different from those who experience it and then choose not to report it. Both of these issues, I think, pose some challenges for estimating the causal effects of exposure to IPV during pregnancy.
Jennifer [00:11:30] Okay, so speaking of data, you use data from New York City to tackle this issue, and it's really an amazing data set that you and your colleagues put together. So tell us about that data and I guess a little bit of a back story on how it all came to be and then yeah, what the data look like.
Maya [00:11:46] Right. So I first of all, I think that my co-author, Mike Mueller Smith, deserves a lot of the credit for putting together the amazing data that we have at our disposal. I know you had him as a guest on your podcast.
Maya [00:11:59] So listeners have already heard about the amazing things he's doing with CJARS, his center for linked administrative data that he's built, I mean, Mike just really has a talent for getting various government agencies to agree to share their data with researchers.
Maya [00:12:19] He was really critical in getting us access to that data set and we really benefited from that. So what we did is we merged essentially three administrative data sets from New York City. So first, we have the universe of reported crimes from the New York Police Department, the NYPD, that concludes all criminal complaints reported between 2004 and 2012. Okay, so this is not arrest, this is not incarceration, this is all reports of criminal complaint. So basically, any time anybody calls the police, that's in our data. These data, importantly for us, have exact longitude and latitude coordinates of where the criminal event allegedly occurred. This is, of course, important for linkage. Our analysis focuses on all reported misdemeanor and felony assaults in New York City over this time period. The second data set that we have is data on the universe of birth records from New York City over the same time period. So we got the restricted version of these data, which means that in addition to kind of standard birth outcomes and pregnancy information, we have information on mother's full maiden names and their dates of birth, as well as their residential addresses that they report at the time of birth. So this enables us to match siblings born to the same mother using her name and date of birth and other characteristics. Then we're also able to merge the crime data to the mother's using their exact addresses. As I mentioned, we have all kinds of information on pregnancy and birth outcomes. So we generate all of our outcome measures using these data. Then we have some demographic information on both the mothers and fathers. Then finally, we have a data set on building characteristics from the NYC Department of City Planning. So these data are important because they allow us to figure out whether a mother's residence is in a single family home or in an apartment in a multifamily home or like a large apartment building in New York. The reason that this is important is because the crime data where we have the exact longitude and latitude coordinates, we essentially know the building in which a crime occurred, but we wouldn't know the exact apartment number if the crime occurred in, say, a large apartment building. So because we want to be able to link mothers to crimes that occur in their homes, specifically assaults that occur in their homes, our main analysis focuses on mothers residing in single family homes where we can be more sure that a reported assault actually occurred at their residence.
Jennifer [00:14:58] Okay, so you use these data to identify women who experienced a violent assault while they were pregnant. But of course, in order to measure the causal effect of that assault, you need a control group, women who look very similar but who did not experience such an assault during pregnancy. This is not easy. And you approach this issue in not one, but three different ways. Tell us about your three empirical strategies and what you see as the benefits of each one.
Maya [00:15:25] Right. So as we discussed earlier, identification is a central challenge for this project, so we use three different strategies, each relies on a different set of assumptions and each is going to have, of course, its limitations. But we think that as a whole, they can paint a pretty convincing picture. Our first strategy is to compare women who experience an assault in their home during pregnancy to those who experience an assault in their home in the nine months following their expected due date. Importantly, we do not compare women who do and do not ever experience IPV.
