How Complexity Shapes Dynamic Purchasing Behavior: The Case of Buy Now, Pay Later

Participants take part in an incentivized online decision-making experiment in which they repeatedly decide whether to not purchase at all, purchase goods outright, or purchase goods using a Buy Now, Pay Later (BNPL) arrangement.

Each participant completes a sequence of dynamic purchasing environments that differ in the granularity of installments and the number of available products. In each environment, participants make repeated purchase decisions over an indefinite horizon with stochastic termination, accumulating product value that partially determines their final earnings.

The order of environments is randomized across participants. One randomly selected environment is paid out for real at the end of the study.

The study is an incentivized online laboratory experiment implemented on Prolific. Participants complete a sequence of dynamic purchasing environments in which they decide in each period whether to (N) not purchase, (O) purchase a product outright, or (B) purchase a product using a Buy Now, Pay Later (BNPL) arrangement.

In each environment, participants manage a wealth account that determines feasibility of purchases and repayment obligations. The level of the wealth account is partially determined by stochastic income shocks that arrive each period. Products yield immediate utility and accumulate in a separate product account that determines final earnings but cannot be used to repay outstanding obligations. Bankruptcy occurs if the wealth account becomes negative, in which case the environment terminates and the participant cannot accumulate further goods.

Environments differ along two dimensions:
(i) the installment granularity of BNPL contracts, with installment lengths 1, 4, or 6, where 1 corresponds to the absence of BNPL;
(ii) the number and composition of available products (either 1, 2, or 3 products are available in a given environment)

The decision problem has an indefinite horizon implemented via a geometric stopping rule with constant per-period termination probability. Participants are informed that each period may be the last with a given probability.

All participants complete all environments. The first environment is fixed across participants as a common baseline. The remaining environments are presented in participant-specific randomized sequences drawn from a pre-generated pool of balanced permutations designed to equalize position effects and transition frequencies.

After completing all environments, one environment is randomly selected for payment. Participants are paid a fixed participation fee plus a performance-based bonus proportional to the total product value accumulated in the selected environment.

In separate parts of the study, we collect background information and survey measures, including demographics, financial experience and familiarity with BNPL products, self-reported financial literacy, and cognitive reflection measures. These data are collected prior to or after the experimental task and are not experimentally manipulated.

2026-01-11

2026-01-19

BNPL take-up share is computed at the participant-by-environment level as the number of periods in which the participant selects the BNPL option divided by the total number of periods in which a purchase decision is feasible.

Bankruptcy incidence is defined as an indicator equal to one if the participant's wealth account becomes negative at any point within an environment and zero otherwise.

Misoptimization rate is computed at the participant-by-environment level as the share of decision periods in which the participant's observed action differs from the action prescribed by the environment-specific optimal benchmark policy, evaluated at the realized state in that period.

All primary outcomes are constructed from participants' state-action histories within each environment.

BNPL take-up share is computed for each participant-environment pair as the number of decision periods in which the participant selects the BNPL option divided by the total number of periods in which a purchase decision is feasible.

Bankruptcy incidence is constructed as an indicator equal to one if the participant's wealth account becomes negative at any point within the environment and zero otherwise.

Misoptimization rate is constructed for each participant-environment pair by comparing the participant's observed action in each decision period to the action prescribed by the environment-specific optimal benchmark policy evaluated at the realized state in that period, and computing the share of periods in which the two differ.

1. BNPL underuse rate: Fraction of decision periods in which the participant does not use BNPL when BNPL is optimal under the benchmark policy.
2. BNPL overuse rate: Fraction of decision periods in which the participant uses BNPL when outright purchase or no purchase is optimal under the benchmark policy.
3. Outright purchase error rate: Fraction of decision periods in which the participant purchases outright when another action is optimal under the benchmark policy.

4. Total product value accumulated within an environment.
5. Survival duration: Number of decision periods completed before stochastic termination or bankruptcy.
6. Self-reported cognitive burden and task difficulty ratings.
7. Environment-level BNPL take-up rate and bankruptcy rate aggregated across participants.
8. Environment-level misoptimization rate aggregated across participants.

9. Demographic characteristics (age, gender, education, income).
10. BNPL familiarity and prior usage history.
11. BNPL vignette.
12. Financial literacy.
13. Debt attitudes.
14. Risk and time preferences.
15. Time preferences.
16. Planning and executive function as measured by the "Tower of London" task.
17. Prolific platform metadata (e.g., completion time, device type, prior study participation).

1.-3. BNPL underuse, BNPL overuse, and outright purchase error rates are constructed from participants' state-action histories by comparing observed actions in each decision period to the action prescribed by the environment-specific optimal benchmark policy evaluated at the realized state. Each error type is computed as a share of decision periods within a participant-environment pair.

