Study design
This was a hospital-based, retrospective cohort study. Recruitment into the study was conducted between September and November 2015.
Study settings
The Lake Zone is situated in the North-West of Tanzania along the Lake Victoria. It comprises of six administrative regions, namely Geita, Kagera, Mara, Mwanza, Shinyanga and Simiyu. The Lake Zone has an estimated total population of about 11.8 million (26.3% of the nation’s population) of which about 11.5% are children below 5 years [16]. By the time of the study, the Zone had a total of 37 Districts (an administrative sub-division of a Region).
In Tanzania, there are public and private health care facilities. The public health care system is layered into five levels, starting with the dispensary level, the health centre, the district, regional and topped with zonal referral hospitals. These levels are differentiated by the staffing, equipment and supplies, range of services available and type and number of patients eligible for admission. Everywhere in Tanzania, including the Lake Zone hospitals, district and regional hospitals are considered referral facilities; getting patients from the lower levels (Health Centres and Dispensaries).
For the purpose of this study, the focus was on three regions (Kagera, Mara and Mwanza) that were among the regions supported by the Tibu Homa Project (THP). Therefore, the three regional hospitals were selected. Also included in the study, were three randomly selected district hospitals; one from each region. These were Ngara (Kagera Region), Tarime (Mara Region) and Magu (Mwanza Region). These health facilities are expected to have reliable services and more experienced medical staff, having advanced examination medical services (basic laboratory examinations and X-ray facilities) and improved pharmacies. Therefore, they are in better position to correctly assess causes of deaths than lower health facilities.
Study population
The study included children aged below 5 years (excluding neonates) hospitalized during the study period. We excluded neonates because in almost all facilities there is lack of neonatal wards and these babies stayed with their mothers in female or maternity wards. We also excluded children who were referred to other facilities because it was not possible to ascertain their treatment outcomes (discharged or died). We collected data among study participants between admission until discharged or death. During the study period, information was retrieved from hospital clinical records and transcribed into study tools.
Sample size and sampling procedure
We estimated a sample of 1155 children based on cluster sampling. Parameters used during the sample size estimation include child inpatient mortality of 60 per 1000 admissions, 95% confidence interval and a 2% margin of error. Furthermore, since the use of cluster sampling does not give equal chance to be selected, we assumed a design effect of 2. We finally adjusted the sample size taking into account non-response that was estimated to be 6%. This sample was split between six clusters. Clusters were public regional and district hospitals. One district hospital was randomly selected per region to make a total of six hospitals. Public regional and district hospitals are facilities with referral medical services. These health facilities are reliable with experienced medical staff, equipped with examination and medical services, (basic examination laboratories and X-ray facilities) and with stocked pharmacies. Based on the hierarchy, regional hospitals are superior to district hospitals. They have more specialized personnel, surgical medical care, equipment and extended capacity to admit. However, all regional and district hospitals have at least one specialist in paediatrics and they also perform laboratory services.
Recruitment procedure
Recruitment was at death or discharge of the child when information was collected on diagnosis and cause of death. Study information was drawn from case notes after obtaining consent from biological parents or legal caregivers.
Process
The study personnel who included, three clinicians and a nurse (staff attached to pediatric wards) were identified from the facility. These are care providers who attend children and had been previously trained by Tibu Homa Project in proper diagnosis and treatment of under-fives using national protocols and guidelines. Complete information about signs and symptoms were obtained from parents or caregivers at admission. Present and past medical histories and detailed family histories of all admitted under fives were recorded. All children were assessed for general danger signs and main symptoms according WHO guidelines. Nutritional status information was available for all children at admission. The information was recorded by measuring weight, length and mid-upper arm circumference and classified using Z-scores. Later on, this information was abridged and compiled for under-fives enrolled to the study.
Specific causes of deaths among children were obtained using patients’ clinical notes and from diagnosis made on death certificates. Most of the variables for the child were based on parent/care-givers report using face-to-face interviews. For example, age, sex, religion; relationship to the child (parents or kin caretaker), education level attained by the parent/care-giver, marital status and occupation. Education level was classified as “never in school”; “some primary” and “above primary”. Occupation was categorized as “salaried”, “unemployed”, “subsistence farming” and “self-employed/wage earners”.
Two independent death auditors reviewed patients’ files and death certificates to ascertain the cause of death using International Classification of Disease 10th Revision (ICD-10). Whenever there was no definite cause of death or when there was disagreement between death auditors, the cause of death was considered “unknown”.
Data processing and analysis
We performed double entry of data to verify and minimize random errors so as to enhance quality of data. Then, syntax programmes and univariate analyses were performed to detect and correct inconsistencies and out-of-range values. Detected inconsistencies and out-of-range values were set to ‘missing’. Frequencies were used to get descriptive statistics. We also performed cross-tabulations to assess the associations between case fatality and pre-determined variables. These variables were included based on theory, research experience and empirical evidence. A cut-off point for a statistical significance was set at a p value of 0.05. Finally, independent factors for inpatient under-five death were determined using binary logistic regression in the multivariable model, controlling for the cluster design effect. Since under-fives were not completely randomly selected, during the analysis confidence intervals and p-values were based on robust estimation of variance by accounting for clustering of study participants within the same facility and their residence. All these procedures were performed using SPSS for Windows (Version 20).