Ethiopian Health and Nutrition Research Institute, P. O. Box 1242/5654, Addis Ababa, Ethiopia
School of Public Health, College of Health Sciences, Addis Ababa University, P.O. Box 9086, Addis Ababa, Ethiopia
Centre for International Health, University of Bergen, Bergen, Norway
Abstract
Background
The highlands of Ethiopia, situated between 1,500 and 2,500 m above sea level, experienced severe malaria epidemics. Despite the intensive control attempts, underway since 2005 and followed by an initial decline, the disease remained a major public health concern. The aim of this study was to identify malaria risk factors in highlandfringe southcentral Ethiopia.
Methods
This study was conducted in six rural kebeles of Butajira area located 130 km south of Addis Ababa, which are part of demographic surveillance site in Meskan and Mareko Districts, Ethiopia. Using a multistage sampling technique 750 households was sampled to obtain the 3,398 people, the estimated sample size for this study. Six repeated crosssectional surveys were conducted from October 2008 to June 2010. Multilevel, mixedeffects logistic regression models fitted to
Results
Overall, 19,207 individuals were sampled in six surveys (median and interquartile range value three). Six of the five variables had about twofold to eightfold increase in prevalence of malaria. Furthermore, among these variables, OctoberNovember survey seasons of both during 2008 and 2009 were strongly associated with increased prevalence of malaria infection. Children aged below five years (adjusted OR= 3.62) and children aged five to nine years (adj. OR= 3.39), low altitude (adj. OR= 5.22), midlevel altitude (adj. OR= 3.80), houses with holes (adj. OR= 1.59), survey seasons such as OctoberNovember 2008 (adj. OR= 7.84), JanuaryFebruary 2009 (adj. OR= 2.33), JuneJuly 2009 (adj. OR=3.83), OctoberNovember 2009 (adj. OR= 7.71), and JanuaryFebruary 2010 (adj. OR= 3.05) were associated with increased malaria infection.
The estimates of cluster variances revealed differences in malaria infection. The villagelevel intercept variance for the individuallevel predictor (0.71 [95% CI: 0.281.82]; SE=0.34) and final (0.034, [95% CI: 0.0020.615]; SE=0.05) were lower than that of empty (0.80, [95% CI: 0.322.01]; SE=0.21).
Conclusion
Malaria control efforts in highland fringes must prioritize children below ten years in designing transmission reduction of malaria elimination strategy.
Background
About half of the total population living between altitudes of 1,500 and 2,500 m above sea level (masl) is at risk of malaria and the areas experience epidemics in Ethiopia
Since 2005, Ethiopia has scaled up malaria control programs using key malaria interventions such as effective case management (artemisinin combination therapy and malaria rapid diagnostic tests), and vector control options (indoor residual spray and longlasting insecticidal nets) in endemic areas (<2,000 masl). Subsequently, the program obtained fruitful results in reducing malaria burden between 2006 and 2008. In addition, the 2011–2015 National Strategic Plan highlights the intent to eliminate malaria in specific geographical areas with historically low malaria transmission; and achieve near zero malaria deaths in the remaining malarious areas of the country
Accordingly, identifying factors influencing malaria infection both at individualand villagelevels (clusters of household) appears useful in guiding targeted malaria interventions at highland areas with malaria low prevalence and at high epidemic risk. In addition, the Ministry of Health of Ethiopia recommended evidenceinformed decision to incorporate high altitude areas (>2,000 masl) in malaria control program
Methods
Study area and study participants
This study was conducted in six rural
Location of the study sites, Butajira area, Southcentral Ethiopia
Location of the study sites, Butajira area, Southcentral Ethiopia.
Malaria is one of the important causes of sicknesses in Butajira area. Between 2004 September and 2010 August, 32.3% (19,923 of 61,654) were microscopically confirmed malaria cases from Butajira and Enseno Health Centres, and Butajira Hospital. On average, more than 10 thousand malaria suspected cases visited these public health facilities between 2002 and 2010. Indoor residual spraying (IRS) operation was performed mainly for epidemic control in the low altitude areas of the present study area. During 2009/2010, a spraying of houses was done to control malaria outbreak in Hobe and Bati Lejano
Sample size calculation
The sample size required for this study was estimated as follows. Estimation of the sample size for malaria prevalence was based on 4.1% prevalence from three Ethiopian regions. The study used a sample size estimated (n=3,398) to measure malaria prevalence in this study area
Study design and sampling procedure
This study used communitybased repeated crosssectional survey. Six rural
Data collection
This study obtained data from household head interviews of sampled households and blood film collection and examinations of family members. Trained data collectors conducted the interviews using a pretested and structured questionnaire to obtain baseline sociodemographic and household characteristics. Standard procedures were followed for blood specimen collection, processing, microscopic examination and reporting of malaria parasites
All positive slides and 10% of slides with negative results were sent to another microscopist blinded to the microscopy results to ensure quality of light microscopy. To ensure maximum response rate of participants, households with absentees were revisited once more. The altitude readings of the sample households were recorded using handheld Global Positioning System (GPS) (Garmin eTrex ®). The principal investigator (AW) and two data collectors (postgraduate students from Addis Ababa University) conducted the GPS recording.
