The Relationship between Lead and Crime

#1587—Jnl of Health and Social Behavior—Vol. 45 No. 2—45207-stretsky

The Relationship between Lead and Crime* PAUL B. STRETESKY Colorado State University

MICHAEL J. LYNCH University of South Florida

Journal of Health and Social Behavior, 2004, Vol 45 (June): 214–229

This study investigates the association between air-lead levels and crime rates across 2,772 U.S. counties. Data for the analysis come from the Environmental Protection Agency, the Bureau of Census, and the Federal Bureau of Investigation. Results suggest that air-lead levels have a direct effect on property and violent crime rates even after adjusting for general levels of air pollution and several structural covariates of crime. We also find that resource deprivation interacts with air-lead levels. The association between air-lead levels and crime rates—property and violent—is strongest in counties that have high levels of resource deprivation and weakest in counties that have low levels of deprivation. This interaction is consistent with arguments and evidence in the health care literature that populations most at risk of lead poisoning are least likely to get the resources required to prevent, screen, and treat the illness. Contemporary research suggests that lead exposure is a potential source of crime and delinquency (Denno 1990; Needleman et al. 1996; Nevin 2000; Pihl and Ervin 1990). However, indicators of the geographic distribution of lead have not been analyzed as a structural covariate of crime. The omission of lead from the analysis of crime rates is unexpected for four reasons. First, lead alters neurotransmitter and hormonal systems in a way that may induce aggressive and violent behavior (Needleman et al. 1996). Moreover, some medical researchers believe that as much as 20 percent of all crime is lead-associated (Needleman 1990:87). Consequently, people living in areas with elevated lead concentrations may be exposed to environmental conditions that possess the potential to stimulate aggressive behaviors such as crime and delinquency. Second, a moderate sized literature exam* Correspondence concerning this article should be sent to Paul B. Stretesky, Department of Sociology, Colorado State University, Fort Collins, Colorado 80523. This manuscript benefited considerably from the comments of Michael Hughes, Thomas Cole, and the anonymous reviews of the Journal of Health and Social Behavior. Any shortcomings, however, are entirely due to the authors.

ines how biological and sociological factors interact to affect crime (e.g., Fishbein 1990; Booth and Osgood 1993). In that tradition, it has long been asserted that the study of crime must be interdisciplinary, combining disciplines such as sociology and biology (Jeffrey 1985). The study of lead and crime at the aggregate level of analysis is consistent with both approaches. Third, there is considerable variation in the geographic distribution of lead and crime across time and space (Nevin 2000; Stretesky and Lynch 2001). It is plausible that the spatial distribution of lead influences spatial patterns of social behavior such as aggression and violence. Fourth, a political economy of lead exists. Various social, political, and economic forces influence the production, consumption, and distribution of lead across the urban and rural landscape (Warren 2000). Thus, while the medical research suggests that lead influences human behavior by altering biological processes, those processes are set into motion by sociological circumstances that cause humans to be exposed to elevated levels of lead. Thus, the relationship between lead and crime must be interpreted relative to sociologically relevant factors that impact exposure to lead such as

Downloaded from hsb.sagepub.com at UNIV OF SOUTH FLORIDA on June 20, 2015

214

#1587—Jnl of Health and Social Behavior—Vol. 45 No. 2—45207-stretsky THE RELATIONSHIP BETWEEN LEAD AND CRIME ethnicity, race, and class. For instance, minorities and the poor are more likely than whites and the affluent to be exposed to environmental sources of lead and less likely than whites and the affluent to be effectively screened and treated if they have lead poisoning (Brody et al. 1994; Pirkle et al. 1994). Clearly there is a need to interpret the relationship between lead and crime sociologically, or as the result of how human society, especially economic production and waste practices, are organized. Is lead an important structural covariate of crime? To answer this question and advance the current knowledge about lead and crime, we draw upon the public health and medical literature to formulate five hypotheses that aid in assessing the potential relationship between county air-lead levels and county crime (violent and property) rates. LITERATURE REVIEW Lead enters the body through a variety of pathways (Mielke and Reagan 1998; Mushak and Crocetti 1989). Numerous behavioral, neuropsychological, and biological studies find that sufficient exposure to lead is associated with brain dysfunction. This explanation—referred to as the “neurotoxicity hypothesis”—suggests that lead exposure alters neurotransmitter and hormonal systems in ways that may induce aggressive and violent behavior (Needleman et al. 1996). Several studies have relied on the neurotoxicity hypothesis to examine the relationship between lead exposure and undesirable outcomes such as impaired school performance and auditory processing, increased distractibility, short attention span, hyperactivity, inappropriate approach to problem solving, and inability to inhibit inappropriate responses (Bellinger, Stiles, and Needleman 1992; Dudek and Merecz 1997; Fergusson and Horwood 1993; Fulton et al. 1987; Gittelman and Eskenazi 1983; Leviton et al. 1993; Lyngbye et al. 1990; Needleman and Gatsonis 1990; Needleman et al. 1979; Oliver, Clark, and Voeller 1972; Oliver et al. 1977; Rice 1996; Thompson et al. 1989; Winneke et al. 1983). During the past decade, a handful of researchers have drawn upon the neurotoxicity hypothesis to examine the relationship between lead and crime. The most notable of these studies have been carried out at the indi-

