Neighborhood Effects on Reproductive Health: How Can We Conceptualize, Operationalize, and Measure Them?

Jennifer Culhane: By way of a brief introduction, as Michael just told you, I study racial and ethnic differences in the rate of pre-term birth and I work exclusively in the city of Philadelphia, and our work is really an integration of several different lines of research. Mainly we look at the role or the effect of individual level of stress. So does a woman perceive her life as very stressful, we call that cognitive appraisal stress. Is she exposed to lots of objectively stressful conditions at the individual level? Has she experienced homelessness? Is she very poor or does she have a lot of material resource deprivation? That’s one area. The second area is the role of neighborhoods and we’re trying to look at whether a stressful neighborhood--and I’ll show you a lot about how we sort of conceptualize what that means--does neighborhood have an independent and significant association to risk of pre-term birth independent of her individual experiences? And then we look to see if these two levels of stress exposure, individual and neighborhood, are ecological. That’s another word I’ll use that means sort of neighborhood. Do these actually product an increased risk of infection in pregnancy and/or exacerbate the consequences of infection in pregnancy that ultimately leads to an increased risk of prematurity? And one of our big areas of interest is if this is a viable pathway, if this is a theory that actually works, then is it true that African-American women are exposed to much higher levels of stress both at their individual level in their own lives and in their neighborhoods? So that’s the sort of gist of our work. What I’m going to spend the next 20 minutes or so doing is talking a lot about the neighborhood piece, and I would like to begin by saying that in the early 1900’s, the notion that ecological factors were related to health outcomes, particularly infant mortality, were observed by researchers that were looking for causes of high infant mortality rates.

And what they concluded, what they found in their study, was that 88 percent of the babies had fathers that earned less than $1,250.00 a year, and they said as the income doubled, the mortality rate was more than halved, which is the more sane and safe conclusion--that 88 percent of these fathers were incorrigibly indolent, or below normal mentally, or that sound public economy demands irreducible minimum living standards, and the study went on to say housing congestion and neighborhood conditions were causal in the production of high infant mortality rates. What’s curious, though, is that the role of ecological or neighborhood conditions in the production of health virtually disappeared for a big chunk of the 1900s, and I think this is due to several important factors. First off, in research there’s what’s called a weariness of ecological data because of something called the ecological fallacy. And this really means is if you measure something at the neighborhood level, let’s say average household income, in a census tract and say it’s $30,000.00 a year, and then there’s an individual that lives in that census tract, and someone concludes they earn $30,000.00 a year. See, there is major problems with that and that’s called an ecological fallacy; attributing characteristics to individuals based on population or neighborhood data. There was also in the last, say, 70 years, huge advances in statistical methods that kept a lot of researchers busy modeling data that is at the individual level. So there was a lot to do, lots of new statistical tools, and so that sort of a force keeping people looking at individual level attributes and their association to disease. I think that we have had a methodological, conceptual, and political individualism, and I think it’s really dominated this country since the 1950s, and there’s been a real disconnect between geographers and epidemiologists and lots of these techniques--looking at ecological data--have come from geographers. But even though we’ve had big growth in statistical methods, and lots of work has been done looking at individual level attributes and their association to disease, we still cannot explain population inequalities in the diseased burden.

These are not fully accounted for by any combination of individual behaviors, and in my opinion, genetic attributes, and so these differences must be attributable to other as yet unmeasured factors. Be them individual, something we haven’t figured out how to measure, or potentially ecological or neighborhood level exposures. Given that I think there’s been a stagnation in perinatal epidemiology--I mean we really haven’t fully been able to explain even or predict why an individual is at risk for pre-term birth or low birth weight, and we certainly can’t explain the factors that underpin the racial and ethnic difference. There’s been a cry or a call for something that’s been termed “The New Public Health” since about the 1990s, and in this sort of new public health there’s a renewed interest in ecological data and the ecological associations--say area level poverty and health outcomes. And this is really looking at the upstream social conditions at the same time simultaneously modeling ecological data and individual data to have more robust explanatory models about things like prematurity, low birth weight, and a whole variety of other health outcomes. So with this renewed interest in ecological data, there’s a lot of studies being published across a broad range of health outcomes. Chronic diseases, even self-rated health, long-term disability, cardiovascular diseases and symptoms, respiratory function, certain health behaviors, domestic violence, low birth weight, and to some extent, prematurity.

There are significant and independent effects of ecological exposures in these health outcomes. I want to just say though that these are largely very small, so the effects are little. They’re not monotonic, which means that they vary a lot by age and gender and race, and so it’s still a very confusing picture about neighborhood conditions and health outcomes. So to begin with, in this sort of idea of the affect of ecological or neighborhood level conditions to these health outcomes, one has to immediately start this sort of line of inquiry by developing theories or by understanding mechanisms, how neighborhood conditions might actually influence health, and there’s a bunch of sort of direct mechanisms and these are generally thought of in sort three broad domains. The first is community social environment, that a neighborhood that’s got high crime rates, high housing, abandonment, etcetera, that this may impinge upon the ability to form social relationships, and that these even weak ties may be how people transmit important information. There’s a sort of a phrase called “the strength of weak ties,” so crummy neighborhoods may impinge upon that sort of transmission of important information. These neighborhoods may also have enough influence on neighborhood cohesion, so that there’s limitations in informal social control. I think this is very important in terms of adolescent behaviors and potentially other sort of behavioral deviance because in these closed loops of social capital, if I see your kid doing something I’ll call you on the phone, and it actually has these ways of controlling behavior but it requires that people actually interact. Shared cultural norms and values was discussed in the previous talk. Potentially there are at norms or values that are maladaptive that can be shared in communities.

