Identifying population segments for effective intervention design and targeting using unsupervised machine learning: an end-to-end guide

Abstract

One-size-fits-all interventions that aim to change behavior are a missed opportunity to improve human health and well-being, as they do not target the different reasons that drive people’s choices and behaviors. Psycho-behavioral segmentation is an approach to uncover such differences and enable the design of targeted interventions, but is rarely implemented at scale in global development. In part, this may be due to the many choices program designers and data scientists face, and the lack of available guidance through the process. Effective segmentation encompasses conceptualization and selection of the dimensions to segment on, which often requires the design of suitable qualitative and quantitative primary research. The choice of algorithm and its parameters also profoundly shape the resulting output and how useful the results are in the field. Analytical outputs are not self-explanatory and need to be subjectively evaluated and described. Finally, segments can be prioritized and targeted with matching interventions via appropriate channels. Here, we provide an end-to-end overview of all the stages from planning, designing field-based research, analyzing, and implementing a psycho-behavioral segmentation solution. We illustrate the choices and critical steps along the way, and discuss a case study of segmentation for voluntary medical male circumcision that implemented the method described here. Though our examples mostly draw on health interventions in the developing world, the principles in this approach can be used in any context where understanding human heterogeneity in driving behavior change is valuable.

Identifying groups of people with similar sociobehavioural characteristics in Malawi to inform HIV interventions: a Latent Class Analysis

“…Improving our understanding of the characteristics of different groups in the population, their risk of being HIV infected and their uptake of HIV testing is essential to design more targeted and tailored prevention and care programs. Clustering analysis, which is an unsupervised machine learning technique, allows us to identify similarity patterns and find hidden sub-groups of people with potentially different risk levels and drivers of having or acquiring HIV[2]. In a previous study, we investigated differences between SSA countries and identified sociobehavioural profiles that coincided with a high HIV prevalence and incidence at the country level[3].…”

Motivation To Test, Treat, and Report Malaria Cases: A Quantitative Assessment Among Private Sector Providers in The Greater Mekong Subregion

“…An understanding of what motivates private sector health providers, the dimensions of their motivation, and how motivation may be associated with current performance and sustained engagement, can be used as a helpful input to the design of interventions that select for or encourage the development of motivation dimensions that are associated with improved care and reporting (32). For example, interventions can use these tools to identify and recruit health providers with speci c motivational attributes to be part of programs that require greater a nity to non-nancial over nancial motivations (33,34) or frame program or network objectives in a way that appeals to providers with speci c motivational attributes (35). Alternatively, programs can take the perspective that motivation is mutable and develop interventions that are meant to enhance motivational dimensions that are associated with desired behaviors (36).…”

Ten-Year Single Institutional Analysis of Geographic and Demographic Characteristics of Patients Treated With Stereotactic Body Radiation Therapy for Localized Prostate Cancer

“…Statistical comparisons for these features and their relation to distance traveled for treatment was performed using the Mann-Whitney U test. After standardizing data, an unsupervised learning algorithm called K means clustering using SPSS version 23 (IBM Corp., Armonk, NY) was performed to identify distinct clusters of patients with respect to median income, racial makeup, educational level and rural residency ( 18 ). Distance traveled from each cluster was reported in miles.…”