Spatial interactions between urban areas and cause-specific mortality differentials in France

Health & Place 24 (2013) 234–241

Contents lists available at ScienceDirect

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

Spatial interactions between urban areas and cause-specific mortality differentials in France Walid Ghosn a,n, Daouda Kassie b, Eric Jougla a, Stéphane Rican b, Grégoire Rey a a b

Inserm CépiDc, Le Kremlin-Bicêtre, France Laboratoire Espace Santé Territoire, Université Paris Ouest, Nanterre, France

art ic l e i nf o

a b s t r a c t

Article history: Received 14 March 2013 Received in revised form 30 September 2013 Accepted 2 October 2013 Available online 12 October 2013

Spatial interactions constitute a challenging but promising approach for investigation of spatial mortality inequalities. Among spatial interactions measures, between-spatial unit migration differentials are a marker of socioeconomic imbalance, but also reflect discrepancies due to other factors. Specifically, this paper asks whether population exchange intensities measure differentials or similarities that are not captured by usual socioeconomic indicators. Urban areas were grouped pairwise by the intensity of connection estimated from a gravity model. The mortality differences for several causes of death were observed to be significantly smaller for strongly connected pairs than for weakly connected pairs even after adjustment on deprivation. & 2013 Elsevier Ltd. All rights reserved.

Keywords: Mortality Spatial interactions Spatial inequalities Migration Gravity model

1. Introduction Although mortality has decreased dramatically in France and other industrialized countries over the last half century, the improvement has been followed by an increase in social and socio-spatial mortality inequalities (Borrell et al., 1997; Leclerc et al., 2006; Leyland, 2004; Mackenbach et al., 2003; Menvielle et al., 2007; Murphy et al., 2006; Pearce and Dorling, 2006; Preston and Elo, 1995; Salem et al., 2000; Shaw et al., 2000; Singh, 2003; Windenberger et al., 2011). In order to explain spatial inequalities, many studies have focused on the social and economic characteristics of each area, as summarized by deprivation indices (Carstairs and Morris, 1989; Pampalon and Raymond, 2000; Rey et al., 2009; Townsend, 1987). In France, in 2001, median income, percentage high school graduates and other measures relating to population composition were associated with spatial inequalities in mortality on various scales (Rey, 2007). Preceding studies of health geography and social epidemiology that have dealt with French mortality inequalities reported large individual socioeconomic differentials (Menvielle et al., 2007) and regional (22 spatial units in France) patterns of mortality (Rican et al., 2009; Salem et al., 2000) that hold even after adjustment for deprivation on the local (commune) scale (Rey et al., 2009; Windenberger et al., 2011). This suggests that individual dimensions such as differences in lifestyle, health habits and diet, alcohol and tobacco consumption play a role as well as contextual driving factors including neighborhood n Correspondence to: Inserm CépiDc, 80 rue du Général Leclerc, Secteur Marron Bâtiment La Force - Porte 58, 94276 Le Kremlin-Bicêtre, France. Tel.: þ 33 149 595337. E-mail address: [email protected] (W. Ghosn).

1353-8292/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.healthplace.2013.10.003

features, broader living environment, health care provision and other unidentified features (Congdon et al., 1997; Scarborough et al., 2011; Stringhini et al., 2011). Among broader contextual factors operating on a larger scale that influence local health, the interactions between spatial units have yet to be investigated in France. Places are considered “a result of endogenous and exogenous processes operating on a variety of spatial scales” (Cummins et al., 2007 p. 1832). When considering spatial interactions, places are defined as nodes in networks rather than as autonomous bounded spatial units (Cummins et al., 2007; Gatrell, 1997). The adjacency between spatial units implies likely interactions, e.g., commuting daily from home to work in a different area. The idea that adjacent units interact with each other and share similarities has also given rise to the use of statistical models that take spatial autocorrelation into account (Besag et al., 1991; Conlon and Waller, 1998). On a local scale, the relative position of a neighborhood in an urban setting, in particular relative deprivation, is associated with health independently of absolute deprivation (Aberg Yngwe et al., 2012; Zhang et al., 2011). A relational approach to spaces involves not only taking adjacency or spatial proximity into account, but also relations in terms of social and population exchanges. Among the relationships that operate between areas, migration between two areas is a marker of socio-economic imbalance at a finer scale (Norman et al., 2005). Micro-economic approaches to migration are based on the idea that an individual compares his current situation (job, income, living environment, etc.) and the situation to which he or she can aspire through a change of residence. This human capital model not only takes into account the economic costs and benefits of migration, but also the psychological costs of leaving friends and family, or the benefits of changing climate

