A better understanding of interaction between malaria parasites and clinical features might lead to a better designing and evaluation of different malaria interventions particularly drugs and vaccines. The main aim of our analysis was to determine the parasite threshold associated with fever in individuals aged 0-19 years. Lowland stratum is characterized as high malaria transmission area and the highlands as the low transmission area [16, 26], and that in the high transmission area, children are the most affected while in the low transmission area all age groups are almost equally affected because of lack of acquired immunity against parasite and clinical disease . The acquired malaria immunity varies with the level of exposure to parasite; hence individuals living in holoendemic areas tend to acquire immunity sooner than those in hyper endemic areas. Older individuals in holoendemic areas have higher immunity than children because of their prolonged exposure to malaria parasites .
Our modeling approach used strata and age of individuals as an important factors in determining parasite threshold associated with fever. We fitted three models (1-3) to determine parasite thresholds. Model (3) is an extension of model (2), where instead of fitting different models for each age group, one model was fitted for the whole data taking age as explanatory variable. A fourth model (4) was fitted for the purpose of assessing whether the first two models fits the data well when compared to the traditional ones. Model (1 and 2) had low information compared to model (4) because of the truncation (i.e parasite below and/or above τ were each categorized in one group), and hence the first model had lowest information.
Comparing the AICs, model (1) performed poorly when compared to the other two models, and this indicates that grouping parasite density into two groups is an oversimplification and might results in poor predictions. It was interesting to note that model (2) performs better than model (4) for most of groups (the exception being lowlands 4-5 and highlands 0-1 years) despite the fact that model (2) was based on a cut-off threshold and therefore truncated data. This probably reflects that quantification of low density parasitaemia by microscopy is difficult and not precise.
Simulation shows that model (2) was better in the estimation of malaria parasite threshold than model (1). The model provides the background level where the probability of fever is considered similar for parasite below the threshold and then probability of fever increases as parasite density increases above the threshold. Model (1) might be simple but this gives only the probabilities below threshold and that above the threshold, and it does not give a flexible way of choosing the threshold; for instance when a strict parasite threshold is desired. Furthermore, the model is weak especially when the slope parameter (rate of change of parasite density) is low, as it can overestimate the threshold parameter, see Table 1.
Model (3) gives a better way of explaining the relation of the thresholds to variable of interest. For example, the age and transmission intensities which are known to be important determinants in malaria transmission. This model provides a way of adjusting for other covariates. However, the problem is the optimization process where computing time grows as number of parameters to be estimated increases. For example, in a computer with Intel Pentium® Dual CPU T3200 2.0 GHz processor and 4 GB of RAM, without sub-sampling bootstrap, it takes few seconds to get the estimates for the parameters in models (1 and 2), while model in(3) where τ was fitted on age and age squared, in a data set with 5050 records it requires about 5 minutes. Threshold parameters were well estimated in model (3) especially for the lowlands strata, as can be seen from Figure 4; that a uniform confidence band will also include the fitted estimates of thresholds in the separate age groups. However, in the highlands, there was much variability in the threshold parameters fitted by model (2), which might be due to the few number of individuals with positive smears in each age group. So, model (3) might be more appropriate since it contains more data. Source of variations in the model fittings can be due to measurement errors in parasite density, which could be due to inaccurate parasite quantification as a result of either poor sample preparation, error in parasite counting and technician performance .
Even though methods used in estimation of parasite threshold and our data set type differ considerably with other studies [10, 28], the findings from this study are still comparable. For example, in children below 5 years, Chandler et., al.  found a parasite threshold of 4000 parasite/μL in those living in the lowlands and 1000 in the the highlands, while Smith et., al.  found a threshold of 5000 in children living in area of high transmission. The parasite thresholds were higher in high transmission area and lower in low transmission area [19, 28]. There was also a difference in the threshold across the age groups where children below five years had higher parasite threshold compared to older individuals, which is similar pattern as found elsewhere . This study shows that children in age group 2-5 years in the lowlands had the highest threshold while individuals in age group 10-19 years had the lowest. The pattern of the thresholds in lowlands, where malaria is endemic shows a similar trend to that of parasite density by age. This suggests that immunity plays a significant role in the threshold, as it has been shown to develop slowly in individuals who are constantly exposed to malaria, where a period of ten years was estimated to be the maximum for full development of immunity .
Results from sensitivity and specificity analysis for case definition using parasite threshold gave results similar to other studies done in lowland areas [10, 28], suggesting that the model can be used in similar settings.