The LMA is the main metropolitan area of Portugal in which over ¼ of the Portuguese population lives. In accordance with other Southern European cities, Lisbon’s population is steadily ageing, particularly within the city centre: the population aged 65 or over in the LMA increased by 43.7 % (1991: 12.8 %; 2011: 18.4 %, according to the Portuguese National Statistics Office–INE). Nowadays, the older population in the Lisbon municipality accounts for 24 % of total population (INE, 2011).

Geographically, the LMA is divided into two main areas by the River Tagus: the northern and southern banks. The centre is the city of Lisbon, surrounded by a highly urbanized urban ring (in north) and a less urbanized urban ring (in south and northern border). Between the 1970s and 90s, the population in the north urban belt has grown very fast, mainly due to migrants from other Portuguese regions and former African colonies. Yet, this growth has not always been accompanied by public services, infrastructures, land-use mix concerns, etc., with consequences that are important to study, mainly related with social exclusion and health inequities [40].

This study follows an ecological design, as defined within the INEQ-CITIES project [8]. The sources of information were mortality registers (aggregated for the period 1995-2008), the 2001 census for population data and socioeconomic indicators, and the 1998 Urban Areas Classification for urbanization data, all from the INE.

The area of analysis was the parish, the lowest administrative level in Portugal. In the LMA there are 207 parishes belonging to 18 municipalities. The parish borders were stable for a long time and during the study period, decreasing the probability of misregistration of death certificates.

Based on an exploratory analysis of sixty causes of death (INEQ-CITIES list), 14 causes of death were selected, representing almost 60 % of total mortality in LMA between 1995 and 2008 (Table 1). The selection was restricted to causes of death previously associated with material deprivation or urbanization level [3, 8, 12, 16, 18–20, 22, 28, 35, 41] and for which numbers can be expected to be large enough to allow small area analysis.Table 1

ICD Codes (for the 9th and 10th Revision) of the causes of death considered in the study

The mortality data by cause of death was aggregated for the period 1995–2008 (N = 355,363), disaggregated by age (<15; 15–24; 25–44; 45–64; > = 65), sex (total, male, female) and parish (small area) of LMA. For reasons of confidentiality and lack of information, our database includes 97.7 % of total deaths from the selected causes of death in LMA, meaning that 13,319 deaths have not been considered (e.g., age, sex and parish are not mentioned on death certificates). The study population consisted of residents of LMA in 2001, stratified by the same sex and age groups as the mortality data.

To evaluate the social and economic conditions of the area of residence, a material deprivation index was built. This is a composite indicator that takes into account three dimensions: education, employment and housing conditions. The chosen indicators (from the 2001 Census) were: 1. Illiteracy rate (population with more than 10 years that does not know how to write and read); 2. Unemployment rate (unemployment among the population between 14 and 65 years), and 3. Substandard housing rate (houses without toilet). The material deprivation index was constructed in accordance with the method used by Carstairs and Morris [4]. The variables were standardised (using the z-score method) so that each variable exerted the same influence upon the final result. This index, and the indicators used to build it, have been already applied in other studies about material deprivation in LMA [12, 40]. The material deprivation index was analysed in terciles (t1: lowest level of deprivation; t3: highest level of deprivation).

Finally, the Classification of urban areas, produced by INE, was used to determine the urbanization level. This indicator takes into account population density and urban land use to categorise the Portuguese parishes in three groups: 1. Predominantly rural area, 2. Medium urban area and 3. Predominantly urban area. For the last one, the criteria was population density higher than 500 inhabitants/km² and more than half the territory classified as urban. Since the focus is only on the parishes of a metropolitan area, the first two have been aggregated to represent the “less urbanized” parishes and the last one representing the “most urbanized” parishes. This has been analysed as a dichotomic variable.

The mortality indicator used for this analysis is the Standardized Mortality Ratio (SMR) for total population of LMA. This variable is dependent on population size since its variance is inversely proportional to the expected values. Thus, areas with low population tend to present estimates with a high variance. When analysing aggregated data from small areas, it is important to consider two sources of variability: first, the spatial dependence between geographical areas, which means that neighbouring areas are more likely to have a similar mortality level than distant areas, according to Tobler’s first law of Geography [42]; second, the non-spatial variability (random variation). In order to take into account this variability we used the hierarchical bayesian model proposed by Besag, York and Mollié obtaining smoothed SMR (sSMR) [43]. This method allows us to produce smoothed estimates, minimizing potential bias while still presenting a valid spatial pattern [3, 44].

