Health aid projects have both expanded and constrained the capacity of health facilities to deliver malaria services to under-five children in Malawi

Conceptual model

The Andersen and Newman Framework of Health Services Utilisation was adapted for this study (online supplementary appendix A).6 The framework treats healthcare use as a result of individual decisions driven by societal resources and service availability. The model states that societal determinants of utilisation affect individual determinants such as predisposition to use services, ability to use them and the need for them both directly and through the health services system.6 This paper adapts this behavioural model of health service utilisation to examine the relationship between both social and demographic factors and the individual utilisation of malaria services (online supplementary appendix A).

Analysis

The analyses used to test the relationships as described in the conceptual model above combine six sources of subnational, spatial-referenced data. Online supplementary appendix B describes each of the variables used in the empirical approach, which construct what the variables relate to in the conceptual model, as well as the data source.

To test the hypothesis that health aid boosts malaria service utilisation indirectly via increasing facilities’ readiness to provide malaria services, this research uses logistic and multinomial logistic regression within a mediation approach. This allows examination of the paths by which aid can plausibly effect utilisation. To assess mediation, I use Baron and Kenney’s four equations, depicted in figure 1. In this figure, the rectangles and the arrows represent the relationship between health aid (X), malaria service readiness (M) and malaria service utilisation (Y).7 This figure also uses notation most widely applied in mediation analysis, with Pathway A representing the relationship between health aid and malaria service readiness (Research Question 2), Pathway B representing the effect of malaria service readiness on malaria service utilisation controlling for health aid (Research Question 3), Pathway C representing the relationship between health aid and malaria service utilisation (Research Question 1) and, finally, Pathway C’ is the effect of health aid on malaria service utilisation controlling for malaria service readiness (Research Question 4). The conceptual model that guides the empirical approach also details and labels these relationships (online supplementary appendix A). Online supplementary appendix C provides the empirical approach to the mediation analysis, including equations.

Request permission

Outcomes

The outcome variables for the logistic analyses are dichotomous, indicating whether or not a child under five who had a fever in the last 2 weeks received (1) advice or treatment, (2) finger or heel stick, (3) malaria artemisinin-based combination therapy (ACT) or (4) any malaria service from the 2012 Malawi Malaria Indicator Survey (MIS). The construction of these outcomes, as online supplementary appendix D describes, are based on guidance outlined in the Household Survey Indicators for Malaria Control, Section 3.3 Case Management among Children under 5 years old.8 The outcome for the multinomial logistic model is an index of malaria service readiness (low, medium, high) that is a quality measure designed to determine the capacity of health facilities to provide malaria services.

Key independent variables

The logistic models use four key independent variables. The first is health aid. The second is an index of facility readiness to provide malaria services. The last two are a subset of facility readiness measures, which include readiness to provide malaria diagnostic services and readiness to provide malaria training. In the multinomial logistic model, the key independent variable was health aid.

Health aid

The first key independent variable is health aid which includes 15 projects sited in 108 locations; totalling USD$120M in cumulative commitments and representing the time period before the Service Provision Assessment (SPA) was completed 2005–2012. This variable was subset from AidData’s Malawi Aid Management Platform (AMP), including data on the locations of aid projects recorded by the Government of Malawi’s AMP and includes roughly 80% of all aid reported to the Ministry of Finance since 2000. Data in Malawi’s AMP are hierarchically coded, where purpose codes represent the overarching purpose of aid projects and activity codes are subcategories of purpose codes indicating specific project activities. Health aid includes only projects that could promote malaria services and utilisation locally, including those with a purpose or activity code from the following aid sectors: basic healthcare, basic health infrastructure, health policy and administrative management, infectious and parasitic disease control, medical education and training or medical services.9 10 In addition, each aid project carries a precision code that indicates the spatial certainty of where the project was allocated. I include all projects with a spatial precision at the district level, excluding projects with spatial precision at the region and country level. This eliminated 27 project locations, leaving 108 in the sample. Projects allocated to a specific location are coded as being allocated to a geographic region, specifically the first-level administrative division, the largest administrative subdivision of a country. Children were coded as living in an area that received aid based on the count of health aid projects allocated to the region in which they resided during the survey. The aid data, while granular, do not distinguish which survey respondents received aid. Therefore, estimates refer to individuals’ residence in areas that received aid.

