Reconsidering the Rural–Urban Continuum in Rural Health Research: A Test of Stable Relationships Using Mortality as a Health Measure
- First Online:
- Cite this article as:
- Cossman, R.E., Cossman, J.S., Cosby, A.G. et al. Popul Res Policy Rev (2008) 27: 459. doi:10.1007/s11113-008-9069-6
- 114 Views
Health researchers commonly use existing rural–urban continua based on population size and adjacency to metro areas to categorize counties. When these continua are collapsed into simple rural-versus-urban aggregations, significant differences within the categories are masked. We show that when the entire range of the 10-category Rural–Urban Continuum Codes (RUCC) is used, the direction of the coefficients may differ and the fit of the model varies substantially across contiguous categories. However, collapsing contiguous categories masks variations within the continuum. To the extent that health policy decisions are made based on such aggregations, inappropriate policy choices may result (e.g., low payments to counties with relatively high needs). Given Congressional calls to address rural health, and the new Office of Management and Budget (OMB) statistical area classification system, debate over appropriate categorizations schemes is timely. We regress age-adjusted all-causes of death on various socioeconomic factors to assess the appropriate use of variants of the rural–urban continuum for health research. Our findings support two main conclusions. First, researchers collapsing urban–rural categorization schemes may be masking important categorical differences, inadvertently influencing policymaking predicated on their work. Second, finer classification of settlements yields uneven results. That is, coefficients shift signs across the continuum, indicating that collapsed models may be inappropriate. Results derived using collapsed variants of the RUCC may be too unstable to use as health research and funding categorization schemes. We suggest that a health status or outcomes categorization scheme is likely to be a more appropriate metric for rural health research.
KeywordsRural–urban continuumHealth policyMortalityInequalityResearch methods
A major difficulty with studying rural health is that “rural” is often defined using various administrative classification scales that make subjective distinctions between counties (Ricketts 1999; Ricketts et al. 1998; Cromartie and Swanson 1996, pp. 31–39). Consequently, it is unclear whether rural residence equals healthy residence. Researchers are typically careful to specify under which conditions their continuum-based research is valid by defining rurality as a nominal categorization of communities and health outcomes. However, there has been a trend toward extending use of the continua beyond their intended uses (Fluharty 2000; Rural Policy Research Institute 2000; Office of Management and Budget 2000, pp. 82228–82238; Ricketts 1999; Ricketts et al. 1998; Hewitt 1992). Policymakers have mistakenly assumed correlations between the continuum and health, thus basing health policy on a flawed categorization scheme (Rural Policy Research Institute 2000; Ricketts et al. 1998; Hewitt 1992). While level of urbanization and adjacency to metro areas can be convenient proxies for factors affecting health status and outcomes, mere population density at the county level is an unlikely determinant of health outcomes. Given Congressional calls to action using collapsed (and potentially flawed) rural–urban continua (Morgan 2002, pp. 382–383; Eberhardt et al. 2001), and issues related to the adoption of the Office of Management and Budget (OMB) authored “core-based statistical area classification” system (Fluharty 2000), exploring the utility of the rural–urban continuum for health research is particularly timely.
Our thesis is that the commonly used Rural Urban Continuum Codes (RUCC) has unstable relationships with some health measures, demonstrated here using one particular measure as an exemplar. We advance other researchers’ theoretical work on the RUCC by providing an empirical demonstration of unstable results when using such classification schemes. We demonstrate that relationships between the RUCC and all-cause age-adjusted mortality rates are neither linear, continuous, nor stable, all conditions necessary for proper model specification. Instead, estimates derived from such categorizations are influenced by the particular categorization schemes chosen. Our methodological work yields two important findings. First, we show that routinely used collapsed rural–urban schemes are misspecified, masking important differences across categories. Thus, a collapsed population-based multi-category continuum may be an inappropriate way to frame health research. Second, we demonstrate that using a more disaggregated scheme with more categories may further confound the interpretation of findings. Disaggregation changes the direction of effects across the continuum, yielding results that are more difficult to interpret. Our position is that these continua are best used as an administrative categorization scheme (as OMB envisioned for their metro/nonmetro classification) (Office of Management and Budget 2000, pp. 82228–82238), or specifically in the research for which they were designed, and not as a basis for programmatic or funding decisions. Understanding that these categorizations relate unevenly to health outcomes will alert researchers to the need to develop more functional continua. We suggest that health analysis is more appropriately based on health status and outcomes among populations, as opposed to the subjective settlement size ranking still used in health research (Ricketts et al. 1994). Future research should develop and test such health-based continua to facilitate more effective targeting and delivery of health services to those with the highest need.