The sSMR were estimated for each cause of death and sex with the following model:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {O}_i\sim Poisson\left({E}_i{\theta}_i\right) $$\end{document}Oi~PoissonEiθi\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \mathrm{l}og\left({\theta}_i\right) = \alpha + {S}_i + {H}_i $$\end{document}logθi=α+Si+Hiwhere, for each small area i, O_i denoted the observed cases of deaths for a particular cause and gender in the small area i, E_i was the expected number of deaths (of each cause and gender) in the small area i and θ_i, the relative risk for each specific area and specific cause of death. ∝ represents the intercept, S_i the spatial random effects and H_i the heterogeneous (non-spatial) effects. The expected numbers of deaths in each area were calculated by indirect standardisation, using the population in 2001 (multiplied by the number of years in the study period: 14 years) and taking as reference mortality rates by gender, age (<15; 15–24; 25–44; 45–64; > = 65) and cause of death in the LMA.

Based on sSMR we measured the probability of excess risk (sSMR > 100), which should also be taken into account when evaluating the statistical evidence provided by estimates of sSMR in each small area.

The geographical distribution of sSMR, calculated through Model 1, was represented using maps of septiles: the dark blue areas have the lowest sSMR and the dark brown ones have the highest. The probability of excess risk was represented using five fixed categories: [0–0.1] (lowest probability sSMR > 100), ]0.1-0.2], ]0.2-0.8], ]0.8-0.9] and ]0.9-1.0] (highest probability sSMR > 100).

The statistical association with the contextual-level variables (material deprivation and urbanization) has been obtained through the application of an ecological regression model that introduces those indicators as explanatory variables.

To evaluate the statistical association between mortality by cause of death and urbanization (Xi) (a dichotomous variable), the regression was formulated as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \log \left({\uptheta}_{\mathrm{i}}\right) = {\upbeta}_1 + {\upbeta}_2{\mathrm{X}}_{\mathrm{i}} + {\mathrm{S}}_{\mathrm{i}} + {\mathrm{H}}_{\mathrm{i}} $$\end{document}logθi=β1+β2Xi+Si+HiWhere exp(β₂) denotes the relative risk of mortality in the more urbanized areas with respect to the less urbanized.

To analyse the relationship between mortality and material deprivation (D_i), we applied a similar model in which material deprivation terciles were introduced as dummy variables:

D_2i = 1 if the small area i is in the second tercile group.

D_2i = 0 otherwise

D_3i = 1 if the small area i is in the third tercile group.

D_3i = 0 otherwise

For this model, called “based”, the regression was formulated as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \log \left({\theta}_i\right) = {\beta}_1 + {\beta}_2{D}_{2i}+{\beta}_3{D}_{3i} + {S}_i + {H}_i $$\end{document}logθi=β1+β2D2i+β3D3i+Si+Hiwhere exp(β₂) (respectively exp(β₃)) denotes the relative risk of mortality in the areas included in the second tercile (respectively third tercile) group with respect to the included in the first tercile deprivation group.

Finally, we estimated the statistical association between material deprivation and mortality adjusted by urbanization level. For this model, called “adjusted”, the regression was formulated as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \log \left({\theta}_i\right) = {\beta}_1 + {\beta}_2{D}_{2i}+{\beta}_3{D}_{3i}+{\beta}_4{X}_i + {S}_i + {H}_i $$\end{document}logθi=β1+β2D2i+β3D3i+β4Xi+Si+Hi

Where exp(β₂) (respectively exp(β₃)) are adjusted by urbanization level (dichotomic variable) and denotes the relative risk of mortality in the areas included in the second tercile (respectively third tercile) group with respect to those included in the first tercile deprivation group.

For these three models, the relative risk (RR) estimates were obtained based on their posterior means, along with the corresponding 95 % credible intervals (95%CI). A RR was considered significantly higher or lower than 1 if its 95%CI did not include 1. The posterior distributions were obtained with the “Integrated nested Laplace approximation” (INLA) method.

