Time Series Analysis of Trends in Malaria Cases and Deaths at Hospitals and the Effect of Antimalarial Interventions, 2001–2011, Ethiopia

Background The Government of Ethiopia and its partners have deployed artemisinin-based combination therapies (ACT) since 2004 and long-lasting insecticidal nets (LLINs) since 2005. Malaria interventions and trends in malaria cases and deaths were assessed at hospitals in malaria transmission areas during 2001–2011. Methods Regional LLINs distribution records were used to estimate the proportion of the population-at-risk protected by LLINs. Hospital records were reviewed to estimate ACT availability. Time-series analysis was applied to data from 41 hospitals in malaria risk areas to assess trends of malaria cases and deaths during pre-intervention (2001–2005) and post-interventions (2006–2011) periods. Findings The proportion of the population-at-risk potentially protected by LLINs increased to 51% in 2011. The proportion of facilities with ACTs in stock exceeded 87% during 2006–2011. Among all ages, confirmed malaria cases in 2011 declined by 66% (95% confidence interval [CI], 44–79%) and SPR by 37% (CI, 20%–51%) compared to the level predicted by pre-intervention trends. In children under 5 years of age, malaria admissions and deaths fell by 81% (CI, 47%–94%) and 73% (CI, 48%–86%) respectively. Optimal breakpoint of the trendlines occurred between January and June 2006, consistent with the timing of malaria interventions. Over the same period, non-malaria cases and deaths either increased or remained unchanged, the number of malaria diagnostic tests performed reflected the decline in malaria cases, and rainfall remained at levels supportive of malaria transmission. Conclusions Malaria cases and deaths in Ethiopian hospitals decreased substantially during 2006–2011 in conjunction with scale-up of malaria interventions. The decrease could not be accounted for by changes in hospital visits, malaria diagnostic testing or rainfall. However, given the history of variable malaria transmission in Ethiopia, more data would be required to exclude the possibility that the decrease is due to other factors.


Background
Malaria affects over 68% of the population in Ethiopia, a country of 94 million people in 2013 [1]. The disease is highly seasonal with varying intensity of transmission owing to altitudinal and climatic variations [2]. Across the nine administrative regions, areas that lie below 2000 m altitude are considered as malarious ( Figure 1) [3]. The occurrence of malaria epidemics in the past indicates that there is little immunity in the majority of the population, owing to few infective mosquito bites per person per year [4,5]. The high transmission season coincides with the cultivation months; hence malaria has a deleterious effect on agricultural production. Plasmodium falciparum accounted for nearly 55% of all malaria cases during 2008-2012 [6].
Malaria control has been one of the major components of the country's National Heath Sector Development and Poverty Reduction Strategy, put forth since 2004 in different phases, the most recent plan being for the period 2010-2015. With investment by the Global Fund to Fight HIV/AIDS, Tuberculosis and Malaria, the USA President's Malaria Initiative, World Bank and other development partners, the government has employed three key antimalarial interventions: i) distribution of long-lasting insecticidal nets (LLINs) through mass campaigns, to the entire population at risk; ii) indoor residual spraying (IRS) in designated epidemic-prone areas; and iii) increased diagnostic testing through rapid diagnostic tests (RDT) and microscopy, along with deployment of artemisinin-based combination therapy (ACT). Distribution of LLINs started in 2005. ACT was introduced to all public health facilities, free-of-charge to all age groups, as the firstline treatment for uncomplicated malaria in July 2004 [7].
Malaria control efforts have benefited from the national health sector development strategy which aims to increase access to effective health care by deploying health extension workers (HEWs) to provide integrated health promotion and treatment at peripheral health posts and at community level. By 2009 over 30,000 HEWs had been deployed. Treatment provided by HEWs targets the leading causes of death in children: malaria, pneumonia, and diarrhoea [8]. For malaria, the HEWs verify ownership and encourage use of LLINs; diagnose malaria with RDTs and treat confirmed cases; then refer suspected severe malaria cases to the health centres and hospitals.
This study aimed to (i) investigate the proportion of the population benefiting from malaria control interventions, (ii) assess trends in malaria cases and deaths in hospitals in malaria risk areas between 2001 and 2011, and (iii) examine the relationship between malaria interventions and the occurrence of malaria cases and deaths during 2001-2011. Attempts were made to take into account other factors which may affect malaria cases and deaths in hospitals and to employ statistical approaches focused on assessing changes in trends of malaria cases and deaths before and after the major scale up of malaria interventions since 2005. The results are discussed in light of international goals and targets for reducing malaria burden.