Index of malaria service readiness

The second key independent variable is an index of malaria service readiness, a quality measure of facilities’ capacity to provide malaria services. Using the Service Availability and Readiness Assessment as a framework and data representing responses to 40 variables included in the 2013/2014 Malawi SPA, an index of malaria service readiness was created using a weighted additive approach11 These 40 variables are categorised into six groups that are outlined in online supplementary appendix E . An illustration of the construction of the total malaria service readiness index is available in online supplementary appendix E. A facility was given one point for reporting yes to each category, producing malaria service readiness scores ranging from 0 to 6. The third and fourth independent variables of interest are a subset of the total malaria service readiness index. Facility readiness for diagnostic capacity combines indicators from rows 1, 2 and 3 (online supplementary appendix E) that include malaria rapid diagnostic testing (RDT) or malaria microscopy diagnostics, either malaria RDT training or malaria microscopy training, and having malaria RDT protocols. Facility readiness for malaria training includes indicators from rows 2, 3, 4 and 6 from online supplementary appendix E: having either malaria RDT or malaria microscopy training, malaria RDT protocols, malaria treatment guidelines and trained personnel. Kernel density estimation (KDE) was used to link the malaria service readiness index to the household indirectly in order to create the index of malaria service readiness. Appendices E and F provide, respectively, further details on construction of the malaria service readiness index and the link (F) between facility-level and household-level data.

Summary statistics

Among the 956 facilities in Malawi in 2013, 49% were health centres, followed by dispensaries (39%) and hospitals (12%). The availability of malaria services varied by type of facility, with hospitals and health centres generally having many of the indicators comprising the malaria service readiness index and dispensaries having few (table 1). Common across all facility types were diagnostic capacity and the availability of first-line medicine to treat malaria; having a staff member recently trained in RDT or microscopy was less common, especially in dispensaries. Hospitals and health centres were similar in the availability of services, except that 79% of hospitals and only 56% of health centres had personnel recently trained in malaria diagnosis and/or treatment. Most dispensaries had RDT and microscopy diagnostics (95%) and first-line medicine (91%), but few had other types of malaria services. Less than one-third of government, Christian Health Association of Malawi and mission/faith-based facilities are malaria-service ready, but only 9% of privately managed facilities are. Readiness was lowest in the southwest (18%) and highest in the central east (35%). Overall, only 35% of facilities had full diagnostic capacity.

Of the 2171 children under five in the sample, 17% received any malaria services, with 7% receiving a finger or heel stick and 10% receiving ACT. Online supplementary appendix H provides descriptive statistics of the individual determinants based on predisposing, enabling and need factors from the conceptual model. Most (75%) of the children in the sample live in rural locations; they were equally distributed across years of age and levels of household wealth and most (71%) lived more than 2 km from a facility. The mean level of the Malaria Ecology Index was 0.310 and was similar in both urban and rural locations with a mean level of risk of 0.274 in urban locations and 0.312 in rural location, consistent with endemic levels of malaria risk across Malawi.

Adjusted associations between health aid and utilisation (Step 1: Pathway C)

Table 2 presents ORs for the adjusted association between health aid and malaria service utilisation. Row 1 presents the adjusted association of health aid across malaria utilisation outcomes and suggests that health aid has a negative or no association with children’s use of malaria services, but these associations are not significant. Using heel stick (Column 2) as an example, the odds (OR=0.967; 95% CI 0.882 to 1.060) of receiving a heel or finger stick are 4% lower for each additional health aid project in the region.

Adjusted associations between health aid and facility readiness to deliver malaria services (Step 2: Pathway A)

Using a multinomial logistic regression model, table 3 presents results in the form of relative risk (RR) ratios, testing adjusted associations between health aid and overall malaria service readiness as well as for diagnostic and training readiness. Columns 1 and 2 present overall facility readiness to deliver malaria services and indicate that increasing the number of aid projects in a region is associated with statistically significant lower levels of malaria readiness. This is also true for malaria training readiness (Columns 5 and 6), but not malaria diagnostic readiness (Columns 3 and 4), where results indicate a positive relationship between increasing aid and diagnostic capacity. More specifically, examining overall malaria service readiness (Column 1) shows that if health aid to a region was to increase by one project, the RR for medium readiness relative to low readiness would likely decrease by a factor of 0.794 (95% CI 0.751 to 0.839). More generally, we can say that if health aid increases in a region by one project, the RR of having a medium level of readiness instead of a low level of readiness likely would decrease by 21%. An additional aid project also decreases high levels of facility readiness (RR=0.954; 95% CI 0.904 to 1.007).