Maya [00:16:02] This is essentially what's done in the prior literature in most of the public health studies that I was mentioning. Instead, our strategy includes women who experience an assault in the months surrounding childbirth and then we leverage variation the exact timing of when this assault occurs relative to the child's expected due date. So then our design relies on an assumption that the time varying determinants of IPV during pregnancy and in the postpartum period are similar, which is actually quite consistent with the existing literature on this topic. Of course, we also do quite a lot of things to try to convince our readers that this assumption is likely to hold. For instance, we show that the characteristics of parents across the two groups, so the group where the woman experiences an assault in her home during pregnancy and the group where the woman experiences an assault in the nine months after pregnancy, those two groups are similar. Both in our administrative data we have a lot of information on things like education level, age, foreign born status, marital status. But then we also use an additional survey data set called the Fragile Families and Child Wellbeing Survey, which over samples sort of relatively low income, disadvantaged families and has a lot more detailed information on a variety of family dynamics, including those related to IPV. We show again that those characteristics are similar in that data set as well. Then finally, we implement a bounding exercise that was proposed in a recent paper by Emily Oster to show that our results from this American strategy hold up even when we account for reasonable degrees of bias due to potential selection on unobservable characteristics that we can't that we can't observe in either our administrative data or in the additional survey data.
Maya [00:17:58] So then our second strategy is to build on the first design, but implement sort of a difference and difference type model. Where essentially we compare the difference between mothers to experience and assault during pregnancy and those who have one in the months after pregnancy, with the analogous difference for mothers who report any type of crime during those two time periods. So this strategy allows for there to be a difference in the reporting rate between women who experience an assault during pregnancy and those who experience an assault post-partum, which is a possible concern for our first strategy. But it assumes that this difference is similar to the difference in reporting rates for other types of crimes. Then the other thing is that the strategy allows us to shed light on the differential effects of assault during pregnancy relative to experiencing other types of crimes that are likely to be stressful, but may not involve direct physical harm to the woman, for instance, like burglaries. Then finally, our third strategy is quite different. It compares siblings born to the same mother where she experienced an assault during one pregnancy and not during the other. By comparing children born to the same mother, the strategy essentially controls for all time invariant, observable and unobservable differences across women who do and do not experience IPV during pregnancy. Then the identifying assumption for the strategy is that time varying determinants of IPV within the same mother are uncorrelated with her children's birth outcomes. Now, of course, this assumption is inherently untestable, but we do have a nice placebo check at our disposal that provides us with a with an indirect test. So what we do is we take mothers who experience an assault in the months after pregnancy. We remove those and who actually experience an assault during pregnancy, like our main treatment group, and then we estimate our same model with maternal fixed effect. If there were some kind of time varying shocks that determine both the timing of IPV in the months surrounding pregnancy and infant health, then we would potentially see placebo effects of exposure to assault after pregnancy on the birth outcomes of that same pregnancy. Reassuringly, we find no evidence of that, suggesting that the identifying assumptions in the maternal fixed effects model are likely to hold.
Jennifer [00:20:23] Yeah, I really like that strategy in the placebo test, I agree it's really compelling. You argue in the paper that the first two strategies where you're essentially comparing outcomes for kids where the mother was assaulted before or after pregnancy will be biased toward finding null effects. So talk us through that. Why should we expect those measured effects to be smaller than the true causal effect?
Maya [00:20:46] Right. So in the first two strategies, remember, the control group includes women who have a reported assault in the months after pregnancy. But we know from a large sociological literature that domestic assaults where the police are called are rarely one off events.
Maya [00:21:05] So in most cases, there's a continuous pattern of abuse before, during, after pregnancy that may, from time to time, culminate in a more serious assault in which law enforcement actually gets involved. So it is likely that all of the women in our sample of analysis in this research design are likely to be in violent relationships and therefore potentially subject to high levels of stress. So essentially, by comparing these two groups, our estimates allow us to capture the effect of the more serious assault itself, the one that leads to the police getting involved. But these models are, ill-equipped, to estimate the effects of the chronic stress from being in a violent relationship, since again, both the treatment and control group in these strategies are likely exposed to such stress. So if we thought that in order to estimate a true causal effect of IPV, we would need to capture both the direct effect of the serious assault that gets law enforcement involved as well as the associated chronic stress, then this means that our results from these two strategies are going to likely underestimate it. But note that in our third strategy, where we compare siblings born to the same mother that does not have this issue. So here we essentially compare women who are subject to IPV during pregnancy to themselves at another point in time where they may have already left or even not yet entered the violent relationship. So we should expect to find larger effect magnitudes when we use the maternal fixed effects strategy than when we use our first two strategies. And that's exactly what we do.