W. Ghosn et al. / Health & Place 24 (2013) 234–241

(Baccaïni, 2006). A long-distance residential migration is an important decision and, in the majority of cases, is related to education or job change (Baccaïni, 2001). Individual migration is linked to individual socio-demographic characteristics as well as to features of the areas of origin and destination (Baccaïni, 2006; Champion et al., 1998). In many countries, mortality is used to address public health issues, because of its exhaustiveness and reliability. With a view to further elucidating spatial health inequalities, several studies have emphasized the need to investigate the link between space and health from the angle of population dynamics (Boyle et al., 2004a; Cummins et al., 2007). In particular, population change and migrations play a role in the mortality rate observed at a given time and in a given place, since migration is a health selective process (Bentham, 1988; Boyle et al., 2004b; Brown and Leyland, 2010; Connolly et al., 2007; Davey Smith et al., 1998; Martikainen et al., 2008; Riva et al., 2011). A study has reported that the largest absolute flow is observed among relatively healthy migrants moving away from more deprived areas toward less deprived areas (Norman et al., 2005). However, empirical research presents contradictory findings. Migrations are not systematically associated with an increase in health inequalities (Jongeneel-Grimen et al., 2011), and can even contribute to reducing the gap between healthy and unhealthy areas (Boyle et al., 2009). The direction and strength of the association may depend on the country, spatial scale, time period, definition of deprivation and health indicator studied (Connolly et al., 2007). All of the studies have considered the association between migration intensities and mortality. However, none of them has specifically considered migration and population exchanges as a spatial interaction marker of connectivity and inter-dependencies as well as similarities that would be associated with health. Indeed, strong migration exchanges are likely to result from a “preference” effect, which may reflect common features and complementarity between the areas. On the contrary, the absence of interaction is a sign of a “barrier” effect, which could be interpreted as a marker of difference and incompatibility. Considering the question from a health perspective, this paper investigates whether residential migration exchanges between pairs of units may be able to measure differentials or similarities between units that are not captured by usual indicators. Could the complementarities or similarities between spatial units characterized by strong balanced exchanges be associated with more similar mortality rates? Equivalently, could weak or unbalanced exchanges be associated with larger mortality differentials? The aim of this paper is two-fold. In the first part of the paper, residential migrations between urban areas are characterized using a spatial interaction model. The gravity model (D'Aubigny et al., 2000) enables estimation of the influxes of people in two spatial units depending on distance and population. Applying this approach allows identifying “preference” (strong influx) and “barrier” (weak influx) effects between the spatial units considered (Courgeau and Pumain, 1996). Then, the connections between pairs of spatial units are categorized by intensity. The second part of the paper focuses on the association between the intensity of connection and mortality differentials. Mortality differences are compared for several causes of death, for men and women, all-age mortality and premature mortality. Since deprived areas are less attractive for migrants than affluent areas, mortality will also be controlled for those dimensions.

2. Methods 2.1. Spatial scale Urban areas (UA, 361 units), as defined in 1990 were used to characterize and study the intensity of migration exchanges between 1990 and 1999 in mainland France. An urban area is

235

defined as a set of communes (smallest administrative unit) that are contiguous and free from enclaves constituted by an urban core and rural communes or urban units (peri-urban ring) and in which at least 40% of the resident working population works in the urban core or its catchment communes (Le Jeannic and Vidalenc, 1997). In 1990, the population of mainland France was about 57 million, with 41 million people living in the 1990 urban areas. The scale of analysis enabled both intra-regional and inter-regional exchanges to be taken into account. The urban areas had a minimum of 8200 people and a maximum of 10,656,900 people with a mean of 117,100 people. 2.2. Demographic data The 1997–2001 mortality data were derived from the InsermCepiDc database for mainland France. The National Institute of Statistics and Economic Studies (INSEE) supplied the population data. The migration data were directly derived from the 1999 census, in which respondents were asked to answer the following question: “Where did you live on 1 January 1990?” Thus, the data only cover people whose 1999 unit of residence was different from that for the preceding census. The data, which were available on the commune scale, were aggregated to the 1990 Urban Area scale. During the 1990–1999 period, 4.6 million people changed their urban area of residence. The total number of possible migratory influxes between urban areas was 361  360 ¼129,960, of which 50,042 were equal to 0. Deaths were grouped in 5-year periods centered on 1999 to avoid small numbers and year-to-year fluctuations. Age and gender standardized mortality ratios (SMR) were calculated, taking national mortality over the period 1997–2001 as the reference. The categories considered were all causes of death for persons and separately for men and women, premature mortality (age less than 65 years) and cause-specific mortality (see Table S1 for ICD codes). 2.3. Spatial interactions model: the gravity model The gravity model was specifically developed to estimate influxes between pairs of units. This model, derived from Newton's gravitation law, was inspired by the empirical observation of migratory exchange summarized by Ravenstein in the 19th century (Ravenstein, 1885):  The population influx for two spatial units depends on their respective populations and decreases with the distance. Each unit has its own “population send capacity” and its “carrying capacity” but the influx depends also on the distance between the two units. Following a modified version of the gravity model (D'Aubigny et al., 2000), the population influx for units i and j, Fij, was expressed as F ij ¼

Pi Pj ; PðDij Þ

with Pi and Pi the population of units i and j and P(Dij) a polynomial function of the distance and with P having a log linear link with the observed influx Fij. The degree of the polynomial function was limited to two because, for higher degrees, P did not monotonically increase over the range of possible distances (0–1000 km). Lastly, the quadratic effect takes into account the decreasing effect of distance on intensity of population exchange (see Appendix for explanation). The influxes Fij were considered to have a Poisson distribution since the probability of migrating from unit i to unit j is relatively low compared to the size of the population. The overdispersion specifically resulting from the

236

W. Ghosn et al. / Health & Place 24 (2013) 234–241

higher than expected occurrence of influxes equal to 0 was taken into account by the following Zero-Inflated Poisson (ZIP) model (Lambert, 1992):

connection, the urban area which received the largest number of migrants was arbitrarily labeled UAi.

