A discrete choice experiment on a digital savings product

Context

A randomised controlled trial answers a definitive evaluative question: "Did this specific version work?" However, it cannot cost-effectively resolve the prior design question: "Which of forty potential product configurations should we build and test?"

A digital savings product delivered over mobile money involves numerous design parameters: monthly account fees, withdrawal liquidity, physical walking distance to the nearest cash-out agent, reminder modalities, and maturity bonuses. Financial service providers and public programmes can realistically afford to field-test only two or three variants in an empirical trial.

A discrete choice experiment (DCE) bridges this methodological gap. By asking respondents to make repeated choices between hypothetical product bundles that vary systematically across key attributes, a DCE identifies the marginal utility of each feature and reveals the trade-offs (marginal rates of substitution) users are willing to make. While standard in health, transport, and environmental economics (see the ISPOR good-practice report on experimental design and Train, Discrete Choice Methods with Simulation, 2009), the method remains underutilised in financial inclusion, where upfront product optimisation is precisely the policy decision at hand.

Methodological Challenges in Low-Literacy Contexts

Applying choice experiments to digital financial services in emerging markets introduces three operational constraints:

1. Literacy barriers: Text-dense choice cards quickly degrade into reading comprehension tests. Every attribute must be depicted using an intuitive, standardised icon accompanied by a scripted oral explanation, while cognitive load must be managed by limiting the number of choice tasks per respondent.
2. Preference heterogeneity around commitment devices: Locked savings features help some individuals shield funds from social and familial claims, but deter households requiring immediate liquidity to buffer emergency shocks. In the Philippines, an exploratory commitment product was adopted by only a minority of clients, though it substantially increased net balances among that subset (Ashraf, Karlan, & Yin, Quarterly Journal of Economics, 2006). An econometric model that estimates only the population-average preference for a lock will obscure this fundamental market segmentation.
3. Hypothetical bias: Stated preferences captured in a DCE measure relative valuations under structured conditions. They identify which product architectures warrant field experimentation, but do not serve as unadjusted accounting forecasts of market uptake.

Attributes and Levels

Attribute	Experimental levels	Econometric coding
Monthly fee	0, 100, 200, 400 LCU	Continuous linear (per 100 LCU)
Withdrawal rule	Unrestricted access; penalty on early withdrawal ("soft lock"); locked until target date ("strict lock")	Indicator variables (reference: unrestricted access)
Walk to nearest cash-out agent	5, 20, 45 minutes	Continuous linear (per 10-minute increment)
Savings reminder channel	None; SMS text message; interactive voice response (IVR)	Indicator variables (reference: no reminder)
Maturity bonus	0%, 3%, 6% of terminal balance	Continuous linear (per percentage point)

Each choice set presents two unbranded product profiles alongside an explicit opt-out alternative ("keep my current way of saving"). Omitting an opt-out forces choice, violates random utility assumptions, and leads to distorted market-share projections.

Experimental Design: Efficiency, Balance, and Blocking

The underlying design is generated algorithmically in Python (NumPy) and documented in the accompanying notebook. For a multinomial logit model with parameter priors $\tilde\beta$, the Fisher information matrix is expressed as:

$I(\tilde\beta) = \sum_s \sum_j P_{sj}\,(x_{sj} - \bar x_s)(x_{sj} - \bar x_s)^\top,\qquad \bar x_s = \sum_j P_{sj}\,x_{sj}$

The overall D-error is computed as $\det\big(I(\tilde\beta)^{-1}\big)^{1/K}$ across $K = 8$ model parameters. Priors are weakly informative and sign-constrained (reflecting the expected disutility of higher fees and longer travel times), consistent with best practice prior to pilot data collection.

Two design algorithms were evaluated:

• An unconstrained coordinate-exchange algorithm minimises the raw D-error (0.178), but collapses the design onto the extreme boundary levels of fees, distances, and bonuses. This structure is statistically efficient only under a strict assumption of linear marginal effects.
• The adopted design applies an attribute-level swapping algorithm (Huber & Zwerina, Journal of Marketing Research, 1996), preserving intermediate attribute levels across alternatives. It achieves a robust D-error of 0.205 (compared to a median of 0.267 across random level-balanced alternatives), delivering a relative D-efficiency of 1.30.

Strictly dominated candidate profiles were systematically purged. The resulting 18 choice tasks are partitioned into three balanced blocks of six tasks, capping respondent burden at six pairwise decisions.

Visual Choice Cards for Low-Literacy Populations

The interactive demo reproduces the visual protocol used by field enumerators:

• Stacked coins representing monthly fees;
• An open, half-closed, or padlocked icon denoting withdrawal liquidity rules;
• A walking figure beside an analogue clock face illustrating travel time to the nearest agent;
• A mobile device displaying either lines of text or an audio speaker to distinguish SMS from voice prompts;
• A wrapped gift box indicating the target-date bonus.

In field deployments, these visual cards are validated during a qualitative feasibility round to confirm that respondents spontaneously understand each graphic. Enumerators read all choice cards aloud in the respondent’s local language. A preliminary warm-up task containing an objectively dominant option screens out participants who fail to engage with the choice protocol.

Econometric Estimation on Simulated Cohorts

The synthetic sample comprises 600 respondents (50% categorised as low-literacy) completing their assigned block of six tasks. Preference structures were calibrated to benchmark estimators against known data-generating parameters: homogeneous disutility for fees and travel distance; normally distributed unobserved heterogeneity for commitment rules ($\text{strict lock} \sim N(-0.30, 1.0^2)$); and divergent valuations for reminder modalities conditioned on literacy status.

