Behavioral Preferences and Smallholder Demand for Digital Traceability Certification in Vietnam

The study administers a stated-preference choice experiment concerning participation in digital traceability certification schemes, a class of institutional arrangement in which producers maintain a smartphone-based production record covering planting dates, input applications, pest and disease treatments and harvest dates, together with photographic verification at specified points, and which downstream buyers and certifying bodies may inspect, and in exchange for which producers obtain either a price premium or access to otherwise closed market segments. Respondents are told explicitly that data entry and photography are self-administered and that no automated monitoring of their plots takes place, a clarification introduced in pilot cognitive interviewing because unprompted respondents tended to construe "digital traceability" as remote surveillance. This misreading would have contaminated the price-premium coefficient with an unmeasured privacy disamenity.
Each hypothetical scheme is characterised by four attributes, each at three levels. Certification type distinguishes OCOP three-star status, which confers access to provincial and regional markets; VietGAP, which opens sales to domestic firms paying above-market prices; and GlobalGAP, which permits export but imposes the most exacting compliance burden. The price premium is expressed as a percentage increase over the price the respondent currently obtains, rather than in absolute currency units, so that the attribute is commensurate across producers of heterogeneous commodities. Practice requirements are specified ordinally as record-keeping alone; restriction of selected inputs; and restriction of inputs combined with alteration of cultivation technique and on-farm inspection by an extension officer. The consequence of failing verification ranges from a warning under which certification is retained, through temporary withdrawal with re-registration permitted at the following assessment cycle, to withdrawal followed by a full re-assessment that may be initiated only after the thirty-six-month cycle has elapsed.
Respondents complete eight choice cards, of which the first two serve as unincentivised practice and are discarded; comprehension of the practice cards is recorded separately so that the analysis may condition on it. Each of the six recorded cards presents two experimental profiles alongside a labelled status quo describing continuation of the respondent's present uncertified practice. Assignment to one of four blocks of six cards, the sequence in which those cards are presented, and the allocation of the two experimental profiles to the left and right positions are randomised independently of one another.
The interview additionally comprises four incentivised decision tasks, administered under a random lottery mechanism disclosed before any decision is taken, which recover risk aversion, ambiguity tolerance, loss aversion and time preference; and a ten-item timed battery of smartphone tasks measuring digital capability under enumerator observation rather than self-report. The behavioural tasks are placed at the end of the instrument so that the risk and loss frames they invoke cannot prime responses to the choice experiment, which is positioned early precisely to insulate the primary outcome from respondent fatigue over a one-hundred-and-five-minute protocol.

2026-08-07

2026-09-30

1. The alternative selected on each recorded choice card, that is, experimental profile A, experimental profile B, or the labelled status quo. This yields 3,000 choice occasions and 9,000 alternative-level observations.
2. Willingness to accept, denominated in percentage price premium, for each non-monetary attribute level: certification type, practice-requirement level, and verification-failure consequence.
3. The Digital Integration Index (DII), measuring the depth of a household's transition to digital channels.
4. The Digital Market Position Index (DMPI), measuring the strength of a household's bargaining position within those channels.

Rationale for the latent-variable treatment
The organising methodological commitment of this study is that elicited behavioural parameters ought not to be entered into a choice model as though measured without error. Single-shot incentivised tasks recover preferences with non-trivial noise arising from comprehension failure, choice inconsistency, and the coarseness of the response grid, which in the risk task admits only eleven points at increments of 20,000 VND. Where such a parameter is interacted with an attribute, classical measurement error in the interacted regressor attenuates the interaction coefficient towards zero, and the attenuation is more severe than in the corresponding main-effect case. A literature that enters raw elicited parameters into conditional logits and reports null interactions is therefore not well positioned to distinguish absence of preference heterogeneity from failure to measure it.
The integrated choice and latent variable framework of Ben-Akiva et al. (2002), developed for ordered attitudinal indicators by Daly et al. (2012) and appraised critically by Vij and Walker (2016), addresses this by treating the preference as a latent construct with multiple imperfect indicators and estimating the measurement and choice components jointly. A second and substantive motivation applies to institutional trust in particular: trust in certifying bodies is multidimensional and no single incentivised task measures it, so a latent treatment is not merely a correction but the only available route to testing hypotheses in which trust appears.

