A comparison of all-cause and cause-specific mortality by household socioeconomic status across seven INDEPTH network health and demographic surveillance systems in sub-Saharan Africa

ABSTRACT Background: Understanding socioeconomic disparities in all-cause and cause-specific mortality can help inform prevention and treatment strategies. Objectives: To quantify cause-specific mortality rates by socioeconomic status across seven health and demographic surveillance systems (HDSS) in five countries (Ethiopia, Kenya, Malawi, Mozambique, and Nigeria) in the INDEPTH Network in sub-Saharan Africa. Methods: We linked demographic residence data with household survey data containing living standards and education information we used to create a poverty index. Person-years lived and deaths between 2003 and 2016 (periods varied by HDSS) were stratified in each HDSS by age, sex, year, and number of deprivations on the poverty index (0–8). Causes of death were assigned to each death using the InterVA-4 model based on responses to verbal autopsy questionnaires. We estimated rate ratios between socioeconomic groups (2–4 and 5–8 deprivations on our poverty index compared to 0–2 deprivations) for specific causes of death and calculated life expectancy for the deprivation groups. Results: Our pooled data contained almost 3.5 million person-years of observation and 25,038 deaths. All-cause mortality rates were higher among people in households with 5–8 deprivations on our poverty index compared to 0–2 deprivations, controlling for age, sex, and year (rate ratios ranged 1.42 to 2.06 across HDSS sites). The poorest group had consistently higher death rates in communicable, maternal, neonatal, and nutritional conditions (rate ratios ranged 1.34–4.05) and for non-communicable diseases in several sites (1.14–1.93). The disparities in mortality between 5–8 deprivation groups and 0–2 deprivation groups led to lower life expectancy in the higher-deprivation groups by six years in all sites and more than 10 years in five sites. Conclusions: We show large disparities in mortality on the basis of socioeconomic status across seven HDSS in sub-Saharan Africa due to disparities in communicable disease mortality and from non-communicable diseases in some sites. Life expectancy gaps between socioeconomic groups within sites were similar to the gaps between high-income and lower-middle-income countries. Prevention and treatment efforts can benefit from understanding subpopulations facing higher mortality from specific conditions.


METHODS
: Rate ratios for cause-specific death rates between SES groups, controlling for year, sex, and age (with 95% uncertainty intervals), 0-2 deprivations as reference group  Not CoD = Not Cause of Death (stillbirths not included as deaths in our analysis)

Multiple imputation methods
We utilized the available data to impute SES indicators in cases of missingness. We were able to use longitudinal SES data in the Nairobi, Karonga, and Manhiça site but SES data from a single year in the other sites. Those different processes are explained separately below. Relatively small proportions of person-time and deaths had no corresponding SES data. In each case, we incorporated uncertainty by creating 20 imputed datasets, carrying forward uncertainty into our model results by conducting analysis on each of the 20 imputed datasets. Every imputation used the Amelia II software package in R [2].
Nairobi: In Nairobi, we utilized the full time series of data during imputation, rather than breaking into two periods as in our results presentation. This allowed for better use of information during imputation. The dataset was formatted to be one observation per household per calendar year. We sought to impute missing values for each of the 7 non-asset indicators as well as each of the individual assets that comprise the asset indicator. In our model, we also included terms for the household, the neighborhood (

*Manhiça missingness reported out of 6 indicators
Harar and Kersa: Living standards data were available for one year in each household-either the beginning of the study period or the year that the household was created if it was created after the beginning of the study period. Education data were available at multiple time points, including 2013 and 2016, the beginning and end years of the study period. First, we imputed education information using several assumptions. We needed information about whether individuals were attending school (if they were of the age range expected to attend up to grade 8). We also needed information on the highest educational attainment. If the person was currently attending school, we made the assumption that they had attended school in previous years (provided data were not available on those years). If they were attending school, we utilized their attainment at the time of the survey and imputed back in time by subtracting a year of attainment for each year back in time. If attainment recorded in previous years was greater than what was imputed, then we used the attainment from the previous years. For adults over age 25, we assumed educational attainment did not change over the time period (a fair assumption in relation to the indicator regarding members of the household attaining 5 years of education). We created the education indicators at the household level. We marked the variables missing if there was insufficient information about the individuals in the household to determine whether that household should be categorized as deprived in the given category. We also created variables for whether there were school-aged children in the household, the total children in the household, and the number of those children missing information in order to inform the imputation model.
We combined the education variables with the living standards dataset. The dataset was formatted to be one observation per household, including data from the beginning year (2013) or the first year of the household's inclusion in the surveillance site. To add information for the imputation model to use about neighboring households, we calculated the proportion of neighboring households with deprivations in each indicator. We sought to impute missing values for each of the 7 non-asset indicators as well as each of the individual assets that comprise the asset indicator. In our model, we also included terms for the household, average proportion of neighboring households deprived in each poverty indicator, reported monthly income, the number of residents leaving the house in that year, the number of residents moving in to the house that year, the number of births in the household that year, the number of deaths in the household that year, the person-time lived in the household that year, the number of residents of various age and gender groups (male under-15, female under-15, male 15-49, female 15-49, male 50-69, female 50-69, male 70+, female 70+), number of years the household existed from the start of the study period, an indicator for households that existed in the first year of the study period, and an interaction of the number of years the household existed since the start of the study period and the indicator for those existing in the first year of the study period. The non-education indicators were missing in about 11% of households in Harar and 4% of households in Kersa, while the education indicators were missing in 8.8% and 17.7% of households in Harar and 37.5% and 24.7% of households in Kersa in after the initial assumptions. Kombewa: In Kombewa, we assumed the SES survey data in the initial year to be representative for the corresponding age and sex groups in the sites to assign person-years to SES groups for denominators. For deaths, we merged SES survey data with the death data by household of residence. We found that 4.0% of deaths had missing SES data associated. We imputed the total number of deprivations (ordinal variable, 0-8) for these deaths using age group, sex, and year. Most of the deaths without SES data collected also did not have verbal autopsy data collected (93%), suggesting these households may have been more difficult to reach or were less likely to participate in surveys. These deaths without SES data or verbal autopsy collected constituted 12.5% of deaths in which verbal autopsy data were not collected.
Cross River: In Cross River, we assumed the SES survey data in the initial year to be representative for the corresponding age and sex groups in the sites to assign person-years to SES groups for denominators. For deaths, we merged SES survey data with the death data by household of residence. We found that 1.7% of deaths had missing associated SES data. We imputed the total number of deprivations (ordinal variable, 0-8) for these deaths using age group, sex, and year.