log ðEðF ij ÞÞ ¼ log ðP i P j Þ þ β 0 log ðDij Þ þ β1 ½ log ðDij Þ2 ( 0 with Prob: φij F ij PoissonðEðF ij ÞÞ with Prob: 1  φij

2.6. Mean square of the mortality differences (MS)

The Vuong statistical test showed a significantly better fit for the ZIP model than for the Poisson model (p o0.01), which resulted in the choice of the former (Vuong, 1989). The model enabled estimation of the expected influx E(Fij), for units i and j, given the size of their populations and the distance between them. 2.4. Characterization of migratory influxes In order to characterize the intensity of migratory exchanges, the ratios of the observed to the expected influxes Rij ¼Fij/E(Fij) were calculated and then classified as “Weak”, “Average”, and “Strong” influxes with population-weighted tertiles as cut-offs: 0.70 (33%) and 1.37 (66%). However, given the marked variations in the ratios resulting from random fluctuations for small expected influxes, the study was limited to ratios with a probability of incorrect classification of less than 0.3, i.e. a minimum expected influx of 14 people (see Appendix for explanations). Thus, out of 129,960 influxes, the analysis was restricted to 23,196 influxes.

The mean square of the log SMR differences was calculated for each category k of the typology of connections for several causes of death as MSk ¼

1 ∑ ðlog SMRi  log SMRj Þ2 ; nk ði;jÞ A k

where nk is the number of pairs in the category k. Then, the ratios between the mean square of log SMR of each category k and that of the “Average – Average” category were considered in order to compare the mortality differences between different causes of death RatiokMS ¼

MSk MSAverage  Average

The statistical significance of the difference between the outcomes and one was calculated with a 1% test, using the delta method and a normal approximation, taking into account the correlation between groups since the same urban area may appear several times in the categories. The 1% significance level was used to minimize the problems associated with multiple testing. 2.7. Mean square of mortality rate differences adjusted on deprivation variables (FDep99)

2.5. Typology of connections between 1990 Urban Areas The final typology of connection was derived from the influx characterizations. Out of 23,196 influxes, the typology was reduced to 23,196/2 ¼11,598 pair-wise connections by gathering the intensity of the influxes from area i to area j (Rij) and the intensity of the influxes from area j to area i (Rji). Then, the pairs of urban areas were grouped in terms of their intensities of connection using two types of categories:  The balanced categories: “Weak – Weak”, “Average – Average”, “Strong – Strong”  The unbalanced categories: “Weak – Average”, “Weak – Strong”, “Average – Strong” For example, the “Weak – Strong” category of connection means that the migration from place i to place j was weaker than expected according to distance and population and the flow from place j to place i was greater than expected. For each pair-wise

The mortality of the urban areas was adjusted based on the variables that compose FDep99, a French deprivation index calculated for the year 1999 (Rey et al., 2009). A Poisson model was used to regress the SMR on unemployment rate, percentage of blue-collar workers in the active population, median household income and percentage of high school graduates in the population aged 15 years and older. About 40–60% of the mortality deviance, depending on gender and age groups, was accounted for by this adjustment. The mean square of the log residual differences, and the ratios with the “Average – Average” category taken as reference, were calculated and compared. The same statistical significance test as for the MS of log SMR was performed. 2.8. Sensitivity analyses As a sensitivity analysis the deprivation index FDep99 was also calculated in each category of connection excluding the pairs that

Table 1 Description of the typology of connection. Categories of connection

Weak–Average UAi

Number of urban areas Population influxn Expected population influxn Number of pairs Average FDep99nn Average FDep99 w/o Parisnn Mortality rate 1990nnn Mortality rate 1999nnn

307 171 321

UAj

308 284 321 2520  0.34 -0.36  0.19 0.18 6.99 7.31 5.91 6.19

Weak–Strong

Average–Strong

UAi

UAj

UAi

193 24 44

214 77 44

274 358 320

865  0.25  0.26 6.78 5.83

 0.26 0.21 7.48 6.36

Weak–Weak

Average–Average

Strong–Strong

Total

352 491 1114 2521  0.36 0.17 7.29 6.16

337 552 547 1649  0.51  0.14 7.00 5.89

345 1769 695 1973  0.54  0.22 6.85 5.82

361 4,318 3726 11598  0.46  0.06 7.06 5.97

UAj

313 592 320 2070  0.3  0.78  0.25  0.09 6.8 6.94 5.81 5.81

For each pair-wise connection, the urban area which received the largest number of migrants is labeled UAi. n In thousands. For example, in the "Weak—Average" category 171,000 people moved from the UAi to the UAj, which was less than expected (321,000) for each pair belonging to the category. The influxes were thus classified as “Weak”. Conversely, in the same category, in all 284,000 people moved from UAj to UAi. Each influx was close to the expected influx estimated by the gravity model so the influxes were classified as “Average”. nn Weighted by the population average French Deprivation Index, 1999, with and without Paris. nnn Mortality rate per 1000 inhabitants standardised on age and gender (with the 1976 European population taken as reference).

W. Ghosn et al. / Health & Place 24 (2013) 234–241

include Paris as a component. This choice was made because Paris exerted a strong influence on this outcome due to the size of the population (11 million people) and high occurrence in all the categories. Table 1. Another point concerned the “Strong–Strong” category. Although it was classified as balanced in the main analyses, in some of its pairs, directional flows are clearly unbalanced. Therefore, as a sensitivity

237

analysis, a more balanced subgroup containing the 20% closest influx intensities in the “Strong–Strong” category was considered to compare mortality differences with the reference category.

