Inclusive Electrification in Yemen: Preferences, Constraints, and Fragility

We conduct a framed field survey experiment in Yemen, focusing on discrete-choice decisions and estimating the marginal impact of loan attributes on loan selection and loan-versus-cash preference. In addition, we are evaluating:

1) constraints (income and expenditure categories), beliefs about loans
2) beliefs (about solar panel breakage, quality, banks, merchants, future events, income volatility)
3) preference parameters elicited regarding intertemporal preferences, present bias, naivete, sophistication, diminishing/increasing sensitivity in gains (risk aversion vs seeking), diminishing vs increasing sensitivity in losses, loss aversion, and social attitudes.

We use randomized constraint framing and randomized placement of the discrete choice experiment within the framed field experiment to evaluate the effects of expenditure and financial salience on preferences for loans versus cash and preferences over loan attributes. We compare these effects as interactions across critical demographic and preference parameters, and evaluate the direct heterogeneous preferences and the direct effects of framing (or mindset) and salience (or related mechanisms). We randomly assign each participant to one of nine conditions in a 3×3 between-subjects design, varying two dimensions that determine the constraint environment in which participants make their choices.

2026-02-11

2026-04-29

This framed field experiment incorporates outcomes from both the discrete choice experiment (focusing on different determinants of variable selection within choice sets and across individuals) and the co-moving mechanism analysis questions.
These outcomes are grouped into seven “primary” outcome families and two “secondary” outcome families; other associated variables are purely exploratory. While this is a substantial set of outcome families, they are necessitated by the diversity of conceptual elements that would not make sense to combine for the sake of power. Four families are “Primary outcomes” in the traditional sense of being explicit endpoints. The first is the set of outcomes related to loan aversion/loan preference, comprised of DCE choices and implicit willingness to pay for cash versus loans. The second is the set of loan attributes and the relative preference for different types of loan products. The third family is simply the binary choice selection decision in the DCE. Together, these provide the key “final” within-survey experimental outcomes.
The first three families, consisting of participant choices in the discrete choice experiment include individual-level valuation outcomes obtained by aggregating choices within an individual, the choice-level marginal probability of selection, and the marginal implied WTP of each attribute. These are obtained by evaluating the choices made in the DCE. Each choice set presents four alternatives that vary in multiple dimensions (e.g., cash versus loan, loan duration, fee/subsidy). All four vary in total cost (300,000–700,000 YER). This wide range is critical for identifying who is price-sensitive and who has preference-based aversion. When a participant chooses 600,000 YER in cash over a 350,000 YER loan, they are paying a 250,000 YER premium to avoid debt (conditional on all other attributes of the loan). Alternatively, some individuals may be willing to pay a substantial premium to secure debt-related financing (which is more typical of standard financing than of subsidized development aid financing). These premiums can then be decomposed into components attributable to the participant’s individual characteristics.
However, we also collect (the fourth primary outcome) a rich set of prior real-world outcomes which should not be affected (except through differential recollection) by the treatments, but are outcomes in the survey (in the sense of being regressed on underlying demographic variables, household status, and other pre-existing conditions) and fulfill triple roles as outcomes, moderators, and the contrast between these real-world outcomes and stated preferences is regressed on treatments to provide deeper insight into using survey experiments as a tool to enrich preference analysis beyond what standard surveys can provide.
Our fifth primary outcome family comprises a central mediator and secondary mediators of importance, including two moderators that are tested for mediation. The central mediator is participants’ self-evaluations of how they answered the DCE, whether according to their own constrained preferences or to deeper underlying preferences that reflect their unconstrained choice behaviour. We include the other suspected mediators in this family (beliefs that loans are Haram, i.e. "prohibited" and perceived repayment likelihood).
In the sixth family, we evaluate a hypothetical outcome with potentially substantial consequences: participants’ willingness to pay for a real solar loan and for the bank to come to them to help them fill out a loan application.
In the seventh, we incorporate all behavioural parameters used for both the reduced-form moderation analysis, evaluating when they must be considered in a dual role as mediators.
The families consist of the following variables:
Loan versus cash preferences (DCE-1)
Selected_Loan
Loan_selection_rate
Is_majority_cash
Always_cash
Always_loan
Loan_aversion_indicator
loan_pref_indicator
Loan_aversion_rate
Loan_aversion_premium
Mean_cost_chosen
WTP_loan_vs_cash
Cond_wtp_cash
Cond_wtp_loan
Family 2) DCE loan attribute selection; DCE-2
Selected_bank
Selected_long
Selected_flex
Selected_simple
bank_share
uncond_Bank_wtp
cond_Bank_wtp
cond_Merchant_wtp
Bank_share_cash_context
Bank_share_loan_context
Duration_12m_share
Duration_wtp
Cost_coefficient
Simple_docs_share
Simple_doc_wtp
Wtp_flex
Family 3) DCE-3
Selected
Family 4) Real world variables
Loan before
Solar ownership
Solar payment
Actual_loan_buyer
Dce_vs_actual_gap
Family 5) Key mediators*
Dce_pref_type
Repayment_beliefs*
Perceived likelihood of negative shock
Haram_beliefs_nointerest*
Family 6) “Real” WTP
Stated_solar_wtp
Stated_bank_wtp
Family 7) Behavioural Parameters
Beta
Delta
Beta_hat
Risk_elicitation_g1
Risk_elicitation_g2
Risk_elicitation_l1
Risk_elicitation_l2
Risk_mixed_1
Risk_mixed_2