Jennifer [00:22:43] So you alluded earlier to trying to, you know, restrict your sample to to really pinpoint the women who were assaulted in their home. Tell us more about the sample restrictions that you make of using your full New York City data set and what the final sample looks like. And the question I have in the back of my mind here is how representative this group is when compared to all women who might experience IPV.
Maya [00:23:07] Yes, that's an important question. So as I mentioned before, the key restriction is that we focus on women who reside in single family homes because that's where we can be more sure that we actually measure an assault occurring in the woman's residence. Then we also limit our sample to women residing in the Bronx, Brooklyn and Queens boroughs of New York City. So we drop Manhattan just because Manhattan has relatively few single family homes. And out of those, relatively very few women actually experience any assaults. So there's just very few observations in Manhattan. Then we also drop Staten Island, it's also relatively small compared to the other boroughs. It just turns out that the demographic characteristics of mothers in Staten Island are much less comparable to those of women in the other boroughs. So we drop them essentially to have a more homogeneous sample. But of course, the cost of making these restrictions is that our sample becomes not less representative of the average women who experiences IPV in the United States or even in New York City. So, for instance, women and single family homes are likely on average to have slightly higher income and just generally be more advantaged than women who reside in large apartment complexes in New York City. You know, our estimates are going to be representative for the group of women that we have in our sample, but can't necessarily speak to what the effects would be for an average woman who is subject to IPV. That said, we do a robustness check where we estimate our main regression models using a sample of women in multifamily homes, although we do still restricted to kind of relatively small multifamily homes just again, because we want to reduce the measurement error associated with assigning assault to occurring in the woman's residence. Our results are actually similar but more muted, which is consistent with the idea that there's more measurement error when we use multi-family homes than when we use single family homes.
Jennifer [00:25:12] You have a whole bunch of outcome measures that are potentially interesting here, both the upside and downside of the rich data set you've constructed. So you combine those various measures into four outcome indices: a severe birth outcomes index, a broad birth outcomes index, a use of medical services index, and a maternal behavioral and wellbeing index. So walk us through this. Some folks out there might not have seen this done before. So first, tell us why you use these indices instead of just looking at the individual outcome measures? And second, what does each of these indices contain?
Maya [00:25:47] Sure. So the reason that we group our outcomes into outcome indices is to mitigate concerns about multiple hypothesis testing. When you have lots of possible outcomes that you could analyze, then by chance you might detect that five percent of your models will yield a statistically significant coefficient. So this means that you might think you have estimated a true effect of exposure to assault during pregnancy, but really, it's just because you ran so many regression models that you were bound to find something statistically significant in a few of them.