215 vidual level of analysis. For instance, Deborah Denno’s (1990) longitudinal study of the link between lead and crime among African American males, which controlled for many other potentially influential independent variables (including environmental and social factors, such as parents’ income and occupation), found that early childhood lead poisoning was one of the most important predictors of three types of delinquent outcomes: (1) disciplinary problems from ages 13–14; (2) juvenile delinquency from ages 7–17; and (3) the number of adult offenses from ages 18–22. Needleman et al. (1996) also found a significant relationship between bone-lead levels and reported and observed delinquency and behavioral problems during the pre-teen years. The researchers concluded that, “lead exposure is associated with increased risk for antisocial and delinquent behavior, and the effect follows a developmental course” (Needleman et al. 1996:363). In another study, Pihl and Ervin (1990) examined the hair of inmates convicted of violent and property crimes to assess whether variations in exposure to toxic substances (including lead) was more strongly related to property crimes or violent crimes (such as murder, rape, and aggregated assault). The researchers found that, despite the inmate’s age, socioeconomic status, months institutionalized, and drug use history, hair-lead levels distinguished the two groups. Specifically, violent offenders were more likely than property offenders to have elevated levels of hair-lead levels. Pihl and Ervin’s research suggests that violent crimes are more likely than property crimes to be the result of elevated lead exposure. The researchers suggest that their results are consistent with the literature concerning the negative outcomes associated with lead, especially those outcomes related to hyperactivity, which is more closely related to violence than property crime. Recent aggregate-level studies also suggest that lead and crime are related. Nevin (2000) studied the lead and violent crime relationship in the United States. between the years 1964 and 1999 by comparing trends in lead use in gasoline and trends in violent crime rates (e.g., murder, rape, robbery, and aggravated assault). He discovered that the per capita consumption of leaded gasoline is strongly associated with all types of violent crime. Stretesky and Lynch (2001) also investigated the association


#1587—Jnl of Health and Social Behavior—Vol. 45 No. 2—45207-stretsky 216

JOURNAL OF HEALTH AND SOCIAL BEHAVIOR

between air-lead levels and homicide rates across U.S. counties and discovered a statistically significant relationship, even after controlling for several important sociological predictors of crime and several different types of air pollution. Hypotheses The medical and public health literature on the relationship between lead and crime suggests three hypotheses: Hypothesis 1: There is a positive association between air-lead levels and violent crime rates; Hypothesis 2: There is a positive association between air-lead levels and property crime rates; and Hypothesis 3: The relationship between violent crime rates and air-lead levels should be stronger than the relationship between property crime rates and air-lead levels.

The three hypotheses reviewed above do not address the sociologically relevant factors that may mediate or moderate the lead-crime relationship. For example, the most recent National Health and Nutrition Examination Survey demonstrates that African Americans and the poor have, on average, higher body burdens of lead than whites and the more affluent (Brody et al. 1994; Pirkle et al. 1994). Several studies suggest that lead exposure patterns are a result of the unequal geographic distribution of lead. For instance, Hird and Reese (1998) used data on industrial releases to examine the distribution of 29 environmental hazards across all U.S. counties. Among other things, the researchers found that air-lead releases are positively related to county racial and economic composition (see also Stretesky 2003). Potential environmental inequality is important to consider in the study of the relationship between lead and crime because industries that release and store lead are more likely to be located in counties that lack social and economic resources. For instance, deprivation influences housing opportunities and choices and therefore influences the relationship between demographic characteristics and proximity to noxious facilities that release lead and other toxins into the environment. Recent

research suggests that the presence of environmental hazards within or near a neighborhood increases negative perceptions of that neighborhood and also reduces property values (Ketkar 1992; Nelson, Genereux, and Genereux 1997). Negative neighborhood perceptions, in turn, interact with race and class to produce residential segregation in lead-prone areas because lower property values attract minority and poor residents that cannot afford to reside in a lead-free environment or are pushed into less desirable neighborhoods because of more direct forms of housing discrimination. Environmental hazards, then, influence levels of neighborhood deprivation over time, increasing both racial, ethnic, and economic segregation (Daniels and Friedman 1999; Stretesky and Hogan 1998). At the same time, more affluent whites “vote with their feet” and move away from lead sources (Bullard 1995). Moreover, black segregation in neighborhoods near environmental hazards such as lead producing industrial facilities is only likely to increase over time because whites are less likely than blacks to move into integrated and black neighborhoods (Massey 1994). Since a myriad of studies in criminology and sociology point to the relationship between levels of deprivation and crime rates (Kovandzic, Vieraitis, and Yeisley 1998; Land, McCall, and Cohen 1990; Messner 1982), it is plausible that lead is related to crime through its influence on levels of deprivation associated with race and class segregation. The literature on the unequal distribution of lead, crime, and resource deprivation suggests the following hypothesis: Hypothesis 4: Resource deprivation mediates the relationship between lead levels and crime rates.