The families consist of the following variables:
Loan versus cash preferences (DCE-1)
Selected_Loan (choice level- chose loan in choice)
Loan_selection_rate (proportion of time loan selected conditional on availability)
Is_majority_cash (indicator -chose cash majority of time available
Always_cash (indicator if chose cash whenever available)
Always_loan (indicator if chose loans whenever available)
Loan_aversion_indicator (indicator at individual level if spent money to avoid loan)
loan_pref_indicator (indicator at individual level if spent money to have a loan)
Loan_aversion_rate (time where spent positive money to avoid a loan)
Loan_aversion_premium (wtp to avoid loans in a decision)
Mean_cost_chosen - average price spent on choice (measure of flexibility of overall payment, important primarily as a moderator)
WTP_loan_vs_cash - unconditional estimates of wtp for loan
Cond_wtp_cash estimated individual wtp for cash conditional on pref for cash
Cond_wtp_loan - estimated individual wtp for loan vs cash conditional on pref for loan
Family 2) DCE loan attribute selection; DCE-2
Selected_bank - choice level, bank selected
Selected_long - choice level, 12 months selected
Selected_flex - choice level decision, flexibility selected
Selected_simple - choice level decision; simple documentation selected
bank_share -Proportion of time bank selected when available choice
uncond_Bank_wtp estimated value of bank versus merchant
cond_Bank_wtp_l - estimated wtp for loan conditional on bank
cond_Merchant_wtp_l - estimated wtp for loan conditional on merchnat
Bank_share_cash_context(Bank cash selected when alternatives are bank cash purchase and merchant cash purchase)
Bank_share_loan_context (Bank loan selected when alternatives are bank loan and merchant loan)
Duration_12m_share (proportion of time longer term loan selected when available)
Duration_wtp (estimated individual-level willingness to pay for loan extensions to 12 months from 6 months, key because it means that cost savings exceed repayment costs)
Cost_coefficient (How responsive are people to cost in selection probability, estimated separately for positively and negative slopes then pooled if approximately linear with no subsidy/cost difference)
Simple_docs_share (proportion of the time that simplicity is available that it is selected)
Simple_doc_wtp (estimated wtp for simpler documentation in purchasing a solar panel)
Wtp_flex (positive/negative value estimated at the individual level for flexibility in the DCE)
Family 3) DCE-3
Selected (a dummy for whether a particular choice was selected within a particular set of four choices in the discrete choice experiment)
Family 4) Real world variables
Loan before (indicator for having previously had a formal loan, with related dummies for having previously had informal loans)
Solar ownership (indicator for owning solar panels)
Binary outcomes for each explanation of how people purchased solar panels
Actual_loan_buyer: Binary measure of actually purchases solar panels with loans previously
Dce_vs_actual_gap_1: Binary measure of whether they prefer a loan in the DCE but do not have a loan in real life, measured by the binary of positive wtp for loan in DCE minus real loan purchases
Dce vs actual_gap_2 - Measure of the opposite (loan in real life but not in survey)

