Experimental Design
We conduct a three-arm randomized controlled trial with approximately 900 entry-level workers in South Africa. Respondents are randomly assigned at the individual level to one of three network elicitation methods. Arm 1 (standard auditory) uses traditional survey questions read aloud by an enumerator, requiring respondents to recall contacts sequentially from memory. Arm 2 (decomposed auditory) scaffolds the recall task by prompting respondents to think through their networks systematically by setting or relationship type before providing counts, delivered verbally. Arm 3 (decomposed visual) uses Network Canvas software, which allows respondents to see their network as they build it with contacts represented as visual nodes, reducing working memory demands while also decomposing the recall task.
The primary comparison of interest is between Arm 1 and Arm 3, which represent the most distinct elicitation approaches. To maximise statistical power for this comparison, these arms are each allocated approximately 375 respondents. The remainder of the sample is assigned to Arm 2, which serves a diagnostic role: by isolating the effect of decomposition from the effect of visual presentation, it helps identify the mechanism through which any differences between Arms 1 and 3 operate — whether differences in network reporting are driven by the structured decomposition of the recall task, the reduction in working memory demands afforded by visual elicitation, or some combination of the two.
All respondents complete a battery of cognitive skills assessments. The primary measure, performed prior to the network elicitation, is the Digit Span subtest from the Wechsler Adult Intelligence Scale, in which participants repeat sequences of numbers in forward and backward order at increasing levels of difficulty, capturing numerical working memory. This is complemented by two subjective measures: a self-reported measure in which respondents rate the ease, comprehensibility, and effort required by the survey task on 5-point Likert scales; and an enumerator-evaluated measure in which field officers independently rate each participant's comprehension and communication. These measures allow us to examine whether treatment effects on network reporting are heterogeneous by cognitive capacity, and whether objective and subjective indicators of cognitive load predict measurement quality differently.