Estimation of the choice outcome
The specification places three latent constructs, namely risk tolerance, ambiguity tolerance and institutional trust, in the deterministic utility component, each interacted with the attribute vector. Each construct is measured by five ordered Likert indicators; risk and ambiguity are measured additionally by their incentivised task outcomes entering as continuous indicators, the CRRA value and the ambiguity difference respectively. The first loading of each construct is fixed to unity and structural variances normalised to one for scale identification. Loss aversion enters instead as an observed interaction with the verification-failure attribute. This asymmetry is deliberate rather than inadvertent: λ is elicited with real money in the loss domain, which is a stronger measurement than any Likert battery could supply, and latent treatment would discard precisely the advantage that incentivised elicitation confers.
Estimation is by simulated maximum likelihood with 500 scrambled Halton draws, with convergence assessed by re-estimation from dispersed starting values and by increasing the draw count to 1,000 to confirm parameter stability. Because the ICLV is demanding on data and may fail to converge with six tasks per respondent, the estimation sequence is pre-registered so that the study's conclusions do not rest on a model that may not estimate: a conditional logit with observed interactions is designated the primary specification, is always estimable, and will be reported irrespective of what follows.

Willingness to accept
WTA for a non-monetary attribute level is the negative ratio of that attribute's coefficient, evaluated at specified values of the latent constructs, to the price-premium coefficient. It is reported at the sample mean of each construct and at plus and minus one standard deviation, with confidence intervals computed both by the delta method and by Krinsky–Robb simulation, the two being reported jointly because the delta method performs poorly for ratios when the denominator coefficient is imprecisely estimated.

Construction of the DII
The DII is constructed from Module H and is intended to capture the share of a household's economic activity mediated by digital channels, on both the output and input sides, together with informational and financial dimensions. Its components are: the quantity of the primary product sold under a price or order agreed by telephone, Zalo, application or online, as a share of total quantity sold; the quantity settled by bank transfer or electronic wallet, as a share of total quantity sold; the value of inputs ordered or arranged digitally, as a share of total input value; the count, across the five information domains of price, inputs, pest and disease, policy and credit, of domains whose principal source is digital, where telephone, social-messaging and application or website sources are classified as digital and face-to-face, broadcast and commune-loudspeaker sources are not; an indicator for maintaining production records in any application or digital system; and the count of digital financial transactions in the preceding thirty days.

Construction of the DMPI
The DMPI is administered only to households reporting positive digital sales, since bargaining position within a channel is undefined for those who do not transact in it; the subsample restriction and its trigger threshold are pre-registered, and selection into the subsample is modelled explicitly rather than ignored. Components, drawn from Module I, are: the number of distinct buyers reached digitally over twelve months, as a measure of outside options; price realisation, computed as the price obtained on the most recent digital sale divided by the commune median price for that commodity in that week, taken from district market monitoring records; the extent of price information held before agreement, on a five-point scale running from no knowledge of prices elsewhere to careful comparison of several sources; the interval in days between delivery and payment, as a measure of who bears working-capital cost; and an indicator for credit or input financing arranged through an application, platform or digital lender.

Outstanding specification. The aggregation rule for both indices, whether equal weighting, standardisation and averaging, or polychoric principal components after the fashion of the asset index, is not fixed in the present version of the research design and must be settled before registration, since an index whose weighting is chosen after the data are seen is not a pre-registered outcome in any meaningful sense.

1. The status-quo constant, interpreted as systematic attachment to current uncertified practice.
2. The elicited behavioural parameters treated as outcomes in their own right: CRRA risk aversion, ambiguity tolerance, the loss-aversion coefficient, the impatience measure, and the dynamic-consistency classification.
3. The digital capability score obtained from the timed smartphone battery.
4. The frequency of corner responses by elicitation task.
5. Where the deferred-payment arm is implemented, the comparison of hypothetical and incentivised time-preference measures.

