Economists typically assume that firms are risk neutral. Yet many enterprises, especially in low and middle-income economies, are small and owner-operated, making household consumption sensitive to business risk. This may cause the risk preferences of firm owners to affect firm decisions. This paper demonstrates that small retailers in Kenya are risk averse, leading them to under-adopt a new product because they face uncertain demand. I develop a model in which risk averse firms learn about demand through stocking decisions, then test its predictions using two field experiments. The first establishes that risk aversion affects the stocking decisions of enterprises. I construct a new test for risk aversion by offering treated firms an insurance contract that lowers expected profits from a new product and reduces the risk of losses, leading to a 50% increase in adoption which implies a rejection of risk neutrality. The second experiment shows that temporarily reducing inventory risk leads firms to permanently stock a profitable new product because they overcome demand uncertainty through learning. These results suggest a departure from perfect competition, since risk averse firms are not competed out of the market. I document information externalities consistent with this prediction: enterprises that are nearby randomly induced entrants stock the new product at a higher rate. These results suggest that risk aversion can impede product diffusion and firm growth, challenging the common assumption that small firms are risk neutral.

Economists often study non-market goods such as health and air quality. This paper introduces a new method to estimate demand for such amenities and applies it to measure the value of a statistical life (VSL) in Kenya. My approach is to update beliefs about the life-saving efficacy of a product (a motorcycle helmet) and elicit product choice. This generates instruments allowing one to use subjective beliefs to estimate demand, rather than assuming rational expectations. This method does not require beliefs to be reported error-free but does require classical mismeasurement. I validate this assumption using features of the experimental design. The estimated VSL is $224, near the left tail of Kenyan estimates. Standard methods for estimating VSL produce skewed results, driven by severe violations of rational expectations. These findings help explain low observed demand for many health products and suggest that directing more development aid towards consumption may increase welfare.

We estimate the intergenerational health impacts of large-scale unconditional cash transfers. One-time transfers of USD 1000 were provided to over 10,500 poor households across 653 randomized villages in Kenya. We collected regional census data on over 100,000 births, including on mortality and cause of death, and detailed data on household health behaviors. In the study's main finding, the cash transfer treatment leads to 55% fewer infant deaths before age one and 49% fewer child deaths before age five. Data on the location of health facilities, as well as the cause of death and transfer timing relative to birth, indicate that unconditional cash transfers and access to delivery services are complements in generating mortality reductions: the largest gains are estimated among households living close to health facilities who receive the transfer around the time of the birth, and treatment leads to a large overall increase in hospital deliveries. Infant and child mortality then largely revert to pre-program levels after cash transfers end. Despite not being the main aim of the original program, we show that unconditional cash transfers in this setting may be a cost-effective way to reduce infant and child deaths.

This paper evaluates a low-cost, customized soil nutrient management advisory service in India. As a methodological contribution, we examine whether and in which settings satellite measurements may be effective at estimating both agricultural yields and treatment effects. The intervention improves self-reported fertilizer management practices, though not enough to measurably affect yields. Satellite measurements calibrated using OLS produce more precise point estimates than farmer-reported data, suggesting power gains. However, linear models, common in the literature, likely produce biased estimates. We propose an alternative procedure, using two-stage least squares. In settings without attrition, this approach obtains lower statistical power than self-reported yields; in settings with differential attrition, it may substantially increase power. We include a “cookbook” and code that should allow other researchers to use remote sensing for yield estimation and program evaluation.