Association between internal migration and epidemic dynamics: an analysis of cause-specific mortality in Kenya and South Africa using health and demographic surveillance data

Carren Ginsburg; Philippe Bocquier; Donatien Béguy; Sulaimon Afolabi; Kathleen Kahn; David Obor; Frank Tanser; Andrew Tomita; Marylene Wamukoya; Mark A. Collinson

doi:10.1186/s12889-018-5851-5

Research article
Open Access
Published: 27 July 2018

Association between internal migration and epidemic dynamics: an analysis of cause-specific mortality in Kenya and South Africa using health and demographic surveillance data

Carren Ginsburg^1,2,
Philippe Bocquier^3,1,
Donatien Béguy⁴,
Sulaimon Afolabi^1,2,
Kathleen Kahn^1,2,5,
David Obor⁶,
Frank Tanser^7,8,9,
Andrew Tomita^7,10,
Marylene Wamukoya⁴ &
Mark A. Collinson^1,2,11

BMC Public Health volume 18, Article number: 918 (2018) Cite this article

1715 Accesses
4 Citations
5 Altmetric
Metrics details

Abstract

Background

Many low- and middle-income countries are facing a double burden of disease with persisting high levels of infectious disease, and an increasing prevalence of non-communicable disease (NCD). Within these settings, complex processes and transitions concerning health and population are underway, altering population dynamics and patterns of disease. Understanding the mechanisms through which changing socioeconomic and environmental contexts may influence health is central to developing appropriate public health policy. Migration, which involves a change in environment and health exposure, is one such mechanism.

Methods

This study uses Competing Risk Models to examine the relationship between internal migration and premature mortality from AIDS/TB and NCDs. The analysis employs 9 to 14 years of longitudinal data from four Health and Demographic Surveillance Systems (HDSS) of the INDEPTH Network located in Kenya and South Africa (populations ranging from 71 to 223 thousand). The study tests whether the mortality of migrants converges to that of non-migrants over the period of observation, controlling for age, sex and education level.

Results

In all four HDSS, AIDS/TB has a strong influence on overall deaths. However, in all sites the probability of premature death (45q15) due to AIDS/TB is declining in recent periods, having exceeded 0.39 in the South African sites and 0.18 in the Kenyan sites in earlier years. In general, the migration effect presents similar patterns in relation to both AIDS/TB and NCD mortality, and shows a migrant mortality disadvantage with no convergence between migrants and non-migrants over the period of observation. Return migrants to the Agincourt HDSS (South Africa) are on average four times more likely to die of AIDS/TB or NCDs than are non-migrants. In the Africa Health Research Institute (South Africa) female return migrants have approximately twice the risk of dying from AIDS/TB from the year 2004 onwards, while there is a divergence to higher AIDS/TB mortality risk amongst female migrants to the Nairobi HDSS from 2010.

Conclusion

Results suggest that structural socioeconomic issues, rather than epidemic dynamics are likely to be associated with differences in mortality risk by migrant status. Interventions aimed at improving recent migrant’s access to treatment may mitigate risk.

Peer Review reports

Background

The classic framework of epidemiological transition describes the shift from infectious diseases to non-communicable diseases (NCDs) as population income increases and mortality levels decline [1,2,3,4]. However, many low- and middle-income countries are facing a double burden of disease with persisting high levels of infectious disease, coupled with an increasing prevalence of NCD [5, 6]. The 2010 Global Burden of Disease study reported that infectious diseases together with maternal, neonatal and nutritional disorders were amongst the leading causes of premature death in sub-Saharan Africa, accounting for 76% of mortality in under 80 year-olds [2]. Cardiovascular disease burden nevertheless increased in sub-Saharan Africa, with death from cardiovascular disease rising by 81% between 1990 and 2013 [7, 8]. Fast-changing health and connected demographic transitions are underway in these settings altering population dynamics and patterns of health and disease [4]. Empirical evidence concerning the dynamics of epidemics and related population trends has highlighted marked regional and national variability [2, 7], however, these processes have been insufficiently documented due to a scarcity of appropriate data.

It is well established that health transitions and epidemic dynamics are influenced by socioenvironmental and economic conditions [9]. Understanding the shifting cause-pattern of disease and mortality [1], as well as the mechanisms through which changing socioeconomic and environmental contexts may influence health is therefore crucial. Migration, which involves a change in environment and health exposure, is one such mechanism. Migration has been recognised as a key risk factor for infectious as well as chronic disease, and operates through complex pathways.

