Neighborhood-level built environment and social characteristics associated with serious childhood motor vehicle occupant injuries

Health & Place 17 (2011) 902–910

Contents lists available at ScienceDirect

Health & Place journal homepage: www.elsevier.com/locate/healthplace

Neighborhood-level built environment and social characteristics associated with serious childhood motor vehicle occupant injuries Glen D. Johnson a,b,n, Xiaoning Lu 1,c a

Office of the Medical Director, Division of Family Health, New York State Department of Health, 2162 Corning Tower, Albany, NY, USA Department of Environmental Health Sciences, School of Public Health, The University at Albany, State University of New York, Albany, NY, USA c Department of Epidemiology and Biostatistics, School of Public Health, The University at Albany, State University of New York, Albany, NY, USA b

a r t i c l e i n f o

a b s t r a c t

Article history: Received 15 June 2010 Received in revised form 18 April 2011 Accepted 21 April 2011 Available online 7 May 2011

The effect of residential neighborhood characteristics on a child’s risk of serious motor vehicle traffic occupant injuries was evaluated in New York State, USA, for the years 1993–2003, with particular focus on the effect of neighborhood walkability. Risk increased significantly (po0.0001) with decreasing street connectivity and as more workers commuted more than 30 min using means other than public transportation, along with more single-parent households and less college attainment in the neighborhood, regardless of whether New York City was in the study. After adjusting for age, gender and socio-economic community factors, the apparent loss of walkability in a child’s neighborhood increases their risk of serious injury as an occupant of a motor vehicle. & 2011 Elsevier Ltd. All rights reserved.

Keywords: Neighborhood walkability Childhood motor vehicle injury New York State Spatial negative binomial regression

1. Introduction Is the risk of serious injury to children and teens from Motor vehicle traffic (MVT) crashes affected by their place of residence? Answering this question is critical for a comprehensive understanding of the causes behind such tragic events. On one hand, the number of motor vehicle-related deaths per vehicle miles traveled and per population has declined in the United States (US) for all ages since the mid-1960s. This decline is associated with ever-improving safety features and more aggressive law enforcement to prevent driving under the influence of alcohol and other drugs (National Highway Traffic Safety Administration, 2004). While on the other hand, the number of vehicle miles traveled continues to increase (National Highway Traffic Safety Administration, 2006; US Dept. of Transportation, 2011), so the overall public health burden has essentially not declined (Richter et al., 2001). Although there is evidence of a recent decline in the number of fatalities (National Highway Traffic Safety Administration, 2010), MVT crashes remain the leading cause of death in the US for ages 3–34 years old (National Highway Traffic Safety Administration, 2008) and overall crash injuries result in about 500,000 hospitalizations and 4 million

n Corresponding author. Permanent address: Department of Health Sciences, Lehman College, City University of New York, 413 Gillet Hall, 250 Bedford Park Blvd West, Bronx NY 10468, USA. Tel.: þ 1 718 960 8775; fax: þ1 718 960 8908. E-mail address: [email protected] (G.D. Johnson). 1 Present address: Division of Clinical Informatics, Beth Israel Deaconess Medical Center, An Affiliate of Harvard Medical School, 1330 Beacon St, Brookline, MA 02446, USA.

1353-8292/$ - see front matter & 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.healthplace.2011.04.009

emergency department visits annually for all ages (Task Force on Community Preventive Services, 2001). There are clearly factors beyond law enforcement and automobile safety features that need to be addressed. If risk is proportional to exposure, regardless of safety features, then risk of serious injury from car crashes is expected to increase with increasing time spent in cars and travel speed, both of which may be affected by one’s residential built environment (Frumkin, 2002, 2006). Fatalities are well known to be much higher among rural (versus urban) car crashes (Jones et al., 2008; Kmet and Macarthur, 2006; National Highway Traffic Safety Administration, 2007); however, these observations are based on location of crashes, not on location of victim’s residences. Meanwhile, a separate line of research has evolved that addresses the effect of one’s residential built environment on physical activity and related health issues (Davis and Jones, 1996; Frumkin, 2002; Frumkin et al., 2004; Handy et al., 2002; Ewing et al., 2006; Frank and Engelke, 2005; Frank et al., 2003, 2004; Nelson et al., 2006; Lopez and Hynes, 2006). These studies support that neighborhoods with well connected local streets and mixed land use, such as older towns and cities, provide a more ‘‘walkable’’ environment than suburban housing developments that are characterized by disconnected streets and homogeneous land use. After adjusting for social and demographic factors, neighborhoods with higher street connectivity and mixed land use tend to have residents who report greater physical activity and less overweight/obesity (Ewing et al., 2006; Frank et al., 2004; Nelson et al., 2006). This effect is especially noted among adults, although it is somewhat limited with childhood ages, which may be from parents driving their children to different