3. Results 3.1. Description of the typology of connection between urban areas In the three unbalanced categories (“Weak–Average”, “Weak– Strong” and “Average–Strong”) the mortality was lower on average in the urban areas that received more population. As was the case for mortality, deprivation was lower in the more attractive areas, in particular when Paris was excluded. For the balanced category (“Weak–Weak”, “Average–Average”, “Strong–Strong”), the mortality was negatively associated with connection intensity. The pairs of urban areas that were strongly or moderately connected were also, on average, less deprived than weakly connected urban areas. Overall, the balanced category had much more population influx than the unbalanced categories. Table 1.

3.0

ratios of MS mortality differences

2.5

2.0

1.5

3.2. Association between intensity of connection and mortality differences 1.0

0.5

0.0

Both

Men

Weak − Weak

Women

Weak − Strong

Average − Strong

Weak − Average

Strong − Strong

99% Confidence Interval Fig. 1. Ratios of Mean Square (MS) of pair-wise mortality differences (each category vs. “Average – Average”)

There was a significant association between mortality differences and intensity of connection. The association was monotonic and steady (Fig. 1). The urban areas that were weakly connected had larger mortality differences than those that were strongly connected. The greatest differences in mortality were observed for the “Weak – Weak” category. The difference decreased for the unbalanced connections “Weak – Strong” and “Weak – Average”, for which the MS were much greater than for the “Average – Average” reference category. The urban area pairs that were the most connected, the “Strong – Strong” category, had significantly smaller differences in mortality than the reference category. The MS of the unbalanced category “Average – Strong” was not significantly different from that of the “Average – Average” category. Overall, the association was as strong and significant

Table 2 Ratio of Mean Square (MS) of pair-wise mortality differences in 1999 by cause of death and by category of connection. Categories of connection Weak – Weak

Weak – Strong

Weak – Average

Average – Average

Average – Strong

Strong – Strong

All-causes Both Women Men Both o 65 years Women o65 years Meno 65 years

2.55n 2.27n 2.47n 2.06n 2.19n 1.88n

2.03n 1.59n 2.17n 2.00n 1.93n 1.88n

1.86n 1.67n 1.82n 1.69n 1.65n 1.60n

1.00 1.00 1.00 1.00 1.00 1.00

1.06n 0.97 1.11n 1.09n 1.05 1.09n

0.77n 0.78n 0.80n 0.84n 1.00 0.86n

By cause of death (both) Ischaemic heart disease Cerebrovascular disease Respiratory disease Cancer Lung Upper airways and digestive tract Breast Prostate Colorectal Bladder Violent deaths Alcoholic cirrhosis and psychosis

1.70n 1.63n 3.05n 1.88n 1.52n 1.71n 1.54n 1.17 1.25 1.21 0.77n 2.07n

1.73n 1.06 1.83n 2.24n 1.33n 2.17n 1.37n 1.12 1.41n 0.99 1.05 2.34n

1.42n 1.23n 1.84n 1.70n 1.29n 1.76n 1.29n 1.16 1.22n 1.12 0.96 1.89n

1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

1.01 0.97 0.99 1.13n 0.98 1.05 1.00 0.99 1.07 1.09 1.00 1.13n

0.87n 0.87n 0.87n 0.76n 0.88n 0.77n 1.02 1.08 0.94 1.17 0.82n 0.77n

For example, in 1999 mean square of mortality differences by cancer of the upper airways and digestive tracts was 71% higher for “Weak–Weak” connected areas than for “Average–Average” connected areas. In addition, mortality differences in “Strong–Strong” connected areas were 29% (0.77) lower than the “Average–Average” reference category. n

Significantly different from the MS of mortality differences in the “Average – Average” reference category (p o 0.01).

238

W. Ghosn et al. / Health & Place 24 (2013) 234–241

Table 3 MS ratios of pair-wise mortality differences in 1999 after adjustment on deprivation variables (Fdep99) by cause of death and by category of connection. Categories of Connection Weak – Weak

Weak – Strong

Weak – Average

Average – Average

Average – Strong

Strong – Strong

All-causes Persons Women Men Persons o 65 years Women o65 years Men o 65 years

1.36n 1.39n 1.41n 1.33n 1.34n 1.27n

1.29n 1.10 1.40n 1.47n 1.36n 1.40n

1.22n 1.20n 1.22n 1.25n 1.18n 1.23n

1.00 1.00 1.00 1.00 1.00 1.00

1.10n 0.96 1.17n 1.17n 1.06 1.19n

1.01 0.92 1.06 1.15n 1.23 1.16n

By cause of death (Persons) Ischaemic heart disease Cerebrovascular disease Respiratory disease Cancer Lung Upper airways and digestive tract Breast Prostate Colorectal Bladder Violent death Alcoholic cirrhosis and psychosis