Life expectancy calculations
We had mortality rates for under-1 year, 1-4 years, and then 5-year age groups to age 85 and over for each site, stratified by three SES groups: 0-2 deprivations, 3-4 deprivations, and 5-8 deprivations on our poverty index. We also calculated n a x (average age of death within each age group) from the datasets [3]. We then utilized n m x and n a x to generate full lifetables for each group and show life expectancy at birth.   Mortality rate ratios in Figure S8 generated by negative binomial regressions stratified by sex, controlling for year and age groups. To test for differences between the estimated SES group mortality rate ratios for men and women, we made pairwise comparisons of 10,000 random draws of the estimated rate ratios for men and women. There were three site-groups with less than 5% of draws of rate ratios (comparing the 3-4 and 5-8 deprivation groups to the 0-2 deprivation group) that were lower in men than women: Nairobi 2010-2015, 5-8 deprivations (0.22%); Nairobi 2010-2015, 3-4 deprivations (4.19%); and Nairobi 2003-2009, 5-8 deprivations (2.09%). While other sites did show some consistent patterns (e.g. male relative rate point estimates and ranges were higher than those in women in the same deprivation groups in Manhiça), there were no other site-groups that met the 5% criterion indicating difference in the degree of inequality by sex.
Estimates from the Harar site for the 5-8 deprivation group omitted because of small numbers of death above age 15 in the 5-8 deprivation group. Mortality rate ratios in Figure S9 generated by negative binomial regressions stratified by broad age groups, controlling for year, sex, and age groups within the broader age ranges. Similar to tests of the difference in the degree of mortality disparities by sex, we examined differences between the estimated SES group mortality rate ratios across age groups by making pairwise comparisons of 10,000 random draws of the estimated rate ratios between each age group. There were two different patterns of differences in different sites. First, some sites showed larger SES-related relative mortality disparities at young ages. In the Harar site, the mortality rate ratio for the 3-4 deprivation group compared to the 0-2 deprivation group was larger under age 15 than in the other two age groups (96.78% of draws higher compared to 15 to 39 age group, 98.59% of draws higher compared to 40 plus age group).
In the Karonga and Manhiça sites, the 5-8 deprivation group showed larger mortality rate ratios compared to the 0-2 deprivation group in the under-15 age group compared to the over-40 age group (95.08% of draws and 99.02% of draws higher, respectively). Additionally, some sites showed lower mortality rate ratios in ages 15 to 39 than in other age groups. In the Nairobi site in 2010-2015, the mortality rate ratios were higher in ages 15 to 39 than under age 15 for both the 3-4 deprivation group (96.42% of draws) and the 5-8 deprivation group (97.55% of draws), compared to the 0-2 deprivation group. In the Nairobi site in 2003-2009, this was true of the 5-8 deprivation group as well (96.64% of draws). Meanwhile, the 15 to 39 age group had higher rate ratios between SES groups than the 40 and older age group in many site-groups:

Sensitivity Analyses
While the sites in Nairobi, Manhiça, and Karonga had SES surveys at multiple time points, the sites in Harar, Kersa, Cross River, and Kombewa only had SES data from a survey at one time point, along with data for those who newly moved in to the site. We decided to use the most detailed data possible rather than using only the data from the first available year for the sites in Nairobi, Manhiça, and Karonga. To build evidence around the differences we might have seen in the results for the sites in Harar, Kersa, Cross River, and Kombewa if we had multiple recurring observations of SES data, we compared the results in Nairobi, Manhiça, and Karonga under two assumptions. In both assumptions, we used the imputation that included longitudinal data. In the first, as is in the main text, we used the estimates of the number of deprivations a household experiences on our poverty index that could vary over years. In the second assumption, we simply took the values of the poverty index from the first year of data for each site (or in the case of Nairobi, for the two time periods we split), and assigned the household those poverty index values for the whole time series. While there were differences, they were not large, and the results were qualitatively similar. When households were assigned the poverty index value from the beginning of the time period, the proportion of person-years (and deaths) in the site lived with more deprivations on the poverty index was higher. This is the result of the gradual reduction in poverty on our index that we observed in these sites over time. The pattern of death rates by SES groups remained similar in each site. The largest difference was in Nairobi 2003-2009, where the disparity grew slightly smaller when using the time-varying SES measures. In Manhiça and Karonga, the disparity in the overall death rate grew slightly larger.
Figure S11: Age-sex-standardized mortality rates by socioeconomic group and cause of death category, comparing SES assumptions