Ethical clearance
Ethical clearance was not required as the retrospective data used from the health facilities were monthly aggregates and anonymous counts of clinical cases and deaths.

Intervention coverage
Information on malaria control interventions (LLINs, IRS, and ACTs) was obtained from national malaria control program (NMCP) records. As the programme started implementation of LLINs in 2005, the number of LLINs distributed to each region during 2005-2011 was recorded. The number of insecticide treated mosquito nets available before 2005 was assumed to be small. Regional population estimates were taken from the 2007 Ethiopian census and adjusted using the population growth of United Nations population estimates for all the study years [9]. The proportion of the population potentially protected by LLINs in a given year was calculated by assuming that each LLIN protected 1.8 persons and lasted three years (number of LLINs distributed during the 3 year period 61.8/population). The proportion of the population protected with IRS nationwide was calculated by dividing the number of persons in households protected by IRS by the population at risk. The availability of ACT treatment was assessed by reviewing records of stock-outs at hospitals that were visited to access trends in malaria cases and deaths. ACTs were counted as available for a given month if ACTs were in stock for more than three weeks in a month.

Malaria cases and deaths
Data on malaria outpatient cases, admissions and deaths were obtained from health facilities that provided inpatient services and were located below 2000 m or served catchment populations predominantly below 2000 m elevation ( Figure 1). Of the total 120 inpatient facilities in Ethiopia, 62 facilities met the above criteria; 55 hospitals were visited, including three located in highland areas that predominantly serve populations from low land areas, while six hospitals were not visited for logistical reasons. Surveillance data were considered adequate if data on all indicators were available for more than 70% of months. Of 56 facilities visited, 41 facilities that had data for .70% of months between 2001-2011 were included in the final analysis.
At each visited hospital, monthly totals of outpatient visits, hospital admissions and deaths, and laboratory tests performed were reviewed. The total numbers of outpatient visits, admissions and deaths for all-causes and for malaria were recorded by two age groups (less than 5 years of age and 5 years old or greater). A suspected outpatient malaria case was defined as a visit in which malaria was the preliminary diagnosis in the outpatient record. A confirmed malaria case was defined if malaria parasites were demonstrated by microscopy (RDT results were not considered as they are used at the lower health facilities and community level by the HEWs). The slide positivity rate (SPR) was defined as the number of microscopy slides with malaria parasites divided by the number of microscopy slides reviewed. Laboratory records could not be broken down by age, nor could they be linked individually to outpatient or inpatient cases. Inpatient malaria cases and deaths were based on the diagnosis at discharge. The non-malaria cases and deaths, used for comparisons, were generated by subtracting malaria from the corresponding all-cause outpatient, admissions and deaths.

Meteorological data
Enhanced National Climate Services (ENACTS) rainfall data that is quality controlled and blended satellite and observational data [10] from the National Meteorological Agency of Ethiopia was obtained by the International Research Institute for Climate and Society, Columbia University, New York, USA. These precipitation data (estimate of rainfall in mm) were aggregated for five major regions of the country and were used as the principal short-term predictor to assess the effect of climate on the trends of malaria cases and deaths [11]. Statistical methods Stata 11 was used to compile data by year and month and perform statistical analysis [12]. Trends in confirmed malaria cases and SPR, malaria inpatient cases, and malaria deaths were assessed by region. Malaria inpatient cases and malaria deaths were assessed by the two age groups. Confirmed malaria cases and SPR were assessed for all ages as these could not be disaggregated by age. Change in indicators over time was evaluated in three The 95% confidence intervals (CI) around the estimates were computed using the CIs around the regression coefficient estimates. A relative percent change in an indicator for which the CI does not include zero was considered statistically significant. In the time-series regression model for (ii) and iii), an Autoregressive Integrated Moving Average (ARIMA) model was used. A correlogram and partial autocorrelation plots were used to guide selection of the best ARIMA model [14].
To provide insight into whether changes in malaria cases and deaths were related to the introduction of malaria interventions in time, a sensitivity analysis was done by dividing the time series into two segments, a pre-intervention and post-intervention period, and varying the dividing point and evaluating the segmented regression model for each break point. The optimal breakpoint, or the month with the maximum change in the trend line, was considered to be the point where the amount of variation accounted for by the segmented regression model (R 2 ) was greatest [15].
Month-to-month variability in indicators was evaluated using monthly values of the indicators during pre-and post-intervention periods using sinusoidal functions. To observe and remove shortterm seasonal fluctuations, the monthly trends were extracted using a Hodrick-Prescott filter with monthly smoothing parameter [16]. Spectral analysis was done to detect the fundamental frequency components of the series. Additionally, month effects were parametrically modeled by using sinusoidal functions (Indirect Discrete Fourier Transform (IDFT).
The time trend in monthly rainfall, the principal short-term predictor of malaria burden [17], was plotted by region.  (Figure 4c, 4d). Overall, malaria inpatient cases in all ages were 54% lower and malaria deaths 68% lower in 2011 than that predicted by trends during 2001-2005, however, these declines were not statistically significant.