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

sporting activities (Nelson et al., 2006). If true, then parents with the best intentions for their children’s well being may also be exposing them to increased risk of a car crash. It is therefore logical to hypothesize that the risk of serious injury or outright death of children of all ages as an occupant in a motor vehicle traffic crash increases as the walkability of a child’s residential neighborhood decreases. The theoretical basis is that more time is spent in cars as neighborhood walkability decreases, thus increasing exposure, especially if day-to-day automobile travel is at fairly high speeds. Therefore, one would expect the population-based rate of serious MVT-occupant injuries to increase as residential walkability decreases. The hypothesis stated above is supported in a study by Ewing et al. (2003), who showed a strong positive association between the degree of ‘‘urban sprawl’’ and both MVT occupant and pedestrian fatalities for all ages, based on a county-level index of sprawl for 448 metropolitan counties in the US. While this was pioneering research, counties are simply too large geographically to represent residential neighborhoods, given the heterogeneity among neighborhoods across most counties. A more detailed study of local authority districts in England and Wales indicated that both fatal and nonfatal injuries decreased with increasing percentage of roads classified as ‘‘minor’’, after adjusting for other multilevel covariates (Jones et al., 2008); however, as the percentage of roads passing through urban areas increased, the risk of fatal injuries decreased while the risk of slight (non-serious) injuries increased. As with the other studies cited above, theirs was based on characteristics of crash site locations, whereas our interest lies with characteristics of where someone lives, particularly children. We evaluated the hypothesis stated above by analyzing neighborhood-level rates of deaths and non-fatal inpatient hospitalizations of children from MVT-occupant injuries using postal ZIP codes of residence in New York State (NYS) to approximate neighborhoods (see Fig. 1). Data were obtained from population databases of deaths and hospitalizations for children covering eleven years in NYS. ZIP code-level rates by age and gender were compared to ZIP code-level variables reflecting neighborhood walkability and commuting behavior, after controlling for select

903

demographic and socio-economic characteristics. Street connectivity provided a proxy for walkability, represented as the density of four-way or greater intersections of local roads/streets (Handy et al., 2002; Ewing et al., 2006; Frank et al., 2004; Nelson et al., 2006), as illustrated in Fig. 2.

2. Methods This cross-sectional study focuses on children in NYS who were 28 days through 18 years old when killed or admitted to a hospital with non-fatal injuries from being the occupant of a motor vehicle that was in a traffic crash from the years 1993 through 2003. The design is intentionally ecologic to evaluate relative risk among neighborhoods defined by different neighborhood-level covariables. The chosen age range stems from a larger project to develop a child death review program (www.child deathreview.org/) in NYS. 2.1. Cases Cases were identified from NYS and New York City (NYC) vital records and statewide hospital discharge records. Access to death records was approved by the NYS Department of Health (DOH) Institutional Review Board, along with the DOH Bureau of Biometrics and Health Statistics and the NYC Department of Health and Mental Hygiene, Bureau of Vital Statistics. Inpatient hospital discharge records were obtained through a de-identified version of the NYS Statewide Planning and Research Cooperative System (SPARCS) database, with approval from the DOH Bureau of Biometrics and Health Statistics. Cases included injuries that resulted in death, as identified from death records, or were non-fatal but required admission to a hospital, as identified by inpatient hospitalizations where the disposition upon discharge was not ‘‘expired’’. A child may be repeated in the database over time since concern is with total injuries, not just total individuals. See the Appendix for codes from the International Classification of Diseases that were chosen

Fig. 1. Location of New York State in the northeast United States, with enlargement of county and postal ZIP code boundaries within New York State.

904

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

Fig. 2. Illustration of different neighborhoods based on street connectivity and population size, based on areas around Albany NY, USA.

for identifying cases of MVT-occupant injuries. The final number of cases for this study equaled 16,286. Cases were aggregated by their residential postal ZIP code, gender and two age groups (28 days through 9 years and 10 through 18 years) to allow analysis of ZIP code-level community variables after adjusting for age and gender. Cases were not aggregated by race and ethnicity categories because these fields are incomplete and internal DOH studies have shown these fields to be unreliable from death and hospital records. This is not considered a limitation since race and ethnicity are not very relevant for this particular outcome, whereas age and gender are well known risk factors (Committee on Injury, Violence and Poison Prevention and Committee on Adolescence, 2006; National Highway Traffic Safety Administration, 2006). 2.2. Consideration of the distance of crashes from home Since the basic theory is that one’s residential environment controls how much time is spent in automobiles on a daily basis, for example to commute to school or work, it is assumed that most childhood MVT-occupant injuries occur ‘‘close to home’’. If many crashes occur far from home, then it would be difficult to associate with the residential environment since most people travel long distance at least occasionally, regardless of where they live. To evaluate the validity of this assumption, hospital records were used to measure Euclidean distance between the hospital and the geographic centroid of a patient’s residence ZIP code as an

indicator of how far from home the MVT crashes occur, assuming that crash victims are taken to an appropriate hospital that is closest to the crash site. The distribution of distance was consistent for five age categories (infants, 1–4, 5–10, 11–14 and 15–18 year olds), and it was clear that the majority of MVT crashrelated hospitalizations occur close to home for all ages (75% o27 km, 90% o54 km). For example, if traveling at a moderate highway speed of 100 km/h, 90% of these hospitalizations occur in less than approximately a half hour drive from home. 2.3. ZIP code management Postal ZIP codes are not the ideal geographic unit for epidemiological investigations (Krieger et al., 2002), since they are rather arbitrary boundaries used for efficient delivery of mail, which are also subject to change over time; whereas a US census tract is intended to be a relatively permanent county subdivision designed to be somewhat homogeneous with respect to population characteristics, socio-economic status and living conditions. There are, however, several reasons why ZIP codes are the only sub-county choice for this study, as they are for many public health studies that use a population database. In order to associate individual residence addresses from death and hospital records with regions like census tracts, one would need to geocode all records to the level of street address, and then capture the geocoded point locations within their respective tracts. Automated geocoding is typically reliable for around 80% of medical

