The trial tested a mobile money training programme with garment workers in Bangladesh. The study was designed to measure financial inclusion outcomes: whether workers who received structured training on mobile wallet use showed increased savings, reduced reliance on informal moneylenders, and greater use of formal financial products. The design was clean — randomised assignment, endline after eight months, attrition low. The primary outcome, savings rate, increased significantly in the treatment group.
The problem emerged when we looked at the qualitative interviews we had run alongside the trial. We had designed them to understand mechanism — to explain why the treatment group was saving more, not just that they were. The answer was not what the programme theory had predicted.
Workers were using mobile wallets to hide savings from family members. Not primarily to access formal financial services, not primarily to build emergency buffers for their own households, but to create a pool of money that was not visible to husbands, brothers, or in-laws. The wallet was privacy infrastructure as much as financial infrastructure.
This is a finding with a complicated relationship to the programme's success narrative. The savings increase was real. The mechanism was real. But the theory of change that justified the intervention — expanded access to formal financial services leading to reduced financial vulnerability — was not the mechanism operating. Something else was operating: a technology that allowed women to exercise unilateral financial decisions in contexts where those decisions were otherwise subject to household oversight.
Publishing a finding framed purely around savings impact would have been accurate and misleading simultaneously. Accurate because savings did increase. Misleading because it would imply that the programme worked for the reasons it intended to work, which it did not.
The question this raises for trial design is whether mechanism matters as much as outcome. One view holds that if the outcome is positive, the mechanism is a secondary concern — programme designers can figure out the mechanism later. We disagree with this view, for two reasons.
First, if you do not understand the mechanism, you cannot predict where the programme will work and where it will not. The hiding-savings mechanism depends on a specific social structure: households where women's financial decisions are monitored by other adults, and where mobile phones are individually owned rather than shared. Change either condition and the mechanism may not operate. A trial result derived from this mechanism cannot be generalised to contexts without it.
Second, some mechanisms have ethical implications that outcomes alone do not surface. Women hiding savings from their households is not inherently problematic — financial autonomy is a legitimate goal, and the women we spoke with described the capacity to save without disclosure as important to their sense of security. But it also implies a household dynamic worth understanding before designing scaled programmes, particularly where household conflict over resources is already a documented risk factor.
The programme funder accepted the finding, and the publication included the mechanism alongside the impact estimate. The process required some negotiation — there was initial concern that a complicated findings narrative would undermine the case for continued investment. The eventual position was that a findings narrative that held up to scrutiny was more valuable for long-term credibility than one that did not.
The methodological lesson is about the design of mixed-methods components in RCTs. Qualitative work that runs alongside a trial without a defined mechanism-testing function tends to produce illustrative case material that is reported in a separate section and read by fewer people. Qualitative work designed explicitly to probe the theory of change — to identify which link in the causal chain is actually operating — produces something that changes how the quantitative findings are interpreted. The two functions require different protocols and different sampling strategies.
Bangladesh financial inclusion data from MicroSave Consulting and BRAC Institute of Governance and Development. The World Bank's Findex provides regional benchmarks on mobile money adoption and savings behaviour.