Continued population growth and increasing urbanization have led to the formation of large informal urban settlements in many developing countries in recent decades. The high prevalence of poverty, overcrowding, and poor sanitation observed in these settlements—commonly referred to as “slums”—suggests that slum residence constitutes a major health risk for children. In this article, we use data from 191 Demographic and Health Surveys (DHS) across 73 developing countries to investigate this concern empirically. Our results indicate that children in slums have better health outcomes than children living in rural areas yet fare worse than children in better-off neighborhoods of the same urban settlements. A large fraction of the observed health differences appears to be explained by pronounced differences in maternal education, household wealth, and access to health services across residential areas. After we control for these characteristics, children growing up in the slums and better-off neighborhoods of towns show levels of morbidity and mortality that are not statistically different from those of children living in rural areas. Compared with rural children, children living in cities (irrespective of slum or formal residence) fare better with respect to mortality and stunting but not with respect to recent illness episodes.
Urban populations have grown rapidly over recent decades in both absolute and relative terms, with more than 50 % of the global population living in urban areas in 2010. This trend is expected to continue, such that more than 60 % of the world’s population is expected to live in urban areas by 2030 (UN 2011). The increasingly large numbers of people living in, and moving to, urban areas in developing countries have led to the formation of large and rapidly growing informal urban settlements, often referred to as “slums.” More than 1 billion people, or about 14 % of the total global population, were estimated to live in slums in 2007 (UN–HABITAT 2007).
The density of slum settlements and poverty of their residents, combined with the general absence of water and sanitation facilities, have raised serious concerns about the health of slum dwellers in general and the health of children growing up in slums in particular. In many respects, modern day slums in sub-Saharan Africa, Asia, and Latin America resemble low-income areas of major European cities of the nineteenth century. Accordingly, it is often assumed that residents of today’s urban slums face the same health disadvantage that residents of European cities faced in the nineteenth century relative to their rural counterparts (Cain and Hong 2009; Preston and Haines 1991; Woods 2003), and that the additional health risks generated by the rapid growth of slums may even delay the mortality transition in low- and middle-income countries (Konteh 2009; Moore et al. 2003; Sclar et al. 2005).
Evidence on mortality differences between rural and urban areas, and of slum areas in particular, is scarce and far from conclusive. Montgomery et al. (2003) analyzed data from 56 Demographic and Health Surveys (DHS) and documented generally lower childhood mortality for urban households in relation to rural populations, while emphasizing that specific urban subpopulations might face higher mortality risks than rural populations. Using data from 43 DHS surveys, Brokerhoff and Brennan (1998) analyzed differences in infant mortality between small towns and large cities. They found that infant mortality was lower in cities in sub-Saharan Africa but lower in towns in Latin America and Asia; they also found that the mortality gap between rural and urban areas has been decreasing. Fotso et al. (2013) used DHS data from African cities supplemented with data from demographic surveillance sites to examine trends in urban child mortality. They found that the pace of decline in urban mortality in sub-Saharan Africa has in most countries not been sufficient to achieve the target of Millennium Development Goal (MGD) 4—that is, to reduce the under-5 mortality rate by two-thirds between 1990 and 2015. They also noted the emergence of intra-urban mortality differences in their analysis. In a study of 18 African countries, Bocquier et al. (2011) found that after they controlled for known demographic and socioeconomic correlates of childhood mortality, urban advantages are greatly reduced or indeed reversed; similar results were found by Van de Poel et al. (2007, 2009). Analyzing 85 DHS surveys, Montgomery and Hewett (2005) found that poor households are often spatially intermingled with well-off households in urban areas, but also that areas of concentrated poverty are generally associated with lower rates of health service usage. Günther and Harttgen (2012) analyzed data from 18 African countries and found that child mortality rates in slum areas are significantly higher than in nonslum urban areas but lower than in rural areas in most countries. Timaeus and Lush (1995) analyzed intra-urban differences in child health for four countries and found that the mortality difference between the urban poor and nonpoor is larger than the difference between rural and urban populations. In terms of health outcomes other than death, Fotso (2007) and Van de Poel et al. (2007) analyzed differences in stunting between rural and urban areas; both studies found that whereas children in urban areas are better nourished than children in rural areas, the gap has been narrowing.