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

905

records, at best (Cayo and Talbot, 2003; Krieger et al., 2001). That means geographic coordinates for the remaining records must be obtained through manual research using a variety of resources, which is too laborious to be practical when considering around 20% of over 16,000 records. Simply excluding records that do not geocode is unacceptable since this will likely impose a strong selection bias on the study (Cayo and Talbot, 2003). Meanwhile, the residence ZIP code is a nearly complete field, with only 0.04% missing among all death and hospital records considered for this study; so is therefore readily available as a sub-county geographic unit of observation that is small enough to at least approximate neighborhoods. This is certainly a large improvement over studies that use whole counties as observational units (Aguero-Valverde and Jovanis, 2006; Ewing et al., 2003, 2006). After summing the cases by age group and gender within each ZIP code, then any ZIP code that was enclosed by another ZIP code, as of the year 2000, was merged into the enclosing ZIP code, as per the year 2000 ZIP code inventory file purchased from GDT CorporationTM. This file is updated annually, but the year 2000 was chosen since population/demographic and socio-economic variables are being drawn from the 2000 census, plus it is within the range of years (1993–2003) used for obtaining the cases. The final data set for analysis consisted of 1548 ZIP codes, each with four demographic cells (2 age by 2 gender categories).

binomial (NB) random variable, where the mean count for either distribution is modeled via a canonical log link to a linear function of the predictors to estimate relative risk among the categories of each predictor. These models were further evaluated with zeroinflation, using the conditional approach (Agarwal et al., 2002). The NB model performed much better than the Poisson for fitting the zero-truncated component of the zero-inflated models, and also fit very well without the need for zero-inflation. Therefore the increased complexity of a zero-inflated model was not supported. The effect of spatial location was also evaluated to account for possibly unidentified spatially varying covariables in the model. For this, a random effect was added for the county within which the majority of a ZCTA lied, resulting in a model with a random addition to the intercept (Waller and Gotway, 2004, p. 384). Therefore, the probability of observing yis cases in a cell defined by the ith ZCTA and s demographic stratum is

2.4. Population and community-level social variables

mis ¼ nis expðx0is b þ lc Þ

Each ZIP code was matched to a US census ZIP code tabulation area (ZCTA), which geographically approximates actual ZIP codes and provides a source of population and social variables from the census SF1 and SF3 files, respectively. The population at risk was defined as the number of children in each ZCTA by gender and age group. The demographic and social variables were chosen to represent different aspects of a community; specifically, income, race/ethnicity, urban status, immigration from foreign countries, transience or domestic migration, education, family structure, household vehicle ownership and commuting behavior, resulting in an initial set of 39 community-level variables. This large set was analyzed both graphically and through principal components and factor analyses (Morrison, 1976) to identify correlations and reduce to a more parsimonious set.

for covariate vector xis, linear coefficient vector b and countyspecific random effect lc, which is assumed to be iid N(0,sc 2 ) for inter-county variance sc 2 . The term a is a scale parameter that is estimable from the data and increases as the variance exceeds the mean, thus allowing more flexible modeling than the Poisson distribution alone, which is a special case of the NB distribution when a-0. Community-level covariables were all categorized into tertiles representing the lower, middle and higher thirds of the distribution of values among the ZCTAs. This approach avoids the need to ensure a linear relationship between each covariate and the response, while also allowing presentation of the results as risk of a ‘‘medium’’ or ‘‘low’’ category relative to a ‘‘high’’ category. Maintaining the covariables as continuous measurements may be potentially more informative, but with the tradeoff of being more difficult to interpret and less robust with respect to ensuring linear relationships. Modeling was carried out with the GLIMMIX procedure (Little et al., 2006) in SAS 9.1.3, using the default approach of pseudolikelihood (Wolfinger and O’Connell, 1993). Overall model fit was diagnosed by comparing the generalized chi-square statistic to the degrees of freedom, which are equivalent when the model is properly specified. Observed values were graphed and regressed against the model-predicted values to assess model predictability. Since the observed number of cases depends on varying underlying population sizes, Pearson-type residuals were analyzed because they are standardized by the square root of the variance of their respective observations. Residuals were assessed for outliers and independence with respect to the model-predicted values, and were also tested for spatial autocorrelation by Moran’s I statistic (Lin and Zhang, 2007) to evaluate the impact of adding a spatial random effect. Separate analyses were applied to the whole state of New York and all of the state outside of NYC. This is because NYC contains approximately 40% of the state’s population within an urban environment that is by far the largest city in the US. Excluding NYC therefore yields results with more external validity since the remainder of New York is a mix of urban, suburban and rural neighborhoods that more likely reflects the greater US.