Researchers have traditionally used the rural–urban continuum concept as a metric for examining human settlements. Two distinctive strands of research have emerged. The major difference between the two strands (for the purposes of this paper) is that the first (e.g., systems of cities and global cities theories (Berry 1962, pp. 147–164; Sassen 1991)) focuses on functional relationships between settlements while the second focuses on ranking and categorizing cities in a non-relationship context. We use the latter, which emphasizes size and class-based rankings of individual settlements based on operationalizations of the rank-size distribution of settlements schemas (Simon 1955, pp. 425–440; Zipf 1949; Rashevsky 1947). Urbanized areas are typically constructed using Christaller’s (1966) Central Place Theory and Losch’s (1954) subsequent extension. Some classifications are dichotomous, including the Bureau of the Census’s tract-based rural versus urban classification (US Bureau of the Census 1995) and the OMB county-based metropolitan versus nonmetropolitan classification (Office of Management and Budget 1995).
The limitations of using collapsed categorization schemes have been recognized, leading to development of several alternative schemes. For example, OMB methodology was modified to account for rural portions of large-area metro counties (Goldsmith et al. 1993), and has been adopted by several federal programs (Ricketts et al. 1998; Hewitt 1992). One scheme developed by ERS-USDA (Economic Research Service of the US Department of Agriculture), the RUCC (a.k.a. Beale Codes), classifies all counties into 10 mutually exclusive categories (Butler and Beale 1994) of metro and nonmetro counties, based on size and proximity to metro areas (Beale 2001); see Appendix, Table A. Another ERS-USDA system, the Urban Influence Codes (UIC), is a nine-class system (Ghelfi and Parker 1995) using two metro county codes (Parker 2001) with the remaining seven categories ranking counties on the basis of metro area adjacency and population size. While all classification systems have relative strengths and weaknesses, we chose the more defined RUCC system to permit us to compare empirical results using collapsed versus expanded continua.
Relating the rank-classification to health care, investigators have recognized that population concentrations facilitate occupational specialization (Britnell 2001, pp. 1–16; Friedrichs 1995; Berliant and Zenou 1995; Desan 1995, pp. 1–15), including the medical field itself (Leeming 2001, pp. 455–485; Gritzer 1982, pp. 251–283; Strauss et al. 1982). Physical distance is typically measured by physical adjacency to a metro area or by population density as an inverse proxy for the physical proximity of specialized care. Thus, residents of metro areas can routinely access medical specialties unavailable to nonmetro residents, with presumably more favorable medical outcomes, though the argument that more care results in better health outcomes is debatable (Fisher et al. 2000, pp. 1351–1362). Such analyses work if relationships between settlement size and health are linear and stable. Of course physical distance to specialized medical care is only one aspect of this issue; economic, social, and institutional barriers also exist.
There are two concerns: measuring health and defining “rural,” both of which determine the results (Ricketts 1999; Miller et al. 1994, pp. 3–26). When the crude death rate for all-causes-of-mortality is calculated, nonmetro areas have higher rates than metro areas (Ricketts 1999; Wright and Lick 1986). However, when standardized for age, sex and race (to permit direct comparison), the urban advantage over rural areas disappears or even reverses, regardless of the continuum used (Miller et al. 1987, pp. 23–34; Clifford et al. 1986; Clifford and Brannon 1985, pp. 210–224; Miller et al. 1982, pp. 634–654). Clifford et al. (1986) found that most rural–urban differences along the UIC in mortality rates resulted from the age structure. When Miller et al. examined mortality rates using the RUCC, they noted that results vary depending on definitions of “rural” (Miller et al. 1982, pp. 634–654).