The migration and health relationship is multi-faceted and bi-directional. Health status can drive a decision to migrate, but such relocation may in turn have an influence on health. A number of hypotheses concerning the relationship between migration and health have been proposed in the literature, each relating to a stage of the migration process [10]. The most utilised hypothesis of selection holds that migrants are often amongst the healthier in the population when they undertake a migration (the “healthy migrant hypothesis”) [10]. A paradox is present since migrants who return to their places of origin have been observed to be amongst the less healthy, resulting in a “salmon bias” effect, first described with respect to the return migration of Hispanic migrants from the United States [11]. Similarly studies of internal migrants in sub-Saharan Africa have described the phenomenon of migrants “returning home to die” as their health in destinations deteriorates [12,13,14].

While migrants may be relatively healthier when they undertake a move, health status may be affected following migration as a result of disruption and stress, behavioural change, unstable employment and possible exposure to disease in a new environment [15,16,17,18]. Migrants may encounter barriers in accessing health care, or be exposed to changing lifestyle factors that negatively affect their health and well-being at destinations [16, 19]. These potential risks may be offset by improved employment and education opportunities, or superior health services and infrastructure at migrant destinations - conditions often associated with urban environments [18]. Nevertheless, the balance between potential negative and positive effects of relocation may vary considerably across settings [20].

Duration of residence in a new location has been found to reduce these effects as migrants adapt to their new environment. The hypothesis of adaptation suggests that over time, migrants’ health and behaviours may assimilate to non-migrants in a destination area, resulting in no discernible difference in health status between migrants and non-migrants [10, 16, 21].

Migration, in particular circular migration, has further been linked to the spread of health conditions or behaviours between origin and destination areas. Studies have revealed how infectious diseases such as HIV may be diffused through a population via migration [22, 23]. This has been attributed to the transmission of infection through physical contact and increased sexual risk behaviours associated with mobility [22, 24]. Non-communicable disease risk has been associated with the process of urbanisation where movement to urban areas may expose migrants to negative lifestyle factors such as unhealthy diets or lower levels of physical activity [19, 25, 26], and these lifestyle behavioural changes may potentially spread to other areas via return migration. The role of migration in the propagation of disease will depend on the stage of the epidemic, the prevalence of the disease in an area, and patterns of movements linking origins and destinations [22, 24]. Evidence on the association between migration and epidemic dynamics, and the pathways linking movement to disease is much needed to improve current understanding of interactions between disease and the social and environmental context [27]. Indeed, intergovernmental organisations and researchers have highlighted the importance of focused regional studies, and the triangulation of varied data sources on burden of disease and surrounding processes such as migration in informing the development of appropriate public health policy and interventions [7, 28].

This paper follows from a previous study of the relationship between internal migration and all-cause mortality in nine Health and Demographic Surveillance System (HDSS) sites in sub-Saharan Africa [21]. The study confirmed that premature adult mortality differed strongly by migration status. It also confirmed the adaptation hypothesis by showing that in the ten-year period after migration, no health differences between migrants and non-migrants were observed in these HDSS populations.

This paper explores the relationship between internal migration and mortality by cause of death through an analysis of four HDSS sites situated in Kenya and South Africa. These are two sub-Saharan African countries that present with high rates of premature adult mortality associated with two primary causes: AIDS/TB and NCDs. The study asks how internal migration status relates to patterns of AIDS/TB and NCD mortality in the context of the double burden of disease in these countries. The paper has two objectives. The first objective is to define for each geographical area, the patterns of AIDS/TB and NCD mortality for a period of nine to fourteen years. The second objective is to test whether the mortality of migrants converges to that of non-migrants as the epidemic evolves over the period of observation.

We hypothesise that health consequences of migration are associated with the extent of diffusion of an epidemic in a particular local population. At the initial phase of the epidemic, we would expect the health of migrants to diverge from that of the host population. As the epidemic is diffused, we anticipate that the health of migrants would converge with that of non-migrants as both origin and destination areas have been affected. This pattern of divergence and convergence would represent an association between migrant status and the dynamics of the epidemic. The alternate hypothesis is that the health consequences of migrants are not associated with the dynamics of the epidemic. This would be observed as a non-convergence of health outcomes between migrants and non-migrants over time. In this instance, we would attribute the persisting differential health consequences amongst migrants and non-migrants to structural determinants that are independent of the epidemic dynamics. These may be individual, household or community factors, or factors relating to health and social systems that result in discriminatory barriers to prevention or treatment.

