The first step to improve the rational use of drugs is to understand prescribing patterns. This paper demonstrates the application of classification tree analysis models a non-parametric modelling methodology to explore factors influencing drug prescription practices in health facilities of rural Tanzania. Classification trees are user friendly and easy to interpret and have been utilized to identify the main risk factors for malaria infection in Burundi and Vietnam [21, 23]. In this analysis, the classification tree method revealed logical results of the relationships between the outcomes of interest (polypharmacy and co-prescription of AL with antibiotics) and the predictor variables.
While multinomial models reveal factors that predict the outcome in the whole population, classification tree analysis helps in detecting population segments that need specific attention in relation to the outcome. Segmenting populations supports decision makers in targeting their efforts to specific subgroups. It is important to note that this analysis does not support any claim of superiority of one methodology compared to the other.
This analysis demonstrated some real-life treatment practices at the facilities. It is common for most patients to report with more than one complaint, which compels the health worker to prescribe more than one medication for each identified illness. The IMCI strategy was introduced by WHO to reduce child morbidity and mortality. Indeed, treatment of childhood illness may also be complicated by the need to combine therapy for several conditions . It is therefore not surprising that the total number of diagnoses was the most important predictor of polypharmacy as revealed by both its ranking in terms of importance and being a major splitter in the classification tree. A plausible explanation is that health workers are insecure about the diagnosis since in most cases the available laboratory services are unable to accurately determine the cause of illness. Therefore, to satisfy the patients, the health worker prescribes more drugs and then justifies this practice by diagnosing several pathological conditions.
Supervision of health workers is another important predictor of polypharmacy. Indeed, polypharmacy was more common among unsupervised health workers. It is worth noting that this study did not comprehensively explore the type of supervision, limiting the conclusions on the possible consequences of not having adequate supervision on prescription practices. Nonetheless, it would be advisable for district health authorities to include drug prescription practices during their routine supervision visits.
Polypharmacy is common in the private sector where individual motivation and incentives may have preponderance over the knowledge and skills of the providers. Health worker age, sex, being trained in IMICI and being supervised in previous six months were the health worker-related variables identified as the other important variables that explain polypharmacy. The observation that patients treated with AL in a clinic that did not experience stock-outs of artemether-lumefantrine are more likely to be prescribed several treatments is expected and could be due to the IMCI strategy, which recommends this practice, especially for children presenting with multiple symptoms.
In general, presumptive diagnosis was common in public facilities while laboratory results were more used in privately owned facilities. In SSA, it is common practice not to use the test result when treating fever cases . The practice of presumptive treatment for malaria has been and is still being practiced in several health facilities, both in rural- and urban-based centres, because the syndromic treatment for febrile illnesses has been standard practice for long time, and clinicians mistrust the laboratory results due to poor quality of the laboratory tests. Patients with a negative malaria test are still treated with an anti-malarial on the grounds that signs and symptoms are compatible with the diagnosis of malaria. This continues even after the introduction of malaria rapid diagnostic tests. 
The recommended first-line anti-malarial drug (AL) was more commonly used in public than in private facilities as the former are supplied with essential drugs directly by the central pharmacy. This may change with the introduction of the Affordable Medicine Facility for malaria (AMFm) strategy in Tanzania whose approach is to supply subsidized AL to the private sector . It will therefore be interesting to look at how these changes in the health system will affect the prescription patterns in Tanzania and other African countries over time.
There was a high level of co-prescription of antibiotics with AL, particularly in children less than five years living around the Ifakara area. A study in Ghana conducted predominantly in government facilities in an urban setting showed that 30.8% of patients were receiving at least one antibiotic in addition to the recommended anti-malarial . Co-prescription with antibiotics is a life saving practice and commonly practiced in health facilities in sub-Saharan Africa since patients can present with multiple illnesses at a single clinic visit. This is why it was promoted under the IMCI strategy. However, this has implications for the patient’s safety as it may increase the risk of drug-drug interactions (D-DI), therapeutic failure, drug resistance and adverse events . If this practice of co-prescription of drugs, which is common in rural health facilities, is not addressed, it may cause a major problem as the risk of adverse drug events (ADE) increases with an increasing number of medicines prescribed [30–33].
Classification tree analysis models are useful in expressing relationships between variables since they do not need to be linear or additive and the possible interactions do not need to be pre-specified or of a particular multiplicative form. Results are presented in the form of a decision tree, a different approach than the standard statistical analysis. The results highlight areas that merit further attention and can act as a guide for further epidemiological and hypothesis-driven research. The classification trees provide a more flexible relationship between variables; missing values of the covariates, multi-colinearity and outliers are taken care of in an intuitively and correct manner . This methodology has proven its usefulness and adequacy in other areas and contexts, for example the bee colony collapse disorder, bovine spongiform encephalopathy and analysis of urban farming systems in central Africa [17, 19, 20]. In malaria, this method has been used for ranking highland malaria risk factors in Burundi and in Vietnam [21, 23]. However, it has a limitation of not providing p-values and standard deviations as in familiar parametric methods. Another limitation is that confounding values make classification tasks more difficult. Although this decreases true positive rates and accuracies, the constructed classification trees are valuable. The benefit of the trees is that they simulate more the real life situation with patients who have confounding attributes. Future work should be aimed at finding different ways to handle confounding values in the reasoning process. Another advantage is that the importance of the variable can still be seen in the variable relative importance.
A variable may be ranked among the top ones for the discriminatory power but may not appear as an important splitter in the classification tree, e g, training in IMCI with anti-malarial component. This happens because it is an important surrogate but not a major splitter. The ranking by overall discriminatory power is determined by the sum across all nodes in the tree of the improvement score that the predictor has when it acts as a primary or a surrogate splitter. Consequently, a health worker having IMCI training with anti-malarial component enters the tree as the top surrogate splitter in many nodes but never as primary splitter.
Initiatives like the INESS Phase IV platform, working within communities through the HDSS system should continue to evaluate the effect of provider practice on new and old products and may be extended to other therapeutic areas such as ARVs, anti-TBs, antibiotics and vaccines. Inclusion of the private sector, e g, private pharmacies, retail shops, mobile drug sellers and even traditional herbalists will provide public health managers with more evidence on which to base their decisions.