There is also evidence suggesting that when minorities and the poor are not disproportionately exposed to lead in the environment, they are still more likely than whites and the affluent to suffer from the ill effects of lead (Kraft and Scheberle 1995; Reed 1992). This is because the poor and minorities have fewer contacts with physicians, and when they do receive treatment, they are more likely than the affluent to receive treatment that is inadequate or incomplete (Smedley, Stith, and Nelson 2002; Williams and Collins 1995). In the specific case of lead poisoning, the resourcedeprived are less likely than those with


#1587—Jnl of Health and Social Behavior—Vol. 45 No. 2—45207-stretsky THE RELATIONSHIP BETWEEN LEAD AND CRIME resources to be screened and effectively treated for lead poisoning (Reed 1992). The General Accounting Office (1999) recently reported that most children receiving federal aid such as Medicaid were not being adequately screened for lead poisoning and that there was little government oversight and enforcement regarding lead poisoning detection and prevention in deprived communities. Thus, more affluent counties do a better job at screening and treating lead poisoning than deprived counties. In addition, poor diet and low psychological well-being (often associated with economic deprivation) are two conditions thought to exacerbate the effects of lead poisoning (Reed 1992). The inequality in health care literature suggests the following hypothesis: Hypothesis 5: Resource deprivation moderates the relationship between lead levels and crime rates so that the effect of lead on property and violent crime rates intensifies as deprivation increases (i.e., lead and deprivation interact).

In this work we set out to determine if resource deprivation is likely to mediate and/or moderate the relationship between lead levels and crime rates. It is also possible that air-lead levels interact with structural covariates of crime other than resource deprivation. For this reason we also conduct exploratory analyses to determine whether there is supporting evidence that any additional relevant structural covariates of crime moderate or mediate the relationship between air-lead levels and crime rates. The next section discusses the data and methods employed to test the five hypotheses and conduct the exploratory research. DATA AND METHODS Unit of Analysis One methodological concern in this study centers on the appropriate unit of analysis. We study counties for two reasons. First, reliable data on the location of lead across the United States can only be obtained at the county level (Rosenbaum, Ligocki, and Wei 1998). Thus, there is no alternative unit of analysis available to study the relationship between lead and crime nationally. Second, it is important that the sample be representative of the country as

217 a whole because the health consequences of airborne lead are not likely to be limited to large cities and metropolitan areas (Baller et al. 2001; Kposowa, Breault, and Harrison 1995). Thus, examining all counties is more likely to produce representative results. Data Sources Three data sources were used in this analysis: The Cumulative Exposure Project (CEP), the Uniform Crime Reports (UCR), and the County and City Data Book. Descriptive statistics and for the variables constructed using these data sources are available in the appendix. Uniform Crime Reports. Our dependent variables, the logged violent and property crime rates, were constructed from the UCR. The UCR contains the only crime data that can be used at the county level of analysis during the time period in question. The statistics and figures produced in the UCR, though widely used by academicians and cited by the media as accurate reflections of crime trends, are not without measurement error, and the validity of the UCR data has been widely debated (Black 1970; Gove, Hughes, and Geerken 1985; Hindelang 1974; McCleary, Nienstedt, and Ervin 1982; Reiss 1971; Skogan 1975). At issue is whether this measurement error is serious enough to alter observed relationships across the large number of ecological aggregates compiled here. To date, there is no firm evidence that systematic measurement error would threaten the validity of the UCR when used in this capacity (Gove et al. 1985). To minimize potential measurement error produced by misclassification and/or instability resulting from yearly fluctuations, we average the UCR data over a three-year period (1994–1996; see Kovandzic et al. 1998; Messner and Golden 1992; Williams and Flewelling 1988). The years 1994, 1995, and 1996 were chosen for temporal and methodological reasons. First, the independent variables used in this analysis reflect conditions during 1990 (the only year for which lead data are available from the Environmental Protection Agency) and it is desirable for the dependent variable to reflect a point in time after 1990. Second, 1994 is the first year that the UCR crime statistics easily allow the user