2.5. Street type and connectivity Street centerline files were obtained from the NYS Office of Cyber Security and Critical Information Infrastructure. Arc ObjectsTM was programmed for using ArcMap 9.0TM to extract all nodes that represent 4-way or greater intersections of local streets/roads (Feature Class Code of A4 includes local, neighborhood, rural roads and city streets). The nodes density (per km2) was then calculated for each ZCTA to represent street connectivity. There are alternative indices of street connectivity that were not obtained since they are highly correlated with node density (Nelson et al., 2006). The proportion of total road length with a Feature Class Code of A4 was also calculated for each ZCTA, since these roads are generally considered more walkable (Nelson et al., 2006), regardless of the presence of sidewalks. 2.6. Modeling Let yis be the count of injuries in each of i¼1, y, n ZCTAs and s equals one of four demographic strata within each ZCTA. As a count of a relatively rare event, yis was considered to be distributed as either a Poisson or the more generalized negative

PðYis ¼ yis 9mis , a, lc Þ ¼

Gðy þ a1 Þ Gðy þ 1ÞGða1 Þ



a1

a1 

a1 þ m

mis

y

a1 þ mis

for a Z0 and G(  ) is the gamma function (Cameron and Trivedi, 1998), where the mean for each cell is linked to the covariables by the log link, after offsetting by the population within each cell, nis, such that

906

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

3. Results After multivariate analyses of the initial 39 community-level census-based covariables, a much smaller, less inter-correlated set was obtained. These were added to the set of built environment covariables for modeling both the whole state and excluding NYC. Any covariable that was significant at the 10% chance of a type 1 error for at least one of these geographic delineations, after adjusting for all other covariables, but not conditioning on county, is reported in Table 1 along with the population and cases within each category. The low, medium and high categories in Table 1 correspond to the lower, middle and upper thirds (tertiles) of the ZCTAs based on the distribution of the respective covariable. Prior to adding the spatial (county) random effect, residual analyses revealed four extreme outlying ZCTAs, which turned out to be either a military installation or a very rural location. These were removed from the study and subsequent models are presented in Table 2, both with and without the spatial random effect. Model diagnostics, listed at the bottom of Table 2, indicate that adding the spatial random effect improves the model with respect to specification (chi-square/degrees of freedom closer to 1) and predictability (higher R2 from regressing observed versus predicted values), while also lowering the residual spatial autocorrelation. Furthermore, differences in the linear coefficients between the two models in Table 2 reveal unmeasured confounders that co-vary spatially, which are somewhat controlled for by the spatial random effect. When the results are conditioned on county, ‘‘median income’’ is rendered insignificant and the effect of ‘‘proportion of local

roads’’ is weakened. The effect of race is strengthened, whereby the risk increases as more of the population claims to be ‘‘white alone’’. The protective effects of a more educated community and less single parent households remain statistically significant. The two census-derived measures of commuting also retain their significance, revealing an increasingly protective effect associated with a decrease in the proportion of workers who drive alone to work or commute more than 30 min using means other than public transportation. Finally, the risk increases significantly as street connectivity decreases. A final model was developed that conditions on county as a random effect and eliminated median household income and the proportion of local roads. This simpler model was then fit to data for the whole state and excluding NYC, as shown in Table 3. As expected, females, and especially younger children, are at less risk than teen boys (National Highway Traffic Safety Administration, 2006); however, it is interesting to see that when NYC is removed from the analyses, the protective effect is slightly increased for younger children, while decreased for older females. The mild protective effect associated with less people claiming to be white alone is consistent regardless of including NYC. As the proportion of single-parent households decreases, there is a protective effect, although its statistical significance is weak. The proportion of workers in a community who drive alone to work also shows a weakened effect after removing NYC, although there is a significantly decreasing risk as this proportion decreases when NYC is included. The remaining covariables reveal a consistent effect with respect to direction, magnitude and statistical significance (po0.0001),

Table 1 Covariables considered for modeling, categorized as low, medium and high according to tertiles (shown in parentheses), with the population of children at risk and cases within each category. Label

Description

Level

Population

Age  gender

Age 28 days–9 years versus 10–18 years by gender

Female o 10 Female 10–18 Male o 10 Male 10–18

1,266,088 1,146,620 1,325,032 1,209,491

1762 5835 2006 8605

Low (1.73–91.75) Med (91.75–97.37) High (97.37–100)

3,395,906 1,132,680 418,645

9830 5514 2864

Community level social and demographic covariables White alone Percent of total population that is white alone (no other additional race)

Cases

Median income

Median household income in 1999 (  $1000)

Low (0–36.52) Med (36.52–48.92) High (48.92–285.7)

1,832,110 1,251,256 1,863,865

5412 4961 7835

College

Percent of the population over age 24 years who have a bachelors degree or higher

Low (0–15.12) Med (15.12–25.54) High (25.54–100)