These convergence or non-convergence patterns may apply to both infectious and non-communicable diseases. Vector-borne diseases may spread more easily through physical contacts but behaviours and differential health care also contribute to the spread of diseases whether communicable or not. We would expect that migrants are more subjected to the divergence-convergence cycle in the case of infectious diseases because of the assumed greater role of the physical environment (transmission of viruses, bacteria, and parasites). However the divergence-convergence cycle may also apply to transmission of health behaviours (e.g.: eating and drinking habits) across social environments through migrants. Conversely the divergence-convergence pattern may not apply at all if migration status is unrelated to physical or social propagation. The epidemic dynamics would then be independent of the migration pattern.

Methods

Study population

The paper employs data from four HDSSs located in Kenya and South Africa. These HDSS centres are members of the International Network for the Demographic Evaluation of Populations and Their Health (INDEPTH), and are part of the INDEPTH MADIMAH project which uses standardised data formats and protocols to analyse prospective longitudinal data on migration and health see [29,30,31]. The HDSS method continuously registers all births, deaths and in- and out-migrations within a geographically defined population. The two South African HDSS sites included in the analysis, Agincourt and the Africa Health Research Institute (AHRI), are located in mostly-rural settlement types in two different provinces of the country. The two Kenyan HDSSs are located in Nairobi and Kisumu. The Nairobi surveillance area consists of two non-contiguous, densely populated urban neighbourhoods, while the Kisumu HDSS is a contiguous, mostly-rural settlement type (see Table 1 for characteristics of the HDSS sites in the sample). The four HDSSs were selected from two countries with high levels of premature adult mortality and a high burden of HIV. In South Africa and Kenya, life expectancy at birth was most recently estimated at 59 years and 61 years respectively, while HIV/AIDS was the leading cause of death in both countries with 54.5 deaths per 1000 attributed to this cause in Kenya and 202.1 per 1000 in South Africa [32]. In 2012, the probability of death between the ages 30 and 70 due to four major NCDs (cancer, cardiovascular disease, chronic respiratory disease and diabetes) was estimated at 27% in South Africa, and 18% in Kenya respectively [32]. The HDSSs included in the study met the following criteria for inclusion in the analysis: cause of death data had been collected, there were a sufficient number of deaths by analysis category, and these sites had undertaken a minimum of nine years of follow-up. Nine years of follow-up was used as an inclusion criterion since it provides a minimum duration of time to analyse migrants who had left and then returned to the HDSS area. In a prior analysis, no significant difference in mortality risk amongst migrants and permanent residents was found ten years following migration due to the effect of adaptation, thus migrants who have been in the HDSS areas for ten years or longer are regarded as permanent residents [21].

Table 1 HDSS sites included in this multi-centre analysis

Full size table

Variables and analysis

Data on causes of death were collected using verbal autopsies that were conducted according to WHO standards [33, 34]. Cause of death assignments based on verbal autopsy data were computed using InterVA4 ver4.02 [35], with cause of death categories corresponding to International Classification of Diseases (ICD 10) [36, 37]. These methods produce standardised data on cause of death across the study locations. In the analysis, causes of death were grouped into a set of broad categories: major risk infectious disease (HIV/AIDS related death or pulmonary tuberculosis); other infections (e.g.: acute respiratory infections, malaria); NCDs (e.g.: diabetes mellitus, acute cardiac disease, stroke) and neoplasms; maternal and neonatal causes; external causes (e.g.: road traffic accidents, assault, self-harm), and unknown or indeterminate causes. The focus of the study is on AIDS/TB and NCD mortality because of the interest in disease dynamics of these two dominant epidemics. Nevertheless, probability of death by cause is presented for all cause categories.

The migration-death competing risk conceptual model employed in the analysis is presented in Fig. 1. Migration in this analysis is defined as a move that crosses the geographical boundary of the HDSS site in an inward or outward direction. Moves that take place within an HDSS area are therefore excluded from the analysis. HDSS sites may apply different time thresholds to define an in- or out-migration, ranging from three to six months residency following a move. In order to standardise the migration definition across the HDSS sites in the study, a six-month residency threshold was applied to determine an individual’s residency status in the surveillance area (see [21, 31] for more details on migration methods).