I think the people that live in very marginalized neighborhoods can have decreased civic participation and that in this instance they’re not demanding public services like they could be. Clearly, there’s limitations to adequate educational opportunities and employment. In terms of direct community services, we know that there are associations between the location of grocery stores and bank machines by sort of neighborhood deterioration. There’s limitations in recreational opportunities, even health care facilities, and retail stores, so that’s sort of a direct effect. And finally, in poor neighborhoods there’s the location of plants such as garbage, incineration, oil refineries. These kind of plants are located in neighborhoods that are poor and disenfranchised, so there’s an increased burden of toxicants, noise, and clearly poor housing. Indirectly, I think as we heard in the previous talk, all these neighborhood conditions can be conceptualized as stressors and they may, in fact, get under your skin in a physiological manner that starts with activation of a stress response, and that this may, in fact, interact with your immune system to produce a generalize susceptibility among people that live in sort high stress neighborhoods. This is currently termed “alostatic load” and what it means is that there’s lots of wear and tear on organisms that are constantly responding to stressful conditions in their neighborhood and the energy that goes into maintaining homeostasis actually can wear out systems more quickly in a person exposed than those that are not exposed.

And indirectly I think that neighborhoods with lots of adverse conditions can have serious implications for mental health, negative emotions, depression, and anger, and hostility. So in terms of this ecological look at the production of health, we have to one, develop the theories for your particular outcome. How does that neighborhood variable influence health? Next we have to understand what we mean by context. Is it purely geographic? Is it physical space--your neighborhood? People spend lots of time at work. Is it work context? Is it religious affiliations? People travel out of neighborhoods to attend churches in different parts of cities all the time. Is it membership in other clubs and other social networks? That’s a very important piece of understanding what we mean by context. We don’t know exactly what contextual level variable should be measured. Currently, almost all of the data looks at area level measures of poverty, average household income, percent of households under federally defined poverty, those are the kinds of data that people are using. What about things like crime or the production of homelessness or positive things like number of grocery stores, access to bank machines, these parks and other things? So what contextual variable should be measured? We don’t know exactly what individual level variable should be measured and then adjusted for in models that simultaneously look at neighborhood and individual because potentially some of the maladapted health behaviors that we’ll see may be what we call intervening variables, and they’re actually produced because of the neighborhood consequences. So if we over adjust and get rid of all the effects of the individual characteristics, we may miss the effect of the neighborhood. That’s called over control. There’s another idea and that is like people live in like neighborhoods, so that people choose and they will then overstate the effect of a neighborhood because similar people have moved there intentionally. That’s an economic theory. It’s called “*indogenaity” and in my opinion, it doesn’t quite function like that in the real world.

I think people’s choices are quite limited. We don’t know what statistical techniques are optimal. I’m going to show you some results from a few, and like we heard, there’s a big limitation in sort of cross-sectional design, so measuring context of a woman in her pregnancy may be the wrong context, and we may need to understand her context from her early life. Okay, so now I want to just talk a bit about measuring context. In these studies mostly what we see is that area level variables are constructed using census levels of aggregation. Okay, these are administrative units. You hear them as blocks, block group census tracts zip codes. There are big limitations to what’s called “vector GIS.” Number one, they don’t actually have a lot to do with lived space. They’re political boundaries, or there’s some sort of barrier reason for cutting a place like a park or a highway, so they’re not actually great in terms of really understanding exposure to a condition. There are other ways, though, to actually use address level data and we call these this raster GIS or kernel density, and I’ll show you a bit about that. And this provides a much more flexibility in understanding the spatial range of a variable: so does high crime matter if it’s on your street? Two streets over? Within a half a mile of your house? That’s what we mean by spatial range. I want to just show you here graphically about the limitation of using administrative units or census tract block group kind of levels of aggregation.

Down there in the green, I guess that will be on your left, you can see what we call the problems of aggregation effect. Those two women are actually--those are dots that represents the address of two women that have participated in some of our studies. They actually live on face blocks; it means they look at each other across the street. And they’re assigned to different block groups, because the street centerline is the dividing line, so we’re attaching them in space to different understandings of exposure. And you can see in that block group there’s women that live much, much further away and we’re calling them equally as exposed to a condition, and we’ll get to that in a minute. And then, like I said, we also have this problem of scale. Maybe a condition only matters in the green circle and all the other women in that block group really aren’t exposed. So administrative units are crude ways of understanding space and they may actually really cause us to miss the important effects of neighborhood. So we’re avoiding this problem in Philadelphia in kind of a unique way, and what it requires that we get public health and administrative date from city agencies. We get data from a whole wide range of city agencies. In this instance what we see is this sort of squared, you know, these sort of dense purple squares going out to white squares represent imminently dangerous structures. And so what we get is the address, the X and Y-coordinate of every imminently dangerous structure that’s been reported to the city of Philadelphia, and what we do then is use a technique that models that as a continuous surface in space. So we can attach for every 100 foot by 100 foot square cell in the city, an actual value for exposure to imminently dangerous buildings based on a whole wide variety of different bandwidths or search radii.