to determine which county law enforcement agencies did not report and which reported for only part of the year. Prior to 1994 there was no adjustment for non-reporting agencies in the county data sets, and crime counts from non-reporting agencies were effectively treated as valid zeros rather than as missing data. Because it is highly unlikely that the source of undercounting is random (e.g., smaller, lessfunded law enforcement agencies in rural areas are the ones most likely to fail to submit crime reports), crime rates in poor counties would be the ones most severely understated. Thus, the problem of undercounting was viewed as more severe than the temporal mismatch created by using the UCR for the years 1994–1996 and estimated data on air-lead levels from 1990. The UCR data and population estimates derived from census data were used to construct two dependent variables: violent crime rate (number of murders and non-negligent manslaughters, forcible rapes, robberies and aggravated assaults per 100,000 residents averaged over 1994, 1995, and 1996) and the property crime rate (number of burglaries, larcenythefts, and motor-vehicle thefts per 100,000 residents, averaged over 1994, 1995, and 1996). Since our models produce residuals that are skewed, which may distort statistical significance tests, we also transform the dependent variable by taking its natural logarithm. This procedure produces a substantial increase in R2 and helps to produce more normally distributed residuals. A total of 339 counties failed to report crime data to the Federal Bureau of Investigation under the UCR program during the years of 1994, 1995, and 1996. Thus, we study 2,772 (or 89%) of the 3,111 counties in the contiguous United States. Cumulative Exposure Project lead measure. Unlike monitoring data, which are only available for a subset of air toxics in relatively few locations, the CEP provides long-term estimates of air-lead levels across all counties in the contiguous United States for the year 1990. Air-lead levels in the CEP were estimated using the Assessment System for Population Exposure Nationwide (ASPEN) that is based on a Gaussian dispersion model (Rosenbaum et al. 1998). This is accomplished in two steps. First, pollutant emissions are estimated based on EPA data on releases from large manufacturing sources (obtained from the EPA’s Toxic Release Inventory), large combustion sources

(e.g., Waste-to-Energy Facilities), small combustion sources (e.g., automobiles and dry cleaners), and EPA inventories of volatile organic compounds and particulate matter. Second, a computer model simulates the impacts of winds and other atmospheric processes on pollutants after they are emitted. Rosenbaum et al. (1998) provides a more detailed description of the estimation process. It is important to note that the ASPEN model does not provide perfect measurement of airborne lead levels because it may underestimate the levels of hazards and indoor air-lead exposure. Nevertheless, an EPA Science Advisory Board review of ASPEN methodology concluded that the model is valid (Woodruff et al. 1998). Currently, the ASPEN data provide the only national level estimate of air-lead levels. Air-Lead Levels are measured in micrograms per cubic meter (␮g/m3). Two states (Alaska and Hawaii) containing thirty-two counties do not have available air-lead level data because the EPA excluded these states from the CEP. Research on the association between airlead levels and blood-lead levels indicates that air-lead is one important source of body lead levels (Brunkenreff 1984; Hays et al. 1994). Other important pathways of lead exposure include soil, food, water, and lead based paint. To determine if CEP estimates of air-lead levels were correlated with blood-lead levels at the county level, we obtained data from the Ohio Department of Health on children under six years of age determined to have elevated blood-lead levels (i.e., >10 ␮g/dL) during 1998. The Centers for Disease Control identifies Ohio’s data as more valid than data from most states because a relatively large proportion of children across all Ohio counties are tested for lead poisoning. We found that across Ohio’s counties, CEP estimated air-lead levels were positively and significantly correlated with the percentage of children (of those children screened) determined to have elevated blood-lead levels (r = .44; p < .001; N = 88). That relationship persisted (r = .48; p < .001; N = 88) despite adjustments for the percentage of pre-1950s housing (1990 census estimate), reinforcing previous medical findings that suggest that lead exposure may result from a variety of contamination sources, including lead in the air. We suspected the relationship between airlead levels and crime rates to be non-linear, given the nature of lead exposure. The natural


