Independent annual community-based, cross-sectional surveys were conducted between 2006 and 2013, between the months of April and July within the Kenya Medical Research Institute and Centers for Disease Control and Prevention Health and Demographic Surveillance System (HDSS) in Siaya County in western Kenya. The HDSS has been described in detail elsewhere.21 22 Briefly, HDSS covers a population of approximately 223 000 people residing in 393 villages located in three of six subcounties of Siaya County, an area of approximately 700 km² along the shores of Lake Victoria. The vast majority of the population are subsistence farmers and fishermen. Health indicators in Siaya County, part of the former Nyanza Province, are poor compared with national standards.23 24 Nyanza Province had the highest rates of child mortality and an estimated 60% of the population lived below poverty level during the survey period.25

A total of 19 315 individuals in 7253 households were surveyed between 2006 and 2013. Overall, 33.9% were children aged <5 years, 26.6% were children aged 5–14 years and the remaining 39.5% were 15 years old adults. Sample size in 2006–2013 were (2006 n=1113; 2007 n=1270; 2008 n=1830; 2009 n=2508; 2010 n=5334; 2011 n=2129; 2012 n=2719; 2013 n=2412) and the mean annual sample was 2414 (table 1).

For each year from 2006 to 2013, different sampling strategies were selected for logistical purposes. Systematic sampling technique was used from a sample frame of eligible households and individuals enrolled into HDSS except in 2009 when a cluster sampling was used (table 2). Households were selected for participation in the surveys if they had at least a child <5 years because many malaria control interventions targeted this age group. In the HDSS, each individual, household, compound and village is assigned a unique number. For the years when systematic sampling was used, a list of households and individuals was made ordered by the unique identifiers and by villages which are spread over the entire study area. Once a sample size for the individuals required in each year, the number of households was estimated assuming a household had an average of five members. The households were then systematically sampled from the list. The individuals sampled were then classified as <5, 5–14 and 15 years and above. In 2009, villages were randomly sampled as clusters and the number of households divided proportionately between the three study areas. Surveys were conducted in Rarieda, Gem and Alego-Usonga subcounties in Siaya County except in 2006 when Alego-Usonga subcounty was not included.

During the surveys, study participants were interviewed by trained staff using personal digital assistants and tablets. Data collected included demographic factors, socioeconomic factors including asset ownership, characteristics and utilities, care-seeking behaviours, history of fever in the 2 weeks before the survey, ITN use and antimalarial medication use both recommended and non-recommended by polices.

During each survey, a blood specimen was obtained from all individuals providing consent in the sampled households using a finger prick and used for measurement of haemoglobin (HemoCue; Ängelholm, Sweden) and to measure malaria parasitaemia by rapid diagnostic test (RDT) (Carestart Malaria HRP-2/pLDH (Pf/PAN) Combo, Somerset, New Jersey, USA). Individuals with a positive malaria RDT were treated in accordance with the Kenya national malaria treatment guidelines.10 12 26 Thick and thin blood smears were obtained for malaria species’ identification and parasite density.

Data coding, recoding, merging and analysis were conducted in Stata V.14. The eight cross sectional surveys were first analysed independently and then as pooled data. The key variables were identified for each year and then appended to each other to form a large dataset. Considering that more one person was selected in households, the analyses have considered clustering. Because these were data taken from different independent samples of the populations over time, there were bound to be missing data. In our analysis we conducted complete case analyses by excluding missing values.27 Trend analysis was conducted using Cochrane trend test.28 29 A generalised linear model, using a Poisson distribution with a log-link function, was used to estimate adjusted prevalence ratios (aPR) accounting for clustering at the household level for each individual year, to address potential section bias. Although these datasets were obtained from independent cross-sectional studies, the pooled datasets combining all the years were analysed using cox proportional hazard models assigning the same follow-up time for each participant via a robust variance estimator to consider repeated measurements of over time. This was because it has been shown that by imposing a constant follow-up time for all individuals, Cox model can be adopted to estimate prevalence rate ratios in cross sectional studies and this addresses selection biases.30 31 Study outcomes included malaria parasitaemia infection, care seeking, medication and ITN use. The independent variables were socioeconomic status (SES), study areas (subcounties), sex and age groups (<5, 5–14 and ≥15 years). SES indices were generated using multiple correspondence analysis using the following variables; occupation of household head, primary source of drinking water, type of cooking fuel, ownership of household assets and ownership of livestock. The households were categorised into five socioeconomic quintiles and then classified into two groups for ease of comparisons. The first three lower quintiles were classified as the ‘poorest’ and the fourth and fifth quintiles classified as the ‘less-poor’.32–34 Backward selection criteria was used to include independent variables in the models and 95% CI of the prevalence rates were estimated in each case. All the analyses were weighted to account for sampling strategies. Sampling weights were created by dividing the population by the sample for each subgroup (age categories and study areas).

