Directly Observed Therapy

Data were double entered and validated using EpiData version 3.1 [22] , and Stata 11.0 [23] was used for data processing and analysis. Analyses accounted for the survey design, adjusting for clustering within health facilities. Sample weights for observations were estimated based upon the sampling probability of each selected health facility [19] . Analyses were stratified by CSRef and collated CSCom as we hypothesised that the systems effectiveness and the predictors of the intermediate processes involved in delivery of IPTp-SP would differ at these different levels of the health system.
Two systems effectiveness analyses were undertaken for IPTp-SP: 1) a cumulative analysis of the overall effectiveness of the intermediate processes for women who reported during exit interview that they were 4 to 8 months pregnant for their 1^stand for their 2^nd visit to ANC, and 2) an assessment of the effectiveness of each individual intermediate process for women 4 to 8 months pregnant for their 1^st, and 2^nd visits to ANC. The effectiveness of each intermediate process in the delivery system effectiveness algorithm was calculated by estimating the proportion of women who successfully reached each step from the previous step [17] (link). Two cumulative delivery system effectiveness estimates were calculated 1) a ‘per policy’ estimate of receiving 3 tablets of SP by directly observed therapy (DOT) on first, and second visits to ANC; 2) a with or without DOT inclusive estimate of the proportion of women either receiving 3 tablets of SP by DOT, or leaving the health facility with 3 tablets of SP and able to report on questioning that they would take all 3 tablets at one time. For ITNs the cumulative effectiveness included only 3 steps which were attend ANC (with no restrictions on gestation), offered an ITN, and given an ITN during ANC consultation.
Principal components analysis (PCA) was used to create an asset index [24] , [25] based upon household characteristics such as source of drinking water, type of toilet facilities, and a range of household assets. All assets were included in the PCA as binary variables [26] (link). The asset index was then used to construct socio-economic quintiles from the poorest households through to the least poor. This method has been validated in household surveys with information on both assets and income or expenditure [27] . In this study these socio-economic quintiles are not representative of the population level, but are a relative score amongst women who attend ANC in Segou District, and a greater proportion of pregnant women from less poor households attend ANC in Mali [28] (link).
Potential predictors of intermediate processes for pregnant women 4 to 8 months pregnant (for IPTp-SP) on their first visit to ANC were assessed using a univariate (unadjusted) logistic regression model. Categories of potential predictors included: socio-demographic; pregnancy factors; health facility factors; and 3 types of process factors which were departments and time; illness or suspected illness; and payments. Individual potential predictors within each of these categories were measured during the structured observations by fieldworkers observing actions, listening to communications, or observing written information (Table 1).
Potential predictors were analysed for those intermediate processes that were found to be ineffective, that is for which less than 80% of women who should have completed the process, did so. These analyses were restricted to women 4 to 8 months pregnant, but gestation was included as a potential predictor. Adjusted Wald tests were used to assess the association between each potential predictor and the outcome of each intermediate process. Predictors with Odds Ratios (ORs) significant at the 10% level (p-values<0.1) were included in multi-variable (adjusted) logistic regression models for each intermediate process outcome in order to determine which potential predictors remained associated with each of the outcomes when adjusted for other predictors. ORs were estimated rather than Relative Risk (RR) due to clustering and stratification of the sample, and therefore our outcomes not being truly representative of the district level population. In the multi-variable models, predictors were considered significant at the 10% level at all stages of model building except for the final model where p<0.05 was used [14] .
The design effect (DE) and intra-cluster correlation coefficients (ICC) were calculated for 6 systems effectiveness intermediate processes defined in the health systems effectiveness algorithm (Figure 1) for aggregated CSRef and CSComs, and for aggregated CSComs alone. The design effect was calculated as the difference between the variance based on the clusters used compared with a modelled variance if a simple random sample was used. A design effect of 1 means that variance is unaffected by clustering and is equivalent to that of a simple random sample. An ICC of 0 equates to a DE of 1 where the variance is equivalent to that of a simple random sample.

We used universal sampling and extracted all cases that fulfilled our study criteria from the MyTB database. Information that was readily available from MyTB database was the age at diagnosis, ethnicity, nationality, gender, education level, smoking status at diagnosis, diabetes and human immunodeficiency virus (HIV) status at diagnosis, sputum AFB load at diagnosis, chest X-ray (CXR) severity at diagnosis, presence of MDR-TB and status of sputum conversion at the end of the two-month intensive phase. Some information that was not available from the MyTB database such as presence of other co-morbidities, alcohol dependence status at diagnosis, duration of symptoms before diagnosis and number of days missing directly observed therapy (DOT) were obtained from the patients’ manual medical records that were kept in the respective health clinics. All data were recorded using data collection form.

The patients were treated using the WHO standard 6-month regimen that included 2 months of isoniazid/rifampicin/pyrazinamide/ethambutol and 4 months of isoniazid/rifampicin (2HRZE/4HR). The directly observed therapy (DOT) strategy was followed in the intensive phase of treatment, while community volunteers were involved in monitoring the patients in the continuous phase. During the initial and follow-up visits, the patients were told to report any symptoms they developed that may indicate the relapse of the disease to their TB unit. Consequently, when the patients returned to their TB unit with any symptoms associated with TB, a chest X-ray was performed, and a sputum acid-fast bacilli smear test was conducted when TB relapse was suspected [25 (link)].

Health facilities served as the entry point for enrolment into the study population. A three-stage sampling strategy was deployed; At the first stage, only health centre level three and above in the study districts were purposively selected. In the second stage, eight villages with similarly high numbers of malaria cases were selected in each health facility catchment area in each district. This was based on malaria data for the same period of the preceding year extracted from health facilities’ outpatient registers. This data was then used to calculate malaria attack rates among children aged 3–59 months based on numbers of confirmed cases presenting at health centres divided by their catchment area populations. In the control district, communities bordering the intervention districts were not selected to avoid ‘leakage’ across district boundaries. In the third stage, all households in each of the eight selected villages were enumerated and twenty-five households with at least one SMC-eligible child were selected in each village using systematic sampling. One SMC-eligible child was then randomly selected from each household. This resulted in a sample of 200 children per district or 600 across the three study districts. To ensure adherence to the full three-day SPAQ regimen, children enrolled into the study in the intervention districts received all doses as directly observed therapy under the supervision of a VHT distributor. In the event of death of child under study, WHO’s verbal autopsy tool was applied to determine the cause of death [25 ]. SMC eligibility was defined according to the standard SMC protocol.