Family 5) Key mediators*
Dce_pref_type (Likert scale from 1-5 indicating whether they answered more according to their constrained preferences or their underlying, unconstrained preferences)
Repayment_beliefs* (Perceived likelihood they will be able to repay their loan)
Perceived likelihood of negative shock (number of months out of 6 where they anticipate a lower income with a fall of at least 25% relative to present)
Haram_beliefs_nointerest* : Proportion of known people who would view a loan with no interest as Haram, i.e. "prohibited" (with a scale of different Haram perceptions about loan as a follow up)
Family 6) “Real” WTP
Stated_solar_wtp: Stated willingness to pay for a solar panel loan
Stated_bank_wtp : Stated willingness to pay for a bank agent to come to them and complete all paperwork for a loan
Family 7) Behavioural Parameters
Beta (inverse of present bias)
Delta (patience)
Beta_hat (sophistication)
All risk_elicitation questions start from bisection for the 100 vs 100,000 gain/loss then ask an exact WTP
Risk_elicitation_g1: Certain gain equal to 50% gain of 100,000 riyals (or 50% gain of 100 riyals)
Risk_elicitation_g2: Certain gain equal to 50% Gain of 250,000 riyals (or 50% chance of gaining 100 riyals)
Risk_elicitation_l1: Certain loss equal to 50% loss of 100,000 riyals (or 50% loss of 100)
Risk_elicitation_l2: Certain loss equal to 50% loss of 250,000 Riyals (or 50% loss of 100 riyals)
Risk_mixed_1: Willingness to accept to participate in a gamble of 100,000 Riyals won/lost (if gamble valuation<0) or Willingness to accept to not participate in a gamble of 100,000 Riyals won/lost
Risk_mixed_2: Willingness to accept to participate in a gamble of 250,000 Riyals won/lost (if gamble valuation<0) or Willingness to accept to not participate in a gamble of 250,000 Riyals won/lost

why_prefer_cash (9 options converted to dummy outcomes)
why_prefer_loan
why_always_12m
why_always_6m
why_always_flex
why_never_flex
why_same_position
dce_difficulty
dce_difficulty_moderator
choice_confidence
months_electricity_savings_40k
solar_fail_1yr
solar_fail_3yr
prob_default_*consequences
debt_attitude
expected_rate_*
expected_inflation_*
quality_diff
failure_rate_bank
failure_rate_merchant
months_electricity_savings_40k

why_prefer_cash (9 options converted to dummy outcomes) -9 binary why_prefer_cash_reason
why_prefer_loan (Similar 9 option binary split)
why_always_12m (Binary outcomes corresponding to follow up asked conditional on selecting 12 month repayment all times available)
why_always_6m (same but for 6 months)
why_always_flex (same but selected flexibility always)
why_never_flex (same but never selected flexibility)
why_same_position
dce_difficulty (asked to respondent - difficulty 1-5 likert scale used to partition respondents who did not respond and also an outcome to test differential focus rates)
dce_difficulty_moderator (same but asked to enumerator)
choice_confidence (asked to participant 1-5 likert scale, again identifying their focus and impact of dce timing through inattention and also used to evaluate more overprecise respondents)
months_electricity_savings_40k (1-12 indicating number of months participant thinks solar panel savings will exceed 40,000 riyals)
solar_fail_1yr (probability a loan will fail in one year)
solar_fail_3yr (same but 3 years)
prob_default_*consequences (a set of 9 probability questions asking likelihood of * consequence between 0 and 100)
debt_attitude (view of installments as equivalent to loans)
debt_aversion (stated aversion to loans)
expected_rate_* (predictions of interest rates, through bank savings, informal lenders, etc)
expected_inflation_* (predictions of inflation/deflation rates from -100 to 500% over the next year) - (outside of the survey we tie these to real market outcomes that we observe with high frequency in each market)
quality_diff (perceived difference in quality of panels sold by banks and merchants)
failure_rate_bank (failure rate conditional on purchase from bank)
failure_rate_merchant (failure rate conditional on merchant)