Module G departs from the prevailing practice of measuring digital capability by self-report, which is contaminated by social desirability and by respondents' limited insight into their own competence. Ten tasks are administered on the enumerator's device under observation and timing: five basic tasks spanning device operation, message retrieval, photographic capture, map navigation and messaging, and five higher-order tasks spanning information search, form completion, unit-price comparison, identification of a probable fraudulent message, and calendar interrogation. Completion and elapsed seconds are recorded for each; an item not completed within 120 seconds is terminated and coded as incomplete. The basic score is the unweighted sum over ten items; the weighted composite assigns weight two to the higher-order items, yielding a range of zero to fifteen. Enumerators are instructed not to coach, and the instruction is enforced through back-checks, since coaching would render the measure a test of enumerator patience rather than respondent capability.
Classification proceeds from the direct comparison of allocations to the earlier date across the two decisions: a smaller earlier-date allocation in Decision 2 constitutes decreasing impatience and is coded present-biased; equality is coded time-consistent; a larger allocation is coded future-biased. The impatience measure is the average share of the endowment placed on the earlier date, bounded in the unit interval by construction and analysed by fractional regression wherever it enters as a dependent variable. The published Vietnamese benchmark reports 21.9 percent present-biased, 21.2 percent time-consistent and 56.9 percent future-biased, a distribution against which the present sample can be assessed.
All four tasks generate corner responses at non-trivial rates, and their treatment materially affects the recovered parameters, so the handling rules are pre-registered in full. A zero allocation in the risk task implies only that risk aversion exceeds the threshold rationalising the smallest positive investment, and is therefore an interval observation on (2.41, ∞); full investment is an interval observation on (−∞, 0]. Acceptance of all six mixed gambles implies λ below 1.00 and rejection of all six implies λ above 3.33, again interval observations. The primary specification treats these as interval-censored and estimates the measurement equation by interval regression, thereby exploiting the information in the corner without imputing a point value it does not contain. A second specification imputes the conventional point values of ρ = 3.0 and 0, and of λ = 0.67 and 3.67, and estimates by ordinary least squares, principally to preserve comparability with the published study. A third excludes corner respondents, reported so that the reader may judge how much the corners matter. The ambiguity difference remains well defined when one component is at a corner provided the corners differ; a respondent at the same corner in both tasks contributes no information on ambiguity, is coded ambiguity-neutral with a flag, and a specification excluding such cases is reported.
Corner frequencies are reported by task in every paper, on the view that the share of respondents at a corner is itself diagnostic of stake calibration. The study commits in advance to the following: if more than sixty percent of respondents are at a corner in any task, the stakes for that task were poorly calibrated, and this will be stated plainly rather than obscured by selective reporting of the specifications that happen to survive.

Approximately 500 smallholder households distributed across six communes of the Northern Midland and Red River Delta regions of northern Vietnam complete a single interview of approximately one hundred and five minutes. The instrument comprises modules on identification and consent; demographic and individual controls; farm, asset and income controls; the discrete choice experiment; attitudinal batteries supplying the ordered indicators of the three latent constructs; a retrospective module on childhood adversity supplying the instruments; a timed digital capability battery; the digital integration and digital market position components; an information-network roster collecting up to five alters across five information domains; and the four incentivised behavioural tasks.
The choice experiment employs a D-efficient design of twenty-four choice sets partitioned into four blocks of six, with three levels on each of four attributes. Each respondent is randomly assigned to a single block and completes two unrecorded practice cards followed by the six recorded cards of that block; the order of cards within the block and the assignment of the two experimental profiles to the left and right positions are randomised independently. Each recorded card offers two experimental profiles alongside a labelled status quo. The resulting data comprise 3,000 choice occasions and 9,000 alternative-level observations.
Module sequencing is itself a design parameter rather than an administrative convenience. The choice experiment is positioned early so that the primary outcome is elicited before fatigue accumulates over a long protocol, and the incentivised tasks are positioned last so that the risk and loss frames they necessarily invoke cannot prime responses to the choice experiment or to the attitudinal batteries.
A pilot of fifty respondents precedes principal fieldwork, administering a smaller orthogonal choice design whose estimates furnish priors for re-optimisation of the main design. The attitudinal batteries are revised before principal data collection should internal consistency fall below the pre-registered threshold, a contingency fixed in advance so that revision cannot be mistaken for opportunistic adjustment.