Maya [00:26:17] In order to address this concern, which I think is a really important one, especially when you use large data sets that have a wealth of outcomes that you could look at, we create for outcome indices, as you mentioned, and this essentially just reduces substantially the number of regression models that we can run. So the indices are constructed in a fairly simple way. We essentially take each outcome we oriented so that they all go in the same direction. So, for instance, for the severe birth outcomes index, we orient things such as such that a higher value means a worse outcome. So like very low birth weight is one for very low birth and zero otherwise, so a higher value means the worse outcome. Then we standardize it by subtracting the control group mean and dividing by the control group standard deviation. And then we just take a simple average of all the standardized outcomes within each index group. So, as you said, we have four of these groups with the severe birth outcomes index includes indicators for very low birth weight, so that's less than fifteen hundred grams, very preterm births, so that's less than 34 weeks gestation, a low one minute APGAR score, that's less than seven and APGAR scores, something that's given by the doctor one minute after birth, NICU admission, admission to the neonatal intensive care unit of the child, an indicator for any abnormal conditions of the child that includes things like the use of assisted ventilation and an indicator for any congenital anomalies of the newborn, and then finally an indicator of whether a death has occurred of the child by the time the birth certificate is filed with the Vital Stats office. Our second index is the Broad Birth Outcomes Index. So this is going to include all of the variables I just mentioned within the severe birth outcomes index and then we also include less severe measures, continuous birth weight in grams, an indicator for low birth weight that's less than twenty five hundred grams, gestation continuous gestation in weeks, and an indicator for preterm births that's less than 37 weeks. Our use of Medical Services Index includes an indicator for first trimester prenatal care initiation, the total number of prenatal care visits, an indicator for induction of labor, an indicator for delivery by C-section, an indicator for any complications during labor delivery. Then finally, our Maternal Behavioral and Wellbeing Index includes indicators for a variety of behavioral channels, things like smoking during pregnancy, using illicit drugs during pregnancy, the mother self reporting being depressed, having too low pregnancy weight gains, that's less than fifteen pounds or too high pregnancy weight gain, that's more than forty pounds. Then an indicator for their mother not receiving WIC benefits, which is the Women, Infants and Children Supplemental Nutrition Program benefits.
Jennifer [00:29:16] That information on the mother and her behaviors that's all on like birth certificate data.
Maya [00:29:22] That's right, yes. So all of that information comes from the birth records data.
Jennifer [00:29:25] It's amazing. Okay, so let's dive into the results for each of your three empirical strategies. What do you find are the results of assault during pregnancy on your outcomes of interest?
Maya [00:29:37] So all three of our strategies consistently demonstrate that experiencing assault during pregnancy has adverse effects on infant health. So first, let's focus on the relatively conservative estimates from the pregnancy versus postpartum assault exposure and the difference in difference models, which actually yield remarkably similar results. So there we find that mothers with assault during pregnancy have a point zero eight standard deviation higher summary index of severe birth outcomes, severe adverse health outcomes compared to mothers who report an assault in the postpartum period, as well as mothers who experience any other crime during either period. Within this index, that result is driven by one point five and two point one percentage point higher rates of very low birth weight births and low one minute APGAR score births. These are actually quite large effects relative to the sample means 61 and 46 percent increases in those outcomes. We looked at how the effects differ in terms of exposure across different trimesters of pregnancy and found that these impacts stem mainly from assaults in the third trimester of pregnancy.
Maya [00:30:54] We also find that assaults in the third trimester of pregnancy are associated with a higher probability of an induction of labor, which we interpret as a medical response to injuries sustained by pregnant victims of abuse. More generally, when we look at the medical services index, our results are less consistent for that one than they are for the birth outcomes, but we have some evidence that victims of assault during pregnancy are more likely to use medical services. And then when we look at the estimates from the maternal fixed effect model where we compare siblings, those are even larger in terms of magnitudes, which is what we expected. Here we find that assault during pregnancy leads to a point three standard deviation, higher summary index of severe birth outcomes and a point to five standard deviation, higher summer index that includes the less severe birth outcome measures. These impacts are driven by increased rates in very preterm births, low one minute APGAR score, as well as admission to the NICU.
Jennifer [00:31:57] I found it really interesting that at least for the first set of results, you find that women who are assaulted are more likely to use medical services. You mention in the paper that that could mean you're actually biasing in the direction of of finding less severe impacts because women could be essentially be compensating for the assault in that way.
Maya [00:32:16] Right.
Jennifer [00:32:16] There's something I wouldn't have thought about ahead of time.
Maya [00:32:20] Yeah, that's right. So we find, for example, that women are more likely to get prenatal care earlier when they're assaulted during pregnancy, which, we interpret as potentially the women are checking on their health of their pregnancy following the assault. So, again, it's sort of evidence of compensating behavior where the effect of the assault could be even larger if a woman didn't have access to that medical service.