This study employs a 3x3 between-subjects factorial design that independently varies two dimensions of constraint salience in a discrete choice experiment (DCE) on solar panel financing in Yemen. Participants are randomly assigned to one of nine cells defined by the cross-tabulation of DCE timing and constraint-framing conditions, using SurveyCTO's random() function.
DCE Structure
Each participant sees 13 choice sets from one of three pre-assigned blocks (39 unique scenarios total). Each choice set presents four alternatives: a cash purchase from a bank, a cash purchase from a merchant, a loan from a bank, and a loan from a merchant. All alternatives vary in total cost (300,000-700,000 YER), and loan options additionally vary in duration (6 or 12 months), flexibility (none or skip one payment), and documentation requirements (simple or full).
Experimental Dimensions
Dimension 1: DCE Timing. The first dimension varies when in the survey the DCE appears, thereby varying the extent to which participants have reflected on their financial situation before making financing choices. The DCE appears either (1) early in the survey, after only demographics and basic solar questions; (2) after a detailed household expenditure module; or (3) after both the expenditure module and a financial assessment module covering income, debt experience, attitudes toward borrowing, and perceived consequences of loan default.
Dimension 2: Constraint Framing. The second dimension varies the instructions given to participants immediately before the DCE. Participants are either (1) instructed to consider their current financial situation when choosing; (2) instructed to imagine they face no financial constraints and to choose based on genuine preference; or (3) given no framing instruction.

Not available

Randomization of which households are surveyed was done using STATA. All treatment randomization is conducted through SurveyCTO after a survey is launched. The randomization within survey occurs with equal probability of all treatment effects and a random selection of one of three sets of 13 DCE questions to be asked.

Individual

Yes

Individual-level clusters 1154-1936 (solely conditional on funding) (no clusters for individual analyses but cluster design for choice-level analysis in the DCE)

1,154 individual participants × 13 DCE choices each = 16,614 choice observations. Participants are drawn from three governorates in south Yemen: Aden, Hadramout, and Taiz using a stratified sampling method (equal numbers in each governorate, equal numbers in each major district, then population-weighted sampling within each district). More participants will be included contingent on funding application approvals up to a maximum of 1,936 participants.

128-216 individuals per treatment arm (each with corresponding 13 DCE choices)

Based on the anticipated minimum of 1,154 individuals (individual-level randomization), we estimate (using a limited pilot study) that the MDE for direct effects on selection at the binary choice level is <0.5 percentage point likelihood of selection, indicating substantive power for both main and interaction effects in family 3. Individual-level outcomes are noisier in the pilot, with MDEs that may be as high as 4 pp for direct effects and 8 pp for interactions. However, the questions have changed substantially from the pilot; therefore, a revised power calculation will be included before analysis, once a partial sample is drawn. Initial pilot results (where irrelevant attributes are incorporated and where the comparisons are less informative) indicate individual-level MDE = 0.054 (5.4 percentage points) at N = 1,154, corresponding to Cohen’s d = 0.20. Unit: proportion (0–1). Standard deviation: 0.271. The study has 80% power at the 5% significance level to detect a 5.4 pp difference in the proportion of choice sets in which a respondent chose a loan alternative, in a pairwise comparison between any two of the three timing arms (N/3 = 385 per arm). The pilot timing effect (10.9 pp, d = 0.41) exceeds this threshold by a factor of two. • MDE = 0.042 (4.2 percentage points) at N = 1,936, corresponding to Cohen’s d = 0.16. With expanded sample: N/3 = 645 per arm. For the Timing × Framing interaction (9-cell, confirmatory): MDE ≈ 0.108 at N = 1,154 (d = 0.40), improving to 0.084 at N = 1,936 (d = 0.31). This is the most demanding test and is marginally powered at the lower sample size. The multiple-hypothesis test adjustments will further reduce this. For choice-level attribute effects (Family 3): the study is very well powered (t > 10 for all attributes except flexibility in the pilot). The power-relevant test is the treatment-arm difference in AMCEs, for which MDE ≈ 0.044 (4.4 pp) at N = 1,154, which is adequate for detecting whether constraint salience meaningfully alters attribute selections and implied preferences.