Short-term Mobility Patterns in Sub-Saharan Africa

Data source. This analysis used data from Demographic and Health Surveys (DHS) in sub-Saharan Africa. DHS are nationally-representative surveys. All women in selected households are invited to participate, while men are recruited in a subsample of randomly selected households. We included surveys that occurred during or after 2011 through 2017 (due to the availability of precipitation data) and that geocoded enumeration areas (EA). This analysis included 23 sub-Saharan African countries (Table 1). We included men and women with full covariate and outcome data.
Table 1

Survey and sample sizes included in analysis

Country	Survey year	Sample size of women	Sample size of men
Angola	2015-16	14,278	5,675
Burundi	2016-17	17,170	7,533
Cameroon	2011	15,153	7,100
Chad	2014-15	17,291	5,155
Côte d’Ivoire	2011-12	9,766	4,963
Democratic Republic of Congo	2013-14	17,314	7,934
Gabon	2012	8,278	5,543
Ghana	2014	9,279	4,316
Guinea	2012	9,122	3,770
Kenya	2014	14,623	12,702
Lesotho	2014	6,609	2,925
Malawi	2015-16	10,411	4,395
Mali	2012–2013	13,704	2,875
Mozambique	2011	24,522	7,461
Namibia	2013	9,895	4,407
Rwanda	2014-15	13,483	6,203
Sierra Leone	2013	16,616	7,250
Tanzania	2015-16	13,259	3,511
Togo	2013-14	9,449	4,460
Uganda	2016	18,103	5,218
Zambia	2013-14	16,287	14,638
Zimbabwe	2015	9,927	8,381

Measures. To quantify precipitation anomalies, we used Climate Hazards Group InfraRed Precipitation Station data, which combines both satellite and station data to interpolate rainfall globally from 1981 to present [23 (link)]. For each unique survey date/EA combination, we summed the quantity of rainfall that occurred over the previous 12 months. We then ranked this quantity relative to the quantity of annual rainfall at the EA level over the 29 previous years and converted this ranking to an empirical percentile. For example, a value of 0.5 represents the median level of annual rainfall over the past 30 years. We used this continuous percentile deviation measure as the exposure variable; lower numbers (approaching 0) therefore represent drier times, and higher numbers (approaching 1) represent heavier rainfall in the year prior to the survey relative to the previous 29 years. This classification of rainfall deviations as a relative percentile is standard [24 (link)–27 (link)].
The outcome was short-term mobility, defined as being away from the participant’s place of residence for longer than one month over the year prior to the survey. Respondents were asked “In the last 12 months, have you been away from your home community for more than a month at a time?” We considered this variable dichotomous [28 (link)].
We adjusted for socio-demographic covariates that may impact short-term mobility. Covariates include age (continuous), wealth quintile (categorical; assessed using an asset index [29 ]), household size (categorical; 1–2, 3–4, 5–7, and 8+), education level (categorical; none, primary secondary, and higher) and binary indicators for urban (versus rural), currently married, and literate. We also included an indicator variable for survey month to adjust for seasonal changes in short-term mobility.
Statistical analysis. We fit multivariable logistic regression models to assess the relationship between rainfall anomalies and short-term mobility. To assess non-linearities, we modeled rainfall deviations using restricted cubic splines, with the number of knots determined using Akaike’s information criterion. Models were fit separately for men and women. All models included country-level fixed effects and standard errors were clustered at the EA level. Because we included country fixed effects, our models are “within” estimators, comparing survey participants with different rainfall exposures within each country. To calculate relative risks, we computed marginal predicted probabilities of short-term mobility among participants living at lower rainfall relative to the prior 29 years (percentile of 0.15) and at higher rainfall relativive to the prior 29 years (percentile of 0.85) and compared these to participants living in the median level of rainfall (percentile of 0.5). We then compared marginal predicted probabilities at the extremes to probabilities at the median. To assess effect modification by marital status, we generated interaction terms between rainfall percentile deviations and a binary variable representing marital status. We considered effect modification present if the spline-marital status interaction terms had p-values that were jointly < 0.05. Analyses were carried out in R-Cran version 3.4 and Stata version 14.2.
Patient and public involvement. Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.