Elsevier

Social Science & Medicine

Volume 53, Issue 12, December 2001, Pages 1711-1719
Social Science & Medicine

Deprivation and mortality: the implications of spatial autocorrelation for health resources allocation

https://doi.org/10.1016/S0277-9536(00)00456-1Get rights and content

Abstract

This paper aims at investigating whether the relationship between mortality and socio-economic deprivation is affected by the spatial autocorrelation of ecological data. A simple model is used in which mortality (all-ages and premature) is the dependent variable, and deprivation, morbidity and other socio-economic indicators are the explanatory variables. Deprivation is measured by the Townsend index; the other socio-economic variables are the median income, unequal income distribution (Gini coefficient) and population density. Morbidity is estimated on the basis of hospital admission rates and overweight prevalence. Spatial autocorrelation is measured by the Moran's I coefficient. All mortality and morbidity variables have significant, positive, and moderate-to-high spatial autocorrelation. Two multivariate models are explored: a weighted least-squares model ignoring spatial autocorrelation and a simultaneous autoregressive model. The paper concludes that spatial autocorrelation has a significant impact on the relationship between mortality and socio-economic variables. Future ecological models intended to inform health resources allocation need to pay greater attention to the spatial dimension of the data used.

Introduction

The high mortality found in small underprivileged areas has been an important issue in health inequalities literature. Studies have repeatedly shown an association between socio-economic deprivation and mortality in small areas (Carstairs & Morris, 1991; Gunnell, Peters, Kammerling, & Brooks, 1995; Jarman, 1984; Phillimore, Beattie, & Townsend, 1994). More recently, this approach has been extended to long-term illness and standardised illness ratios (Bentham, Eimermann, Haynes, Lovett, & Brainard, 1995; Boyle, Gatrell, & Williams, 1999; Shouls, Congdon, & Curtis, 1996a). Such work has had a significant influence on the allocation of health resources. Deprivation indices are now taken into account in the allocation of health care resources. In the UK they have been used since 1990 for the distribution of resources to general practitioners through the Jarman index (Department of Health, 1990). Using data at the ward level, the latest revision of the Resources Allocation Working Party (RAWP) formula has led to the inclusion of socio-economic indicators in the allocation of resources to Regional Health Authorities in the UK (Carr-Hill et al., 1994a). In this paper, we shed light on a specific issue related to such ecological work: spatial autocorrelation.

In recent years there have been important methodological improvements in deprivation indices. New indices (such as the Townsend and Carstair indices) have been created in order to correct some biases in the Jarman index, such as the underestimation of deprivation in rural areas (Boyle et al., 1999; Moore, 1995; Talbot, 1991); the socio-economic variables and their weighting scheme in such indices have been revised (Gilthorpe, 1995; Gordon, 1995) and have been applied to smaller areas (Carr-Hill & Rice, 1995; Hastings, 1996; Majeed, Martin, & Crayford, 1996; Moore, 1995).

Despite these improvements, analysis of the relationship between deprivation and health has paid little attention to the spatial properties of the data, with the exception of the data aggregation problem (Martin, Senior, & Williams, 1994). In particular, critics have pointed out that it is not reasonable to assume that mortality counts at a small area level can be treated as independent events (Lawson, 1996). In fact, there is an increasing and recent body of evidence showing that health events are affected by spatial autocorrelation (Clayton, Bernardinelli, & Montomoli, 1993; Tiefelsdorf, 2000). Spatial autocorrelation indicates whether observations which are geographically close are related to each other, that is, are not statistically independent of one another. Positive spatial autocorrelation, meaning that nearby areas have similar levels of mortality, has been shown for cancer incidence in particular (Stam-Moraga et al., 1998; Thouez, Emard, Beaupre, Latreille, & Ghadirian, 1997; Walter et al., 1994). Recent research has extended spatial correlation to other health problems such as paediatric lead poisoning (Griffith, Doyle, Wheeler, & Johnson, 1998) and asthma (Hsiao, Tzeng, & Wang, 2000). Negative spatial autocorrelation means that areas with high mortality are contiguous to areas with low mortality.