Interventions
Concomitantly, non-malaria outpatient cases, inpatient cases, and deaths either increased or remained unchanged (Table 1, (Table 1), and lower by 81% (CI, 47%-94%) than the level predicted for 2011. In this agegroup, malaria deaths were 73% (CI, 48%-86%) lower in 2011 than predicted by previous trends. Among those 5 years and older, malaria inpatient cases and malaria deaths were lower in 2011 than predicted based on 2001-2005 trends, however these changes were not statistically significant. The proportions of inpatient cases due to malaria (of all-cause inpatient) in children under 5 years fell from 23% in 2001-2005 to 11% in 2011; and malaria deaths fell from 17% to 8% of all-cause deaths during the same period. The proportion of inpatient cases due to malaria in 5 years and above fell from 15% to 6% and the proportion of malaria deaths decreased from 19% to 5%.
The observed number of confirmed malaria cases in 2011 was lower than the mean number of confirmed cases during 2001-2005 in all 10 regions ( Table 2). The decrease in confirmed cases varied among regions from 36 to 99%. Similarly, the SPR and the number of malaria deaths were lower in all regions, while malaria inpatient cases were higher in 2011 in one region, Gambella. Confirmed malaria cases were lower in 2011 compared with 2001-2005 in 33 of 41 hospitals; inpatient cases were lower in 34 hospitals and malaria deaths were lower in 37 hospitals.

Seasonality
Using the Hodrick-Prescott filter method to remove short-term (monthly) fluctuations and determine the long term time-series over multiple years, a decrease in all malaria indicators is seen ( Figure 5). Confirmed malaria cases and SPR decreased through 2011, while malaria inpatient cases and malaria deaths decreased until 2008, followed by stability between 2008 and 2011. The spectral analysis showed that the fundamental frequency components of these series is one or annual.
Monthly mean confirmed malaria cases, SPR, malaria inpatient cases, and malaria deaths were lower for all months during 2006-2011 (post-intervention period) compared to 2001-2005 (preintervention period (Figure 6 a, b, c, d) Linear regression on confirmed cases and SPR by region showed no significant effect during the pre-intervention period in all regions except in SNNP region (where rainfall had significant increasing effect on both). During the post-intervention period, rainfall had a significant increasing effect on confirmed cases and SPR in all the regions except in Oromia, where rainfall had significant decreasing effect (p,05) and in Tigray with no significant effect.