Jennifer [00:32:43] Right. And so then you do a bit to figure out which mechanisms seem to be driving the effects, as you've already talked about a little bit. So more specifically, you're looking at you're trying to figure out whether the results are the direct effect of the physical assault itself or the indirect effects of related stress and any coping mechanisms by the mother. So what do you test for there and what do you find?
Maya [00:33:07] Right. So, you know, our birth records data do allow us to look at a variety of these kind of behavioral channels, as we talked about. We looked at the behavioral index and we looked at the individual outcomes within there, so smoking cigarets, illicit drug use, too low or too high weight gain, lack of WIC benefit received, self reported depression, and we just find no evidence that assault during pregnancy affects any of these behaviors. As we mentioned, if anything, there's some evidence of compensating behaviors in terms of prenatal care. As a result, we think that the main channel driving the adverse effect of assault during pregnancy on infant health is the direct physical channel.
Maya [00:33:48] So pregnant the victims of assault may be more likely to go to the hospital because of the resulting physical trauma, where they may need to have their labor induced quite early prematurely and therefore deliver very preterm and very low birth weight babies that may as a result, for instance, need to be admitted to the NICU. You know, as I mentioned, our main first two regression model strategies are less equipped to identify the effects of sort of the chronic stress associated with violence and a violent relationship. So we're not saying that that type of stress is not there and that that's not an important channel it's just that we don't have a very good way of capturing that with these estimates.
Jennifer [00:34:32] Right. And I think your third estimate, actually, because it's so much larger, suggests that stress probably is a big factor.
Maya [00:34:38] That's right, yeah.
Jennifer [00:34:41] Okay, so then you you do a bunch of robustness checks as as all these papers do to make sure you're isolating the effects of violent assaults and not just picking up underlying differences between the types of women who do and don't experience violence during pregnancy. So talk us through those checks and how you're able to convince yourselves that you've measured the causal effect of IPV.
Maya [00:35:03] Right. So we've already discussed some of the checks that we do. For instance, as I mentioned, we show that a wide range of maternal and paternal characteristics are very similar across our treatment and comparison groups. Because the key sort of threat to us being able to estimate the causal effect is that somehow the types of women that experience assault during pregnancy are just systematically different from the types of women who experience an assault after pregnancy, at least this is a threat to our first to identification strategies. So, you know, like I said, we can look at all of the observable characteristics that we have in our administrative data and in the survey data. We don't find any evidence that those are different, but you might still worry that they're sort of unobservable out there that we just can't capture the data. So this is where we turn to this really nice bounding exercise that was proposed by Emily Oster in a recent paper that essentially allows us to do a calculation that assesses the degree of bias that is introduced, if there is, in fact, any selection on unobservable characteristics. So the basic idea behind this exercise is to calculate how large would that selection on unobservables is have to be relative to any selection on observables in order to make the treatment effect that we estimate go away. So we do this calculation and we find that the selection on unobservable would have to be more than one point three times higher than selection on observables in order to completely explain away our effect. We think that this is unlikely to hold off for a couple of reasons. The first, we have a lot of observable characteristics in our data, so just the set of potential unobservable characteristics becomes relatively small. We've also shown that there's no evidence of selection on observable characteristics.
Maya [00:36:50] So it just seems quite unlikely that there would be this high degree of selection on unobservables. And in Oster's paper, she herself proposes a threshold of one, which is that, you know, we worry about selection on unobservables if that selection is less than equal to selection on observable. So the ratio is less than one, we instead find a ratio of 1.3, which I think is quite reassuring.
Jennifer [00:37:19] All right, so then you use your results to estimate the total social costs of IPV, at least in terms of the negative effects on birth outcomes when the victim is a pregnant woman. So walk us through what you include in that calculation. So that's a tough one to do and what you find, right?