For readiness to provide diagnostic services, results indicate a positive association between aid and diagnostic capacity (Columns 3 and 4). Results indicate that the relative likelihood that a facility has a medium level of diagnostic readiness increases by 12% (RR=1.118; 95% CI 1.060 to 1.179) and the relative likelihood that a facility has a high level of readiness increases by 23% (RR=1.230; 95% CI 1.161 to 1.303). However, malaria training readiness results (Columns 5 and 6) indicate a negative association between aid and training readiness. Specifically, an additional aid project in a region does not increase facilities’ likelihood of having a high level of training readiness (RR=1.016; 95% CI 0.959 to 1.076) and actually decreases their relative likelihood of having a medium level of training readiness by 8% (RR=0.925; 95% CI 0.879 to 0.974).

Adjusted associations between health aid, malarial readiness and utilisation(Steps 3 and 4: Pathways B and C’)

Tables 4A–C are one table split into three tables for clarity. They indicate that when both malaria service readiness and aid were taken into account, there is little association with utilisation. That is, results show an insignificant, though positive, correlation between facilities with a medium level of overall diagnostic capacity and use of malaria services for children. Results indicate a weak significant association (p<0.10) and the odds of children receiving a heel or finger stick was 59% higher in facilities with a medium level of readiness compared with low readiness facilities (OR=1.598; 95% CI 0.915 to 2.793). However, this weak association was not apparent in facilities with a high level of readiness. Among facilities with a high level of readiness the odds of receiving ACT was 45% lower than in those with low levels of readiness (OR=0.559, 95% CI 0.310 to 1.007), but again, this association was weak (p<0.10). Turning to the other independent variable of interest, health aid, results showing a negative association between health aid and all utilisation outcomes are also not statistically significant.

Discussion of results

According to the Baron and Kenney mediation approach, Step 1 typically would necessitate a significant association between health aid and malaria service utilisation to justify continuing the mediation analysis. In recent years, the literature has supported dropping the Step 1 requirement that Pathway C or the ‘total effect’ is significant to assess mediation.14 Recent studies have demonstrated that mediation analyses provide explanatory value and should be pursued even when the total effect is not statistically significant.14–16 In addition, if the ‘total effect’ of health aid on utilisation of malaria services is not significant, there may be practical or theoretical reasons for estimating mediating effects.17 In the case of this study, identifying the association between health aid and malaria service readiness still has scientific merit in order to understand which components of the malaria healthcare system need to be strengthened or require improved measurement.17 Referring to the Baron and Kenney model, results from Step 2 indicate a significant relationship between aid and diagnostic service readiness, but not overall malaria readiness or training readiness. Further, these results appear to present conflicting evidence of the association between health aid and facility readiness such that aid appears inversely associated with facility readiness overall but positively associated with diagnostic capacity. This finding is interesting for several reasons. First, it indicates that health aid is translating to increased malaria diagnostic capacity. Early and accurate diagnosis of malaria is essential to lower mortality and stem the spread of disease.18 Diagnosis is also important, because not every fever contributes to malaria. Due to increasing concerns related to drug resistance, presumptive treatment is no longer recommended unless diagnostic tests are not accessible.18 Second, the finding that aid is translating into diagnostic capacity, but not overall facility readiness highlights a gap that could significantly impact the quality and accessibility of malaria diagnosis. Referring back to the Baron and Kenny model for steps 3 and 4, findings indicate that the coefficient for Pathway B was significant in two cases of overall malaria service readiness. This result is not surprising given the fact that 96% of health facilities in Malawi had full capacity to provide first-line treatment medicine. Individuals may have obtained effective treatment from facilities with a lower level of readiness. The results also suggest that both Pathways C and C’ are similar and not statistically different from zero, suggesting there is neither a direct nor indirect (ie, mediated via facility readiness) association between health aid and malaria service utilisation. Therefore, mediation was not formally tested using a Sobel test.19