Changing to a greater degree of data aggregation (i.e., fewer categories of counties), Hayward et al. (1997, pp. 313–330) measured social structural and institutional factors for mortality, using three categories of counties, finding a slight (but statistically insignificant) rural mortality advantage. A similar study (McLaughlin et al. 2001, pp. 579–598) found mortality and inequality relationships strong in nonmetro areas, but not in metro counties. Standardized by age, sex and race and using a two-class metro–nonmetro scheme, mortality rates for nonmetro counties were slightly lower than metro counties.
To summarize, the relationships between population concentration and health outcomes are inconsistent partly because rate standardization is not uniform. Differences in comparisons of age-standardized versus age/sex/race standardized rates will further confound the results. A related explanation is that results are dependent on the operationalization of the rural–urban continuum. Still unresolved is the debate of how fine or complex a working rural–urban continuum should be. Miller et al. (1994, pp. 3–26) argued in favor of a 17-class scheme as a means of gaining new insights into the nature of rural America. The Community Health Status Indicators Project’s website stratifies counties into 88 groups as a means of making county-to-county comparisons (Health Resources and Services Administration 2000). Meanwhile Goodall et al. (1998, pp. 101–111) have put forth an even more complex population-based classification scheme for counties. Complicating matters further, other models have found a multitude of additional correlated variables (Health Resources and Services Administration 2000; Hayward et al. 1997, pp. 313–330). The findings indicate that model specification (i.e., selection of variables) is not the fundamental issue but rather the issue is one of a stable classification scheme.
We argue that existing county classification schemes, as currently used in collapsed form, mask the diversity among US counties, particularly rural counties. Our empirical demonstration shows that, depending on how counties are parsed, different statistical relationships are revealed because the relationships are not stable; findings may be misleading. While this masking has health policy implications across the entire rural–urban continuum, it may be more misleading in rural or nonmetro areas.
The optimal level of data aggregation for analysis of structural conditions and health outcomes is the subject of debate (Lynch and Kaplan 1997, pp. 297–314), as are problems of ecological bias that occur when aggregated group-level data are used to infer behaviors or outcomes for individuals (Morgenstern 1998; Greenland 1992, pp. 1209–1223). This research empirically assesses the problematic statistical relationships between aggregate health outcomes and aggregate county traits.
Using the Compressed Mortality File (CMF) from the National Center for Health Statistics (NCHS), we calculated age-adjusted mortality rates to the year 2000 standard million using the direct standardization method, averaged over the period of 1993 to 1997(National Center for Health Statistics 2000, 2001a, b; Anderson and Rosenberg 1998a, b; Feinleib and Zarate 1992; Shryock et al. 1976) using county as the unit of analysis. Five-year averaged rates provide more rate stability for sparsely populated counties. The Compressed Mortality File contained 3,143 records. We combined independent cities into their respective counties and eliminated counties with zero population and Alaska, reducing the working file to 3,072 records.
The human settlement categorization scheme we use is the ERS-USDA’s RUCC, which ranks all counties in a 10-category scheme based on metro status and adjacency to a metro area (Beale 2001; Parker 2001; Ghelfi and Parker 1995). Comparisons with the nine-category UIC yielded no significant differences in the results, thus we report results based only on the RUCC.
In this paper, we split the data by RUCC in order to show how the coefficients for a standard mortality equation vary across the type of place. We recognize that it is more statistically sensible to include the RUCC codes as dummy variables and to interact them with each independent variable; however, this method would result in 80 coefficients—a far less parsimonious model that could not be interpreted easily. We use this initial step of breaking the file by type of place as one step in the direction of understanding the importance of place on mortality risk for a population.
Variables for our model were chosen to test various patterns of income inequality. We use 1989 county-level Gini quartiles calculated from Census STF 3B and median household size from the Area Resource File as the first series of independent variables. Gini coefficients are measures of relative distribution of income, ranging from 0 (perfect equality in the distribution of wages) to 1 (total inequality with only one family receiving all the wages). In our model the Gini coefficient ranged from 0.2534 to 0.5399. Because community characteristics affect families unevenly across the wage continuum we divided the Gini coefficient into quartiles (McLaughlin et al. 2001, pp. 579–598; LeClere et al. 1997, pp. 169–198). The variable Gini1 category, containing all counties with a Gini coefficient in the lowest 25%, is the reference category for the other three Gini categories. This approach permits examination of the effects of family wage inequality on county mortality rates.
