Whose choices matter? Rethinking peer influence in microfinance

Letting the data shape the social norm

Models of peer effects describe how other people influence an individual’s decisions. Many assume that influence depends on the average behaviour of someone’s social contacts. Chadha’s study asks what happens when that assumption is wrong—and whether a more flexible model can identify whose behaviour matters.

The paper, posted to arXiv on November 9, 2026, is a preprint rather than a journal publication. It combines mathematical results, computer simulations and an analysis of microfinance adoption in rural Karnataka, India.

Here, a “social norm” is a mathematical summary of perceived peer behaviour, not necessarily an explicit community rule. The model uses a constant elasticity of substitution, or CES, function. Its adjustable curvature allows the norm to resemble an average, emphasise the neighbour most likely to adopt, or emphasise the least likely adopter.

Crucially, the model combines beliefs about neighbours’ adoption probabilities—not their observed yes-or-no choices. A maximum-like norm therefore gives greatest influence to the neighbour believed most likely to borrow. A minimum-like norm remains low when even one neighbour is unlikely to participate. This belief-based formulation is an assumption, not a direct measurement of households’ perceptions.

Separating two kinds of influence

The model distinguishes complementarity, meaning that adopting becomes more rewarding as peers are expected to adopt, from conformity, meaning a preference for matching the perceived norm. Complementarity also scales with the number of social connections, helping distinguish the two mechanisms.

Households make binary decisions with incomplete information: they know their own private preferences but not everyone else’s. An equilibrium occurs when adoption probabilities agree with the beliefs guiding decisions. The paper establishes conditions for an equilibrium to exist and be unique, and conditions under which observed choices, characteristics and network connections can identify the model’s parameters.

Estimation uses nested pseudo-likelihood, which jointly recovers parameters and unobserved choice probabilities, implemented with a spectral numerical algorithm. Simulations compare estimates with known values used to generate artificial data. The flexible model generally improves as samples grow, while imposing an average-based norm can leave persistent bias. However, extreme curvature is difficult to estimate precisely, and small samples or low adoption can weaken recovery.

Evidence from village networks

The empirical analysis uses 6,733 households in 29 villages, drawing on earlier microfinance research. Participation comes from lender administrative records; 17.4% of households participated. Social connections were measured before the lender entered, using reported visits between households.

The estimates include household characteristics, village-specific effects and a separate baseline adjustment for households without network connections.

Across the pooled sample, the flexible model favours a strongly maximum-like norm. A likelihood-ratio test rejects the average-based restriction with a reported p-value of 0.031. The fitted curvature is about 491.5, but the likelihood becomes nearly flat near the maximum limit: that number should not be read as a precisely measured behavioural quantity.

Relative to the flexible fit, the average-based model estimates complementarity more than three times as large and conformity about one-quarter smaller. These are differences between model estimates, not independently observed measurements of the mechanisms.

Splitting villages by caste composition and wealth proxies reveals further variation. Using bed ownership, poorer segments favour minimum-like norms and richer segments maximum-like norms. Results change when latrine ownership measures wealth, however, and several smaller segments do not reject the average-based restriction.

Different models, different subsidy recipients

The paper then allocates a hypothetical fixed subsidy budget to maximise expected adoption, accounting for both recipients’ responses and network feedback. The budget is expressed in utility-index units, not currency.

Using their respective estimates, the flexible and average-based models fund 77 and 132 households, with only 19 recipients in common. Among households funded exclusively by the flexible model, roughly 62% are designated village leaders, compared with about 12% among those funded exclusively by the average-based model. Leaders included teachers, shopkeepers and self-help-group heads who were initially informed about the programme.

The models also predict very different adoption increases. These are counterfactual forecasts under different fitted models—not results from a subsidy trial or proof that one allocation delivers a measured real-world advantage.

Important boundaries

The analysis treats the network as exogenous and programme awareness as given. It assumes neighbours’ characteristics affect a household through neighbours’ adoption, without separate direct effects. Removing the adjustment for isolated households drastically changes the estimated conformity and norm.

Some village segments may have multiple equilibria. Moreover, the subsidy objective is not globally concave, so numerical optimisation does not guarantee a unique global optimum. Finally, maximising borrowing is not the same as improving welfare, which the exercise does not evaluate.

The central finding is therefore conditional but consequential: how a model combines peer influence can change both its account of behaviour and the households it recommends targeting.