The main objective of this article is a comprehensive assessment of the child morbidity and mortality differences across residential areas, with a particular focus on health disparities between poor urban slum areas and rural and better-off urban areas. To do so, we expand the scope of the existing research in four directions. First, rather than working with selected countries, we combine all data available from DHS into a pooled data set covering 1.38 million children under the age of 5 across 191 surveys in 73 low- and middle-income countries.
Second, we extend the analysis of the health effects of urban and urban slum residence beyond child mortality: we analyze episodes of diarrhea and acute respiratory infection as measures of acute health conditions, and stunting as a measure of cumulative exposure to infectious disease or undernutrition. Although a large literature has documented the impact of poverty and lack of access to water and sanitation on child health and child development (Fink et al. 2011; Grantham-McGregor et al. 2007; Walker et al. 2011), no study has (to our knowledge) systematically assessed the association between urban (slum) residence and such a broad array of child health outcomes.
Third, we improve upon most of the existing literature by reviewing the slum concepts that are commonly used, and introduce a community-based measure for “slum residence.” Although household-based definitions are common in the literature following the UN–Habitat (2007) definition, we argue that defining a household’s residents as slum dwellers based on the lack of infrastructure of their own household alone, with no neighborhood context, is counterintuitive and likely to miss a key aspect of slum residence: the large negative effects on children growing up in environments with low hygiene and generally poor security standards.
Last, in an effort to align our study more closely with public perceptions of slum settlements, we explicitly distinguish slums in small- and medium-sized towns from those in large (metropolitan) urban areas, on the grounds that the notion of “slum” is generally linked to informal settlements within major cities (such as the slums in São Paulo, Brazil; Mumbai, India; or Nairobi, Kenya). Previous studies have also argued (but not tested) that health disadvantages should be largest in (the slums of) major cities (Brokerhoff and Brennan 1998; Cain and Hong 2009; Gould 1998).
Data and Methods
The data used in this article are from Demographic and Health Surveys (DHS), nationally representative population-based surveys that have historically had a primary focus on fertility and reproductive health but also cover a large array of child health outcomes and household characteristics beyond standard child survival measures. Largely funded by the U.S. Agency for International Development (USAID), more than 230 standard DHS surveys have been conducted in 86 countries since 1986.1 The individual-level data sets are generally publicly available. Some of the earlier surveys do not include information on household spatial characteristics needed for the classification of urban residence, resulting in a final sample of 191 surveys across 73 countries. Figure 1 shows the global distribution of these surveys. About one-half of the surveys cover sub-Saharan African countries, with the remainder roughly equally distributed between Asia and Latin America. A full list of all surveys is provided in Online Resource 1.
Classification of Towns and Cities
The principal geographic identifier contained in all DHS is the region where the person interviewed resides (administrative level 1 units corresponding to states or regions), which is captured in the DHS standard recode variable hv024. In addition to the region identifier, DHS data sets contain a variable capturing the “type” of residence (hv025), which primarily distinguishes between rural and urban areas; they also contain a variable capturing the “place” of residence (hv026), which provides a slightly less coarse division of households into rural, small town, and larger city residence. We stress that definitions of what constitutes an urban area are not consistent across countries, and the DHS uses whatever classification a particular country has adopted. Given that the common perception of slum settlements is generally linked to major urban settlements (such as the slums in Accra, Ghana; Mumbai, India; Nairobi, Kenya; and São Paulo, Brazil), we make an explicit distinction between informal settlements within large urban agglomerations and those in smaller urban areas. We classify urban settlements as “cities” if the United Nations population estimate for 2010 exceeds 1 million (UN 2011), and we refer to smaller urban settlements (and slums within these settlements) as “towns.”
Distinguishing towns from cities in DHS surveys requires some additional coding. Large urban agglomerations, and especially capital cities, in many cases constitute separate administrative regions (that correspond to one particular region of hv024). All households in such regions could be classified as “city households.” However, in some cases, such regions may also include small rural settlements. To ensure that these boundary households do not affect the estimation, all households classified as “rural” according to hv025 within larger metropolitan areas were coded as “rural households.” For 14 countries (Brazil, Colombia, Egypt, Honduras, India, Indonesia, Morocco, Nigeria, Pakistan, South Africa, Ukraine, Turkey, Vietnam, and Yemen), large metropolitan areas do not overlap perfectly with the administrative units provided by DHS. For surveys of these countries, the “type of place” variable (hv026) is used to categorize households into rural, town, and city residence. Given that DHS surveys do not follow a standardized definition of “city,” some households that should be classified as “town” households according to our criteria may be misclassified as “city” households when using this approach.