When one is interested in relating mortality to various explanatory variables through a multivariate ecological model, spatial autocorrelation is a crucial issue. If found, it suggests that many statistical tools and inferences are inappropriate: correlation coefficients or ordinary least-squares (OLS) estimators are biased and overly precise. They will be biased because the areas of greater concentration of events will have a larger impact on the model estimates; and they will overestimate precision because, since events tend to be concentrated, there are actually fewer independent observations than are being assumed (Cliff & Ord, 1981; Richardson, 1996; Tiefelsdorf, 2000). Cliff and Ord (1981) showed that positive autocorrelation, which occurs for mortality and socio-economic factors, leads to overestimated precision, although the impact of positive spatial autocorrelation depends on the spatial layout (Anselin, 1990).

In fact, various important formulae dedicated to health care allocation have been elaborated using ecological multivariate models (Carr-Hill et al., 1994a; Jarman, 1984). None has taken spatial autocorrelation into consideration (although multilevel modelling has been used to control for inter-regional variation). Hence it is valuable to assess how far the relationship between deprivation and mortality is affected by incorporating the spatial autocorrelation of the data.

In this paper, we measure the spatial autocorrelation of health and socio-economic variables in Belgium; we then compare two multivariate models of all-causes mortality: one ignoring spatial autocorrelation, the other one incorporating it. Our aim, here, is to investigate whether the relationship between mortality and socio-economic deprivation is affected by spatial autocorrelation of the ecological data. This will help us assess the validity of the various ecological models which have driven health care allocation in various OECD countries.

Section snippets

Model

Here we examine a simple problem with mortality as the dependent variable, while deprivation, other socio-economic variables, and morbidity are used as explanatory variables. There are two reasons for considering morbidity together with deprivation variables. First, as morbidity and deprivation are only partially correlated, this avoids taking deprivation as a poor proxy for morbidity (Carr-Hill & Sheldon, 1992). Second, ignoring morbidity will produce biased estimates to the extent that

Results

Table 1 gives the weighted mean, the standard deviation and the Moran's I for each variable. Population density has a high spatial autocorrelation (I=0.67) meaning that dense municipalities tend to conglomerate. We observed positive and significant spatial autocorrelation for deprivation (I=0.40), income inequality (I=0.53), and median income (I=0.59). This means that wealthy municipalities tend to be located close to well-off municipalities, and, conversely, places with low income tend to be

Discussion

This study shows that mortality, morbidity, and socio-economic status all show moderate-to-high spatial autocorrelation. Incorporating such spatial autocorrelation in regression models has a significant impact on the apparent relationship between mortality and deprivation. The failure to take into account the spatial structure of the data can produce biased results and thus lead to erroneous conclusions about the relationship between mortality and deprivation — a relationship which occurs

References (67)

  • M Benzeval et al.

    The determinants of hospital utilization — implications for resource-allocation in England

    Health Economics

    (1994)
  • J.F Bithell et al.

    Controlling for socioeconomic confounding using regression methods

    Journal of Epidemiology and Community Health

    (1995)
  • F Carlsen et al.

    More physiciansimproved availability or induced demand?

    Health Economics

    (1998)
  • Carr-Hill, R., Hardman, G., Martin, S., Peacock, S., Sheldon, T., & Smith, P. (1994a). A formula for distributing NHS...
  • R Carr-Hill et al.

    Is enumeration district level an improvement on ward level analysis in studies of deprivation and health?

    Journal of Epidemiology and Community Health

    (1995)
  • R Carr-Hill et al.

    Rationality and the use of formulae in the allocation of resources to health care

    Journal of Public Health Medecine

    (1992)
  • R Carr-Hill et al.

    Allocating resources to health authoritiesDevelopment of method for small area analysis of use of inpatient services

    British Medical Journal

    (1994)
  • V Carstairs et al.

    Deprivation and health in scotland

    (1991)
  • D.G Clayton et al.

    Spatial correlation in ecological analysis

    International Journal of Epidemiology

    (1993)
  • A.D Cliff et al.

    Spatial processesModels and applications

    (1981)
  • M Colonna et al.