Limitations and strengths

Several limitations should be acknowledged when interpreting these results. First, it is not possible to directly link the MIS to the facility. Therefore, KDE was used to assign a density of malaria service readiness to each child. Although this method allows for the incorporation of distance decay as well as the possibility to assign an aggregate measure of readiness to each child, it does not reflect the true travel distance to the clinic. Second, since the KDE disperses the effect of a facility across space by facility characteristics; the malaria service readiness index was derived using an equal weighting additive approach for interpretability. One could argue that the items in the index have different clinical values and thus should be weighted differently. Since a weighted additive index has not been validated within the context of KDE, there are still remaining assumptions that the dimensions do in fact carry equal weights.20 Third and related to the creation of the index, the malaria service readiness index ranked facilities as high, medium or low quality based on the distribution of the score. This was done in order to examine the differences among facilities while retaining a more intuitive understanding, Although it is a common practice in social science to dichotomise continuous variables, it may result in a loss of statistical power and result in an ability to detect significant differences.20–22 For this reason, I categorised the continuous variable into terciles. Discretisation in this way allows for the comparison of the highest and lowest groups and results in less efficiency lost.20 23 Fourth, this work is based on the assumption that the readiness of the facility to provide malaria services influences the decision to use a facility. Parents may choose to use facilities for reasons unconnected to service readiness.13 Fifth, the SPA and MIS data were collected within 1 year of each other. The malaria service utilisation data preceded the service provision data used to determine service readiness. Therefore, this research assumes that the health facility service provision capacities did not change much between the two surveys. Although this follows DHS guidelines for linking SPA and MIS data, data on malaria service availability collected during the SPA survey could have changed after the MIS survey.24 Sixth, the MIS data describe the 2-week period of prevalence of fever among children under five. While a history of fever is often strongly associated with malaria parasitaemia in malaria-endemic countries, children who had a fever may not have had malaria. Seventh, I restrict the AidData dataset to health aid projects that could feasibly have local impacts on malaria services and utilisation. However, I cannot subset the data to only include malaria-specific funding. In addition, AidData does not assign financial amounts to individual project locations. Many researchers choose to divide the total aid amount equally across all activities. In order to test the sensitivity of my results to other definitions of aid, I repeated the analysis defining health aid in terms of the dollar amount across each region. The results of this analysis were qualitatively similar to the more rigorous definition of aid based on counts of health aid projects that were used in this study. Also, this research uses date of planned completion to subset the timing of aid distribution. However, from this dataset, it is not possible to know exactly how much of this aid was actually transferred to the areas in question. A long lag between aid and the period the surveys were conducted may attenuate the association between aid and the outcomes of interest. In addition, enough time may not have elapsed for changes to occur in facility readiness that influence utilisation. In order to test the sensitivity of my results to temporality and endogeneity, I examined aid in two bins of early and late aid. Results of this analysis are located in online supplementary appendix G and indicate that it is appropriate for future research to consider binning aid into early and late periods. Eighth, this analysis assumes that health aid mainly influences the supply side of malaria care. It is possible that aid is used to create demand for malaria care and subsequently increase service use. Although, in theory improved quality of care can stimulate demand for care, in a resource constrained country like Malawi, cost associated with seeking care are formidable and unless these are addressed, improving facility readiness may not stimulate service use.25 Ninth, this analysis does not explicitly control for government and private investments in health that foreign donors do not fund, which could bias estimates by underestimating the aid amount. However, the Malawi government’s contribution to health expenditure has been consistently below the 15% recommended in the Abuja declaration, estimated at 7.2% in 2011.26 Consequently, any bias from excluding government funded interventions should be minimal. Tenth, the cross-sectional nature of both the SPA and the MIS prevent more rigorous estimates of the influence of service readiness on service utilisation. Specifically, I cannot examine whether changes in health aid in a geographic region cause changes in malaria facility readiness or utilisation.

Despite these limitations, this study has several noteworthy strengths. First, this paper linked facility census data and nationally representative household data to provide a more robust assessment of whether and how health aid, facility readiness to provide malaria services and the use of malaria services relate.13 The use of KDE to join these data provides a methodological improvement in the way the service environment was constructed and linked to household survey data. Similar to Wang, but with a different empirical approach, my analysis incorporates distance decay to measure the effect of the malaria service environment instead of only looking at the closest facility, reducing the likelihood that individuals were misclassified to a specific facility.13 Second, this is the first study to my knowledge to combine subnational aid flows with individual-level and facility-level measures of malaria service availability and use. Third, the mediation approach was innovative in its attempt to provide evidence on potential mechanisms by which health aid and malaria utilisation relate and identifies a need to strengthen or improve measurement of programme components.