#1587—Jnl of Health and Social Behavior—Vol. 45 No. 2—45207-stretsky THE RELATIONSHIP BETWEEN LEAD AND CRIME log transformation of air-lead levels supports this assumption. In these data it appears that there is a threshold effect for lead exposure on property crime and violent crime rates, though there is still evidence of an association between lead and crime when air-lead is not transformed. This transformation of the lead variable is consistent with a portion of the medical literature that has examined lead exposure at the individual level of analysis (Banks, Ferretti, and Shucard 1997). Census data. Several independent variables identified by Land et al. (1990) as important structural covariates of crime are included in this analysis of air-lead levels and crime rates to reflect previous research on the sociological study of crime rates. The variables are computed using data obtained from the County and City Data Book, 1990 (U.S. Bureau of the Census 1994). Resource deprivation is a principal component index that consists of the percent black; the natural log of the median family income; and a Gini index of family income inequality, percent of families below poverty, and the percent of families that are female headed (Baller et al. 2001; Land et al. 1990). Population structure is the principal component index that consists of the natural log of the population size and density. We also include the percentage of unemployed persons aged 16 to 64 years old (% unemployed), the percentage of the county’s population that is aged 15 to 29 years old (% 15–29 years old), the percent divorced (% divorced), and South, a dummy variable that indicates whether the county is located in a state that was a part of the Confederate South (see Land et al. 1990). Cumulative Exposure Project air pollution measure. To assess the likelihood that our results are not simply due to the effect of air pollution (or things associated with air-pollution) we include an indicator of air pollution in the model. The air pollution variable is an additive index and is composed of all 147 pollutants (except air-lead) that were estimated in the CEP (Rosenbaum et al. 1998). This variable is measured in micrograms per cubic meter (␮g/m3). Among the various pollutants measured are three air pollutants (benzene, ethylene, and ethylene dichloride) that are constituents of gasoline (as lead was prior to its discontinued use in that capacity); one pollutant (ethylene glycol) that is used in antifreeze and brake fluids; and three pollutants (trifluralin, naphthalene, and ethylbenzene) that are

219 estimated by the CEP to be the most highly concentrated in urban areas. One potential problem with including pollutants other than air-lead as an independent variable is that the included pollutants might be indirect indicators of lead levels. This may be a problem because many of the underlying causes of higher lead levels would also be related to elevated levels of other pollutants (e.g., manufacturing processes that emit lead and other pollutants as well). Moreover, there is sufficient justification for excluding a general indicator of air pollution as a covariate of crime, as there is no existing evidence that these pollutants affect crime rates. However, there is also reason to suspect that the association between air-lead levels and crime will be exaggerated if air pollution is not controlled. There is enough indirect evidence suggesting that air pollution may be associated with manufacturing processes that are also proximate to socially disorganized (e.g., crime elevated) areas (Shaw and McKay 1969). We therefore feel it necessary to include a general indicator of air pollution in the models we estimate. RESULTS The five hypotheses derived from the lead exposure literature are examined by means of the SpaceStat software package developed by Anselin (1999). To account for potential spatial dependence associated with clustering of similar counties, we introduce a spatial lag into each model (Baller et al. 2001). That lag is the weighted average of the crime rates (property or violent) of the five nearest counties. According to Baller et al. (2001) the spatial lag “incorporates the influence of the unmeasured independent variables and stipulates an additional effect of neighbor’s . . . [crime rates] (p. 567).” Langrange Multiplier diagnostic tests indicate that the inclusion of the spatial lag adequately accounts for spatial dependence in the model (available from the authors upon request). Hypothesis 1 Table 1 shows the coefficients and standard errors used to test the hypothesis that there is a positive association between air-lead levels and violent crime rates across counties. Model




TABLE 1. Maximum Likelihood Coefficients from the Spatial-Lag Regression of Violent Crime Rates on Air-Lead Levels Variables

Model 1

Model 2

Model 3

Air–Lead Levels

159*** .— .082*** [.319] .— [.164] (19.379) .— (4.329) Air Pollution .— .006 .004 .— [.028] [.017] .— (1.405) (0.918) Resource Deprivation .— .158*** .158*** .— [.134]– [.134] .— (7.004) (7.051) Population Structure .— ..235*** .071 .— [.199] [.060] .— (8.950) (1.562) % 15–29 Years Old .— .015*** .014** .— [.054] [.048] .— (3.323) (3.045) % Divorced .— .161*** .154*** .— [.273] [.252] .— (15.161) (14.431) % Unemployed .— .029* .024 .— [.033] [.027] .— (2.131) (1.684) South .— .003 .040 .— [.001] [.018] .— (.075) (.939) Rho .500*** .350*** .345*** (24.462) (16.354) (16.147) Intercept 3.469*** 1.862*** 2.386*** (25.930) (11.661) (11.887) .317 .440 .444 R2 Log–Likelihood –3790 –3560 –3543 N 2,772 2,772 2,772 * p < .05; ** p < .01; *** p < .001. two-tailed tests. Note: Unstandardized regression coefficients, standardized regression coefficients in brackets, and t-ratios in parentheses.