So we can create a density based on a 1/16 of a mile, and 1/8 of a mile, a 1/2 of a mile, and then we can see the surface of that exposure in this instance, imminently dangerous buildings. Then we can locate women on that surface and we can assign them a value for exposure to the condition of interest. I know that might sound a little bit too complicated, but maybe these examples will help. So for instance, these are our different search radii that are overlaid for the traditional administrative units. Behind in green you see blocks and block groups and census tracts, and above that you can see how we will search out with this flexible spatial range. These are what these densities actually look like. Right here’s really a raster density or a kernel density map based on a half a mile search radius for STD’s. This is gonorrhea and chlamydia from ‘98 to 2000. And the city of Philadelphia was kind enough to geocode all those addresses to the X and Y-coordinate and provide me with the “D” identified data to create that raster surface. All the black dots actually represent women in a study of ours. Here is imminently dangerous buildings. Again, you can see that there’s clearly high-density neighborhoods--they’re the dark purple--and here’s homelessness. In Philadelphia we have a centralized shelter intake system and women have to report their previous address to get emergency shelter, and again, the office of emergency shelter geocoded those records and we’ve created these surfaces of sort of homelessness production by neighborhood. The dark yellow areas are high-density neighborhoods in terms of producing homeless people. All right, so given that we have ecological data that we’ve measured in several ways; the traditional block group census tract way, which is used mostly because that’s the data we can get from the census, right?

You need these point data or address level data to do the kind of continuous surface stuff that we’re doing. But we have now data that we’ve derived using the densities and from the administrative units, and we’ve started looking at them from a descriptive point of view, using those ecological variables and logistic regression models, and we’ve used them in multi-level models, and I’ll show you some of that data. And we’ve also now launched into a really interesting use of all these data in techniques that people might know as “data mining.” They’re actually not hypothesis driven; the answers just come out of the information, and I’d like to show you a little of that. In terms of description, one of our main points is I sort of began this talk with is trying to see if these exposures co-vary a lot by racial and ethnic group membership because they have to if they’re involved in the racial and ethnic gap. And the first thing we did is we took a year of birth certificate data and we linked it into five years of the shelter intake data, and then we looked at the prevalence of homelessness by maternal race, maternal education, and maternal parity. So if you see on the far left the big green bar that’s over 40 present, that represents African-American women that have 12 years or less of education where the index birth represents their fourth or greater birth. About 42 percent of those women in Philadelphia have requested an emergency shelter stay over a five-year period.

You can see that this is very, very different. Well, let me just mention: you see the next set of bars. Those are African-American women with more than 12 years of education and although there is an education gradient, still almost 30 percent of those women have requested an emergency shelter stay. For the same profile, except changing race--so white, 12 years of education or less, fourth or more birth, the homeless shelter request rate is only about 5 percent, so there’s huge racial and ethnic differences in very unstable housing measured by requesting emergency shelter. What we see here is that by block group, black women live in blocks that have much higher rates of assault, of tax delinquency, and of shelter requests. These are not them; these are the people that live in their block group. We were looking for a good measurement of neighborhood--of sort of crime or neighborhood danger, and we looked at the address on the birth certificate, and we drew a concentric circle around that address. And what we found was that about 35 percent of the black infants were discharged to homes that were within 500 feet of a gun related death as opposed to about 5 percent of the white infants. Actually, this is probably the last thing we’re going to get to because of a time constraint. What we’re interested in is seeing if these neighborhood conditions actually begin to explain some of the racial and ethnic difference. In this regression, our outcome is whether the woman has bacterial vaginosis or not in pregnancy, not prematurity. We haven’t gotten there yet. And what we see is in the un-adjusted model, black women have about 3.3 times the risk of having BV at their first antenatal visit in prenatal care.

When we adjust that model for individual income and education, we see no reduction at all in the black-white BV odds ratio. When we add in health behaviors that have been associated with BV--sexual practices, feminine hygiene practices, and perception of stress, we don’t see any reduction in the black white BV odds ratio. When we add in objective level stressors at the individual level--when we ask women, “Have you ever been homeless? Do you have enough money?” We see that we reduce the black-white BV odds ratio a little. Then when we add in at neighborhood or ecological level data, and this is crime, homelessness, and tax delinquency, we see the first real big bite out of this ratio, so we go from about 3.3 to 2.4. Although there’s still a huge gap, at least it tells us that maybe ecological level exposures are really related to risk, be it by production of susceptibility to infection or other mechanisms that may increase women’s likelihood of adverse reproductive outcomes, and I’m going to have to stop there. Thank you.