Per capita income from 1990 is a measure of absolute income level, used to capture level of wealth in the county. The percent black in the county, also from 1990, is a measure of race-based disadvantage. Percent rural (1990) permits us to assess the potential effects of rurality even within urban areas. Median household size (1990) is a control permitting assessment of the dilution of income among households of different sizes.
Rank-order settlement size is typically operationalized by dichotomizing counties into rural versus urban, or metro versus nonmetro categories. We use the latter approach because it is consistent with other categorizations we are testing (described in more detail below). The threshold for 1990 metro areas was 50,000–100,000; details available from the Census Bureau (US Bureau of the Census 1995). The threshold for rural was “places (incorporated or unincorporated) with fewer than 2,500 residents and open territory” (Cromartie 2001). Our decision to split the file into metro/nonmetro was a strategic one. Had we used urban/rural for comparison purposes 2 urban and 8 rural categories would have been collapsed. By using metro/nonmetro, 4 metro categories and 6 nonmetro categories were collapsed, minimizing problems associated with skewness of the distribution.
Morbidity and mortality are expressed through multiple pathways (Committee on Health and Behavior 2001; Evan et al. 1994). One pathway focuses on place of residence interacting with wage inequality. We use an existing form of a mortality correlates model modified from McLaughlin et al. (2001, pp. 579–598), using 5-year averaged age-adjusted mortality rates, centered on 1995, as the dependent variable. To account for variability in the mortality rates among sparsely populated counties we use a weighted least squares regression equation (McLaughlin et al. 2001, pp. 579–598; Lynch et al. 1998, pp. 1074–1080). Using the inverse of the variance in the 5-year rates, we give more weight to counties with less variance in their rates; despite using a 5-year rate, variance in rates for small population counties is larger than the variance for large counties.
Descriptive statistics for mortality and socioeconomic variables for US counties by rural–urban continuum
All counties (N = 3,072)
RUC#0 (N = 168)
RUC#1 (N = 132)
RUC#2 (N = 314)
RUC#3 (N = 198)
RUC#4 (N = 132)
RUC#5 (N = 107)
RUC#6 (N = 607)
RUC#7 (N = 644)
RUC#8 (N = 248)
RUC#9 (N = 522)
>1 million central
>1 million fringe
Age-adjusted death rate
Gini coefficient, 1989
Gini 1 (lowest value, 0–25%) Mean (% of county in category)
Gini 2 (26–50%) Mean (% of county in category)
Gini 3 (51–75%) Mean (% of county in category)
Gini 4 (76–100%) Mean (% of county in category)
Per capita income, 1990
Percent black, 1990
.02 – 63
Percent rural, 1990
Median household size, 1990
The Gini1 variable demonstrates concentration of low-inequality counties. Sixty-three percent of counties within RUC#1 fall within the lowest quartile of income inequality, while RUC#5 has the smallest proportion of counties with low income inequality (11%). At the other extreme RUC#7 has the highest percentage of counties with high income-inequality scores (31%), while RUC#1 has the lowest (5%). There are significant differences in income inequality between contiguous areas. For example, RUC#0 has 12% fewer lowest inequality counties than RUC#1 and 7% more of the highest inequality counties, indicating a significant difference in overall income distribution between two contiguous categories.
Per capita income, an indicator of county level wealth, is not linear across the rural–urban continuum. Per capita income is highest in RUC#0 ($20,452) and drops steadily across the continuum, reaching a low point in RUC#8 ($13,886), but in contiguous RUC#9 has rebounded to nearly the level of RUC#4 ($14,962 versus $15,156). Plotted, this variable would resemble a long ski jump with a lift at the end.
The percent black ranges from a low of 4.99% in the most rural category (RUC#9) to 12.25% in the most urban grouping (RUC#0), however, the relationship is not strictly linear: the contiguous rural category (RUC#8) has a black population twice that of RUC#9 (11.5% versus 4.99%), suggesting spatially isolated predominantly black communities may be overrepresented in RUC#8. A comparable difference can be found between contiguous categories RUC#0 and RUC#1 (12.25% versus 6.96%). Collapsing either of these two sets of contiguous categories together would mask differences in the racial composition of the counties between the categories. Given the importance influence of race on mortality, this alone indicates one of the risks associated with dichotomous rural–urban comparisons.