1 of Table 1 shows the effects of air-lead levels and the spatial lag (rho) on violent crime rates. That model explains nearly 32 percent of the variance in violent crime rates. Moreover, the effect of air-lead levels in model 1 is rather strong; a one standard deviation change in airlead is associated with a .32 standard deviation change in the violent crime rate. When we introduce additional covariates into the model, R2 increases from .32 (model 1) to .44 (model 3), and the effect of air-lead decreases from .16 to .08 (unstandardized coefficient). However, the standardized coefficients in model 3 of Table 1 demonstrate that the air-lead is still substantial when compared to more traditional explanatory variables. For instance, in model 3 a one standard deviation change in air-lead is associated with a .16 standard deviation change in the violent crime rate. Only the standardized coefficient for percent divorced is stronger (beta = .25). Also important is the finding that air pollution is not statistically significant (p = .36), suggesting that there is

something unique about the association between air-lead levels and violent crime rates that cannot be captured by air pollution measures in general. Hypothesis 2 Table 2 contains the coefficients and standard errors used to test the hypothesis that there is a positive association between air-lead levels and property crime rates. Again, the effect of air-lead is substantial. A one standard deviation change in air-lead is associated with a .45 standard deviation change in the property crime rate. When we introduce additional covariates into the model, R2 increases from .29 (model 1) to .34 (model 3), and the effect of air-lead decreases from .16 to .12 (unstandardized coefficient). Unlike the violent crime model, however, the standardized coefficient of air-lead in model 3 of Table 2 suggests that air-lead is the strongest predictor in the model


#1587—Jnl of Health and Social Behavior—Vol. 45 No. 2—45207-stretsky THE RELATIONSHIP BETWEEN LEAD AND CRIME

221

TABLE 2. Maximum Likelihood Coefficients from the Spatial-Lag Regression of Property Crime Rates on Air-Lead Level Variables Air-Lead Levels

Model 1 Model 2 Model 3 .162*** .— .115*** [.450] .— [.319] (25.242) .— (7.645) Air Pollution .— .002 –.001 .— [.013] [–.006] .— (.555) (–0.296) Resource Deprivation .— .057** –.057** .— [–.067] [–.067] .— (–3.113) (–3.165) Population Structure .— .223*** –.006 .— [.262] [–.007] .— (10.571) (–0.169) % 15–29 Years Old .— .015*** .013*** .— [.074] [.064] .— (4.141) (3.669) % Divorce .— .122*** .114*** .— [.287] [.266] .— (14.443) (13.334) % Unemployed .— .024* .016 .— [.038] [.025] .— (2.214) (1.447) South .— –.117*** –.067* .— [–.068] [–.039] .— (–3.520) (–2.001) Rho .279*** .213*** .204*** (11.603) (8.747) (8.398) Intercept 6.383*** 4.852*** 5.631*** (31.411) (23.933) (24.875) Log–Likelihood –2998 –2928 –2889 .289 .329 .343 R2 N 2,772 2,772 2,772 * p < .05; ** p < .01; *** p < .001. two-tailed tests. Note: Unstandardized regression coefficients, standardized regression coefficients in brackets, and t-ratios in parentheses.

(followed by percent divorced [beta = .23]). A one standard deviation change in air-lead is associated with a .32 standard deviation change in the property crime rate. Again, air pollution is not statistically significant in any of the models we estimate, suggesting that the association between air-lead levels and property crime rates is not the result of air-pollution in general.

provide little support for Pihl and Ervin’s findings. The partial effect of air-lead levels is much stronger in the property crime models (Table 2, models 1–3) than in the violent crime models (Table 1, models 1–3), suggesting that cross-county variation in air-lead levels is likely to produce a greater standard deviation change in the property crime rate than the violent crime rate.

Hypothesis 3

Hypothesis 4

According to Pihl and Ervin (1990) lead exposure has a stronger potential impact on violent crime than on property crime. If Pihl and Ervin’s research can be generalized to aggregate level studies of crime, the relationship between violent crime rates and air-lead levels should be stronger than the relationship between property crime rates and air-lead levels. However, the coefficients in Tables 1 and 2

The literature suggests that deprivation mediates the relationship between air-lead levels and crime rates. Logically, there should be empirical evidence that resource deprivation is an intervening variable that aids in explaining why air-lead levels are associated with crime rates (Baron and Kenny 1986). Table 3 shows the air-lead coefficients for property crime and violent crime with and without resource depri-




TABLE 3. Maximum Likelihood Coefficients from the Spatial-Lag Regression of Crime Rates on Air-Lead Levels and Resource Deprivation Violent Crime

Property Crime

Variables Air-Lead Levels

Model 1 Model 2 Model 1 Model 2 .159*** .178*** .162*** .162*** [.319] [.356] [.450] [.449] (19.379) (22.404) (22.242) (25.326) Resource Deprivation .— .279*** .— .027* .— [.236] .— [.032] .— (15.502) .— (1.980) Rho .500*** .434*** .279*** .282*** (24.462) (21.073) (11.603) (11.795) Intercept 3.469*** 3.910*** 6.383*** 6.375*** (25.930) (29.346) (31.411) (31.400) R2 .317 .385 .289 .290 Log-Likelihood –3791 –3668 –2998 –2996 N 2,772 2,772 2,772 2,772 * p < .05; ** p < .01; *** p < .001. two-tailed tests. Note: Unstandardized regression coefficients, standardized regression coefficients in brackets, and t-ratios in parentheses.

