Resource allocation for biomedical research: analysis of investments by major funders

Grants data for 2016 were collected using the export function of the World RePORT online platform, complemented, where available, with grant abstracts collected directly from each funder’s website and mapped to the exported World RePORT database using the unique grant identifier number.

The World RePORT data include information on direct (primary) grants provided to recipient institutions as well as collaborations with other institutions resulting from these grants (indirect grants administered by recipient institutions).

The analysis first explored the distribution of direct grants according to the parameters below and then explored the nature of collaborations between institutions that resulted from those direct grants. The following questions were explored (the analysis is also available in interactive data visualisations from the WHO Global Observatory for health R&D, which enables exploration of several of these parameters in relation to each other [12, 13]):

The data on funding amounts for 2016 was also explored but, since they have not been complete or harmonised yet for 2016, they were not considered for this analysis.

Data checks for consistency and internal validity were performed using Microsoft Excel software. These included internal validity such as valid range of years or uniform country names.

Regional classification follows the WHO regional groupings [14]. Country income group classification is based on the world development indicators of the World Bank [15]. When the country or area was not included in the World Bank income classification list (2% of the data), we performed an online search of the most recent and reliable data on gross domestic product per capita for these areas and applied the cut-off point for income groupings proposed by the World Bank to classify them into one of the four income groups [16].

To determine the type of grant, we searched for existing taxonomies, glossaries or categories of the type of grants from the websites of major health research funders (such as National Science Foundation’s glossary and NIH’s glossary and acronym list) and contacted the focal points of each the World RePORT platform funders for any unpublished sources. The lists we retrieved generally included long lists of keywords not appearing to belong to an intentional classification of projects by type (e.g. outcomes, software, database, evaluation, anthropology). We therefore developed our own synonyms list to capture the various terms used to refer to the following categories that emerged from the data: core institutional funding, training (e.g. postgraduate degrees), capacity strengthening (e.g. fellowship, prize), meetings and networking. All other grants falling outside of these categories were classified as research. The categories and list of synonyms for each category were refined and expanded in various iterations during data cleaning and analysis. This was done by reviewing the grant titles and searching for various ways of expressing the category in a snowball manner, including language variations. The search continued until no further synonyms were found.

Automated data-mining techniques were used to classify grants by health category. JavaScript and Microsoft Excel were used for this analysis.

First, a comprehensive list of disease synonyms was compiled using the following sources: the Unified Medical Language System, the 10th version of the International Classification of Diseases (ICD-10) and the WHO Global Health Estimates disease list. The list was then complemented by synonyms found in the text fields (titles, keywords, abstracts) of the various databases used by the WHO Global Observatory on Health R&D such as the WHO International clinical trials registry platform, the World RePORT and the AdisInsight database for product pipeline analysis [17‐19]. The list also includes abbreviations or language variations as well as misspellings.

Next, a code for an automated algorithm to classify the grants into health categories was written in Java to screen two textual data fields, the grant’s title and the abstract for a match with the synonyms list. The algorithm was constructed to screen the title first; if a match was found, the algorithm stopped, if not, the abstract field was searched next. The algorithm stopped when the first match closer to the beginning of the text field was found. This avoids the assignment of more than one disease. This method was developed and verified using at least five random samples of 100 records from the data to test and refine the comprehensiveness of the synonyms list, which confirmed that the primary disease focus of the grant was almost always the one first mentioned in the text-based field. This was particularly consistent in the title field. It is possible that a grant has more than one disease focus; this is not captured by this algorithm, but its significance (frequency of occurrence) was tested in the sensitivity analysis.

As shown in Table 1, a total of 69,420 grants were provided by the 10 funding organisations in 2016. The United States of America’s NIH funded the greatest number of grants (52,928; 76%) and had the longest average grant duration (6 years and 10 months). Out of the total number of grants, 70.4% were for research (48,879), followed by training (13,008; 18.7%) and meetings (2907; 4.2%) (Fig. 1).

Of grant recipients by income group, high-income countries received 98.9% of all grants, whereas low-income countries received only 0.2% (165) (Table 2). Among the 450 grants received by African countries (Table 3), South Africa (upper–middle-income country) received the highest number of grants (156; 34.7%) and was the fifth on the list of top 10 countries that received the highest number of grants. The remaining 9 countries were in the European (7) and the Americas regions (2) (Table 3).

Almost three-quarters of all grants were for non-communicable diseases (72%; 40,035), followed by communicable, maternal, perinatal and nutritional conditions (20%; 11,123) and injuries (6%; 3056) (Table 4, Fig. 2).

Among non-communicable diseases, 24% (9483 grants) were for malignant neoplasms, followed by mental and substance use disorders (15%; 5945), neurological conditions (12%; 4981), and cardiovascular diseases (11%; 4473). Among communicable, maternal, perinatal and nutritional conditions, nearly 80% of grants (8826) were for infectious and parasitic diseases, followed by respiratory infections (7%; 738), nutritional deficiencies (6%; 651) and neonatal conditions and maternal conditions (both at 4%; 496 and 412, respectively) (Table 4).

Looking at select health areas of global importance, analysis of grants for neglected tropical diseases show that they represented 1.1% (792) of all grants, of which dengue (16%; 125 grants) and leishmaniasis (13%,102 grants) were the two individual diseases that received the highest number of grants. Similarly, 0.4% (274) of all grants were for one of the priority diseases on the WHO list of highly infectious pathogens (R&D blueprint pathogens); 83% of these were for Ebola virus disease (43%; 117), Zika virus disease (32%; 89) and severe acute respiratory syndrome (8%; 21).

Around 10% (6918) of direct grants resulted in collaborations with other institutions, which did not always translate into a transfer of funds from the primary recipient to the collaborating institutions; 96.4% (6669) of these direct grants had been awarded to recipients in high-income countries (Table 5) and 75.8% (14,619) of the collaborations resulting from these grants were with others in high-income countries. In fact, for each income group, collaborations were most likely to be with others in the same income group, followed by institutions in high-income countries. For example, grant recipients in low-income countries (66) collaborated most with institutions in low-income countries (88), followed by institutions in high-income (78), lower–middle-income (11) and upper–middle-income (8) countries (Table 5).

Table 6 describes the sample size for the sensitivity analysis and the percentage of each funder’s contribution to the total number of direct grants (69,420) in 2016. The sample consisted of 107 records, after rounding up of percentage figures.

Table 7 shows that, out of a random sample of 107 grants, 81% were assigned to a health category and, in 91% of the cases, the classification was accurate. Classification accuracy was 98% when the title was used compared to 84% when the abstract was used. However, classification based on abstract contributed around 50% of classified grants, hence its usefulness. In 40% of the cases when a grant was not classified, no abstract was available. In the 28 cases where grants were misclassified, the main reasons were unspecific or very technical language used with no disease mentioned (11; 39%), general topic not linked to a specific disease focus (7; 25%), or new synonyms were discovered that could have allowed a classification to be made (9; 32%).

Overall, applying a data-mining algorithm that selects the first mention of a disease in the title or, failing this, the abstract, appears to yield reliable results; only in 1% of all classified grants (1/87) was the primary disease not the first mentioned in the title or abstract. In this case, the attributed disease was associated with the primary disease topic of the research.