For all models, an intrinsic conditional autoregressive prior distribution (ICAR) was assigned to the spatial effect, which assumes that the expected value of each area coincides with the mean of the spatial effect of the adjacent areas and has variance of σ_s², while the heterogeneous effect is represented using independent normal distributions with mean 0 and variance σ_h² [43]. A half-normal distribution with mean 0 and precision 0.0001 was assigned to the standard deviations σ_s and σ_h. A vague prior distribution was assigned to the parameters β_1, β₂ and β₃ [45].

These models were developed using the INLA library (version 3.0.1) and the R statistical package (version R.2.15.2) [46].

As far as we know, this paper is the first one that aimed at measuring and identifying the association between material deprivation, urbanization and mortality within a metropolitan area. Further, it is the pioneer research in Portugal that uses small area data to present mortality inequalities.

However, there are a number of limitations which may impact on the findings presented. First of all, mortality due to Symptoms, signs and abnormal clinical and laboratory findings cause of death is high, representing 5.8 % of total mortality. As a consequence, the other causes here studied may be under-represented, especially Suicide, Diabetes mellitus [57] and Cancer. Second, due to statistical confidentiality, the National Statistics Office only gave access to aggregated mortality data from fourteen years. This time-aggregation was imposed to have access to space-disaggregation data and did not allow to apply time-series cross-sectional analysis. Third, cause-specific mortality maps can only be used to indicate potential problems in material deprivation and urbanization level at small area level, which then have to be studied with more specific information and better local data on the relation between health determinants and health outcomes. The fourth limitation is related with population mobility. As we only have access to the deceased person’s last place of residence, we do not know how long s/he had been living there and how long s/he had been exposed to material deprivation. Furthermore, material deprivation is defined in the period 1995–2008 in the same way as in 2001. However, there were changes over these fourteen years in unemployment and in the number of substandard housing, although the geographical pattern had not changed. Fifth, we were not able to explore if there was an interaction effect between urbanization and material deprivation. Although statistically this could be done, the small number of less urbanized areas does not give us enough sample power to estimate these interactions. Finally, in terms of methodology, there are two main issues: (i) the standardization of mortality data took into account a structure of four age groups, which does not entirely remove the confounding effect of age; and (ii) the existence of statistical associations between the characteristics of place of residence and mortality patterns may be carefully interpreted in terms of causality [58].

Our findings can extend current knowledge by showing spatial patterns of cause-specific mortality in the LMA, identifying small areas with an excess risk of mortality associated with material deprivation and thereby pointing at problematic areas that could potentially benefit from public policies addressing specific causes of death and the effect of social inequalities. These results highlight the need to implement effective policies to reduce inequalities, namely through the intervention of government institutions (local and regional) on specific areas within the metropolitan area [59, 60]. Physical and social environments in neighbourhoods can be overtly hazardous. For instance, evidence about local risk factors (unemployment, illiteracy and poor housing conditions) associated with the Infectious and parasitic diseases, Chronic liver disease and Diabetes mellitus, within LMA, will potentially support the development of local interventions addressing those social and material conditions. Local governments are in a better position to tackle some of these health determinants, by implementing social programmes and built environment interventions aiming to reduce poverty and to improve constructed features that encourage healthy behaviours. Policy measures tackling unemployment and poverty include strategies reducing supply-side unemployment (e.g., education and training schemes, self-employment assistance), the number of families at risk of poverty (e.g., social benefits for low-income individuals and families, local council taxes reduction, affordable housing, access of disadvantaged population groups to health services, lifelong learning actions). Interventions targeting the built environment, urban design and planning in economically disadvantaged small-areas includes various interventions related with physical surroundings (e.g., buildings, green urban spaces, schools, road systems and other infrastructures), housing conditions (e.g., rehousing, refurbishment and community regeneration) and food environment (e.g., increasing the availability of healthy food choices, activities to encourage families to purchase healthier food options) [40].

As so, our results must be transferred to the local stakeholders, especially from sectors such as urban planning, culture, leisure, education, environment, social services and housing, due to their ability to exacerbate or reduce intra-urban health inequalities [61, 62].