vation in the model. Those results suggest that resource deprivation does not mediate the effect of air-lead on violent or property crime rates. The relationship between violent crime, property crime, and air-lead levels is not diminished when resource deprivation is entered into the model (see Baron and Kenny 1986). Thus, we find little support for the hypothesis that resource deprivation mediates the effect between air-lead levels and crime rates. Hypothesis 5 Research on deprived populations and access to health care suggests that resource deprivation is likely to interact with air-lead levels. To test for an interaction effect we add the interaction term between air-lead level and resource deprivation into the model. The results of these tests are shown in Table 4 and clearly indicate that air-lead interacts with resource deprivation. Air-lead appears to explain a greater percentage of the variation in property and violent crime rates in counties with higher levels of resource deprivation than in counties with lower levels of deprivation. The interaction can be interpreted as suggesting that living in an economically deprived county aggravates the potential effect of airlead on crime while living in a less economically deprived county mitigates this effect. Figure 1 shows the strength of the association between violent crime rates, property crime rates, and air-lead levels for six different

levels of resource deprivation (based on equal groupings). The trends in Figure 1 reveal that the interaction is almost linear and the difference in the association between air-lead and crime (property and violent) increases equally over all levels of deprivation. The statistically significant interaction term supports the hypothesis that deprivation moderates the relationship between air-lead levels and crime rates and might indicate that residents in the most deprived counties are the least likely to afford or have access to health care that might prevent (screen and treat) lead poisoning. Both race and income are part of the resource deprivation factor. It is not possible, then, to determine if race and income are equally responsible for the observed interaction effect identified in Table 4 and in Figure 1. To expand upon this important issue we examine the size of the air-lead coefficients for five different groups of counties based on quintile scores for percentage African American and median family income. The results are presented in Table 5. The quintile scores for median family income and percentage African American indicate trends in coefficients that are consistent with the resource deprivation hypothesis and our finding that there is an interaction between resource deprivation and air-lead levels. That is, for median family income the unstandardized coefficient is largest for the first quintile (i.e., the sample of low income counties) and smallest for the fifth quintile (i.e., the sample of high income counties). Moreover, with one exception the relationship between air-lead levels and crime



223

TABLE 4. Maximum Likelihood Coefficients from the Spatial-Lag Regression of Crime Rates on Air-Lead Levels And Lead-Deprivation Interaction Violent Crime Rates

Property Crime RatesVariables

Model 1 Model 2 Model 1 Model 2 .183*** .083*** .170*** .118*** [.254] [.167] [.474] [.329] (22.989) (4.421) (26.852) (7.963) Air Pollution .— .001 .— –.006 .— [.004] .— [–.038] .— (0.230) .— (–1.883) Resource Deprivation .471*** .313*** .323*** .227*** [.554] [0.265] [.380] [.267] (10.837) (6.914) (9.508) (6.334) Population Structure .— .089 .— .027 .— [.075] .— [.032] .— (1.944) .— (0.742) % 15–29 Years Old .— .014** .— .014*** .— [.051] .— [.069] .— (3.140) .— (3.958) % Divorced .— .151*** .— .108*** .— [.256] .— [.254] .— (14.124) .— (12.888) % Unemployed .— .026 .— .020 .— [.030] .— [.032] .— (1.839) .— (1.828) South .— .040 .— –.061 .— [.017] .— [–.036] .— (0.941) .— (–1.849) Lead*Deprivation .039*** .031*** .059*** .057*** [.248] [.142] [.375] [.362] (4.849) (3.917) (9.458) (9.125) Rho .442*** .353*** .293*** .216*** (21.598) (16.537) (12.398) (8.973) Intercept 3.893*** 2.392*** 6.321*** 5.618*** (29.494) (11.952) (31.708) (25.085) Log-Likelihood –3656 –3543 –2951 –2858 .389 .447 .311 .362 R2 N 2,772 2,772 2,772 2,772 * p < .05 ** p < .01 *** p < .001. two-tailed tests. Note: Unstandardized regression coefficients, standardized regression coefficients in brackets, and t-ratios in parentheses.b * p < .05; ** p < .01; *** p < .001 (two-tailed tests). Air-Lead Levels

rates consistently decreases as median family income increases. In the case of percentage African American, the evidence of an interaction between race and air-lead levels is less pervasive, but still persistent. For instance, the coefficient for the fifth quintile—the group of counties with the highest percentage of African American residents—is the largest unstandardized coefficient (b = .248). The next largest coefficient, however, is observed for the sample of counties in the first quintile—the sample of counties with the lowest percentage of African Americans. The likely explanation for this finding is that racial segregation concentrates most African Americans in a relatively small number of counties and there is not much distinction in terms of overall racial inequality

between the four groups of counties that are less than 16 percent African American. Indeed, unlike the findings concerning median family income, the trends in the air-lead coefficients over the first four quintiles of African American counties is rather small. Exploratory Analyses Air-lead levels might interact with structural covariates of crime besides resource deprivation. To examine the potential for both mediating and moderating effects, we first refer to the information presented in Table 1 and Table 2. According to those results, the addition of air-lead levels to the analysis has little impact on regression coefficients for air pollution lev-