Calculation of household wealth index
This study used a dataset of relative household wealth index data computed in another study
Suitability of the data for factor analysis was assessed before performing PCA. Thus, both the KaiserMeyerOklin (KMO) value and Bartlett’s Test of Sphericity results were supporting the factorability of the correlation matrix. PCA result was repeated with alterations until the resulting model was suitable for the survey data. Finally, 11 indicators were selected to run the final PCA. PCA revealed the presence of two components with eigenvalues above 1, explaining 36.0% and 9.8% variance in the dataset, respectively. The first principal component (with eigenvalues of 5.04) represented 36.0% of the variance in the sample and was used to generate the wealth index of the study households.
The 11 variables with greatest weights were loaded on the first principal component: possession of motorcycle (0.937), sewing machine (0.869), truck (0.869), television (0.809), grainmill (0.646), lanternkerosene (0.610), phone (0.575), electricity line (0.568), bicycle (0.423), types of sleeping places (0.356) and cart (0.355). The wealth index varied from −0.256 to 13.27. Then, all households were ordered into three wealth groups: the “lowest” ranked group (30.9%, n=228), followed by the “middle” ranked group (35.7%, n=264), and finally the top third in the “higher” ranked group (33.4%, n=247). Data collectors assessed physical condition of houses and recorded the data. Training manual was distributed to help in categorizing the houses into three including dilapidated, houses with their walls allowing mosquito entry and good condition.
Outcome and predictor variables
Finding
Variables were categorized and coded as follows. Age grouped into four classes and coded as <5, 5–9, 10–14, and ≥15 years. Altitudinal location of households was classified into three including low (1,8001,899 m), midlevel (1,9001,999 m), and high (2,0002,300 m) altitudes. There were three wealth groups including lowest, middle, and higher; and three house status categories such as dilapidated, walls with holes, and good condition. The survey seasons are OctoberNovember 2008, JanuaryFebruary 2009, JuneJuly 2009, OctoberNovember 2009, JanuaryFebruary 2010, and June 2010. In all cases the last categories were considered as reference groups.
Data management and analysis
Data entry and cleaning was done using Epi Info version 6 (Centers for Disease Control and Prevention (CDC), Atlanta, Georgia (USA). Descriptive statics was performed using IBM® SPSS® Statistics version 20.0. Descriptive statistics was performed to describe characteristics of predictor variables. Multicollinearity was checked using linear regression as recommended and no multicollinearity was evident. Multivariate analysis was done using STATA version 11.0 (College Station, Texas, USA). Mixedeffects logistic regression was fitted using selected independent variables to estimate individual
Multilevel analysis is a statistical tool applied to data with nested sources of variability, which involve units at lower level nested within units at a higher level
Intraclass correlation coefficients (ICC), median odds ratios (MOR) and 80% interval OR (IOR80) were computed to estimate villagelevel variance in
MOR was computed to translate the area level variance in the widely used odds ratio scale. The MOR is defined as the median value of the odds ratio between the area at higher risk and the area at lowest risk when randomly picking out two areas the MOR can be conceptualised as the increased risk that (in median) would have if moving to another area with a higher risk
In order to integrate the area level fixed effect and the random residual variations using the 80% interval odds ratio (IOR80%) is suggested
Assessing models
To assess whether a model predict the outcome variable beyond what would be expected by chance, the familiar ChiSquare likelihoodratio test of a difference between models is used. The degrees of freedom for the ChiSquare are the differences in the number of parameters for the models being compared
In this study, individuallevel predictor model (with −2 Log Likelihood value of 969.7091 and 4 df) compared against the full model (with −2 Log Likelihood value of 886.97871 and 15 df). By subtracting, the difference (969.7091 and 886.97871) is 55.9 and this value showed statistically significant with (15–4) = 9 df, so the full model leads to prediction that is significantly better than chance.