1.18n 1.30n 1.90n 1.20n 1.11 1.27n 1.48n 1.28 1.13 1.18 0.83n 1.37n

1.25n 0.89 1.19n 1.57n 1.05 1.52n 1.35n 1.06 1.23 0.97 1.01 1.55n

1.06 1.13n 1.28n 1.21n 1.03 1.39n 1.29n 1.13 1.08 1.16 0.87n 1.35n

1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

0.93 0.97 1.04 1.09 0.98 1.05 1.03 0.99 1.02 1.02 0.99 1.14n

0.90 0.94 0.99 0.90 0.91 1.02 1.07 1.20 1.02 1.05 1.00 1.08

n

Significantly different from the MS of residual differences in the “Average – Average” reference category (p o 0.01).

for premature mortality, except for women, in the strongly connected pairs of urban areas. Table 2. The causes of death monotonically associated with connection were all cancers, upper airways and digestive tract cancers, and alcoholic cirrhosis and psychosis. Ischaemic heart disease, cerebrovascular disease, and lung and breast cancer were partially associated with greater mortality differences, specifically for the “Weak – Weak”, “Weak – Strong” and “Weak – Average” categories. The remaining causes of death were not significantly associated with connection category. 3.3. Association between intensity of connection and mortality differences adjusted on deprivation variables Overall, after adjusting for the deprivation, the associations weakened. The MS differences of the residual mortality were still significantly greater for the “Weak – Weak”, “Weak – Strong” and “Weak – Average” categories (at least one-way weak connection). The mortality differences in the remaining categories “Average – Strong” and “Strong – Strong” were not different from the reference category: “Average – Average”. But, when the subgroup of the most balanced Rij  Rji intensities in the “Strong–Strong” category was considered, the MS was significantly smaller than for the reference category (0.90, p o0.01 result not shown). There was still a significant association between the categories with at least one-way weak connection and larger MS differences in the residual mortality for both genders, men, mortality under 65 years, all cancers, upper airways and digestive tract cancer, and alcoholic cirrhosis and psychosis. There was no longer any association for lung cancer. Table 3.

4. Discussion 4.1. Summary The association between connection and mortality differences was significant, monotonic and steady. The differences in pair-wise urban area mortality rates were smaller when the spatial units were more strongly connected. The presence of at least a one-way

weak influx was associated with an increasing difference in mortality between urban area pairs. The strongest associations were observed for the causes of death closely related to socioeconomic level and risk behaviors (smoking, alcohol). After controlling for deprivation, the association between connection and mortality was reduced. The strong connection was no longer associated with smaller differences in mortality except for the pairs with the most similar intensities of connection. The “Weak– Weak” connections were generally associated with significantly greater differences in mortality, even higher than the “Weak– Strong” category, the most unbalanced connections. 4.2. Interpretation The interpretation of the association between connection and mortality differences should first consider the chosen definition of connection. Residential migration and more generally demographic changes are not only deprivation markers but are also independently associated to mortality (Ghosn et al., 2012). However, in France, socioeconomic factors were not found to be associated with regional attractiveness (Audureau et al., 2013; Baccaïni, 2007). During the 1990–1999 period, the border region of the Atlantic ocean, Mediterranean sea and hinterland as well as South-west France have experienced the highest net internal migration rate (Baccaïni, 2001). This study highlights the associations between intensity of connection and all-causes and cause-specific mortality differentials. However, these associations, especially the large mortality differences observed in the unbalanced categories, may be influenced by population dynamics which have been highlighted as important in previous publications (Boyle et al., 1999; Brown and Leyland, 2010; Davey Smith et al., 2001; Ghosn et al., 2012; Molarius and Janson, 2000). In the unbalanced connection, after controlling for deprivation, mortality differences remain significantly higher than average for connections with at least one-way weak influx. In addition, adjusting on population growth rate had no impact on mortality differentials (result not shown) which suggests that the results are not due to population change in either the origin or destination area. It is also unlikely that health-selective migration is an important factor. The numbers of people migrating between pairs of urban areas represent only a very small fraction of the whole population and therefore are