1,364,513 1,790,363 1,792,355

4962 7089 6157

Single parent

Percent of total households that have at least one child under 18 years old and only one parent at home

Low (0–21.95) Med (21.95–29.15) High (29.15–89.32)

1,413,325 1,055,283 2,478,623

5822 4849 7537

Drive alone

Percent of workers 16 years and over who do not work at home who drive alone to work

Low (0–73.05) Med (73.05–80.70) High (80.70–100)

2,950,648 984,082 1,012,501

7884 5079 5245

Long commute

Percent of workers 16 years and over who do not work at home that travel at least 30 min to work by means other than public transportation

Low (0–25.67) Med (25.67–37.48) High (37.48–81.67)

2,638,239 1,590,559 718,433

6896 7068 4244

Low (0–70.96) Med (70.96–81.00) High (81.00–100)

2,135,446 1,950,847 860,938

6573 7622 4013

Low (0–0.13) Med (0.13–1.30) High (1.30–110.41)

272,070 975,607 3,699,554

2046 5275 10,887

Built environment covariables Local roads Local and secondary streets/roads (class A4) as a proportion of total road length Street connectivity

Density of 4-way or greater intersections of local and secondary streets/roads, per square kilometer

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

907

Table 2 Model results, presented as coefficient estimates and corresponding p-values with and without a spatial random effect (whole state). Variable

Level

Without spatial effect Coefficient

With spatial random effect p-Value

Coefficient

p-Value

Age  gender

Female o 10 Female 10  18 Male o10 Male 10–18

 1.678  0.329  1.587 Ref

o 0.0001 o 0.0001 o 0.0001

 1.658  0.328  1.568 Ref

o 0.0001 o 0.0001 o 0.0001

White alone

Low Medium High

 0.128  0.001 Ref

0.0023 0.7561

 0.118  0.068 Ref

0.0790 0.0574

Median income

Low Medium High

 0.097  0.131 Ref

0.0378 0.0002

 0.033  0.056 Ref

0.4611 0.1250

College

Low Medium High

0.274 0.190 Ref

Single parent

Low Medium High

 0.095  0.005 Ref

0.0270 0.8700

 0.139  0.016 Ref

0.0010 0.5993

Drive alone

Low Medium High

 0.301  0.085 Ref

o 0.0001 0.0063

 0.176  0.106 Ref

o 0.0001 0.0006

Long commute

Low Medium High

 0.453  0.078 Ref

o 0.0001 0.0060

 0.335  0.112 Ref

o 0.0001 o 0.0001

Local roads

Low Medium High

 0.098  0.064 Ref

0.0014 0.0309

 0.043  0.058 Ref

0.1724 0.0411

Street connectivity

Low Medium High

0.334 0.220 Ref

o 0.0001 o 0.0001

o 0.0001 o 0.0001

0.233 0.144 Ref

0.380 0.254 Ref

Overdispersion parameter

0.169 (0.01)a

0.121(0.009)a

Chi-square/df

1.32

1.26

R2 obs. versus pred.b

0.71

0.77

Moran’s I of Pearson residualsc Female o 10 Female 10–18 Male o10 Male 10–18

0.07 0.08 0.06 0.12

b c

o 0.0001 o 0.0001

0.048 (0.011)a

County variance

a

o 0.0001 o 0.0001

(p o 0.0001) (p o 0.0001) (p o 0.0001) (p o 0.0001)

0.05 0.04 0.06 0.07

(p ¼ 0.0023) (p ¼ 0.0056) (p ¼ 0.0004) (p ¼ 0.0001)

Standard error in parentheses. R-squared from regression of observed versus predicted values. Global Moran’s I based on first order queen contiguity weight (Lower coefficients and higher p-values indicate less clustering).

regardless of whether NYC is included. Risk increases as either the community-level college attainment decreases; or as the proportion of workers who commute over 30 min by means other than public transportation increases; or as the street connectivity decreases, regardless of including NYC.

4. Discussion Childhood risk of serious injury from being an occupant in a motor vehicle traffic crash appears to increase as the walkability of a child’s residential neighborhood decreases. This observation is supported by evidence from New York State where children of the same gender, age group and similar socio-economic status of their neighborhood reveal significantly increasing risk as their neighborhoods have decreasing connectivity of local streets, along with an increasing proportion of workers who have long commutes by means other than public transportation. Note that risk also

increased significantly as ‘‘household vehicle ownership’’ increased, but this census-derived variable was not included in our model because it was very highly linearly correlated with the proportion of workers who have long commutes by means other than public transportation. We further experimented with adjusting for total population density (people per km2) since this is highly correlated with local street connectivity. While significant, population density did not affect the direction, magnitude or significance level of street connectivity. A closer investigation revealed that many ZIP codes in New York State with over 90% local roads were actually very rural; therefore, both urban and rural areas can have similarly high proportions of local roads, although their walkability and MVT crash risk are very different. This explains why the proportion of local roads was not a strong predictor of MVT-occupant injury risk when evaluating all of New York State; however, it may be valuable for comparing different neighborhoods that are

908

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

Table 3 Final model, including a spatial random effect. Relative risk and associated 95% confidence intervals are presented for fixed effects, along with other parameter estimates for the whole state and excluding NYC. Variable