Most developing countries have only one or two large metropolitan areas, one of them usually being the capital city. In the data set that we analyze, two countries—Brazil and India—differ markedly from this pattern, with more than 20 agglomerations with populations in excess of 1 million inhabitants. The table in Online Resource 1 shows the complete list of metropolitan areas in the sample analyzed.
Classification of Slums
The only formal definition available of a slum is the one proposed by UN–Habitat (2007), which defines a household as a slum household if it lacks any one of the following five elements:
Access to improved water (access to sufficient amount of water for family use, at an affordable price, available to household members without being subject to extreme effort)
Access to improved sanitation (access to an excreta disposal system, either in the form of a private toilet or a public toilet shared with a reasonable number of people)
Durability of housing (permanent and adequate structure in nonhazardous location)
Sufficient living area (not more than three people sharing the same room)
Security of tenure (evidence of documentation to prove secure tenure status or de facto or perceived protection from evictions)
Two features of this definition stand out: (1) the inclusion criteria are very broad, and (2) neighborhood characteristics are absent. Although all five elements listed are undoubtedly part of proper living conditions, immediately classifying a household not meeting any one of these items as a slum household is counterintuitive because it implies that, for example, all households in an urban area without (or with expensive) local water supply would be considered slum households. Moreover, the exclusive focus in this definition on household characteristics appears inconsistent with the more commonly or colloquially used concept of a slum: an area, a neighborhood, or a group of buildings with certain characteristics (e.g., densely populated, run-down, impoverished)—not as a single household lacking specific amenities.
Following the approach proposed by Günther and Harttgen (2012), we use a neighborhood-based rather than a household-based definition of slums in this article. To operationalize this definition, we take advantage of the cluster sampling approach used by DHS surveys, which typically use a two-stage sampling approach: the first stage being the sampling of census enumeration areas (EAs), and the second stage being the sampling of households within the selected EAs. Census EAs are generally defined by national statistical offices to have clearly defined boundaries and to contain a number of households that a single census interviewer can cover in one day. Given these features, DHS sample clusters will often (although not always) fall within a neighborhood instead of straddling two or more neighborhoods. We thus base our definition of slum neighborhoods on the characteristics of sample clusters. A typical EA comprises approximately 200 households, of which 20 to 25 are selected at the second sampling stage. Thus, a full set of household characteristics is available for 20 to 25 households in each sampled neighborhood.
In detail, we use the following approach. The UN Habitat slum classification involves five characteristics of households: water, sanitation, building material, crowding, and security of tenure. DHS surveys collect detailed data only on the first four of these. The fifth, tenure security, is generally hard to measure in developing countries, where property rights are often poorly defined and frequently assigned through communities rather than through formal legal systems. For this reason, we restrict our coding to the first four criteria (water, sanitation, building material, and crowding).2 In a first step, a count variable is generated for each household to capture the total number of the four dwelling characteristics that failed to meet the UN Habitat standards. Households with a count of zero satisfy all UN criteria; households with a score of 4 lack all four characteristics analyzed.3 In a second step, these household variables are aggregated at the neighborhood level.
To provide a better sense of how frequently households lack these four basic conditions, as well as the associated difficulty with generating a rule for defining slum neighborhoods, we show some basic classification results in Fig. 2. Column 1 of Fig. 2 shows the results for the coding proposed by UN Habitat, which is based on including households lacking any of the four basic dwelling features described earlier. On this basis, more than 70 % of urban households in our sample would be classified as slum households. This is not surprising and highlights the general absence of improved sanitation as well as the frequency of improper building materials in urban areas of developing countries (Günther and Harttgen 2012). An even larger fraction of households would be classified as slum households if a simple majority rule were applied at the neighborhood level (column 2), classifying all households in a sample cluster as slum households if at least 50 % of them lack at least one of the four housing characteristics. Even if the threshold were increased to 75 % of cluster households lacking at least one amenity, 60 % of urban households would be considered as slum households (column 3). We regard it as implausible to classify the majority of urban households in the developing world as slum households, so a series of more restrictive definitions is considered.