Migration status is defined as either first time in-migrant to the HDSS area, return migrant to the HDSS or permanent resident (individuals who have not migrated). In-migrants are individuals who have not previously resided in the HDSS surveillance area, while return migrants are former residents who have temporarily relocated (generally to take up employment). In the case of in- and return migrants, the analysis further discriminates risk of mortality by the duration of time since entry into the HDSS area. The effect of duration was observed to be significant in previous work [21]. The models used in this study control for three categories of duration following an in- or return migration to the HDSS area: six months to two years; two to five years and five to nine years. For in-migrants and return migrants, the reference category employed in the models represents the most recent migrants (migrants who have been in the HDSS area for between six months and two years). Migrants who have been in the HDSS areas for ten years or longer are regarded as permanent residents, while very recent in-migrants and return migrants are excluded from the analysis due to the six-month residency threshold described above. For return migrants, the length of time spent outside the HDSS is also controlled for in the models to represent the migrant’s exposure to the destination area prior to return home (with a longer duration of more than three years contrasted with a shorter duration of less than three years). The models therefore control for a net effect of duration of residence, as well as duration of exposure outside the HDSS.

The analysis controls for calendar effects in order to capture the dynamics of the AIDS/TB and NCDs epidemics. Time is divided into three-year periods, which ensures that trends in mortality over time are captured for the different diseases and short-term fluctuations reduced. The four HDSS sites contribute different periods to the analysis depending on the length of time since inception. All sites contribute data from the year 2004 onwards. Models include an effect of period where its coefficients represent the average period effect for permanent residents.

In order to isolate a migration effect by period, the models control for the interaction between period effect and migration status. These terms represent in-migrants and return migrants in the respective periods who have resided in the HDSS area for between six months and two years following entry. The reference category represents permanent residents for each corresponding period. These terms allow us to test the hypothesis of convergence of in-migrants and return-migrants to that of permanent residents in the population over the observation period. Convergence is observed for a specific disease if the difference in mortality risk between migrants and non-migrants declines over time, implying that migration status has become less relevant to the dynamics of the disease. Conversely, the difference in mortality risk by migrant status can increase over time (diverge), or remain stable, showing no relationship to the epidemic.

All analyses are performed separately for males and females because of the difference in migration and mortality patterns between the sexes. Finally, the models control for the following sociodemographic characteristics: age (limited to a 15–60 year age range in accordance with the definition of premature adult mortality) and education level which is time-varying (standardised across the four sites to contrast no formal education with primary, secondary and tertiary-level education).

The Fine and Gray statistical model is used for estimation and is based on the cumulative incidence function that does not assume independence of cause of death [38]. This is a superior approach to the regular Cox proportional hazards model which makes the assumption of independence of different causes of death in the analysis of mortality. The use of this method showed only slight differences in the results compared with the method based on simple hazard rates.

The statistical model for each large cause of death, and separately for males and females, may be written as follows:

$$ {\displaystyle \begin{array}{c} logH\left[t,X(t)\right]={logH}_0(t)+{\beta}_1\ast period\\ {}+{\beta}_2\ast inmigrantperiod+{\beta}_2\ast retmigrantperiod\\ {}+{\alpha}_1\ast inmigrantduration+{\alpha}_2\ast retmigrantduration\\ {}+\gamma \ast retmigrantexposure+\delta \ast educationlevel\end{array}} $$

Where:

“period” is a set of indicators defining each 3-year period (2010–2012 is the reference category),
“inmigrantperiod” and “retmigrantperiod” are two sets of indicators defining periods for in-migrants and return migrants respectively; to note, their effects are measured net of, and add on to the main effect of period,
“inmigrantduration” and “retmigrantduration” are two sets of indicators defining duration since migration into the HDSS for in-migrants and return migrants respectively (with reference category “6 months to 2 years”),
“retmigrantexposure” is an indicator defining the duration spent outside the HDSS for return migrants only (with reference category less than or equal to 3 years),
“educationlevel” is an indicator defining the individual’s broad level of education (the reference category is no formal education)

H₀ is the baseline cumulative incidence function for all the above indicators set to zero (i.e.: representing mortality over the 2010–2012 period for non-migrants with no formal education).

Results

Descriptive results

The characteristics of the sample are presented as an additional table (see Additional file 1: Table S1), which details the person years for all variables in the analysis, by sex. All variables included in the models have person-years at risk (PYAR) in excess of 500. Additional file 1: Table S1 indicates that in- and return-migrants in their first ten years of residence represent between 59% (females in Nairobi) and 27% (males in Kisumu) of the total PYAR during the study period. Rates of in- and out-migration in these HDSS sites were reported on in an earlier publication see [31]. The number and percentage of adult deaths due to AIDS/TB and NCDs, by HDSS site and sex are presented in Additional file 2: Table S2. In all sites, AIDS/TB accounts for the largest proportion of total deaths between ages 15 and 60 (ranging from 36% of deaths among males in the Nairobi HDSS to 71% of deaths among females in the AHRI HDSS).