The most striking nonlinear distribution is that of percent rural. As expected, RUC#8 and RUC#9 are 99% rural; however, 63% of RUC#1 (metro fringe) residents are considered rural, only slightly less than RUC#6 (nonmetro adjacent) at 68%. Some of the least rural counties (in terms of percent rural population) are also relatively low on the continuum. RUC#5 (nonmetro, nonadjacent) is only 39.5% rural while RUC#2 (medium metro) is 41% rural. There is a nonlinear relationship between percent rural and the rural–urban continuum order.
Finally, for median household size, RUC#1 has the highest value (2.41 median persons per household) and the largest standard deviation (0.49), while RUC#9 has the lowest value (2.05 persons per household) and standard deviation (0.25). In general, these descriptive statistics suggest there is more variation among these categories than generally appreciated. Because of this, we should determine the capacity of ecological level mortality regression equations to account for these differences in health research. We test one regression model using county-level measures of inequality and demographics to estimate age-adjusted mortality rates, demonstrating that results vary substantially across the urban–rural continuum, suggesting that standard ecological mortality regressions may be inappropriate applications for much health research.
Model Regression Results
Weighted least squares regression analysis of mortality for US counties: metro and nonmetro
1990 Per capita income
Percent black population 1990
Median household size
Weighted least squares regression analysis of mortality for US counties by Rural–Urban Continuum (RUC) codes
1990 Per capita income
Percent black in 1990
Median household size
In Table 2, a three-category model is presented: metro versus nonmetro adjacent to a metro county versus nonmetro nonadjacent to a metro county, addressing the role that adjacency to a metro area may play. Table 3 shows the full 10-category RUCC model.
In calculating the various references equation we also performed a Condition Index (CI) test for multicollinearity. Initially we tested “All Counties” and the popular dichotomy of metro versus nonmetro. The “All County” model had a CI of 1,755, the “Metro” model had a CI of 1,653 and the “Nonmetro” model had a CI of 18,999. A CI score above 30 indicates severe multicollinearity in the equation (Dillon and Goldstein 1984; Belsley et al. 1980). Among the problems in interpreting results in the presence of multicollinearity are that the magnitude or sign of estimated regression coefficients may change, providing inaccurate interpretations. Multicollinearity also inflates estimated variance, eventually resulting in a t-statistic value very close to zero. Thus, it might not be possible (appropriately) to reject the null hypothesis if important variables are collinear. A related problem is relatively high adjusted R2 for the entire equation (indicating relatively “good” fit), but with statistically insignificant individual variables, misleading as to the importance of individual variables (Pfaffenberger and Patterson 1987; Lomax 1992). When reconfigured to “Nonmetro, Adjacent to Metro”, and the “Nonmetro, Nonadjacent to Metro”, both models have CI scores of 1.41. Likewise, when the model was expanded to the 10-category RUCC the CI dropped to an acceptable 1.42 for each classification. Having dummy variables in the model may somewhat inflate the CI; therefore, we also tested using the VIFs. The VIF for each independent variable indicates that only the Gini dummies had occasional high inflation factors (above 2).
When the model is divided into three categories (metro, nonmetro adjacent to a metro area, nonmetro nonadjacent to a metro area scheme) interesting differences begin to emerge (Table 2). Metro counties have the lowest intercept, indicating that, controlling for each of these variables, metro counties have lower age-adjusted mortality rates of the three categories. However, there are important differences among the independent variables.
Income inequality appears to have only weak effects in metro areas. Two of the three inequality measures are not statistically significant, in contrast to the nonmetro counties. For metro counties, being in the highest inequality quartile (Gini4) is not statistically correlated with mortality rates (referenced to Gini1). However, the nonmetro nonadjacent category shows the highest level of wage inequality is associated with 110 more deaths per 100,000—double the effect seen in nonmetro adjacent counties (46 per 100,000). This may indicate a lack of an effective social safety net in those counties, with either far fewer adjacent alternatives for medical care or possibly friction of distance issues (e.g., transportation issues, physical distance, physical barriers, climate, or weather).