Level

Whole state

Excluding NYC

RR (95% CI)

RR (95% CI)

Age  gender

Female o 10 Female 10–18 Male o 10 Male 10–18

0.190 (0.178, 0.202) 0.720 (0.687, 0.755) 0.208 (0.196, 0.221) Ref

0.178 (0.163, 0.194) 0.761 (0.718, 0.808) 0.199 (0.183, 0.216) Ref

White alone

Low Medium High

0.893 (0.820, 0.972) 0.937 (0.873, 1.004) Ref

0.858 (0.783, 0.942) 0.951 (0.914, 1.021) Ref

College

Low Medium High

1.272 (1.190, 1.360) 1.158 (1.096, 1.225) Ref

1.329 (1.220, 1.450) 1.141 (1.057, 1.231) Ref

Single parent

Low Medium High

0.857 (0.799, 0.919) 0.990 (0.934, 1.049) Ref

0.863 (0.785, 0.947) 0.984 (0.914, 1.057) Ref

Drive alone

Low Medium High

0.856 (0.793, 0.924) 0.914 (0.864, 0.969) Ref

0.975 (0.894, 1.063) 0.932 (0.872, 0.996) Ref

Long commute

Low Medium High

0.710 (0.662, 0.761) 0.895 (0.847, 0.945) Ref

0.750 (0.690, 0.815) 0.903 (0.841, 0.970) Ref

Street connectivity

Low Medium High

1.489 (1.349, 1.642) 1.295 (1.209, 1.389) Ref

1.560 (1.393, 1.745) 1.397 (1.277, 1.529) Ref

County variance

0.048 (0.010)a

0.019 (0.006)

Overdispersion parameter (a)

0.123 (0.009)

0.111 (0.013)

Diagnostics Chi-square/df R2 obs. versus pred.b

1.260 0.770

1.31 0.73

a b

Standard error in parentheses. R-squared from regression of observed versus predicted values.

otherwise similar, such as variations of suburban neighborhoods (Nelson et al., 2006). The complex dependence of risk on both the built environment and local culture is seen by the increased risk for children from neighborhoods with less college attainment and a higher proportion of single-parent households. This highlights the importance of adjusting for variables reflecting community composition since two neighborhoods with similar built environment characteristics may nurture different levels of overall risky behavior that is associated with different socio-economic conditions. This evidence should be considered in light of a 114% increase in the per capita vehicle miles traveled by the US population from 1966 to 2006 (National Highway Traffic Safety Administration, 2006), which is offsetting the gains made from ever-improving automobile and traffic safety features since the mid-1960s (National Highway Traffic Safety Administration, 2004) when the National Highway Traffic Safety Administration was established. Automobile dependence has steadily risen with increasing suburban housing developments that arose since the 1950s. Residents in these suburban ‘‘neighborhoods’’ may live very close to retail shopping, schools, etc., yet cannot safely walk or bicycle to these destinations and may have limited or no public transportation alternatives, unlike residents of older cities and towns where land use is mixed, streets are well connected and public transportation is readily available (Frank et al., 2004). While physical inactivity and overweight/obesity are increasingly tied to living in car-dependent neighborhoods, serious injury and outright death from MVT crashes are obviously much

more critical public health concerns. This is especially true when children are the victims and MVT crashes are the leading cause of death and serious injury for all ages up to the mid-thirties. Interventions to reduce automobile-dependence for routine day-to-day use may require modifications of the built environments where people reside. This starts with urban and regional planning that acknowledges the actual environmental and public health impacts of car-dependent living (i.e. Urban Land Institute, 2010; Heishman and Dannenberg, 2008). An advantage of this study is that it evaluates a population database, instead of a small sample, spanning eleven years. Also, while ZIP codes may not be the ‘‘best’’ way to represent local neighborhoods, they present much greater explanatory power compared to using whole counties, as has so far been the finest geographic resolution for this type of study in the USA. While the results show evidence that a child is at lower risk if he/she resides in a neighborhood with higher local street connectivity, this study cannot directly attribute this observed effect to less time spent in cars and more walking, bicycling and use of public transportation. However, these findings agree with surveys that show individuals do spend more time in cars as local street connectivity decreases (Frank et al., 2004).

5. Recommendations for future research This research would be improved by incorporating covariables derived from real property data that characterize land use mix as