W. Ghosn et al. / Health & Place 24 (2013) 234–241

unlikely to have a significant impact on mortality differentials (Verheij et al., 1998). In addition, studies of all-causes mortality in Britain (Brimblecombe et al., 1999) and cardiovascular mortality in England and Wales (Strachan et al., 1995) found that selective migration had no significant impact on regional geographic variation in mortality. In general, researchers have observed a statistically significant health selection effect on geographical mortality inequalities only at finer geographical scales (Brimblecombe et al., 2000). This study did not analyze selection effects directly. Overall, the approach used in this study is not suitable for testing the selection hypothesis, because the same urban area can be both an attractive component of one migration pair and a repulsive component of another. In the case of balanced connections (“Weak – Weak”, ”Average – Average”, ”Strong – Strong”), the pairs of connected urban areas may have similarities that are less marked for un-connected cities. The stronger the connection was, the more similar were the mortality levels. Similarities in characteristics of the areas, such as socio-cultural context, amenities, natural environment and climate, could result in more balanced population exchanges. However, these factors were not investigated directly. Starting from an individual point of view, the propensity to migrate is related to age, socio-demographic characteristics and life-events (studying, employment, marriage, children, retirement) (Baccaïni, 2006; Champion et al., 1998). On the other hand, the choice of the destination is influenced by the information the individual has collected through his or her personal relationship and broadcasting channels but is also subject to comparison with his own living environment (Baccaïni, 2006). Given that long-distance migration is a purposeful decision, mutual attractiveness between areas might reflect their complementarities more than their similarities: for example, a person may migrate from a southern to a northern city for a higher paying position while another person might migrate in the reverse direction in order to live in a place with milder weather. Strongly connected urban areas have been reported least deprived on average. This might highlight that areas with high socio-economic levels are attractive to each other so that their mortality levels are similar. After controlling for deprivation, the association between strongly connected areas and smaller mortality differences vanished. However, the large heterogeneity of the “Strong – Strong” category may have influenced this outcome. The elements above suggest that, for the strongest reciprocal exchanges of population, similarities in mortality that go beyond socio-economic differentials are at work. Moreover, diseases related to risk behaviors such as alcoholic cirrhosis and psychosis (alcohol consumption) and upper airways and digestive tract cancers (tobacco smoking and alcohol consumption) were the diseases with the strongest association between connection categories and mortality differentials. The reciprocal exchanges may be a marker of connection that implies not only residential mobility but also social networks and socio-cultural similarities that may influence social norms and lifestyle, and thus health-related behaviors. Finally, the lack of connection (“Weak – Weak”) may be interpreted as a marker of distinction related to incompatibility and differences between spaces that are supposed to interact (with the definition of connection used). A highly unbalanced connection is associated with a closer mortality level in comparison to unconnected urban areas. This result is a major finding since it emphasizes that cities that are not related, even in a unilateral exchange, face greatest distance in terms of mortality. Although this result vanished after controlling for deprivation, one interpretation of this result is that one side of these weak connections may face complete isolation from the regional and indirectly from the national dynamic. Further studies should explore this issue by quantifying the connectivity of each urban

239

area as in Dennett and Stillwell (2011) and identifying factors associated with this potentially adverse additional effect. 4.3. Deprivation, migration and mortality Deprivation is likely to contribute strongly to the attractiveness of spaces. For that reason, deprivation was considered a confounding factor in the association between connection and mortality differences in this study. However, migration dynamics can also affect deprivation. Considering this reversed scenario, adjusting mortality for deprivation would be erroneous if the association between connection and mortality differences would have been interpreted as causal. The link between migration and deprivation is complex and the two phenomena are likely to be selfreinforcing. In further studies, a more detailed and dynamic analysis, taking into account changes in deprivation, connection and mortality levels, would undoubtedly enhance the interpretation of the results. 4.4. Reflections One limitation of the study relates to the choice of distance “as the crow flies” for the gravity model. An alternative distance between the centres of urban areas could have been travel time since it is shorter between bigger cities. However, travel time may also depend on traffic density, or the choice of means of transportation, which may vary with the socioeconomic level of the population. Thus, selecting travel time would require additional assumptions. Moreover, the use of a large aggregated scale implies that travel time is not relevant since it might require more than an hour to go from an inner extreme point to another in the same unit. The scale of analysis, the 1990 Urban Area, omitted rural areas and areas that belong to several urban cores. However, using this scale enabled most of the pair-wise connections to be characterized with sufficient precision while regional and national exchanges and mortality differentials were also captured. 4.5. Future directions This study focused on connections between pairs of units. This straightforward approach contributes to describing some of the underlying dimensions of spatial inequalities that are not measured by usual indicators. It would be interesting to investigate the common features shared by strongly connected areas. Also, it would be of interest to investigate whether a strong connection with a healthy area would influence mortality independently of the socioeconomic changes. A network approach that considers cities as nodes in a network and connections as the edges may warrant investigation. As a first stage, spatial units could be characterized through their degree of connection, identifying areas that are isolated from their local network. Maintaining relatively weak population exchanges with surrounding spatial units doubtless reflects a specificity or singularity related to individual and/or contextual features that may influence health. Secondly, the area (node) could be categorized by the characteristics of the areas that are connected to it taking into account the direction and strength of these links (the edges). In order to characterize the nodes, information on the incoming and outgoing population should also be taken into account (age groups, socio-professional categories, etc.). A further step would be to use graph theory algorithms and classification algorithms to draw a typology of the various networks, for example: centered versus connected continuous urban structure. Focusing on network forms and distributions should also help identify isolated nodes, and linked and fully connected

240

W. Ghosn et al. / Health & Place 24 (2013) 234–241

nodes. Moreover, the network to which a spatial unit belongs could be determined without using administrative boundaries.

5. Conclusion This paper emphasizes the connections between areas in order to reveal common features that are associated with mortality. In the context of increasing spatial mortality inequalities, future studies should focus on spatial interactions and explain their role. An investigation of the non-measured features that are likely to be responsible for the correlation of mortality between urban areas would be necessary. In addition, failing to account for spatial similarities in mortality that are unrelated to spatial proximity may result in overestimating the statistical power of associations in ecological studies. This exploratory study renews the idea that spatial interactions may be a line of investigation worth exploring with a view to further elucidating the mechanisms that underlie and perpetuate spatial mortality inequalities.