The probability of death by cause amongst adults aged between 15 and 60 years (i.e.: 45q15, noted “q” for simplicity in the following) is presented in Figs. 2, 3, 4 and 5. These probabilities are calculated based on the cumulative incidence function meaning that they take into account the competing risk. These graphs depict the transition in mortality for each cause of death over the period of analysis. Note that across the four HDSS sites, indeterminate and unknown causes of death comprised between 5 and 15% of all deaths.

Agincourt HDSS, South Africa

The length of mortality monitoring in the Agincourt HDSS captures the rise and fall of the AIDS/TB epidemic in this area. Amongst males and females in the Agincourt HDSS, mortality due to AIDS/TB increased in the early 2000s and, following an inverse-U shape, began to decline in 2006 with the roll-out of antiretroviral therapy (ART). Males experienced slightly higher mortality due to AIDS/TB as compared with females (in the 2010–2012 period q = 0.29 for males and q = 0.19 for females). The probability of mortality due to NCDs began to increase from 2004 for males, reaching 0.16 in the most recent period, and is gradually increasing amongst females (q = 0.11 in 2010–2012). The two main causes of death display different trends, one decreasing and the other increasing.

Africa Health Research Institute HDSS, South Africa

The rapid rise in AIDS/TB-related mortality in the AHRI HDSS is not captured on the graph due to population surveillance starting in the year 2000 after the epidemic was well established. The charts do however capture the decline in AIDS/TB mortality. The probability of death due to AIDS/TB is slightly lower amongst females in the 2010–2012 period (q = 0.23) as compared with males (q = 0.35). With regards to NCDs, there was a slight increase until 2009, but the probability of premature adult mortality due to NCDs decreased in the recent period (q = 0.07 for both males and females).

In both the AHRI and Agincourt HDSS populations, high levels of external causes of death are apparent amongst males (around q = 0.10).

Kisumu HDSS, Kenya

Over the nine years of data available for Kisumu, the decline of the AIDS/TB epidemic is apparent as the probability of death amongst males and females decreased to q = 0.21 and q = 0.15 in the 2010–2012 period respectively. The probability of death from NCDs for both sexes in Kisumu decreased in this same period (q = 0.09 for males and q = 0.06 for females).

Nairobi HDSS, Kenya

The nine years of data from the Nairobi HDSS indicate lower AIDS/TB mortality as compared to the other three sites, albeit higher amongst females (q = 0.11 for males in the 2010–2012 period, and q = 0.14 for females). In these urban neighbourhoods, the trends in NCD mortality have remained stable over the time frame, but higher amongst females (for males q ≈ 0.04 and females q ≈ 0.09). To note that the Nairobi HDSS experienced an upward trend in external deaths amongst males (q = 0.10) in 2010–2012.

In summary, the trends with respect to the four HDSS sites suggest similarities between the sexes within each site. Nevertheless, in cases where sex differences are observed, males tend to experience higher levels of premature adult mortality as compared with females (the exception being in the case of AIDS/TB and NCD deaths in Nairobi).

Results of the competing risk models

The competing risk models are presented as additional tables (see Additional file 3: Table S3, Additional file 4: Table S4, Additional file 5: Table S5 and Additional file 6: Table S6). The coefficients by migration status and period (and corresponding sub-hazard ratios on the right-hand scale) are displayed graphically by HDSS site, cause and sex in Figs. 6, 7, 8 and 9. These graphs allow us to examine how migration status contributes to the trends in mortality over time, after controlling for calendar and duration of residence (adaptation) effects. In all sites, the models reveal a decline in the risk of mortality with increasing duration of residence since entry into the HDSS area. The in-migrants and return migrants represented are those who have resided in the respective HDSS areas for between six months and two years following migration (the reference category in the models). For return migrants, the trends represent those who spent less than three years outside of the HDSS (the reference category). In each graph, permanent residents are represented by the value zero on the coefficient axis and the value one on the sub-hazard ratio axis. Where the graphs by migration status present as a horizontal line, it indicates that the effect of migration status has not changed over time of (i.e.: the period trends observed in Figs. 2, 3, 4 and 5 hold for the full population, regardless of migration status). Else, the graphs may reveal a pattern of convergence or divergence between migrants and non-migrants.