    Détection de l’autocorrélation spatiale du risque de cancer dans le cas où la densité de population est hétérogène

    Revue d’épidémiologie et santé publique

    (1993)
  • F Cowell

    Measuring inequality

    (1995)
  • Department of Health (1990). General practice in the national health service: 1990 contract. London:...
  • Department of Health and Social Security (1976). Sharing resources for health in England: Report of the resource...
  • S.A Dolan et al.

    Measuring disadvantageChanges in the underprivileged area, Townsend, and Carstairs scores 1981–91

    Journal of Epidemiology and Community Health

    (1995)
  • J Eachus et al.

    Deprivation and cause specific morbidityEvidence from the Somerset and Avon survey of health

    British Medical Journal

    (1996)
  • M.S Gilthorpe

    The importance of normalisation in the construction of deprivation indices

    Journal of Epidemiology and Community Health

    (1995)
  • D Gordon

    Census based deprivation indicesTheir weighting and validation

    Journal of Epidemiology and Community Health

    (1995)
  • D.A Griffith

    Spatial regression analysis on the PCSpatial statistics using SAS

    (1992)
  • D.A Griffith et al.

    A tale of two swathsUrban childhood blood-lead levels across Syracuse, New York

    Annals of the Association of American Geographers

    (1998)
  • D.J Gunnell et al.

    Relation between parasuicide, suicide, psychiatric admissions, and socioeconomic deprivation 581

    British Medical Journal

    (1995)
  • A Hastings

    Deprivation payments should be based on enumeration districts

    British Medical Journal

    (1996)
  • C.K Hsiao et al.

    Comparing the performance of two indices for spatial model selectionApplication to two mortality data 1

    Statistical Medicine

    (2000)
  • Cited by (73)

    • The spatially and temporally varying association between mental illness and substance use mortality and unemployment: A Bayesian analysis in the contiguous United States, 2001–2014

      2022, Applied Geography
      Citation Excerpt :

      For example, a county-level analysis of a panel study in the United States reported the increase in mortality rates in response to the increase of unemployment rates (Halliday, 2014). This could be a big miss since empirically, both mortality and unemployment are highly and spatially autocorrelated (Lorant et al., 2001; Molho, 1995; Patacchini & Zenou, 2007). Populations in different areas might respond differently to unemployment due to variations in local contexts (Shoff et al., 2012; Trgovac et al., 2015).

    • The fiscal impact of health care expenditure: Evidence from the OECD countries

      2020, Economic Analysis and Policy
      Citation Excerpt :

      The geographical concentration of health is well documented in the literature. Health outcomes, conditions and pathologies have been examined from a spatial point by several studies (Alexander, 1993; Gatrell and Whitelegg, 1993; Thouez et al., 1997; Hsiao et al., 2000; Lorant et al., 2001; Agovino et al., 2019). These concentrations appear in a very localised form.

    • Environmental and economic assessment of traffic-related air pollution using aggregate spatial information: A case study of Balneário Camboriú, Brazil

      2019, Journal of Transport and Health
      Citation Excerpt :

      The clustering of similar values of a variable in adjacent spatial units indicates the presence of positive spatial autocorrelation; when geographic areas tend to be surrounded by neighbors with very different values, there is strong evidence for the presence of negative spatial autocorrelation (Khomiakova, 2008). The study of Lorant et al. (2001) has shown that the use of spatial autocorrelation in regression models may affect the relationship between pollutant emissions and traffic zones. To that end, the possibility of autocorrelation likely underpinning the spatial structure of the data should be investigated in order not to lead to erroneous conclusions.

    • The association between medical care utilization and health outcomes: A spatial analysis

      2019, Regional Science and Urban Economics
      Citation Excerpt :

      Note that in our risk adjustment, we have already adjusted for the patient's ZIP code, so this aggregate zip code could capture an agglomeration effect if (for example) higher-quality physicians locate in regions with higher incomes and more amenities. A wide literature points to the existence of geographical concentration in population health and health care services (Rushton, 2003; Lorant et al., 2001; James et al., 2004; www.dartmouthatlas.org). Yet nearly all studies assume a zero correlation with regard to shocks affecting nearby regions, whether hospitals, states, or HRRs.

    View all citing articles on Scopus
    View full text