Given that most households in developing countries lack at least one amenity, we define in a second classification substandard households as households lacking two or more of the key household characteristics. Considering enumeration areas with a majority (≥50 %) of substandard households under this classification, 35 % of all urban households would be coded as slum households (Fig. 2, column 5). If the slum threshold is increased to require at least 75 % substandard households, the fraction of urban households considered slum households drops to 19 % (Fig. 2, column 6).
Because the primary focus of this article is the health conditions of the poorest urban neighborhoods, we adopt the most restrictive coding rule for most of the analysis: namely, categorizing as slum households all the households in clusters in which at least 75 % of households lack two or more amenities. This coding assumption is conservative and implies that our empirical models estimate health disparities between the worst quintile of urban neighborhoods and the remaining 80 % of households, some of which may also be considered poor. To show how results change with less-conservative coding, we estimate additional empirical models as robustness checks. The use of cluster-level characteristics for the residential coding has important implications for the interpretation of the estimated coefficients on the slum variable. In our definition, households that lack two or more basic facilities but that are not located in a “slum” cluster are not defined as slum households. On the other hand, households that are not “substandard” but are located in a cluster where 75 % of other households do lack basic infrastructure are considered to be slum households.
Child Mortality and Morbidity Outcome Measures
To provide a comprehensive assessment of the effect of slum residence on child health, a range of mortality and morbidity outcomes are analyzed. Childhood mortality outcomes are divided into three age ranges, reflecting phases of early childhood with different patterns of risk factors. Neonatal mortality is defined as any child death occurring during the first month after birth. In this phase of childhood, risk factors associated with pregnancy and delivery may be expected to dominate. Postneonatal mortality is defined as any death occurring between the first and twelfth months of a child’s life; the greatest risks are likely to be associated with infectious diseases. Child mortality is defined as the death of a child between the ages of 1 and 5 years (between the 12th and 59th months). In this age range, infectious diseases will remain important risks; in addition, the risk of injuries increases with children’s ability to independently explore their environment. Given concerns about high mobility of urban populations and that the DHS data have only limited information on prior residence, we deviate from the usual five- or 10-year reference periods in our mortality outcome analysis and focus instead on a shorter two-year period preceding the interview. For neonatal mortality, children born in the month of the interview are excluded. For postneonatal mortality, only children between 1 and 12 months of age over the two-year period prior to the survey are included. Similarly, for child mortality, all children of ages 12–59 months at any time in the two-year period are included in our analysis. As a robustness check—and to exclude any self-selection into certain types of residence—we limit our sample to children of mothers who have never moved for the subset of DHS surveys for which this information is available.
To capture morbidity effects, three additional variables are analyzed: the prevalence of acute lower respiratory infection (ALRI) in the two weeks prior to the interview, the prevalence of diarrhea in the two weeks prior to the interview, and stunting at the time of the interview. A child is coded as a case of ALRI if the mother reported an episode of both coughing and short, rapid breaths during the two weeks preceding the interview. A child is coded as a case of diarrhea if the mother reported that the child suffered diarrhea during the two weeks before the interview. Stunting is defined as height-for-age that is ≥2 standard deviations below the WHO reference median (WHO Multicentre Growth Reference Study Group 2006), and can be interpreted as the cumulative outcome of lifetime disadvantage (generated by poor nutritional intake and illness episodes).
In a second step, in order to assess the degree to which residential health differences can be explained by biological and socioeconomic differences, we include child characteristics (sex, whether a twin, birth order, and preceding birth interval), mother’s education, and a household asset indicator capturing household access to electricity as well as ownership of radio, TV, motorcycle, car or truck, and refrigerator. In the third and last step, two additional variables to control for health access are included to assess the degree to which residential health differences can be explained by access to health services. Specifically, we include maternal responses to two DHS questions: (1) “Many different factors can prevent women from getting medical advice or treatment for themselves. When you are sick and want to get medical advice or treatment, is getting money needed for treatment a big problem or not?” and (2) “When you are sick and want to get medical advice or treatment, is distance to the health facility a big problem or not?” We code the answers to both questions into binary indicator variables (1, if the woman considers money/distance a big problem; 0, otherwise). Obviously, these two variables are imperfect proxies for health access. We considered including immunization coverage or the prevalence of institutional deliveries as alternatives; however, both variables are a combination of access and usage and would hence measure parental preferences in addition to access to services.