Appendix A. Supplementary material Supplementary data associated with this article can be found in the online version at http://dx.doi.org/10.1016/j.healthplace.2013. 10.003. References Aberg Yngwe, M., Kondo, N., Hägg, S., Kawachi, I., 2012. Relative deprivation and mortality – a longitudinal study in a Swedish population of 4,7 million, 1990– 2006. BMC Public Health 12, 664. Audureau, E., Rican, S., Coste, J., 2013. From deindustrialization to individual healthrelated quality of life: multilevel evidence of contextual predictors, mediators and modulators across French regions, 2003. Health Place 22C, 140–152. Baccaïni, B., 2001. Les migrations internes en France de 1990 à 1999: l'appel de l'Ouest. Econ. Stat. 344, 39–79. Baccaïni, B., 2006. Observation et concepts en matière de migrations internes. Démogr.: Anal. Synth. 8, 65–88. Baccaïni B., Les flux migratoires interrégionaux en France depuis 50 ans, in: INED, Population 2007 n°1, 143–160. Bentham, G., 1988. Migration and morbidity: implications for geographical studies of disease. Soc. Sci. Med. 26, 49–54. Besag, J., York, J., Mollie, A., 1991. Bayesian image restoration with two applications in spatial statistics. Ann. Inst. Stat. Math. 43, 1–59. Borrell, C., Plasencia, A., Pasarin, I., Ortun, V., 1997. Widening social inequalities in mortality: the case of Barcelona, a southern European city. J. Epidemiol. Community Health 51, 659–667. Boyle, P., Norman, P., Rees, P., 2004a. Changing places. Do changes in the relative deprivation of areas influence limiting long-term illness and mortality among non-migrant people living in non-deprived households? Soc. Sci. Med. 58, 2459–2471. Boyle, P.J., Exeter, D., Flowerdew, R., 2004b. The role of population change in widening the mortality gap in Scotland. Area 36, 164–173. Boyle, P.J., Gatrell, A.C., Duke-Williams, O., 1999. The effect on morbidity of variability in deprivation and population stability in England and Wales: an investigation at small-area level. Soc. Sci. Med. 49, 791–799. Boyle, P.J., Norman, P., Popham, F., 2009. Social mobility: evidence that it can widen health inequalities. Soc. Sci. Med. 68, 1835–1842. Brimblecombe, N., Dorling, D., Shaw, M., 1999. Mortality and migration in Britain, first results from the British Household Panel Survey. Soc. Sci. Med. 49, 981–988. Brimblecombe, N., Dorling, D., Shaw, M., 2000. Migration and geographical inequalities in health in Britain. Soc. Sci. Med. 50, 861–878. Brown, D., Leyland, A.H., 2010. Scottish mortality rates 2000–2002 by deprivation and small area population mobility. Soc. Sci. Med. 71, 1951–1957. Carstairs, V., Morris, R., 1989. Deprivation: explaining differences in mortality between Scotland and England and Wales. Br. Med. J. 299, 886–889. Champion, A.G., Fotheringham, A.S., Rees, P., Boyle, P., Stillwell, J., 1998. The Determinants of Migration Flows in England: A Review of Existing Data and Evidence: A Report Prepared for the Department of the Environment, Transport and the Regions. Department of Geography, University of Newcastle upon Tyne, Newcastle upon Tyne. Congdon, P., Shouls, S., Curtis, S., 1997. A multi-level perspective on small-area health and mortality: a case study of England and Wales. Int. J. Popul. Geogr. 3, 243–263.