Ethical consideration
Ethical approval of the study was obtained from the Faculty of Medicine at Addis Ababa University, and the Ethiopian Ministry of Science and Technology. Individual informed consent was obtained from adults, and from the parents or guardians of children aged less than 18 years. In addition, minors gave verbal assent. Blood specimens were collected as recommended using an alcohol swab and disposable blood lancets by trained staff
Results
Characteristics of study participants
Overall, 19,207 individuals were sampled in six surveys (median and interquartile range value of three). Most of the participants were 15 years old and above with a mean (±SD) age of 20.5 (±17.2), and the range was between one month and 99 years. Above half (51.3%) of the participants were females. A total of 3,416 participants were included in the baseline survey conducted during OctoberNovember 2008. In the consecutive five followup visits, there were 3,205 (JanuaryFebruary 2009), 3,227 (JuneJuly 2009), 3,210 (OctoberNovember 2009), 3,127 (JanuaryFebruary 2010), and 3,022 (June 2010) participants sampled (Figure
The number of study participants in baseline and followup surveys, Butajira area, Ethiopia, 2008–2010
The number of study participants in baseline and followup surveys, Butajira area, Ethiopia, 2008–2010.
A recent study estimated 0.93% (178 of 19,207) of the participants had malaria infection. Most of the infections were due to
Factors
Total examined, n
Positive, n (%)
ChiSquare
Total
19,207
178 (0.93)
Age category
70.8
<0.001
<5
3,042
54 (1.77)
59
3,513
59 (1.68)
1014
2,702
20 (0.74)
≥15
9,942
45 (0.45)
Gender
3.4
0.06
Male
9,347
99 (1.06)
Female
9,852
79 (0.80)
Total
738
44 (6.00)
Wealth status
70.0
<0.001
Lowest
6,379
111 (1.74)
Middle
6,419
40 (0.62)
Higher
6,401
27 (0.42)
House status
39.0
<0.001
Dilapidated
3,195
21 (0.66)
Holes
4,844
81 (1.67)
Good
11,160
76 (0.68)
Altitudinal strata
106.5
<0.001
1,8001,899 m
5,547
107 (1.93)
1,9001,999 m
2,034
29 (1.43)
2,0002,300 m
11,618
42 (0.36)
Seasons
86.5
<0.001
OctoberNovember 2008
3,416
20 (0.58)
January.February 2009
3,205
11 (0.34)
JuneJuly 2009
3,227
27 (0.84)
OctoberNovember 2009
3,210
72 (2.24)
January.February 2010
3,127
35 (1.11)
June 2010
3,022
13 (0.43)
Univariate analyses
In the univariate logistic regression, children aged below five years (unadjusted OR= 3.71), children aged five to nine years (unadj. OR= 3.40), low altitude (unadj. OR= 5.12), midlevel altitude (unadj. OR= 3.63) and houses with holes (unadj. OR= 1.57) had increased risk of having
Fixedeffects
Unadj. OR (95% CI)
Adj. OR (95% CI)
*
Age groups
<5
3.71 (2.495.52)**
3.62 (2.435. 40)**
59
3.40 (2.305.02)**
3.39 (2.305.01)**
1014
1.48 (0.872.51)
1.49 (0.882.53)
Gender
Male
1.33 (0.991.79)
1.24 (0.921.67)
Altitudinal strata
1,8001,899 m
5.12 (3.297.98)**
5.22 (2. 67–10.22)**
IOR80%
(3.747.24)
1,9001,999 m
3. 63 (2.026.52)**
3.80 (2.096.91)**
IOR80%
(2.725.26)
Wealth group
Lowest
2.02 (1.033.94)
0.75 (0.371.53)
Middle
1.32 (0.782.22)
1.00 (0.591.69)
House status
Dilapidated
1.00 (0.601.66)
1.00 (0.601.67)
Holes
1.57 (1.112.22)*
1.59 (1.122.26)*
IOR80%
(1.142.22)
Seasons
OctoberNovember 2008
7.95 (3.9416.01)**
7.84 (3.8915.81)**
IOR80%
(5.6410.91)
January.February 2009
2.35 (1.055.25)*
2.33 (1.045.21)*
IOR80%
(1.683.16)
JuneJuly 2009
3.76 (1.947.30)**
3.83 (1.977.43)**
IOR80%
(2.745.31)
OctoberNovember 2009
7.68 (4.2513.88)**
7.71 (4.2613.93)**
IOR80%
(5.5310.70)
January.February 2010
2.93 (1.545.54)*
3.05 (1.615.77)*
IOR80%
(2.184.22)
Parameters/Models
Empty [95%CI] (SE)
Individualpredictor [95%CI] (SE)
Final [95%CI] (SE)
Fixedeffects
Village intercept
0.81 [0.491.31] (0.25)
0.11 [0.060.18] (0.28)
0.01 [0.0060.03] (0.37)
Village intercept variance
0.80 [0.322.01] (0.21)
0. 71 [0.281.82] (0.34)
0.034 [0.0020.615] (0.05)
Random effects
ICC (%)
19.5
17.7
1.0
MOR
2.34 (0.21)
2.23 (0.34)
1.19 (0.05)
PCV (%)

11.2
95.7
Multivariate, multilevel models
Multilevel logistic regression of the fixed effects showed that age, altitudinal location, house status, and seasons were related to higher risk of malaria infection. These variables had about twofold to eightfold increase in prevalence of malaria. Furthermore, among these variables, OctoberNovember survey seasons of both during 2008 and 2009 were strongly associated with increased prevalence of malaria infection (Table
The estimates of cluster variances (or random effects) revealed differences in malaria infection. The villagelevel intercept variance for the individuallevel predictor (0.71 [95% CI: 0.281.82]; SE=0.34) and final (0.034, [95% CI: 0.0020.615]; SE=0.05) were lower than that of empty (0.80, [95% CI: 0.322.01]; SE=0.21). The ICC value for the final model was 95.7%. Moreover, the MOR values for the empty (2.34±0.21), individuallevel predictors (2.23±0.0.34) and final (1.19±0.0.05) models were large (Table