Because some of these variables were not collected in all 191 surveys, we show separate results for the (full) sample where outcome variables are available as well as for the restricted (full information) sample where all covariates of interest are available.
Table 1 shows the means for all variables of interest for each of the five residential areas (city nonslum, city slum, town nonslum, town slum, rural) analyzed. Note that the measures in Table 1 are derived from unweighted observations and thus should not be interpreted as representative for low- and middle-income countries at large. In general, each DHS tends to be about the same size, giving individuals in small countries or individuals from countries with multiple DHS surveys a higher probability of being sampled. In theory, we could construct weights that would make our sample representative of the developing countries that have (multiple) DHS. However, not all developing countries are included in the DHS (e.g., China), so we could never account for the whole population of developing countries; in addition, there are strong theoretical arguments against weighting in regression analysis (see, e.g., Deaton 1997).
A few patterns are worth highlighting. First, unconditional on any covariates, slum mortality is higher than mortality in other urban areas but generally lower than mortality in rural areas. The same ordering also holds for stunting as a measure of cumulative exposure to childhood adversity. The picture is less clear for short-term acute illness: with respect to incidents of diarrhea, children living in slums appear to fare worse than children living in nonslum urban areas and children in rural areas. The prevalence of ALRI appears to be similar across all residential groups.
With respect to the dynamics of childbearing, the main difference between rural and urban areas is family size, with an average of 4.1 children born to mothers in rural areas but only 3.2 and 3.0 children to mothers in towns and cities, respectively. Interestingly, with respect to family size, slum areas appear to fall roughly halfway between nonslum urban areas and rural areas. Lower rates of fertility are also reflected in child birth order, with rural areas having a lower fraction of children who are first-borns, and a higher fraction of children of birth order five and higher. Somewhat surprisingly, no systematic differences in birth spacing are found, with 16 % to 18 % of children born within less than two years of the preceding birth across all birth orders and residential areas.
Large differences by residence are apparent with respect to maternal education and household wealth. Urban mothers have, on average, completed more than twice the number of years of schooling of rural mothers; similarly, the average asset score of rural households is 1.16 (of a maximum score of 6), which is less than one-half the asset score of urban households. Once again, town and city slums fall roughly halfway between rural and nonslum urban areas. With respect to access to health services, stark differences across residential areas are apparent. Although only 20 % of mothers in urban areas considered distance to a health facility as an obstacle to treatment seeking, 50 % of mothers in rural areas considered distance to be an obstacle. Financial constraints also appear more challenging for rural residents, with 60 % of mothers reporting insufficient financial resources as a reason for not seeking treatment, compared with 40 % in nonslum urban areas, and 46 % and 57 % in city and town slums, respectively.
Last, nonslum urban areas also have large advantages over both urban slums and rural areas with respect to water, sanitation, dwelling, and space. In comparison with urban areas, more than twice as many children in rural and slum areas lack access to sanitation, and more than three times as many children lack access to an improved water source. Urban slums fare somewhat better on water supply and floor material than rural households, but do slightly worse with respect to sanitation. The only domain where slum dwellers fare substantially worse compared with rural residents is space, with more than three-quarters of slum dwellings being categorized as overcrowded—that is, with more than three persons per habitable room.
The Cox and logit models shown in Table 2 report unconditional associations between residence and child health: the only covariates included are the residential indicators and a vector of survey fixed effects to control for unobserved factors specific to a given DHS. The top panel of Table 2 (panel A) shows the estimated associations for the full sample; the bottom panel (panel B) shows the same associations in the restricted sample for which all covariates of interest (see Table 1) are available. Even though the restricted data set is substantially smaller, the estimated associations change only marginally, suggesting that the more restricted sample is fairly representative of the larger DHS sample.
