A malaria seasonality dataset for sub-Saharan Africa

Categories: Disease & Virus

October 29, 2025

Dataset structure

The malaria seasonality data are stored in an instance of a relational database system that we export in tabular form for use by researchers and make publicly available at https://doi.org/10.6084/m9.figshare.28879805. Each row in the seasonality dataset consists of a single malaria seasonality record for a given time and place. Columns in the dataset are described in the Supplemental Information and can be grouped into the following categories: (1) geographic location, (2) dates of data collection, (3) the malaria metric captured, (4) the source of the data (e.g., publication citation), and (5) the observed values at a monthly resolution. To maximise flexibility for future analyses, the dataset was designed to include any data that could contain information relevant to analyses of malaria seasonality, including anecdotal mentions of seasonal timing published in peer-reviewed journals. To bridge anecdotal and empirical data, each row also contains a simplified monthly timeseries with ‘peak’ months codified with a 1, second peak (if present) with a 0.5, and other months codified with a zero. For anecdotal data, the peak months were taken directly from the text, while the simplified summaries derived from the empirical data were processed by assigning the highest value from the timeseries as the peak. Contiguous months exceeding one standard deviation (SD) from the monthly mean value were then used to define the extent of each peak period. Two further rules where applied. First, when there was a fragmented peak (such as peak spanning three months with the middle month not exceeding 1 * SD) the peak was considered to include the apparent gap. Second, in the absence of monthly values exceeding 1 * SD from the mean, peaks were still identified if the timeseries exhibited a continuous period of at least three months with values exceeding 0.5 * SD over the mean. We used this approach to define “in season” as there is no universally agreed upon definition for this metric. Our choice was guided by the need for a measure that was independent of scale as we were attempting to compare data for varied metrics and from areas with vastly different burdens of malaria. The use of standard deviation satisfies the requirement of scale independence. We also needed an approach that could distinguish between places with seasons and places without seasons, which ruled out simpler definitions of peak such as observing more than 1/12^th of the annual cases in a month. Lastly, the peak months defined by the anecdotes aligned strongly with geographically coincident measured malaria timeseries in terms of the timing of the defined peaks. However, the anecdotal peaks tended to be longer than the peaks derived from the observed timeseries. We attributed the narrower peaks among the data rows to greater geographic specificity, as a much higher proportion of the data rows were from measurements collected at point locations. In contrast, the anecdotal rows were frequently attributed to national or the largest subnational units. Also relevant for explaining the longer peak lengths among the anecdotal data is interannual variability, with anecdotes generalised across many years and thus including transmission seasons starting early, lasting longer than normal, etc.

Data gathering

Literature review and data extraction

We carried out a literature review of papers published from January 2000 to December 2022, using the following broad keyword search “malaria AND list of any endemic Country in sub-Saharan Africa”. Due to the large number of results returned from the PubMed query, we developed a seasonality literature review classifier (SLRC) that utilised natural language processing, a form of machine learning, to identify potentially relevant sources from our initial PubMed search results. The SLRC was initially trained using the results of an ad-hoc malaria seasonality literature review, and it was iteratively trained thereafter using the results of the manual abstract classification process (see section Technical Validation for further details). The iterative nature of this training process allowed us to continuously retrain and improve the accuracy of the SLRC model over time. Researchers manually screened all abstracts deemed to be relevant by the SLRC and a full-text review was conducted for all papers with abstracts that had passed both the natural language processing and manual assessment. The full-text review process applied the inclusion/exclusion criteria of our study protocol (see Supplemental Information), and all malaria timeseries were extracted from the text, figures, or tables, for entry into our dataset.

Routine surveillance dataset

The Malaria Atlas Project (MAP) has curated an extensive, annually updated dataset of geolocated clinical case incidence and mortality data collected from all malaria-endemic countries. The dataset contains over 340,000 national and subnational values from 1980 to 2023 from relevant sources that have been systematically identified, assessed for quality and consistency, and ingested into the routine surveillance dataset. Most of these routine data have an annual temporal resolution, and only 236 timeseries from sub-Saharan African countries contained weekly or monthly timeseries for a complete year and were thus included in the seasonality dataset. Only publicly available routine surveillance data were included, and the data source for each malaria timeseries is identified in the dataset. We excluded any sources for which we did not have permission to publicly share the information.

Incorporating existing compiled data

We utilised the PANGAEA dataset (https://doi.pangaea.de/10.1594/PANGAEA.892682) dataset as a reference list for sources containing malaria seasonality timeseries. PANGAEA contains malaria entomological inoculation rate data at monthly temporal resolution, with durations of at least one year, from sub-Saharan Africa, and collected between 1968 to 2013¹². To ensure consistency with our dataset, we reextracted the entomological inoculation rate timeseries using our protocols for the sources within PANGAEA. Because the PANGAEA timeseries predates our 2000–2022 collection, we extracted timeseries identified by PANGAEA only from 2000–2013.

