Fig. 4: Lognormal hurdles models correcting for estimated yield gaps following Ponisio et al.¹⁵.

Correcting for yield gaps does not affect the first hurdle (a), but does shift the coefficient estimates in the second hurdle up (b) relative to the unadjusted model (Fig. 2). Figure details are otherwise the same as Fig. 2. The x-axis indicates different pesticide use outcomes: kg ha⁻¹ active ingredients (AI), kg ha⁻¹ products (Prd), kg ha⁻¹ of products targeting insect pests only (Insect), kg ha⁻¹ of products with a propensity to drift (Drift), kg ha⁻¹ products of potential hazard to fish and bees (Fish, Bee), as well as products of higher (EPA signal word 1–2) and lower (EPA signal word 3–4) acute human toxicity (High, Low). Symbols indicate point estimates (mean) and error bars represent the 95% CI. All models include cluster robust standard errors clustered at the farm-by-crop family level. For the second hurdle (b), percent change is calculated from the log-level model as \(100(e^{\beta }-1)\) and standard errors are derived using the delta-method implemented with the nlcom function in Stata. All models include covariates for field size, farm size, and soil quality as well as farm-by-crop family random effects. N = 91,926 for all specifications in the first hurdle (a) and N = 68,704 (AI), N = 68,816 (Prd), N = 52,606 (Insect.), N = 67,988 (Drift), N = 60,653 (Fish), N = 48,254 (Bee), N = 61,883 (High), and N = 65,593 (Low) in the second hurdle (b). Coefficient estimates for all covariates are provided in Supplementary Table 9.