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

a key indicator of neighborhood walkability (Frank et al., 2004; Handy et al., 2002). Such data is readily accessible for most of New York State; however, there are a few highly populated counties that are exceptions. This makes it difficult to study the entire state’s population, although a sample of communities could be evaluated. Aerial imagery can be analyzed in conjunction with real property data to characterize land use mix. Imagery can also be used to identify sidewalks, an obviously strong indicator of walkability that is not available through other data sources. Analyzing imagery, however, may be laborious since images are ‘‘dumb’’ in that they can only be viewed but not queried like other GIS layers. Also, a rather simple approach was taken to control for residual spatial autocorrelation by adding counties as a random addition to the intercept. The statistical model may be improved by a more refined adjustment of residual spatial autocorrelation, such as through a conditional autoregressive model (Besag, 1974) where parameters are described by posterior distributions from a fully Bayesian solution (Johnson, 2004) and even further through fully Bayesian modeling that incorporates space and time interaction (Aguero-Valverde and Jovanis, 2006). This study focused on children 28 days through 18 years old, as part of a larger project to develop a child death review program in NYS that applies to this age group. A logical next step is to focus on all ages from early childhood to the early thirties since this is the overall highest risk age group (National Highway Traffic Safety Administration, 2008). Finally, the effect on MVT pedestrian injuries requires investigation. We purposely excluded pedestrian injuries since they may be associated with different risk factors and should therefore be studied separately. When New York City is compared to the rest of New York State over the years 1993–2003, the fatality rate for pedestrians and bicyclists who are struck by motor vehicles is very similar for young children (o10 years old) and higher outside of NYC for adolescents and teens (10–18 years old). Since one may intuitively expect higher rates in NYC where both automobile and pedestrian volume are high and in close proximity, then much remains to be learned about why the rates are actually lower in NYC for older children and about equal for younger children when compared to NYS outside of NYC.

Acknowledgments This work was supported in part by the US Department of Health and Human Services, Health Resources and Services Administration through the Maternal and Child Health Block Grant to New York State. The content is solely the responsibility of the authors and does not necessarily represent the official views of the Health Resources and Services Administration or the New York State Department of Health. The authors appreciate the assistance of Lifeng Guo in the early stages of this research, and of Peter Herzfeld in the New York State Department of Health, Bureau of Biometrics and Health Statistics for expeditiously providing electronic death records after receiving approval. We are also grateful for the substantial help of an anonymous peer reviewer, along with Dr. Lance Waller of Emory University for advice and valuable conversation about the statistical modeling methods applied herein.

Appendix External injury codes of the International Classification of Diseases (ICD) that were chosen for specifically identifying cases of MVT-occupant injuries.

909

For death records from 1993 to 1998, and hospital discharge records for all years, ICD version 9 codes were: E8100, E8101, E8102, E8103, E8110, E8111, E8112, E8113, E8120, E8121, E8122, E8123, E8130, E8131, E8132, E8133, E8140, E8141, E8142, E8143, E8150, E8151, E8152, E8153, E8160, E8161, E8162, E8163, E8170, E8171, E8172, E8173, E8180, E8181, E8182, E8183, E8190, E8191, E8192 and E8193. For death records from 1999 to 2003, ICD version 10 codes were: (V30 through V79 with fourth character subdivisions .4 through .9) or (V83 throughV86 with fourth character subdivisions .0 through .3) or (V20 throughV28 with fourth character subdivisions .3 through .9) or (V29 with fourth character subdivisions .4 through .9). References Agarwal, D.K., Gelfand, A.E., Citron-Pousty, S., 2002. Zero-inflated models with application to spatial count data. Environ. Ecol. Stat. 9, 341–345. Aguero-Valverde, J., Jovanis, P.P., 2006. Spatial analysis of fatal and injury crashes in Pennsylvania. Accid. Anal. Prev. 38, 618–625. Besag, J.E., 1974. Spatial interaction and the statistical analysis of lattice systems (with discussion). J. R. Stat. Soc. B 36, 192–236. Cameron, A.C., Trivedi, P.K., 1998. Regression Analysis of Count Data. Cambridge University Press. Cayo, M.R., Talbot, T.O., 2003. Positional error in automated geocoding of residential addresses. Int. J. Health Geogr. 2, 10 /http://www.ij-healthgeo graphics.com/content/2/1/10S. Committee on Injury, Violence and Poison Prevention and Committee on Adolescence, 2006. The teen driver. Pediatrics 118, 2570–2581. Davis, A., Jones, L.J., 1996. Children in the urban environment: an issue for the new public health agenda. Health and Place 2, 107–113. Ewing, R., Brownson, R.C., Berrigan, D., 2006. Relationship between urban sprawl and weight of United States youth. Am. J. Prev. Med. 31 (6), 464–474. Ewing, R., Schieber, R.A., Zegeer, C.V., 2003. Urban sprawl as a risk factor in motor vehicle occupant and pedestrian fatalities. Am. J. Public Health 93, 1541–1545. Frank, L.D., Andresen, M.A., Thomas, L., Schmid, T.L., 2004. Obesity relationships with community design, physical activity and time spent in cars. Am. J. Prev. Med. 27 (2), 87–96. Frank, L.D., Engelke, P., 2005. Multiple impacts of the built environment on public health: walkable places and the exposure to air pollution. Int. Reg. Sci. Rev. 18 (2), 193–216. Frank, L.D., Engelke, P.O., Schmid, T.L., 2003. Health and Community Design: The Impact of the Built Environment on Physical Activity. Island Press. Frumkin, H., 2006. Cities, suburbs and urban sprawl. In: Freudenberg, N., Galea, S., Vlahov (Eds.), Cities and The Health of the Public. Vanderbilt University Press, pp. 143–175. Frumkin, H., Frank, L.D., Jackson, R.J., 2004. Urban Sprawl and Public Health: Designing, Planning, and Building for Healthy Communities. Island Press. Frumkin, H., 2002. Urban sprawl and public health. Public Health Rep. 117, 201–217. Handy, S.L., Boarnet, M.G., Ewing, R., Killingsworth, R.E., 2002. How the built environment affects physical activity, views from urban planning. Am. J. Prev. Med. 23 (2S), 64–73. Heishman, H., Dannenberg, A.L., 2008. Influencing the built environment in your community. J. Environ. Health 71 (4), 66–67. Johnson, G.D., 2004. Small area mapping of prostate cancer incidence in New York State (USA) using fully Bayesian hierarchical modelling. Int. J. Health Geogr. 3, 29 /http://www.ij-healthgeographics.com/content/3/1/29S. Jones, A.P., Haynes, R., Kennedy, V., Harvey, I.M., Jewell, T., Lea, D., 2008. Geographical variations in mortality and morbidity from road traffic accidents in England and Wales. Health and Place 14, 519–525. Kmet, L., Macarthur, C., 2006. Urban–rural differences in motor vehicle crash fatality and hospitalization rates among children and youth. Accid. Anal. Prev. 38, 122–127. Krieger, N., Waterman, P., Chen, J.T., Soobader, M.J., Subramanian, S.V., Carson, R., 2002. ZIP code caveat: bias due to spatiotemporal mismatches between ZIP codes and US census-defined geographic areas—the Public Health Disparities Geocoding Project. Am. J. Public Health 92, 1100–1102. Krieger, N., Waterman, P., Lemieux, K., Zierler, S., Hogan, J.W., 2001. On the wrong side of the tracts? Evaluating the accuracy of geocoding in public health research. Am. J. Public Health 91, 1114–1116. Lin, G., Zhang, T., 2007. Loglinear residual tests of Moran’s I autocorrelation and their applications to Kentucky breast cancer data. Geogr. Anal. 39, 293–310. Little, R.C., Milliken, G.A., Stroup, W.W., Wolfinger, R.D., Schabenberger, O., 2006. SAS for Mixed Models. SAS Institute Inc. Lopez, R.P., Hynes, H.P., 2006. Obesity, physical activity and the urban environment: public health research needs. Environ. Health 5, 25 /http://www. ehjournal.net/contents/5/1/25S. Morrison, D.F., 1976. Multivariate Statistical Methods. McGraw-Hill. National Highway Traffic Safety Administration, 2004. Lives Saved by the Federal Motor Vehicle Safety Standards and Other Vehicle Safety Technologies, 1960– 2002: Passenger Cars and Light Trucks. US Department of Transportation (Publication No. DOT HS 809 833). /http://www-nrd.nhtsa.dot.gov/Pubs/ 809833PART1.PDFS.

