Lack of ownership of mobile phones could hinder the rollout of mHealth interventions in Africa

Abstract

Mobile health (mHealth) interventions, which require ownership of mobile phones, are being investigated throughout Africa. We estimate the percentage of individuals who own mobile phones in 33 African countries, identify a relationship between ownership and proximity to a health clinic (HC), and quantify inequities in ownership. We investigate basic mobile phones (BPs) and smartphones (SPs): SPs can connect to the internet, BPs cannot. We use nationally representative data collected in 2017-2018 from 44,224 individuals in Round 7 of the Afrobarometer surveys. We use Bayesian multilevel logistic regression models for our analyses. We find 82% of individuals in 33 countries own mobile phones: 42% BPs and 40% SPs. Individuals who live close to an HC have higher odds of ownership than those who do not (aOR: 1.31, Bayesian 95% highest posterior density [HPD] region: 1.24-1.39). Men, compared with women, have over twice the odds of ownership (aOR: 2.37, 95% HPD region: 1.96-2.84). Urban residents, compared with rural residents, have almost three times the odds (aOR: 2.66, 95% HPD region: 2.22-3.18) and, amongst mobile phone owners, nearly three times the odds of owning an SP (aOR: 2.67, 95% HPD region: 2.33-3.10). Ownership increases with age, peaks in 26-40 year olds, then decreases. Individuals under 30 are more likely to own an SP than a BP, older individuals more likely to own a BP than an SP. Probability of ownership decreases with the Lived Poverty Index; however, some of the poorest individuals own SPs. If the digital devices needed for mHealth interventions are not equally available within the population (which we have found is the current situation), rolling out mHealth interventions in Africa is likely to propagate already existing inequities in access to healthcare.

Keywords: Africa; epidemiology; global health; human; inequities; mHealth; medicine; mobile phones; smartphones; telemedicine.

Figure 1.. Basic mobile phone and smartphone ownership in 33 African countries.

(A) Boxplots show the probabilities of not owning a mobile phone (cream), owning a basic mobile phone (BP; orange), or owning a smartphone (SP; red). Country-level probabilities (dots) are overlaid and jittered to reduce overlap. (B) Barplot shows the country-level probabilities of BP ownership (orange) and SP ownership (red) ordered by SP ownership. Geographic distribution showing probabilities of (C) BP ownership and (D) SP ownership in 33 Afrobarometer countries at the sub-national level.

Figure 1—figure supplement 1.. Flow diagram of participant sample sizes.

Shown for 33 countries that collect phone ownership data as part of the Afrobarometer Round 7 (R7) survey.

Figure 1—figure supplement 2.. Internet usage by type of phone owned.

Afrobarometer Round 7 (R7) surveyed individuals on how often they used the internet. Individuals were not asked to specify how they used the internet (i.e. they did not necessarily have to use or own a phone to access the internet).

Figure 1—figure supplement 3.. Phone ownership in the poorest individuals.

(A) Country-level basic mobile phone (BP) ownership in individuals with Lived Poverty Index (LPI) = 3. (B) Country-level smartphone (SP) ownership in individuals with LPI = 3.

Figure 2.. Proximity to health clinics and mobile phone ownership in 33 African countries.

(A) Boxplots show the proportion of individuals who live in close proximity to a health clinic (HC) based on whether they do not own a mobile phone (mean 0.49), own a basic mobile phone (BP; mean 0.56), or own a smartphone (SP; mean 0.66). Country-level probabilities (dots) are overlaid and jittered to reduce overlap. (B) Scatterplot shows the country-specific proportions of individuals who live in close proximity to an HC amongst individuals who: do not own a mobile phone (white dots), own a BP (red dots), or own an SP (black dots).

Figure 3.. Phone ownership by age and gender in 33 African countries.

Population pyramid displays the distribution of the population stratified by gender and 5 year age groupings (with the exception of the 18–20 age class) by ownership of a mobile phone: no mobile phone (cream), a basic mobile phone (BP; orange), or a smartphone (SP; red).