As anticipated, per capita income is negatively correlated with mortality rates across the three-categories. The percent black coefficient is highest in metro counties (4.69), while the effect is substantially reduced in nonmetro counties—whether adjacent or not (3.26 and 2.94, respectively). Finally, for each additional percent rural in metro areas, there is a slight increase in the mortality rate (0.47), while for nonmetro this effect is not statistically significant. Median household size is not statistically significant in the three-category model.
In Table 3, the fully specified 10-class RUCC model clearly illustrates the diversity among categories of relationships between the county variables and mortality rates in three significant ways: (1) reading across RUCC categories (rows), the variables differ widely in magnitude and statistical significance, (2) perhaps more importantly, the direction of the relationships reverses across categories belying its status as a “continuum,” and, finally, (3) reading down columns, the fit of the model, indicated by the number of statistically significant variables and the explanatory power of the model, differs across the RUCC. In theory, a properly specified model would demonstrate consistency in results across a data continuum if the underlying statistical relationships did not vary across categories. Obviously, these relationships do vary, sometimes dramatically, across the RUCC. Our contention is that representing the RUCC as an ordinal construct masks important categorical differences captured by the 10-category approach, particularly problematic when research or policy is predicated on collapsing contiguous categories.
Specifically, reading across categories, three of the seven independent variables (Gini2, percent rural and median household size) are statistically significant in half or fewer of the 10 RUCC categories. The top quartile Gini (Gini4) is significant in six of the 10 rural–urban categories. One variable (Gini3) is significant in eight categories, and one (per capita income) is significant in nine categories; the only variable statistically significant in all 10 RUCC categories is percent black.
Reading across categories, Gini3 (the third highest inequality quartile) is statistically significant across the lower metro and most of the nonmetro categories (RUCC#2, 3, 4, 6, 7, 8, 9); however, Gini4 (the highest quartile of inequality) is statistically significantly different from Gini1 (the referent category) only in the most urban (RUC#0) and the more rural areas (RUC#6–9). Signs for significant coefficients change direction across the RUCC. Gini2 (the second lowest quartile) is positively linked to age-adjusted mortality rates in RUC# 3, 8, and 9 (medium metro, and most rural), but negatively related in nonmetro, nonadjacent areas. While this is a statistically weak reversal, it does highlight the variability among categories. For another example, we turn to the proportion of the county population living in rural areas, which is positively associated with mortality in RUC# 0, 4, 5 and 7, but is negatively correlated with the most rural category (RUC#9). Given that RUC#9 is almost exclusively rural (see Table 1), this would indicate that for every one percent increase in rural population, mortality rates would decline by 3.43 deaths per 100,000. Rather than an important finding of a relationship, however, this may suggest the redundancy of both a percent rural variable and a 10-category RUCC scheme. But, we believe including a measure of the concentration of rurality is important in light of variation in rurality across the RUCC (Table 1). Likewise, median household size is negatively correlated to mortality rates in one metro (RUC# 3) and two nonmetro categories (RUC#6 and 7) and positively correlated with mortality rates in the largest nonmetro category (RUC#4) and the most rural category (RUC#9). These nonlinear patterns of relationships would be unexpected if there was an inherent ordinality to the RUCC.
Finally, there is variety in the model’s fit across the RUCC. Whereas the most populated category (RUC#0) has the highest adjusted R2 (0.62) and the least populated has the lowest adjusted R2 (0.40), there is not a constant pattern across the 10-class continuum. When graphed, the result can best be described as three quarters of a “W,” further suggesting unevenness across the RUCC and nonordinality among contiguous categories.
The research and policy implications of this volatility in understanding mortality may be substantial. One risk is that broad-based health policy decisions based on an inappropriately specified RUCC model might have effects opposite to those intended. To illustrate the potential for such an outcome, consider that median household size is negatively related to mortality rates in several categories (Table 3). In RUC#6 (small nonmetro, adjacent), a one-person increase in median household size results in a reduction in mortality rates by 45 per 100,000 deaths; however, in the most rural category (RUC#9), a one-person increase in median household size results in an increase in mortality rates by 126 deaths per 100,000. Thus, the same change in household appears to have opposite effects on mortality at different levels of the RUCC. Reading down the columns, the differences are equally striking when comparing the number of statistically significant variables in individual RUCC categories. Only two of seven independent variables contribute to explaining mortality in the metro fringe (RUC#1), while all seven independent variables are statistically correlated with the dependent variable in the most rural category (RUC#9). Though this is not a statistically valid test of the model, the observation suggests substantially varying fit across this particular population-based categorization.