Discussion
This study reflects that malaria transmission is highly seasonal and influenced by age of children and altitudinal location as well as poor housing condition at highlandfringe area in rural setting of southcentral Ethiopia. The months preceded by main rainy season is found to be consistently a good predictor of increased malaria infection in the present longitudinal study. Multilevel mixedeffects logistic regression analysis found increased malaria risk in children aged below five years, five to nine years, low altitude, midlevel altitude, poor housing condition, and survey seasons in Butajira area, Ethiopia. Most of the clusterlevel variance was explained and strong enough using the variables measured. More interestingly, the present communitybased longitudinal survey revealed seasonality of malaria transmission that overlapped with abnormal weather condition such as below average annual rainfall. However, health facilitybased past studies found an initial reduction of malaria burden following the largescale interventions in progress since 2005
This study has got some limitations. The present study used longitudinal parasitological data for the study participants. However, concurrent ITN possession data was limited to household level and baseline survey period. Household spray status was also incomplete. In both situations the absence of complete data on vector control can be considered as setback of this study. In the present multilevel analysis, household data was aggregated to villagelevel, which is believed to unnecessarily introduce statistical problem. As missing to followup is an inherent problem in longitudinal study, this study might also suffer from this problem. Despite these shortcomings, the present study has contributed to improving sampling problems in which some of the studies have focused on peak malaria transmission season and missed the rest of the seasons. In contrast, this study employed repeated crosssectional surveys during various seasons with different prevalence. This study also recruited more participants from high altitudes with expected low prevalence to increase the probability of finding malaria positives.
The finding of more malaria infection in children aged five to nine years is consistent with a study conducted in highlands of Ethiopia, Kenya and Uganda
It is not clear why agedependent malaria risk is observed in such lowendemic highlands such as Butajira area, where malaria risk is expected to be uniform across all age groups. Obviously, children aged below five years are vulnerable to malaria infection. The possible explanations could be lowest vector control coverage including low insecticidetreated bed net (ITN) coverage and poor ITN conditions in the study area
The finding of high malaria risk in low altitude can be explained by the presence of suitable high ambient temperature and topography that favours mosquito abundance
The increased risk of malaria in houses with their walls having holes can be due to increased access of mosquitoes to bite humans. Housing conditions allowing mosquito entrance were indicated as malaria risk factors
In conclusion, this finding showed increased malaria infection is associated with survey seasons, age of participants, altitudinal location and housing conditions in highlandfringe areas with low transmission settings. The current malaria control efforts could benefit through application of targeted interventions to villages of high malaria cases by prioritizing children aged below ten years in highlandfringe areas of low endemicity. Subsequently, seasonal transmission reduction could be operational in low transmission like Butajira area. Future studies should consider designing more frequent observations and incorporate household spray status and ITN use concurrently.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
AW contributed to conception and design, acquisition of data, analysis and interpretation of data, and drafting the manuscript. WD, AA and BL substantially contributed to conception and design of the study and reviewing the manuscript, revisiting it critically for important intellectual content. BL, AA, WD and AW reviewed the paper and all authors approved the final version.
Acknowledgements
The authors are grateful for study participants involved in the study. Authors also thank data collectors and malaria microscopist for their strong commitment during the field data collection. This work was supported by the Norwegian Programme for Development, Research and Education (NUFU) and the University of Bergen, Norway for funding.