Figure 4.. Country-level gender and urban-rural effects.

(A) Posterior distributions of the country-level effect on mobile phone ownership of being male (compared to female), sorted by median. (B) Posterior distributions of the country-level effect of living in an urban area (compared to living in a rural area), in the same country order as (A). Both (A) and (B) are on the logit-scale – and should be viewed respectively as country-specific adjustments to the population-level effect of (A) being male or (B) living in an urban area.

Figure 4—figure supplement 1.. Model 1 posteriors – (country-level) intercept.

Posterior distributions (medians and 95% highest posterior density [HPD] regions) of $u_{0j}$ , the country-level intercept, for country j.

Figure 4—figure supplement 2.. Model 2 posteriors – (country-level) intercept.

Posterior distributions (medians and 95% highest posterior density [HPD] regions) of $u_{0j}$ , the country-level intercept, for country j.

Figure 4—figure supplement 3.. Model 2 posteriors – (country-level) urban/rural.

Posterior distributions (medians and 95% highest posterior density [HPD] regions) of $u_{1j}$ , the country-level effect of living in an urban area, for country j.

Figure 4—figure supplement 4.. Model 2 posteriors – (country-level) gender and proximity I.

Posterior distributions (medians and 95% highest posterior density [HPD] regions) of $u_{2j}$ , the country-level effect of being female in close proximity to a health clinic (HC), for country j.

Figure 4—figure supplement 5.. Model 2 posteriors – (country-level) gender and proximity II.

Posterior distributions (medians and 95% highest posterior density [HPD] regions) of $u_{3j}$ , the country-level effect of being male not in close proximity to a health clinic (HC), for country j.

Figure 4—figure supplement 6.. Model 2 posteriors – (country-level) gender and proximity III.

Posterior distributions (medians and 95% highest posterior density [HPD] regions) of $u_{4j}$ , the country-level effect of being male in close proximity to a health clinic (HC), for country j.

Figure 4—figure supplement 7.. Model 1 Markov Chain Monte Carlo (MCMC) diagnostics and ROC curve.

(A) Violin plots of the log posterior (top) and No-U-Turn Sampling (NUTS) acceptance statistic (bottom). There are no divergent transitions. To further assess chain convergence, we used ${\displaystyle \hat{R}}$ , the potential scale reduction factor (Gelman and Rubin, 1992). All values of ${\displaystyle \hat{R}}$ are nearly equal to 1 ${\displaystyle (|\hat{R}|<1.005)}$, indicating all chains have converged. (B) NUTS energy plot (Betancourt, 2017) for all four chains. The transition distribution (darker histogram in each plot) and target distribution (lighter histogram in each plot) are moderately well aligned, indicating efficient sampling. (C) The effective sample size $N_{eff}$ from the posterior distribution of each parameter was used to gauge the level of autocorrelation in each sample. ${\displaystyle N_{eff}/N>0.1}$ for all model parameters indicating independent draws within each sample. (D) ROC curve used to assess the diagnostic capability of fitted Bayesian multilevel logistic regression (BMLR) model. The AUC was 0.79 indicating good predictive accuracy.

Figure 4—figure supplement 8.. Model 2 Markov Chain Monte Carlo (MCMC) diagnostics and ROC curve.

(A) Violin plots of the log posterior (top) and No-U-Turn Sampling (NUTS) acceptance statistic (bottom). There are no divergent transitions. Additionally, all values of ${\displaystyle \hat{R}}$ are nearly equal to 1 ${\displaystyle (|\hat{R}|<1.005)}$, indicating all chains have converged. (B) NUTS energy plot for all four chains. The transition distribution (darker histogram in each plot) and target distribution (lighter histogram in each plot) are moderately well aligned, indicating efficient sampling. (C) ${\displaystyle N_{eff}/N>0.1}$ for all model parameters indicating independent draws within each sample. (D) ROC curve used to assess the diagnostic capability of fitted Bayesian multilevel logistic regression (BMLR) model. The AUC was 0.78 indicating good predictive accuracy.