To briefly summarize, we find unstable regression results using county-level metro–nonmetro classifications and examining age-adjusted mortality in an inequality model. This is evident for both collapsed (i.e., dichotomous or three category) and fully expanded (10 category) categorization schemes. We also examined an intermediate classification scheme. We re-categorized the RUCC codes into five classes to model the scheme used in Urban and Rural Health Chartbook, Health, United States, 2001 (Eberhardt et al. 2001). As anticipated, collapsing from 10 to 5 categories suppressed significant differences, as well as dramatically decreasing the magnitude of the coefficients in the model (analysis available from the authors). Using this particular scheme masks important differences between categories.
Using a population-based classification scheme and an inequality model we have demonstrated four important points. First, using collapsed (dichotomous) categorization models masks important variations among county classes. The same is true for intermediate (five-category) schemes. Using the full 10-category classification scheme demonstrates instability and problematic results. Thus, it is not a matter of finding the “right” number of categories. Second, simply changing the model’s specifications cannot solve these twin issues of masking and instability. Notwithstanding the theoretical implications of changing the specifications, models with different variables (both dependent and independent) demonstrate similar masking and instability (McLaughlin et al. 2001, pp. 579–598; Mansfield et al. 1999, pp. 893–898; Lynch et al. 1998, pp. 1074–1080; Hayward et al. 1997, pp. 313–330; Lynch and Kaplan 1997, 297–314; LeClere et al. 1997, pp. 169–198; Miller et al. 1994, pp. 3–26; Miller et al. 1987, pp. 23–34; Wright and Lick 1986; Clifford et al. 1986; Clifford and Brannon 1985, pp. 210–224; Miller et al. 1982, pp. 634–654). Third, it is unlikely that changing the scale of data collection (i.e., from county-level to census block- or tract-level (Morrill et al. 1999, pp. 727–748; Cromartie and Swanson 1996, pp. 31–39) will alleviate this, since it does not address the fundamental problem: the basis for the classification scheme is flawed. Fourth, we conclude that population-based continua are inappropriate proxies or classification schemes for health research, based on the demonstrated instability of the continuum. Our conclusion is similar to Miller et al. (1994, pp. 3–26) when they observed that, to examine the rural health disadvantage, “…any standard that emerges must extend beyond the reductionist approach of classifying counties merely as having met or not having met proximity population counts…” (p. 24). Researchers should consider using spatially specific health status and/or health outcome measures as a means of place categorization for health research, not just as the predicted value of their models.
One recommendation is that researchers embark on more meaningful categorizations to understand the spatiality of health in the US, per Ricketts call to identify “nascent regional systems” (Ricketts 2002, pp. 140–146). Our work in progress focuses on developing health related measures as a means of categorizing settlements. Preliminary findings indicate that communities may be ranked or categorized directly by some health measures. For example, when counties are ranked by all-causes-of-mortality and then mapped, spatial clusters of “healthy and unhealthy counties” are evident (Cossman et al. 2003). Additionally, when these rankings are mapped across time there is evidence of spatial persistence in mortality rates (Cossman et al. 2002). Rankings based on health status and spatial stability warrants further investigation. This approach may improve our capacity to identify the determinants of health that are ripe for intervention and the delivery of needed health services.
This study was funded by a grant from the Office of Rural Health Policy of the Department of Health and Human Services (Grant 4-D1A-RH-00005-01-01) through the Rural Health Safety and Security Institute, Social Science Research Center (SSRC), Mississippi State University (MSU). Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the Office of Rural Health Policy. The authors thank Carol Campbell, graduate assistant, for data analysis, Debra Street (FSU), Isaac W. Eberstein (FSU), Troy Blanchard (MSU), and Wesley James (MSU) for helpful comments on earlier drafts.