Why Don't Farmers Insure? Sunk-Cost Perception and the Demand for Agricultural Insurance

The intervention is the offer of a risk-management tool, with four variants tested within-subject. Each participant — a smallholder grape grower in the district of Cascas, region La Libertad, Northern Peru — faces a sequence of simulated two-season crop production decisions under weather risk. In each round, the participant is shown the probability of a good harvest year and is offered one of four tools, which they may accept or decline:

Insurance (I). Premium α is deducted from harvest revenue; an indemnity is paid if the year is bad.
Emergency Loan (EL). A loan is disbursed if the year is bad; it is repaid with interest δ in the following season (which is deterministically good).
Cheap Insurance (CI). Identical to Insurance, but the premium is reduced to Jα with J = 1/2, framed to the participant as a 50% government subsidy on the premium. Indemnity unchanged.
Cheap Emergency Loan (CEL). Identical to Emergency Loan, but the interest is reduced to Jδ with J = 1/2, framed as a 50% subsidy on the interest. Loan amount unchanged.
The proportional cost reduction (J = 1/2) is identical across CI and CEL by design, so the expected-cost reductions induced by "cheapening" each tool are matched. This proportional symmetry is what identifies the sunk-cost mechanism: if the adoption gap between Insurance and Emergency Loans is driven by the sunk-cost perception of a paid premium, the gap should shrink when both tools are made cheaper by the same proportion.

Participants receive a fixed show-up payment plus a performance-based payment equal to a pre-specified scaling factor γ times the profit earned in a single round drawn uniformly at random from all rounds played. Working capital is provided at the start of each round, so liquidity is never a binding constraint. The instrument is a custom Spanish-language HTML survey that runs fully offline on Android tablets.

2026-05-23

2026-07-31

A binary indicator, recorded by the tablet at the moment of decision, equal to 1 if the participant accepts the risk-management tool offered in that round and 0 otherwise. Each participant contributes 20 such observations: 5 rounds in each of the 4 within-subject blocks.

Derived condition-level statistic. For each of the four conditions (Insurance, Emergency Loan, Cheap Insurance, Cheap Emergency Loan), the participant-level adoption rate is the share of rounds within that condition's block in which the participant accepted the tool. Population-level adoption rates for each condition are estimated by the pooled regression described below.

Two pre-registered hypothesis tests:

H1 — Adoption gap. The difference in population mean adoption rates between Emergency Loan and Insurance. Tests whether the emergency loan is adopted at a higher rate than insurance.
H2 — Sunk-cost contrast. The change in the adoption gap when both tools are made cheaper by the same proportion (50% reduction in the priced dimension of each tool). Defined as the gap between Cheap Emergency Loan and Cheap Insurance, minus the gap between Emergency Loan and Insurance. A negative contrast is consistent with sunk-cost perception of the insurance premium driving the adoption gap.
Both quantities are estimated from a pooled linear-probability regression of the binary accept indicator on condition dummies (Insurance is the omitted reference), with individual and round fixed effects, and standard errors clustered at the participant level. Inference is two-sided. Directional alternatives are pre-committed for interpretation but not used to relax the rejection criterion. No family-wise error correction is applied to the two primary tests (the hypotheses target substantively distinct claims — existence of a gap, and a specific mechanism for it); a Bonferroni-adjusted p-value (multiplied by 2) is reported as a supplementary column for transparency.

Within-subject crossover lab-in-the-field experiment. The four conditions described under Intervention are the four levels of a within-subject factor; each participant is exposed to all four.

Block structure. Each participant plays R = 20 rounds, organized into K = 4 blocks of r = 5 consecutive rounds. Within a block the offered tool is held fixed; between blocks it changes. The order of the four blocks varies across participants (see Randomization Method below).

Within-block variation. Within each block, the five rounds vary the probability of a good harvest year over the grid θ^H ∈ {0.5, 0.6, 0.7, 0.8, 0.9}, presented in randomized order. Probability is shown to the participant before each decision. Probability is a source of within-subject, within-block variation; it is not a treatment dimension.

Round structure. Each round has a two-season horizon. Season-1 yield is high (Y^H) with probability θ^H and low (Y^L) with probability 1 − θ^H. Season-2 yield is high deterministically. The Season-1 outcome and the round selected for the performance-based payoff are realized using physical randomization devices (coin/marked die for yield; numbered tickets drawn from an opaque bag for the paid round) that the participant can observe — to support trust in the randomness of payoff-relevant events.

Identification. The within-subject design absorbs all time-invariant individual heterogeneity through individual fixed effects. Round fixed effects absorb fatigue/learning across the global round counter. The Williams-balanced block-order assignment (see Randomization Method) ensures that block-position effects and first-order carryover are orthogonal in expectation to the treatment contrasts. The primary analysis is a pooled linear-probability regression of the binary accept indicator on condition dummies (Insurance is the omitted reference), with individual and round fixed effects, and standard errors clustered at the participant level. Inference is two-sided. The full pre-analysis plan — including hypothesis statements, robustness checks, and a pre-registered heterogeneity analysis by elicited risk aversion — is archived as a companion document.

Not available

All randomization is performed in advance by a computer, in the office, using a fixed seed; the seed and the generating script are committed to the project repository so the full assignment is reproducible. No randomization is performed in the field. Two levels of randomization are pre-generated and keyed to each participant's anonymous, pre-printed card_id:

Block (treatment) order. Each participant's sequence of the four conditions (I, EL, CI, CEL) is assigned from a Williams-balanced Latin square for K = 4 conditions. The Williams construction yields 4 sequences and balances both (a) treatment assignment across block positions and (b) every ordered adjacent pair of conditions, so first-order carryover is differenced out across participants. Sequences are assigned to participants in equal proportions (≈ 37–38 per sequence), with overflow allocated uniformly at random.
Within-block probability order. For each (participant, block) cell, the five rounds use an independent uniform random permutation of the good-year probability grid {0.3, 0.4, 0.5, 0.6, 0.7}.
The pre-generated assignments are loaded onto tablets before deployment. When an enumerator scans a card at the start of a session, the tablet retrieves that card_id's block sequence and within-block probability permutations.

Two payoff-relevant events that occur during the session — the Season-1 yield realization in each round and the single round drawn at the end of the session for the performance-based payment — are realized using physical instruments visible to the participant (a coin or marked die for yield; numbered tickets drawn from an opaque bag for the paid round). These are randomization events but they are not assignment to treatment; they are recorded on the tablet by the enumerator after they occur.

The individual participant is the unit of randomization for treatment assignment. Each participant receives an independently drawn block order from the Williams-balanced Latin square, and independently drawn within-block probability orders for each of the four blocks.

There is no clustering of treatment assignment above the participant: communal centers and session time slots are used as logistical groupings only, with each participant within a session receiving their own randomly assigned block sequence. No higher-level (community, session, household) randomization is used.

Yes

N = 150 participants for the main experiment. A separate pilot of n = 10 is conducted before main fieldwork; pilot data are not pooled with the main analysis.

3,000 observations (150 individuals and 20 rounds for each one)

150 individuals (each participant goes through all treatment arms)