Maya [00:37:37] Right. So first, just the reason that we want to do this type of calculation is because it struck us that most of the estimates of the social cost of crime, which is like a huge literature trying to come up with these estimates because they're very important in any type of cost benefit analysis that you do on policies related to crime or even policies not related to crime, like early childhood education, that end up finding some sort of effect on crime. They rely on some estimate of what's the social cost of crime to do their own cost benefit calculation, so these are really important numbers. In the existing literature, these numbers are typically done through one of two ways. So one is kind of jury award estimates, which is basically based on actual cases, looking at how much juries award to various victims of crimes, it's obviously a relatively selected sample of of who is included in those estimates. Then also things like hedonic kind of pricing models where all of these rely on the assumption that the full costs of the crime on the victims are fully known in observable either to the juries or people answering surveys about how much they're willing to paid to avoid crimes or are fully capitalized on things like housing prices. So we think that this is unlikely to hold, especially because we have very few estimates of the effects of all kinds of different crimes, including, for instance, intimate partner violence, and especially on various outcomes that people maybe haven't thought about in these social cost literatures such as birth outcomes. We thought that including this number would be important for this literature. What we do is we take our estimate of the effect of assault on very low birth weight births, and then we consider the best available evidence on the cost associated with having a very low birth weight child arising through six different channels. First, very low birth weight children have high rates of infant mortality. They have higher medical costs at and immediately following birth, they also have potentially increased rates of childhood disability, which itself entails cost. Then there's research showing that birth weight is correlated with adult income and medical wellbeing in adulthood. We think about future decreases in adult income and increased medical costs associated with adult disability, as well as reductions in life expectancy. So when we do this, we generate an average social cost of thirty six thousand eight hundred fifty seven dollars per assault's during pregnancy. This is using our most conservative estimate of the effect of assault during pregnancy on the likelihood of very low birth weight. If we instead use the larger maternal fixed effect estimate, we actually get an average social cost of eighty five thousand nine hundred ninety nine. Then with an average of three thousand one hundred seventy seven pregnant women between 2004 and 2012 in New York City who suffered from physical abuse based on available data, our estimates imply the total social cost that were previously unaccounted for in just New York City alone, are somewhere between hundred seventeen to two hundred and seventy three million dollars per year. Then if you aggregate this under some assumptions of the rate of abuse during pregnancy for the entire United States, you get an annual social cost of around three point eight to eight point eight billion dollars. This is again under the assumption of the best available nationwide victimization estimate for pregnant women, as well as the fact that there are approximately three point nine million births per year. I want to emphasize that after we do all of these calculations, these numbers likely completely underestimate the full social cost of assault on pregnant women for at least five reasons. The first, we have measurement error in our crime data. That's just because, you know, not all assaults get reported to the police, so we're only picking up the effects of those that do get reported. We don't know anything about assaults that are not reported to the police. This is related to also underreporting of IPV to the police more generally, I should say. Also, the measurement error in our crime data also stems from the fact, again, that we we have to assume that an assault occurring in the home of a woman actually affects the woman herself. Then we also, of course, do not measure the effects of the assault on the mother's own well-being. Everything that we've done so far relates to the costs associated with her unborn child's well-being, but the mother herself could be affected, in fact, there's really great research showing that women who are victims of crimes experience, for instance, reduced labor market earnings in the years following the crime. There's also, as we discussed, potential compensatory responses on the part of the mother's that could reduce the damage to the fetus from the assault that should be taken into account. Then, as we are already discussed, that women are subject to IPV are likely to be living with high levels of stress, which in itself is also known to affect fetal development. So all taken together, we think that we have a decent estimate of the social cost of physical assault during pregnancy based on the health of the infant, but we really underestimated, for the reasons I just mentioned and, you know, full cost calculation should really take into account. Also, the well-being of the mother herself.
Jennifer [00:43:22] It really highlights how complicated these cost benefit analysis are. The students always give when I'm reading these papers in class, students always give the authors a hard time for saying back of the envelope cost analysis and discussion. Yeah, but it's likely to be and has to be back of the envelope because they really are so complicated. Yeah, we do love that phrase.