The research questions of this study were informed by patient’s priorities, experience and preferences and public were fully involved. Malaria disease is considered a priority to patients in this study areas because it can cause disabilities and deaths among patients. Similarly, poverty is a known problem that hinder many patients from accessing and utilising health interventions. Hence, examining the trends in burden of malaria in population subgroup is key to informing policies that reduce the burden and improving access to interventions and at the same time ensuring equity. Ethical considerations in this study required that a rigours community mobilisation be done through their advisory committees, meetings were held with health management teams in the local areas, participants were assured during consenting processes that patients who were found to have malaria parasites would be treated. For data collection we recruited field assistants from the same communities where we did our study and also with help of community health volunteers. Before conducting these surveys, we did not know who was positive for malaria and hence no patients conducted the recruitment. Results of this study will be shared with the Siaya county health management team for policy considerations and with the Kenya national malaria control programme who are charged with responsibilities of identifying priorities areas for interventions. Results will also be shared in workshops involving community members.

Overall and in the pooled dataset, prevalence of malaria parasitaemia identified using microscopy was 36.5% with substantial variation between age groups (38.2% in children <5 years; 56.8% in children 5–14 years; 20.9% for adults ≥15 years). The prevalence of malaria parasitaemia was relatively stable between 2006 (38.3%) and 2011 (39.8%), but reduced from 36.3% in 2012 to 34.5% in 2013. The proportion of individuals who received the first-line antimalarial medication, AL, in the 2 weeks prior to survey increased from 0% in 2006 to 44.0% in 2013 (table 1).

In the pooled data (n=11 383), prevalence of malaria infection was significantly higher among poor individuals compared with less-poor overall (39.9% vs 33.5%; aPR=1.17; 95% CI 1.11 to 1.23). The prevalence of malaria infection was also significantly higher in poor individuals in each age group (children <5 years: aPR=1.20 (95% CI 1.11 to 1.31); children 5–14 years: aPR=1.13 (95% CI 1.06 to 1.21)); adults ≥15 years: aPR=1.18 (95% CI 1.05 to 1.33)). There was no clear trend in malaria prevalence by SES either overall or stratified by age group over time for the pooled analysis (table 3). For the pooled data, there was no significant difference in the proportion of individuals who sought care for illness between poor and less-poor households (61.1% vs 62.5%, aPR=0.99 (0.95 to 1.03)) overall or by age group and year (table 4). Overall, medication use was similar among the poorest individuals and less poor (73.2% vs 76.2%, aPR=0.95 (0.92 to 1.00)). However, poorest individuals were less likely to use a recommended first-line antimalarial medication (ie, AL or quinine for pregnant women) among those reporting fever in the 2 weeks prior to survey (18.8% vs 22.1%, aPR=0.81 (0.72 to 0.91)). Poorest households were slightly less likely to report ITN use the night prior to the survey (55.2% vs 57.8%, aPR=0.95 (0.91 to 0.99)).

Trends analysis for the period 2006–2013, showed non-significant change in parasitaemia (overall trend p=0.2560), among poorest (p=0.235) or among less poor (p=0.254) over time. However, among children 5–15 years the burden significant reduced among wealthier individuals (trend test p=0.007) but not among poorest individuals (p=0.158). Care seeking for fever among poorest individuals did not change (p=0.059) but significantly increased among less poor individuals over time (p=0.012). Overall ITN use significantly increased between 2006 and 2013, and also increased among poorest individuals (p<0.001) and among those less poor (p<0.001). Utilisation of medication for malaria increased in both the poorest and less poor individuals (p<0.001) overtime. ITN use also significantly increased over time in both groups and the gap were narrower over time (p<0.001) (table 4).

The main strength of this paper is use of 8 years of pooled data which provided more power to assess socioeconomic inequalities and equity. For lack of recent data, these historical data provided an opportunity to monitor socioeconomic inequalities and equity effect of interventions. There did not exist enough studies assessing socioeconomic inequalities over time and progress towards achieving SDG goals by 2030. The study had three main limitations. First, the findings were based on data from cross-sectional surveys preventing any evaluation of cause-and-effect of SES on malaria indicators over time. However, robust statistical analysis including accounting for households clustering. Second, only households with children <5 years were included in the surveys based on protocol-specific objectives. Although all children <5 years in a household were surveyed every year, only a small proportion of persons ≥5 years were included in the survey samples and lastly these results are generalisable to study area and not nationally. The difference in sampling techniques over time whereby in 2009, cluster sampling was used instead of systematic sampling may have in selection bias and may confound the interpretation of results.