Discussion
In this study reviewing data from hospitals in Ethiopia, malaria decreased between 2001 and 2011 by several measures. Among all age groups, the number of confirmed malaria cases, malaria inpatient cases and deaths declined by more than half, while malaria inpatient cases and malaria deaths among children under 5 years old decreased more than two-thirds from the numbers predicted if levels at the start of the decade had persisted. The observed declines in malaria cases and deaths were seen across all regions where malaria transmission occurs and could not be accounted for by changes in outpatient visits or by the number of diagnostic tests performed. During this time, measured rainfall was sufficient to support malaria transmission and levels of rainfall could not be consistently correlated with changes in malaria cases. Counts of cases and deaths were used in this analysis as the catchment population of each hospital was not known. Given the Therefore this could be seen as the most conservative approach to this time series analysis. In addition to the decreasing trends seen in annual cases and deaths, lower numbers of malaria cases, admissions and deaths were seen when comparing the mean for each month during the postintervention period to the same month in the pre-intervention period. Seasonal peaks also appeared lower in the post-intervention period compared to pre-intervention period.
The hospital data analysed for this study indicate that the decreases in cases and deaths coincided with the scale-up of malaria interventions. Analysis of the breakpoints in the time series of malaria cases and deaths were consistent with a change in the trend approximately when interventions were either introduced, or increasing during 2005 and 2006. The proportion of population potentially protected with LLINs increased from negligible levels   Assessment of trends in malaria cases and deaths and the effect of malaria control activities reported previously by Ethiopia, and other sub-Saharan African countries (Zanzibar [18,19], Rwanda [20], The Gambia [21] and Sao Tome and Principe [22]) have utilized both facility data and household survey results and have shown significant decrease following scale up of interventions. Compared to the 2008 study in Ethiopia [23], the present study benefits from three additional years of data and attempts to address concerns regarding analysis of trends given large variability in malaria transmission based on historical data in Ethiopia and a short, possibly spurious baseline in available facility data.
This study is subject to several limitations. A key challenge in assessing the trends in malaria in Ethiopia is the historically high variability in baseline malaria transmission. An attempt was made to account for this in the analysis; however, the time period reviewed may have been too short to reveal longer term trends in malaria transmission that may have been unrelated to malaria interventions. Visiting hospitals and reviewing data at hospitals helped in minimizing reporting bias present in routine surveillance. Even so, excluding visited hospitals with incomplete data introduces its own potential bias; this is likely to be small as the excluded hospitals represent a small fraction of the hospitals visited. The main reason for the incomplete data at the hospitals was disappearances of record books and monthly reports (for some months or years particularly the early years) and improper handing over of records during change of health staff. The study did not include lower level health facilities and it is unclear if trends there would be the same as those seen in the hospitals considered. There were no data on the volume of patients cared for or the number of malaria cases identified through community case management in the Health Extension Programme. The effect of this program on trends in hospital malaria cases appears to be limited as the decline in malaria cases and deaths occurred two years before the scale-up of community case management in 2009 and, given health care seeking patterns in Ethiopia, only a third of patients treated in the community were likely to have been seen at hospitals [24]. In addition, significant decrease in malaria cases and deaths were observed against the background of increased allcause outpatient and all-cause inpatient cases and deaths in the hospitals. Information on testing practices at the district hospitals was limited. The number of suspected cases presenting to health facilities that should be tested for malaria was not recorded in standard reporting forms and therefore, it was difficult to assess whether the proportion of suspect patients tested changed over time. The measures of malaria control intervention coverage in this study were crude estimates of these interventions in the community however, they did reflect the time period during which interventions were available and were adequate to denote large changes in intervention coverage. More precise measures of LLIN coverage and ACT treatment would be needed to more closely relate changes in coverage to changes in malaria occurrence. Given the very high ecological variations within a region, annual average rainfall for the major regions may not be conclusive to accurately assess the climatic effect on trends of malaria cases and deaths. In a separate analysis of rainfall data by the International Research Institute for Climate and Society, Columbia University, New York, USA, using Enhancing National Climate Services (ENACTS) (not shown here), for most regions there was lower rainfall in the pre-intervention (baseline) period (2000)(2001)(2002)(2003)(2004)(2005) relative to the post-intervention period (2006)(2007)(2008)(2009)(2010). In addition, analysis of temperature indicate that the climate of Ethiopia has been getting warmer over the last 30 years -particularly in highland areas. Higher temperatures and rainfall in the intervention period suggest that the climate suitability for malaria from 2006-2010 was above that of the baseline period. Any large-scale declines in malaria cases and deaths during 2006-2010 are unlikely therefore to be as a result of a drier or cooler climate and an associated reduction in transmission suitability. The results of health facility-based changes in disease occurrences should be cautiously interpreted, as they may be influenced by changes in service utilization, testing rates, reporting, and other factors [25]. To the extent that these are addressed, review of hospital data from Ethiopia shows that malaria cases and deaths have decreased during the second half of the study period and these changes may be related to the increase in malaria control interventions. Given the history of malaria transmission in Ethiopia, strengthening of the surveillance system to allow regular monitoring of the disease trends is crucial to the success of the programme. Malaria in Ethiopia is highly heterogeneous and despite the recent reduction in malaria, the risk of resurgence and epidemics is still very real [5]. Intensive malaria control activities guided by epidemiological stratification with high resolution and adequate funding are therefore needed to reduce malaria transmission further and remove the risk of resurgence. In moderate-high transmission areas with considerable number of outpatient malaria cases, national malaria control programmes should consider surveillance on aggregates of outpatient cases and use line-listing of inpatient cases for assessing the use of interventions and causes of continued transmission in some areas. In low transmission areas where number of cases are manageable, line-listing of outpatient cases should be considered for in-depth geo-referencing and investigation of foci [26].
It has been difficult to discern trends in malaria cases and deaths from national aggregates reported in Ethiopia because of inconsistent and incomplete reporting [6]. This study was undertaken to attempt to account for reporting inconsistencies. Similar studies from sample hospitals will likely be required until the health information system becomes fully functional at national level to provide facility-level quality data to allow reliable assessment of trends of malaria cases and deaths; and other related indicators. Such studies would be stronger if complemented by trend analysis at levels of health centres and health posts to confirm if trends observed in hospitals are consistent with that of lower level facilities.