Jennifer [00:43:43] Okay, so you released this research as a working paper earlier this year. So has any other research come out since then that is relevant to this research topic that we should talk about?
Maya [00:43:54] Yeah, so I mean, I just mentioned actually in the discussion of the social cost. So there's a really interesting recent working paper by Anna Bindler and Nadine Ketel. So what they do is they use Dutch administrative data that is unique in that it links victims of various crimes to their labor market outcomes. They use a really nice study design where they essentially find that victims of crimes and especially victims of domestic violence experienced large earnings losses and increases in public benefit receipt that last up to eight years following the crime. So I think these findings are really important because they really complement the evidence that we're showing for infant health and they suggest that in order for us to approach some estimate of the full cost of IPV during pregnancy, which, as you mentioned, is very hard to do, we are at the very least have to add all of these costs together to try to get a sense of both the cost to the mothers and the infants themselves.
Jennifer [00:44:52] So putting it all together, the results of this study and the other studies we've talked about, what are the policy implications of this work?
Maya [00:45:00] Yeah, so I think that our results, as well as those from the other studies, have important implications for thinking about persistence and inequality, which is where we started this discussion earlier.
Maya [00:45:10] So poor pregnant women are much more likely to be victims of assault than their more advantaged counterparts. The majority of all violence against women is perpetrated by domestic partners. So our results suggest that intra family conflict might be an important and previously understudied mechanism through which early life health disparities perpetuate persistent economic inequality across generations. So we're practically, you know, interventions that reduce violence against pregnant women and there's a number of these out there, you know, of course, that have varying degrees of effectiveness, but to the extent that they have meaningful consequences for the women and their partners, potentially, that our results imply that they could actually have benefits for the next generation and just the broader society as a whole.
Jennifer [00:46:01] And what's the research frontier here? What are the next big open questions in this area that you and others will be thinking about in the years ahead?
Maya [00:46:10] Right. So I think there's still a lot of work to be done on the topic of domestic violence and economics. So to any students that are out there listening to this, that are looking for research ideas, I really think that this is an area that's understudied within the economics field and in particular when it comes to understanding the effects of domestic violence on the victims. So there's, you know, some studies that think about bargaining intra household models and think about the determinants of domestic violence, but there's really quite a bit less on the effects of the violence itself. So, for instance, as we already mentioned, we have very little evidence on the causal effects of intimate partner violence on the mother's well-being, kind of rigorous causal evidence on their mental health, their physical health, their labor market trajectories, their future family formation. Also, you know, we have shown that prenatal exposure to IPV has large adverse effects on infant health, but, you know, do these effects translate into later childhood development? Do these effects persist into adulthood as other early childhood shocks have been shown to? These are still outstanding questions that we don't have answers to. So I think they're all very fruitful areas for future research.
Jennifer [00:47:22] Yeah, and I'll add to that, we also know very little about just policies, what interventions we could we could put in place to try to mitigate some of those costs.
Maya [00:47:32] Absolutely. That's right. Yes. I think there's a variety of interventions out there, but a lot of the research uses fairly small samples, sort of isolated case studies and really having a more comprehensive evidence base where we can think about what works and what doesn't in this area, I think is would be really valuable.
Jennifer [00:47:51] My guest today has been Maya Rossin-Slater Stanford University. Maya, thanks so much for talking with me.
Maya [00:47:57] Thank you so much for having me.
Jennifer [00:48:04] You can find links to all the research we discussed today on our website, probablecausation.com. You can also subscribe to the show there or wherever you get your podcasts to make sure you don't miss a single episode. Big thanks to Emergent Ventures for supporting the show and thanks also to our Patreon subscribers. This show is listener supported, so if you enjoy the podcast, then please consider contributing via Patreon. You can find a link on our website. Our sound engineer is Caroline Hockenbury with production assistance from Elizabeth Pancotti. Our music is by Werner, and our logo is designed by Carrie Throckmorton. Thanks for listening and I'll talk to you in two weeks.