Not available

Randomisation is conducted in the office by computer in advance of fieldwork. Respondent identifiers drawn from commune household listings are assigned to one of four choice-experiment blocks by a seeded pseudo-random number generator, stratified by commune to guarantee balance of blocks across sites; the block is pre-printed on the questionnaire at item A6 before the enumerator reaches the household. The order of cards within block and the allocation of profiles to the A and B positions are generated by the same seeded procedure and likewise pre-printed. The enumerator therefore exercises no randomisation discretion in the field, which forecloses the principal channel through which stated-preference designs are compromised in practice. Seeds and assignment files are archived for replication.

Selection for payment proceeds by public lottery in the field: on completion of all interviews in a commune, one participant in ten is drawn, and the task to be played for real is determined by a draw or coin flip conducted in the respondent's presence, with the outcome recorded at items K5 and K6 and payment made within forty-eight hours against a signed receipt.

The individual respondent, corresponding to the farming household, is the unit of randomisation. Block assignment, card ordering and profile position are assigned at the individual level, stratified by commune. Commune enters the design as a stratum and as a fixed effect, and serves as the level at which standard errors are clustered for inference, but is not itself a unit of treatment assignment.

A second randomisation operates at a higher level of aggregation: whether the time-preference task is administered hypothetically or with deferred payment by post-office voucher is determined at the site level, recorded at item K4f, and implemented only if the pilot establishes that voucher delivery is workable.

No

Six communes function as strata for assignment and as the level of clustering for inference. Randomisation is conducted at the individual level within these strata; communes are not units of treatment assignment.

500 farming households, generating 3,000 choice occasions at six recorded cards apiece and 9,000 alternative-level observations at three alternatives per card. A further 50 households are interviewed in the pilot; pilot observations furnish design priors and are not pooled with the principal sample.

500 farmers in total, allocated at random across the four blocks of the choice design: approximately 125 farmers to Block 1, 125 to Block 2, 125 to Block 3, and 125 to Block 4. There is no untreated arm in the conventional sense, since every respondent completes six recorded choice cards; the labelled status quo alternative present on every card discharges the function of a no-participation option within each choice occasion, and the status quo constant recovers the systematic component of attachment to it.

Minimum detectable effects are computed at 80 percent power and a 5 percent significance level, with commune-clustered standard errors. For a choice-experiment attribute main effect, the MDE is 0.09 standardised units, obtained by the simulation procedure of Rose and Bliemer (2013) under 500 respondents, six recorded tasks and three alternatives per task, with priors drawn from Nguyen-Anh, Nguyen-Van and To-The (2026). For a latent-construct interaction in the choice equation, the MDE is 0.16 standardised units. The distinction between this figure and the preceding one warrants comment, since it drives the design. Interactions between attributes and latent constructs are identified overwhelmingly from between-person variation in those constructs, which by construction do not vary within a respondent. The binding constraint on power for the interaction tests is therefore the number of respondents, not the number of tasks each completes. This is the reason six recorded tasks per respondent suffice for the hypotheses the study is designed to test, while remaining marginal for unrestricted random-coefficient specifications. That same constraint is why mixed logit is placed among the robustness exercises rather than in the primary sequence, and why the study advances no claim to recover a full random-coefficient distribution. For the Digital Integration and Digital Market Position indices as dependent variables, the MDE for main effects is 0.15 standard deviations under 500 farmers and six commune fixed effects. Power for the DMPI is lower than this figure implies in practice, since the index is defined only on the subsample reporting positive digital sales; the realised subsample share is not known ex ante and will be reported.