Inclusion/exclusion criteria

The final dataset was created by employing the following study eligibility criteria: i) contains malaria seasonality information; ii) reported data were able to be geolocated to either a point location with an associated latitude and longitude, or to a national or subnational administrative unit polygon; iii) available timeseries span at least 12 consecutive months; and iv) if the full text for the study was obtainable. Additionally, we included anecdotal information on seasonal timing when the peak month or period was clearly defined in the text. All study designs from published papers were deemed eligible unless the timeseries measurements were collected as a component of a study quantifying the impact of an intervention regime change.

Data and anecdotal information

Malaria transmission in sub-Saharan Africa is often seasonal and can vary considerably between locations and years, even within the same country². We developed this dataset to support analyses of this complex aspect of malaria epidemiology. The resulting dataset is the most comprehensive collection of data quantifying seasonal patterns of malaria in sub-Saharan Africa. A novel aspect of the dataset is that it contains both empirical data and anecdotal information from published articles and reports. The anecdotal information, which represents 14% of the rows in the dataset, was included to capture the knowledge of local experts when available in published reports, and to the best of our knowledge, this type of data has not previously been systematically recorded. By including both types of information, our dataset may support analyses exploring the differences between perceived and measured seasonal patterns.

The anecdotal data were potentially imprecise relative to the timeseries. As such, anecdotal data were included using the following protocol. First, we only included the information when it clearly stated the peak month or series of months for malaria transmission, malaria cases, or vector density (e.g., “The peak of malaria transmission in district X is between July and October”). Notably, we did not include tangential metrics such as “rainy season” unless they were directly linked to malaria within the same publication. Second, if a paper cited a reference with an anecdotal statement about malaria seasonality, we reextracted that information from the original source when it was available if it was suitable per our inclusion criteria. Because we recognise that anecdotal information may be biased, our dataset clearly defines which rows contain timeseries and which contain only anecdotal information. As such, researchers using the dataset can easily choose which type(s) of data they use for their analyses.

Dataset description

The traditional literature review conducted using PubMed identified 32,574 potentially relevant peer-reviewed or pre-print articles that were published between 2000 and 2022. Some of these papers included data collected earlier, which resulted in extracted timeseries stretching from 1964 to 2021, but timeseries dating before 2000 were not gathered systematically and should be considered opportunistic inclusions to the dataset. The SLRC reduced this initial set to 5,742 sources it deemed relevant (Fig. 1) and thus eligible for full-text review. The full-text review yielded 790 new sources, to which we added 71 papers identified by the PANGAEA study that our literature review initially failed to capture. The 861 literature sources resulted in 4,110 unique records of malaria seasonality in sub-Saharan Africa. These were combined with 236 observations from MAP’s routine surveillance database to produce a final dataset containing 4,346 records (rows) (Table 1). The resulting dataset includes information on the source, the start and the end of the study period, the population study group (e.g., only females), and the location of the study. Full descriptions of the data columns are provided in the Supplemental Information. Location information is provided for each row at the lowest identifiable administrative level, or using latitude and longitude for points. Fourteen percent of the dataset consists of rows derived from anecdotal descriptions, while 86% of the rows were supported by data. Among the data rows, 78% contain a numerical timeseries, 5% are associated with plots from which numerical data are impossible to extract but we could ascertain peak periods, and 3% are associated with written descriptions of seasonal timings that were observed in unpublished data (i.e., the reported seasonality was based on data and therefore not considered anecdotal). The most common metrics are incidence (73%), entomological (20%), prevalence (4%), and mortality (2%). Fifty percent of the malaria seasonality observations are available at point level and have longitude and latitude attached, followed by the administrative levels two (23%), one (13%), zero (i.e., national) (8%), and three (6%) (Table 2). Information is available from 47 sub-Saharan African countries (Fig. 2), 43 of which have at least one timeseries available (Fig. 2). However, the density of data is asymmetrically distributed across the continent and seasonality may be poorly characterised in places with very few associated rows. Of particular concern are countries on or very near the Equator that lack regular peaks in malaria transmission. Figure 3 shows a map of the mode primary peak month for malaria incidence seasonality from sources within our dataset. To illustrate the heterogeneity present in the dataset, even within single countries, the map in Fig. 3 is displayed hierarchically such that higher resolution administrative units overlay lower-level ones (e.g., subnational units are on top of countries). Features of this map include a primary peak between September and October for the Sahel Region, and peaks between March and May for Southern Africa. This map should be interpreted cautiously, however, as it reflects data from only the primary peak (or period) of sampled locations within the dataset. As such, it may reflect inaccuracies related to small sample sizes and/or timeseries collected with localised epidemiological settings that differ from general trends within the country.