Conlon, E., Waller, L.A., 1998. Flexible Neighborhood Structures in Hierarchical Models for Disease Mapping. University of Minnesota Biostatistics Research Report, No. rr98-018. Connolly, S., O'Reilly, D., Rosato, M., 2007. Increasing inequalities in health: is it an artefact caused by the selective movement of people? Soc. Sci. Med. 64, 2008–2015. Courgeau, D., Pumain, D., 1996. Analyse Spatiale de Données Biodémographiques. John Libbey Eurotext. Cummins, S., Curtis, S., Diez-Roux, A.V., Macintyre, S., 2007. Understanding and representing ‘place’ in health research: a relational approach. Soc. Sci. Med. 65, 1825–1838. D'Aubigny, G., Calzada, C., Grasland, C., Robert, D., Viho, G., Vincent, J.M., 2000. Approche poissonnienne des modèles d'interaction spatiale. Cybergeo: Eur. J. Geogr. , http://dx.doi.org/10.4000/cybergeo.4357. Davey Smith, G., Shaw, M., Dorling, D., 1998. Shrinking areas and mortality. Lancet 352, 1439–1440. Davey Smith, G., Shaw, M., Dorling, D., 2001. Population change and mortality in men and women. J. Epidemiol. Community Health 55, 9. Dennett, A., Stillwell, J., 2011. A new area classification for understanding internal migration in Britain. Popul. Trends 145, 146–171. Gatrell, A.C., 1997. Structures of geographical and social space and their consequences for human health. Geogr. Ann.: Ser. B, Human Geogr. 79, 141–154. Ghosn, W., Kassié, D., Jougla, E., Salem, G., Rey, G., Rican, S., 2012. Trends in geographic mortality inequalities and their association with population changes in France, 1975–2006. Eur. J. Public Health , http://dx.doi.org/ 10.1093/eurpub/cks078. Jongeneel-Grimen, B., Droomers, M., Stronks, K., Kunst, A.E., 2011. Migration does not enlarge inequalities in health between rich and poor neighbourhoods in The Netherlands. Health Place 17, 988–995. Lambert, D., 1992. Zero-inflated Poisson regression, with an application to defects in manufacturing. Technometrics 34, 1–14. Le Jeannic, T., Vidalenc, J., 1997. Pôles urbains et périurbanisation. Le zonage en aires urbaines. INSEE Première 516, 1–2. Leclerc, A., Chastang, J.F., Menvielle, G., Luce, D., 2006. Socioeconomic inequalities in premature mortality in France: have they widened in recent decades? Soc. Sci. Med. 62, 2035–2045. Leyland, A.H., 2004. Increasing inequalities in premature mortality in Great Britain. J. Epidemiol. Community Health 58, 296–302. Mackenbach, J.P., Bos, V., Andersen, O., Cardano, M., Costa, G., Harding, S., Reid, A., Hemström, O., Valkonen, T., Kunst, A.E., 2003. Widening socioeconomic inequalities in mortality in six Western European countries. Int. J. Epidemiol. 32, 830–837. Martikainen, P., Sipila, P., Blomgren, J., van Lenthe, F.J., 2008. The effects of migration on the relationship between area socioeconomic structure and mortality. Health Place 14, 361–366. Menvielle, G., Chastang, J.-F., Luce, D., Leclerc, A., 2007. Changing social disparities and mortality in France (1968–1996): cause of death analysis by educational level]. Rev. Epidemiol. Sante Publique 55, 97–105. Molarius, A., Janson, S., 2000. Population change and mortality in men and women. J. Epidemiol. Community Health 54, 772. Murphy, M., Bobak, M., Nicholson, A., Rose, R., Marmot, M., 2006. The widening gap in mortality by educational level in the Russian Federation, 1980–2001. Am. J. Public Health 96, 1293–1299. Norman, P., Boyle, P., Rees, P., 2005. Selective migration, health and deprivation: a longitudinal analysis. Soc. Sci. Med. 60, 2755–2771. Pampalon, R., Raymond, G., 2000. A deprivation index for health and welfare planning in Quebec. Chron. Dis. Can. 21, 104–113. Pearce, J., Dorling, D., 2006. Increasing geographical inequalities in health in New Zealand, 1980–2001. Int. J. Epidemiol. 35, 597–603. Preston, S.H., Elo, I.T., 1995. Are educational differentials in adult mortality increasing in the United States? J. Aging Health 7, 476–496. Ravenstein, E.G., 1885. The laws of migration. J. Stat. Soc. London 48, 167–235. Rey, G., 2007. Surmortalité Liée Aux Vagues de Chaleur: Facteurs de Vulnérabilité Socio-démographique et causes Médicales de Décès. Université Paris XI. Rey, G., Jougla, E., Fouillet, A., Hémon, D., 2009. Ecological association between a deprivation index and mortality in France over the period 1997–2001: variations with spatial scale, degree of urbanicity, age, gender and cause of death. BMC Public Health 9, 33. Rican, S., Jougla, E., Roudier-Daval, C., Gancel, S., Gourdon, G., 2009. Atlas de la Mortalité par Cancer en France Métropolitaine, Tendances 1970–2004 (Atlas of Cancer Mortality in France. Trends 1970–2004), 1st ed. INCa. Riva, M., Curtis, S., Norman, P., 2011. Residential mobility within England and urban–rural inequalities in mortality. Soc. Sci. Med. 73, 1698–1706. Salem, G., Rican, S., Jougla, E., 2000. Atlas de la Santé en France. Vol. 1: Les Causes de Décès (Atlas of Health in France. Vol. 1: the Causes of Death). John Libbey Eurotext, Paris. Scarborough, P., Allender, S., Rayner, M., Goldacre, M., 2011. An index of unhealthy lifestyle is associated with coronary heart disease mortality rates for small areas in England after adjustment for deprivation. Health Place 17, 691–695. Shaw, M., Gordon, D., Dorling, D., Mitchell, R., Smith, G.D., 2000. Increasing mortality differentials by residential area level of poverty: Britain 1981–1997. Soc. Sci. Med. 51, 151–153. Singh, G.K., 2003. Area deprivation and widening inequalities in US mortality, 1969–1998. Am. J. Public Health 93, 1137–1143. Strachan, D.P., Leon, D.A., Dodgeon, B., 1995. Mortality from cardiovascular disease among interregional migrants in England and Wales. Br. Med. J. 310, 423–427.

W. Ghosn et al. / Health & Place 24 (2013) 234–241

Stringhini, S., Dugravot, A., Shipley, M., Goldberg, M., Zins, M., Kivimäki, M., Marmot, M., Sabia, S., Singh-Manoux, A., 2011. Health behaviors, socioeconomic status, and mortality: further analyses of the British Whitehall II and the French GAZEL prospective cohorts. PLOS Med. 8, e1000419. Townsend, P., 1987. Deprivation. Int. Soc. Policy 16, 125–148. Verheij, R.A., van de Mheen, H.D., de Bakker, D.H., Groenewegen, P.P., Mackenbach, J.P., 1998. Urban–rural variations in health in The Netherlands: does selective migration play a part? J. Epidemiol. Community Health 52, 487–493.

241

Vuong, Q.H., 1989. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica: J. Econometric Soc. 57, 307–333. Windenberger, F., Rican, S., Jougla, E., Rey, G., 2011. Spatiotemporal association between deprivation and mortality: trends in France during the nineties. Eur. J. Public Health 22, 347–353. Zhang, X., Cook, P.A., Jarman, I., Lisboa, P., 2011. Area effects on health inequalities: the impact of neighbouring deprivation on mortality. Health Place 17, 1266–1273.