Statistical analysis

To correct plate-to-plate variations, the dilutions of the PNG adult pool were fitted as plate-specific standard curves using a 5-parameter logistic regression model [22, 24]. For each plate, median fluorescence intensity (MFI) values were interpolated into relative antibody units based on the parameters estimated from the plate’s standard curve.

Associations with age, exposure and correlations between antibody levels of different subclasses and/or different antigens were determined using Spearman’s rank correlation, and differences by infection status using Mann-Whitney U tests. Generalized estimating equation (GEE) models with exchangeable correlation structure and semi-robust variance estimator were used to analyze the relationship between antibodies to PvRBPs and prospective risk of P. vivax episodes (defined as axillary temperature ≥ 37.5°C or history of fever in preceding 48 hours with a concurrent parasitaemia >500 P. vivax parasites/μl) over the 16 months of follow-up [22, 25]. For this, antibody levels were classified into tertiles (cut-off values are shown in Table 1), and analyses were done comparing the incidence rate ratio (IRR) of clinical malaria in those with medium and high versus low antibody levels. Children were considered at-risk from the first day after the blood sample for active follow-up was taken. For each child, the molFOB was calculated as the number of new blood-stage genetically distinct P. vivax clones acquired/year-at-risk, and square-root transformed for better fit [20]. Adjustments were made for seasonal trends, village of residency, age, and molFOB. In order to study the breadth of anti-PvRBP antibodies, for each antigen antibody levels stratified into tertiles were scored as 0 for low, 1 for medium and 2 for the high tertiles, respectively. Scores were then added up to reflect the breadth of anti-RBP antibodies, yielding a median score of 6 (IQR 2–9).

Abbreviations: IQR = Interquartile range

*Values in arbitrary units. Values were interpolated from standard curves using a 5PL logistic regression model.

All analyses were performed using STATA version 12 (StataCorp) or R version 3.2.1 (http://cran.r-project.org).