Conditional Logit Model. Choice probabilities follow the standard formulation $P_{nsj} = \exp(x_{sj}^\top\beta) / \sum_l \exp(x_{sl}^\top\beta)$. The log-likelihood function is maximised using analytic gradients in SciPy, with cluster-robust sandwich standard errors grouped at the respondent level to account for repeated individual observations. Interacting reminder modalities with literacy status shows that the coefficient for voice reminders rises from 0.35 among literate respondents to 0.56 among low-literacy respondents, while the SMS coefficient falls from 0.28 to 0.07. A likelihood-ratio test confirms the joint significance of these interactions ($\chi^2(2) = 10.9$).

Mixed (Random-Parameters) Logit Model. To accommodate unobserved preference heterogeneity without imposing the restrictive Independence of Irrelevant Alternatives (IIA) property, coefficients on commitment rules are specified as random normal distributions: $\beta^{\text{lock}}_n \sim N(\mu, \sigma^2)$. The multidimensional integral is approximated via simulated maximum likelihood using 200 scrambled Halton draws per individual.

The mixed logit recovers a strict-lock mean of $-0.22$ ($\text{SE} = 0.09$) and a standard deviation of $0.97$ ($\text{SE} = 0.10$), closely matching the data-generating parameters ($-0.30$ and $1.0$). This delivers a crucial strategic insight: approximately 41% of respondents attach a positive valuation to a strict lock, despite a negative population-average taste parameter. The standard conditional logit model—reporting an average coefficient of $-0.02$—entirely conceals this viable target segment.

Willingness to Pay (WTP)

Monetary willingness to pay for attribute $k$ is calculated as the ratio of marginal utilities: $\text{WTP}_k = -100 \times (\beta_k / \beta_{\text{fee}})$ LCU per month. 95% confidence intervals are estimated using both the delta method and Krinsky–Robb parametric bootstrapping (10,000 draws; Krinsky & Robb, Review of Economics and Statistics, 1986):

Evaluated attribute	Mean WTP (LCU/month)	Delta method [95% CI]	Krinsky–Robb [95% CI]	Benchmark value
Maturity bonus (per percentage point)	39	[32, 47]	[32, 47]	43
Agent proximity (per 10-min additional walk)	−30	[−39, −22]	[−39, −22]	−34
Voice reminder (low-literacy subgroup)	169	[121, 218]	[122, 219]	171
SMS reminder (low-literacy subgroup)	20	[−28, 67]	[−28, 67]	14
Voice reminder (literate subgroup)	107	[53, 160]	[55, 163]	57

These estimates indicate substantial, policy-relevant trade-offs: for a low-literacy saver, receiving a voice prompt carries the same perceived value as a four-percentage-point increase in financial return, whereas the value of an SMS notification is statistically indistinguishable from zero.

To confirm finite-sample properties, the simulation pipeline was replicated across 100 independent draws: empirical confidence intervals contain the true parameter value in 92% to 99% of replications across continuous and discrete attributes.

Market-Share Simulation

The interactive simulator allows practitioners to specify competing product configurations and predict market capture relative to the baseline opt-out, either across literacy sub-segments or the pooled market.

Under default baseline assumptions (Product A: 100 LCU fee, unrestricted access, 20-min walk, SMS, no bonus; Product B: 200 LCU fee, strict lock, 5-min walk, voice prompt, 6% bonus):

• The conditional logit predicts 15.1% share for Product A, 35.6% for Product B, and 49.3% opt-out;
• The mixed logit predicts 15.0% for Product A, 34.5% for Product B, and 50.5% opt-out.

Discrepancies between the two models widen when configurations diverge on commitment rules—the domain where preference dispersion is highest and where the proportional substitution assumptions of the standard multinomial logit are least defensible.

Methodological Limitations

1. Synthetic data calibration: These numerical outputs demonstrate algorithmic validity; they do not represent real-world market distributions.
2. Hypothetical bias: Stated preferences routinely overstate real-world uptake, especially for fee-bearing services. The market simulator should be interpreted as an ordinal ranking of product concepts rather than an accounting forecast of customer acquisition.
3. Price sensitivity assumptions: The marginal utility of money is specified as constant; in real populations, price sensitivity varies with income, requiring models that accommodate random or income-segmented price coefficients.
4. Functional forms: Linear specifications for continuous variables (fees, travel times, bonuses) must be validated against piece-wise dummy specifications on empirical field data.

Strategic Implementation Sequence

For international donors and programme steering committees, embedding a DCE into the product design lifecycle follows an efficient, six-stage sequence:

1. Qualitative Discovery: Identify critical pain points, attribute lists, and realistic levels
2. Feasibility Piloting: Audit pictorial comprehension, literacy barriers, and task complexity
3. Experimental Design: Generate D-efficient candidate arrays using weakly informative priors
4. Choice Survey: Administer visual choice sets across a representative sample of clients
5. Advanced Econometrics: Estimate mixed logit models to uncover latent preference segments
6. Powered Field Trial: Deploy randomised trials (RCTs) exclusively on the 2–3 leading designs

This de-risking framework ensures that the substantial financial capital required for large-scale field evaluations is committed only to product architectures with documented, evidence-based user demand.