Overall, the estimated associations displayed in Table 2 closely mirror the basic patterns outlined in Table 1. Mortality rates are lowest in nonslum cities, where (in the unrestricted sample) children face between 35 % (neonatal) and 57 % (child mortality) lower hazards of dying compared with the rural reference group. The protective effect of nonslum town residence is smaller, with an estimated 10 % reduction in the hazard of neonatal mortality and an estimated 33 % reduction in the hazard of child mortality. The estimated risks for children in city slums are higher than for those in nonslum areas of cities but appear to be fairly similar to the risks faced by children in nonslum areas of towns. The mortality risks faced by children in town slums appear similar to those in rural areas.
The differences across residential areas with respect to stunting—the most commonly used broader measure of child development—are generally even larger than for mortality. Moreover, town and city slums appear similar with respect to stunting, with an estimated 20 % and 30 % reduction, respectively, in the odds of stunting relative to rural areas. Nonslum city residents fare best, with an estimated odds ratio of 0.48, compared with 0.57 for town residents. For the two acute morbidity measures, the protective effect of town and city residence is weaker than for mortality or stunting: children in nonslum areas of towns and cities have somewhat lower risks of experiencing both diarrhea and ALRI than rural children, but children in slum areas have marginally lower risks of ALRI and no significant difference in risk of diarrhea in comparison with rural children.
In Table 3, we show how the estimated associations change after child, mother, and household characteristics are controlled for. In the top panel of Table 3 (panel A), we replicate the estimates from Table 2 (panel B) but add controls for children’s age (for morbidity outcomes), sex, twin births, parity, and birth spacing. The results are almost identical, which confirms the observation from Table 1 that little relation exists between residence and birth spacing. In the bottom part of Table 3 (panel B), we include the full set of controls as shown in Table 1; specifically, we control for maternal education, household asset ownership, and health access in addition to the child characteristics described earlier.
The inclusion of the full set of covariates (Table 3, panel B) reduces the estimated associations between residence and child health substantially. After all characteristics are controlled for, significant risk reductions for town residence are found only for stunting (nonslum and slum). Conditional on all covariates, town slum residence is now even associated with increased risk of postneonatal mortality and diarrhea. City residence continues to be associated with reduced risk for all the mortality outcomes and for stunting, but the estimated protective effect is reduced by 30 % to 50 %, depending on the outcome. Conditional on all covariates, (nonslum and slum) city residence no longer significantly decreases the risk of diarrhea, and even increases the risk of ALRI in nonslum areas.
To provide a better sense of the factors driving the differences observed between Table 2 and Table 3, we estimate a range of partial models in which we sequentially include groups of covariates. More specifically, we estimate five models: (1) a model that controls for child characteristics only; (2) a model that controls for child characteristics and maternal education; (3) a model that controls for child characteristics and household wealth; (4) one that controls for child characteristics and the two health access variables; and (5) one that controls for all variables simultaneously. All models control for survey fixed effects. To be able to provide absolute mortality differences, we use simple multivariate linear ordinary least squares (OLS) regression models to estimate the adjusted associations of interest. The results for these models are summarized in Fig. 3 for childhood mortality and Fig. 4 for morbidity. All controls generally weaken the associations between residence and child health in broadly similar ways. Household wealth and maternal education reduce the basic associations between long-term health outcomes and urban residence by between 30 % and 50 %; the reductions in the estimated associations between residence and health appear somewhat smaller for health access. The effects of wealth, education, and health access on recent disease episodes tend to be much smaller, on average.
One of the concerns surrounding the associations between residence and health outcomes is the frequently short-term nature of household residence. Although we try to minimize estimation bias generated by migration by restricting our analytical time horizon to the two years preceding the surveys, the estimated associations could still be biased if households self-select into neighborhoods based on children’s health. To ensure that our estimates are not affected by residential mobility and selection, we estimate a set of models restricted to mothers without recent moves. DHS surveys do not collect detailed migration histories; they do, however, ask mothers about the duration of current residence (standard recode variable v104). We use this variable to divide the sample into women with and without a history of migration (never moved), and show separate estimates for the “never moved” sample in Table 4. As in Table 3, we show residence associations conditional only on child characteristics and the mother’s fertility history (panel A), and then we show results for the fully adjusted models (panel B). Even though the restriction on nonmigrating mothers reduces the sample by more than 50 %, the overall results change only marginally.
