Study population
This study is based on data from the Nouna HDSS which had 80,000 individuals at the end of 2007. It is run by the Centre de Recherche en Santé de Nouna [
39] and located in the North-West of Burkina Faso in a rural area predominated by a sub-Sahelian climate with one rainy (June–October) and dry (November–May) season per year [
3,
40]. This study covers the observation period from 1 January 1998 to 31 December 2007. Data for Nouna town could only be analysed since 1 January 2000 after Nouna town was integrated into the study area.
To record COD data, the Nouna HDSS has been applying the VA method since 1993. After a death has occurred, trained interviewers visit the household of deceased people in the study area to conduct the VA interview using a standardized questionnaire after obtaining oral informed consent. Most VA interviews are carried out between 3 and 6 months after death allowing for the mourning period [
41].
For this study, the InterVA-3 method was applied here to derive the most likely cause of 35 possible COD groups [
42]. A detailed description of the InterVA model has been given elsewhere [
43]. In short, the InterVA model defines the probability of a cause for a particular death given the presence of a specific disease indicator or symptom using an automated Bayesian model [
26]. It displays up to three probable CODs and their corresponding likelihoods. In order to consider local epidemiology for important diseases in the Nouna HDSS region, the malaria and HIV/AIDS prevalence was set to “high” for malaria and to “low” for HIV/AIDS [
44]. Results will be compared with the previously most often used method to derive the cause of death from VA data, i.e. physician coding.
Data analysis
All individuals registered in the Nouna HDSS within the study period were included in the analysis except a few individuals (N = 97) for whom no information on month of death were available. To calculate monthly mortality rates, the population by month, year, sex, age group, and area (Nouna town and rural) was estimated as the average of the population at the beginning and end of a month (mid-month-population). Age groups were defined as follows: infants (<1 year), children (1 to <5 years), adolescents (5 to <15 years), adults (15 to <60 years) and the elderly (60+ years). Likewise, the total number of deaths in these categories was calculated. For malaria-specific analysis, the groups “malaria”, “other causes”, and a third category containing missing causes either due to missing data or causes that could not be determined by either method (PCVA or InterVA) were considered. Only CODs with the highest likelihood as estimated by InterVA, were considered as COD for an individual in the analysis.
The monthly mortality rates μ per 1,000 were calculated as μ = (D/M) × 1,000 in which D denotes the number of deaths in the respective month. M denotes the approximate estimate of the person-years, estimated by dividing the mid-month-population by 12.
To analyse malaria mortality, an imputation procedure was used to consider deaths with missing verbal autopsy questionnaire. The probability that a death with missing questionnaire was due to malaria was estimated by a log-linear model depending on age group and month. The number of deaths with missing VA questionnaire, multiplied with this probability is then the expected number of additional malaria deaths in a given month and age group.
For graphical assessment of seasonal variations and long-term trends, a weighted 5-month moving average (MA) was used according to
$${\rm M}{\rm A}_{\text{month}} = \,0.4 \, \times \, \upmu_{\text{month}} \,\,+\,\, 0.2 \, \times \, (\upmu_{{{\text{month}}+1}} + \upmu_{{{\text{month}}-1}} ) \,\, + \,\, 0.1 \, \times \, (\upmu_{{{\text{month}}+2}} \,\, + \,\, \upmu_{{{\text{month}}-2}} ).$$
For assessing the relative monthly effect on overall and malaria mortality, age group-specific Poisson regression models were fitted according to
-
Model I
$$\ln \, [\upmu \left( {{\text{x}}_{ 1} , {\text{ x}}_{ 2} , {\text{ x}}_{ 3} , {\text{ x}}_{ 4} } \right)] = \upbeta_{1} {\text{x}}_{1} + \upbeta_{2} {\text{x}}_{ 2} + \upbeta_{3} {\text{x}}_{ 3} + \upbeta_{4} {\text{x}}_{4})$$
and
-
Model II
$$\ln \, [\upmu \left( {{\text{x}}_{ 2} , {\text{ x}}_{ 3} , {\text{ x}}_{ 4} } \right)] = \upbeta_{0} + \upbeta_{2} {\text{x}}_{ 2} + \upbeta_{3} {\text{x}}_{ 3} + \upbeta_{4} {\text{x}}_{ 4}$$
Model I is defined as a model without intercept, where x
1 is a vector with binary dummy variables for each month, x
2 is a continuous variable for calendar year running from 1 (year 1998) to 10 (year 2007) to investigate an overall change in rates over time, x
3 represents sex and x
4 area. Model I has no intercept and calculates an estimate for each month. In Model II, an intercept β
0 instead of a monthly effect is estimated. The relative monthly effect on mortality was calculated by the difference β
1–β
0 of the monthly effect of Model I and the overall effect of model II.
To further assess the seasonal effect on malaria mortality, rate rations (RRs) were estimated using a Poisson regression model with a continuous function of month of death. For this, a sine-function was used of the form g
1(x
1) = sin (x
1 × π/6) and a cosine-function of the form g
2(x
1) = cos (x
1 × π/6) in which x
1 adopts a value between 1 and 12, corresponding to the months January to December. This resulted in the model
-
Model III
$$\ln \, [\upmu \left( {{\text{x}}_{ 1} , {\text{ x}}_{ 2} , {\text{ x}}_{ 3} , {\text{ x}}_{4} } \right)] = \upbeta_{0} + \upbeta_{11} g_{1} \left( {{\text{x}}_{1} } \right) + \upbeta_{12} g_{2} \left( {{\text{x}}_{1} } \right) + \upbeta_{2} {\text{x}}_{2} + \upbeta_{3} {\text{x}}_{3} + \upbeta_{4} {\text{x}}_{4}.$$
From the regression parameters β
11 and β
12 the amplitude is calculated as
\(\sqrt {\beta_{11}^{2} + \beta_{12}^{2} }\) and the phase φ as arctan(−β
11/β
12) which determines the day of the year with the highest rate. To test the strength of evidence, the difference of deviances of Model II and III was calculated which is asymptotically χ2-distributed with two degrees of freedom since two parameters (β
11, β
12) are estimated. Effects are calculated as logarithmic RR.
Since Nouna town was encompassed in the study area in 2000, 432 out of 480 observations, determined by all possible cross-classifications of the variables year, sex and area for which people were observed, were included in each model. For every model 48 observations were set missing, because the number of individuals in these observations was zero. Data analysis was carried out with SAS, 9.2. Poisson regression used the SAS-procedure PROC GENMOD.