FIGURE 1. Association between Crime Rates and Air-Lead Concentrations by Level of Resource Deprivation

Note: Coefficients are adjusted for Air Pollution Levels, Resource Deprivation, Population Structure, % 15–29 Years Old, % Divorced, % Unemployed, and South.

els, resource depravation, percent divorced, percentage of the population aged 15 to 29 years old, and the Southern dummy indicator. It is unlikely, then, that air-lead levels play an important mediating role between those sociological variables and violent or property crime rates. It is, however, important to point out that the addition of air-lead levels to the crime models does reduce the coefficients for population structure and percentage unemployed. In the case of percentage unemployed, the coefficients are marginally significant in Table 1 (violent crime) and Table 2 (property crime) prior to the addition of air-lead levels, and they are insignificant once air-lead is controlled. However, it is important to point out that the direct effect of unemployment on crime is small, as indicated by the standardized coefficients. Moreover, the most conservative mediation tests suggest that coefficient differences are not statistically significant (see MacKinnon et al. 2002 for several of these tests).

In the case of population structure—which is highly significant and relatively important when air-lead levels are not in the model—the results are much more definitive. It appears that air-lead levels intervene between population structure and violent and property crime rates. When we add air-lead levels to the model, population structure is not statistically significant in either the property crime model or the violent crime model. Baron and Kenny (1986) refer to this condition as complete mediation. These results suggest that the large slope associated with population structure absent controls for air-lead may be largely due to the indirect effect of population structure on crime through air-lead emissions. In short, our exploratory analysis suggests that the production and concentration of air-lead may be one of the underlying mechanisms through which population structure influences crime rates. We found very little evidence that any structural covariates of crime examined in this research mediate the relationship between air-



225

TABLE 5. Maximum Likelihood Coefficients from the Spatial-Lag Regression of Crime Rates on Air-Lead Levels by Median Family Income and Percentage African American 1st Quintile [< 0.1%] 0.8%] .146*** (8.807) .347*** (6.925) 5.832*** (13.749) .209 554 [16.9%] 4.1%] 16.9%] Air-Lead Levels .125*** .092*** .130*** .248*** (8.846) (8.645) (12.780) (13.323) Rho .440*** .482*** .215*** .105 (9.087) (10.642) (4.001) (1.778) Intercept 5.010*** 4.496*** 6.883*** 8.337*** (12.265) (11.864) (15.117) (16.134) R2 .232 .274 .325 .294 N 555 554 554 555 Median Family Income [$8,401– [$10,332– [$12,060– [>$15,026] $10,331] $12,059] $15,026] Air-Lead Levels .237*** .163*** .177*** .138*** .101*** (8.378) (10.993) (12.681) (11.165) (11.461) Rho .244*** .321*** .336*** .318*** .360 (4.249) (6.332) (6.991) (6.246) (7.453) Intercept 7.050*** 6.237*** 6.132*** 6.079*** 5.590*** (13.860) (14.113) (14.809) (14.150) (13.550) R2 .151 .317 .343 .256 .264 N 551 552 551 553 550 * p < .05; ** p < .01; *** p < .001. two-tailed tests. Note: Unstandardized regression coefficients and t-ratios in parentheses.

lead levels and crime rates. Only the addition of percent divorced impacted the association between air-lead levels and crime rates by an appreciable amount, reducing the coefficient for air-lead levels from .159 to .101 in the violent crime model and from .162 to .130 in the property crime model (available from authors upon request). Although this difference might be interpreted as substantial, there is little theoretical support in the literature for the position that divorce rates mediate the relationship between air-lead levels and crime rates. Still, it is possible that air-lead levels disrupt interpersonal relationships which, in turn, may impact crime (Nevin 2000). It must be noted, however, that the lead to divorce to crime proposition is highly speculative and that differences in the air-lead coefficients are much smaller when other structural covariates of crime are controlled. To determine if air-lead levels interact with structural covariates of crime other than resource deprivation, we estimate several additional violent crime and property crime models to test interactions between air-lead levels and (1) air pollution, (2) population structure, (3) percent 15-29 years old, (4) percent divorced, (5) percent unemployed, and (6) South (available from authors upon request). The results of these exploratory analyses provide little support for lead interaction effects, as none of the

six interactions we tested are statistically significant in the violent crime model (p