A second concern regarding our main results relates to the specific slum definition chosen. As Fig. 2 shows, the slum definition used for the purpose of this article corresponds to the most restrictive option among the possible options considered, which results in slightly less than 20 % of urban households being classified as slum households. By choosing a highly restrictive definition, we expect to identify only the worst-off urban areas and thus maximize the likelihood that we would find slums associated with adverse outcomes for children. To address this concern, we reestimate our empirical model with a more inclusive definition; here, we define slums as neighborhoods where at least 75 % of households lack at least one (and not two) of the critical amenities. As shown in column 3 of Fig. 2, this leads to a classification whereby 60 % of urban residents are being considered slum dwellers. The empirical results based on this alternative classification are displayed in Table 5. Because we now include somewhat “better-off” neighborhoods in the slum category, the observed protective effects of slum residence (relative to rural areas) increase slightly. The overall pattern observed is clear: given that urban areas are characterized by a rather pronounced socioeconomic gradient, more stringent definitions of slums mechanically make slums look more similar to the rural reference group. To the extent that the definition chosen for our main result is more stringent than commonly used definitions in the existing literature, the main effects reported here can be considered as lower bounds of the true protective effects of slum residence (in comparison with rural areas).
As a last robustness check, we report results from models where we directly control for slum characteristics at the household level in addition to neighborhood deprivation. This means that we include households’ lack of water and sanitation, poor housing quality, and overcrowding as additional controls. Hence, we test whether health differences across residential areas are driven only by differences in (individual) housing characteristics or whether instead neighborhood externalities exist in addition to individual household effects. The results from these models, reported in Table 6, are very similar overall in terms of direction and significance to the results reported in Table 3.
The objective of the analysis in this article is to assess empirically the validity of the popular perception of slums, and particularly slums in large urban agglomerations (such as São Paulo, Brazil; Mumbai, India; or Nairobi, Kenya) as extremely unhealthy environments for children. Using data from 191 highly standardized, nationally representative DHS carried out in 73 low- and middle-income countries, we estimate models with (1) no controls (which identify differences in health outcomes by type of place of residence alone) and (2) adjusted models (through which we try to assess the effect of slums conditional on child, maternal, and household characteristics). Overall, the patterns observed are rather clear: compared with children living in rural areas, children in slums have significantly better health, even though they generally fare worse than children in better-off neighborhoods of the same urban settlements. Our results suggest that a large fraction of the observed residential health differences is explained by differences in maternal education, household wealth, and access to health services. However, some urban and urban slum advantages persist even when all three factors are controlled for, especially in large urban agglomerations.
For mortality, independent of the age interval analyzed, children from nonslum city neighborhoods face the lowest risk, followed by city slum residents, nonslum towns, and town slum residents. After we control for maternal education, household wealth, and access to health services, the ordering of type of urban residence remains mostly the same, but the differences to rural children become substantially smaller and mostly insignificant for town residents living in slums or better-off neighborhoods. For stunting as a cumulative measure of children’s nutrition and health experiences, significant protective effects are found for all urban dwellers, with the largest protective effects for nonslum city and town residence even when we control for a range of control variables. Although we cannot directly assess the causal mechanisms underlying these urban advantages, consistent access to (a wider variety of) food appears to be a plausible explanation for the observed gaps.
Different results were found for acute morbidity measures based on maternal reports. In the uncontrolled models, all urban children have an advantage over rural children in terms of experiencing ALRI, but only nonslum urban children have a health advantage in diarrhea incidence. In general, the urban health advantage is lower for short-term morbidity than for mortality or stunting. After controls are introduced, the protective effect of urban residence on having diarrhea and ALRI disappears; children in town slums are now even at higher risk of being reported to have had an episode of diarrhea in the past two weeks than rural children; and surprisingly, nonslum city children are now at higher risk of experiencing ALRI than other groups, which one could interpret as evidence of the adverse effect of increased exposure to poor air quality but which may also reflect reporting differences across residential areas. Given a growing literature questioning the reliability of mother-reported morbidity measures (see, e.g., Schmidt et al. 2011; Talley et al. 1994), the differences between the stunting and acute morbidity results might at least partially be explained by the lower overall quality of acute morbidity data as well as the associated measurement error.