910

G.D. Johnson, X. Lu / Health & Place 17 (2011) 902–910

National Highway Traffic Safety Administration, 2006. Traffic Safety Facts 2006. US Department of Transportation (Publication No. DOT HS 810 818). /http:// www-nrd.nhtsa.dot.gov/Pubs/TSF2006FE.PDFS. National Highway Traffic Safety Administration, 2007. Traffic Safety Facts 2007 Data, Rural Urban Comparison. US Department of Transportation (Publication No. DOT HS 810 996). /http://www-nrd.nhtsa.dot.gov/Pubs/810996.PDFS. National Highway Traffic Safety Administration, 2008. Motor Vehicle Traffic Crashes as a Leading Cause of Death in the United States, 2005. US Department of Transportation (Publication No. DOT HS 810 936). /http://www-nrd.nhtsa. dot.gov/Pubs/810936.PDFS. National Highway Traffic Safety Administration, 2010. Highlights of 2009 Motor Vehicles Crashes. US Department of Transportation (Publication No. DOT HS 811 863). /http://www-nrd.nhtsa.dot.gov/Pubs/811363.PDFS. Nelson, M.C., Gordon-Larsen, P., Song, Y., Popkin, B.M., 2006. Built and social environments: associations with adolescent overweight and activity. Am. J. Prev. Med. 31 (2), 109–117.

Richter, E.D., Barach, P., Ben-Michael, E., Berman, T., 2001. Death and injury from motor vehicle crashes: a public health failure, not an achievement. Inj. Prev. 7, 176–178. Task Force on Community Preventive Services, 2001. Recommendations to reduce injuries to motor vehicle occupants. Am. J. Prev. Med. 21 (4S), 16–22. United States Department of Transportation, 2011. Key Transportation Indicators, February 2011. US Department of Transportation, Research and Innovative Technology Administration. Bureau of Transportation Statistics. /http://www. bts.gov/publications/key_transportation_indicators/S. Urban Land Institute, 2010. Moving the Needle: Regional Coalitions as Catalysts for Sustainable Development. Urban Land Institute. Waller, L.A., Gotway, C.A., 2004. Applied Spatial Statistics for Public Health Data. John Wiley and Sons Inc. Wolfinger, R., O’Connell, M., 1993. Generalized liner models: a pseudo-likelihood approach. J. Stat. Comp. Simul. 48, 233–243.