4. Total product value is defined as the sum of product values accumulated in the product account within an environment.
5. Survival duration is defined as the number of decision periods completed before (stochastic termination or) bankruptcy.
6. Cognitive burden and task difficulty are measured using post-environment self-report scales.
7.-8. Environment-level BNPL take-up, bankruptcy, and misoptimization rates are computed by aggregating the corresponding participant-level outcomes within each environment.

9. Demographic characteristics (age, gender, education, income) are collected using a background survey administered outside of the experimental task.
10. BNPL familiarity and prior usage history are collected using a background questionnaire.
11. The BNPL vignette consists of a short hypothetical scenario designed to elicit participants' stated beliefs and attitudes toward BNPL.
12. Financial literacy is measured using a standardized financial literacy questionnaire.
13. Debt attitudes are measured using a self-report survey scale and incentivized task.
14.-15. Risk and time preferences are elicited using incentivized choice tasks administered outside of the main experimental decision task (strictly concave budget restrictions).
16. Planning and executive function are measured using the Tower of London task.
17. Prolific platform metadata (e.g., completion time, device type, prior study participation) are recorded automatically by the platform.

All background and diagnostic measures are not experimentally manipulated and are used for descriptive analysis, heterogeneity analysis, and mechanism validation.

The study is an incentivized online laboratory experiment in which participants complete a sequence of dynamic purchasing environments. In each environment, participants repeatedly decide whether not to purchase, to purchase a product outright, or purchase using a BNPL arrangement.

Environments differ in the structure of BNPL contracts and the number of available products. The decision problem has an indefinite horizon implemented via a stochastic stopping rule.

All participants complete all environments. The order of environments is randomized across participants. One environment is randomly selected for payment at the end of the study.

Apart from the main decision tasks described above, the study elicits various background characteristics used for descriptive analysis, heterogeneity analysis, and mechanism validation.

The study is an incentivized online laboratory experiment implemented on Prolific. Participants complete a sequence of dynamic purchasing environments in which they repeatedly decide whether to (N) not purchase, (O) purchase a product outright, or (B) purchase a product using a Buy Now, Pay Later arrangement.

Each environment consists of a sequence of decision periods. In every period, participants receive a stochastic income shock and update a wealth account that determines purchase feasibility and repayment obligations. Product purchases generate immediate utility and accumulate in a separate product account that determines final earnings but cannot be used to repay outstanding obligations. An environment ends if the participant's wealth becomes negative (bankruptcy).

Environments vary systematically along two dimensions: (i) installment structure of BNPL contracts, and (ii) the number and composition of available products. The decision horizon is indefinite and implemented via a geometric stopping rule with a constant per-period termination probability. Participants are informed that each period may be the last but are not told the expected duration.

All participants complete all environments. The first environment is identical for all participants and serves as a common baseline. The remaining environments are presented in participant-specific sequences drawn from a pre-generated pool of randomized permutations constructed to balance environment positions and transition frequencies. Participants are randomly assigned to one of these sequences at the start of the study.

At the end of the study, one environment is randomly selected for payment. Participants receive a fixed participation fee and a performance-based bonus proportional to the total product value accumulated in the randomly selected environment.

Randomization is implemented by computer prior to data collection. A pool of environment sequences is generated using randomized permutations subject to balance constraints on environment positions and transition frequencies.

The unit of randomization is the individual participant. All participants complete all environments, but the order of environments is randomized across participants. Treatments are assigned at the individual participant level. There is no group-level or cluster-level randomization.

No

Same as number of participants: 150–175 participants (individual-level randomization; no clustering).

150-175 participants (individual-level randomization; no clustering). We gather around 3,600 to 4,200 participant-environment observations (150-175 participants x 24 environments per participant). However, in the main analyses, we will use 1,800 to 2,100 observations from pre-determined core environments. The remaining 1,800 to 2,100 observations will be used for auxiliary analyses (in isolation and pooled).

This is a within-subject experimental design. All participants complete all environments. There are no between-subject treatment arms. Variation is generated by within-participant exposure to multiple environments with randomized order.

Power calculations are conducted for the primary outcomes (BNPL take-up share, bankruptcy incidence, and misoptimization rate) and are based on the number of participants. Under a conservative variance bound of 0.25 and standard normal approximations (two-sided test, significance level 0.05, 80% power), a sample size of N=150 (N=175) participants yields a minimum detectable effect size (MDE) of approximately 11.4 (10.6) percentage points for tests against a benchmark proportion (e.g., whether an outcome exceeds a pre-specified threshold). For comparisons between two environments, the corresponding MDE is approximately 16.2 (15.0) percentage points. These MDEs are conservative upper bounds given the within-subject design.