In terms of the relative effects of different types of control variable, it is interesting to note that fertility dynamics have little to do with urban-rural differences in child health (compare panel B of Table 2 with panel A of Table 3) despite the fact that urban fertility is usually substantially lower than rural fertility (see Table 1). The effect modification estimates shown in Figs. 3 and 4 suggest that maternal education and household wealth are equally important in protecting urban children from health risks. The effect of access to health services, on the basis of mother’s responses as to whether they regard distance to a health facility or cost of services to be barriers to seeking service, is more modest and must be interpreted with caution, given that the two variables used are based on maternal self-reports and are thus unlikely to fully capture differences in actual access to health care services.
This study has three remaining limitations. First, the results presented are neither representative of a specific country nor of low- and middle-income countries more generally, but rather are average estimates based on the pooled and unweighted sample of all DHSs. Although the sample is very large and we consistently use survey fixed effects in all our models to capture differences in data collection or response rates across countries, across time, and across surveys, the population analyzed may differ from the average developing country population today. From a global low- and middle-income perspective, the most obvious and important country missing in the data set is China, where megacities are numerous, and the estimated associations between residences and child health might look different from the results presented here.
Second, the analysis that we present is descriptive: the reported point estimates should not be interpreted as causal effects of randomly moving a child from one type of residence to another, but rather as a description of actual differences (controlling for some observable characteristics). All the reported estimates are conditional associations, which may, but do not necessarily, represent causal effects of the residential variables analyzed.
A third limitation of the article is the uncertainty surrounding the correct definition and empirical identification of slums. Although we follow the approach suggested in Günther and Harttgen (2012) of focusing on neighborhoods rather than households, defining slum neighborhoods is complex and might have an impact on the magnitude of the relationships estimated. A related critical issue is whether DHS sample clusters can be regarded as neighborhoods. There is no definitive answer to this concern, but given that sample clusters are normally based on census EAs that are quite small and have clearly defined boundaries, we believe that a sample cluster can be regarded as a neighborhood in most cases. Of course, there is wide heterogeneity of built environments in urban areas, and some sample clusters will include small areas of disadvantaged households mixed in with more affluent households; we will fail to classify such households correctly, but we argue that the effects of such a failure are likely to be negligible because such households are unlikely to fare worse than those in larger areas of concentrated disadvantage.
In this article, we present evidence from a pooled sample of 1.3 million children across 73 low- and middle-income countries to show that children in slums face higher health risks compared with urban residents not residing in slums, but lower risks compared with children living in rural areas. A large share of these observed differences is explained by differences in maternal education, wealth, and health access. After these factors are controlled for, children in slums of smaller urban areas (less than 1 million population) and even in nonslum neighborhoods of such towns do not have health advantages over rural children except with respect to stunting. Children in cities—clearly for nonslum neighborhoods and to some extent also for slum neighborhoods for most outcomes—continue to have an advantage in mortality and stunting even after controls are added for observable factors. For short-term morbidity, any health advantages also disappear for city residence. Three points stand out: (1) children in cities have relatively good health outcomes regardless of the nature of their neighborhood; (2) children in smaller urban areas—and especially in slums of such towns—fare less well, and scarcely better than rural children on most outcomes; and (3) the indicator showing the most consistent urban advantage is stunting.
Numbers reflect list as of November 2012 (http://www.measuredhs.com).
Households are considered without access to safe water if the household does not have access to a private or public pipe, bore hole, or a protected well or spring. Households are defined as being deprived of basic sanitation if they either rely on open defecation or use an unimproved pit latrine. Shared sanitation facilities are considered as basic sanitation if they provide access to a flush toilet or ventilated improved pit latrine. A dwelling is considered as overcrowded if there are more than three persons per habitable room. If the floor material of a house is made of earth, dung, sand, or wood, its structure is considered inadequate. The information in the DHS thus allows us to closely approximate the UN Habitat criteria.
In some of the DHS, information on one of the four housing characteristics (mostly crowding) was missing. To maintain as large a sample size as possible, in such cases we imputed a positive value: that is, that the household was not overcrowded, or had good water supply or sanitation. As a result, we may have somewhat underestimated the share of slum dwellers. For our regression results (Table 2), this means that we might have underestimated the difference in health between urban nonslum areas and slum areas on the one hand, and nonslum urban areas and rural areas on the other hand.