🌐 English · Español

Credit Choice Experiment — Discrete Choice Modeling

A discrete-choice analysis of how the visual salience of credit terms in digital advertising affects consumer choices over loan offers. Participants in a controlled experiment were randomly assigned to four ad-design conditions and asked to choose between credit alternatives across multiple scenarios.

The analysis combines descriptive EDA, conditional logit (mlogit), mixed logit with unobserved heterogeneity, and a machine-learning model comparison (CART, SVM, KNN, Random Forest) to ask both why people choose what they choose and how well we can predict it.

→ Read the full rendered report — every plot, table, and model output, no R installation required.

---

Research Question

Does the way credit information is visually displayed in digital ads change which offers consumers choose? And does it change the consistency of those choices relative to a baseline where the same offers were shown without advertising material?

The motivating policy question: should regulators force credit advertisers to display key financial information (APR, total cost, installment value) in a more prominent, standardized way?

---

Experimental Design

Participants were randomly assigned to one of four conditions:

Group	Description
GC (Control)	Typical layout used today in the Chilean credit market
T1	Total cost of credit (CTC) is visually emphasized
T2	Both total cost (CTC) and installment value (VC) are emphasized
T3	All financial information is emphasized vs. other visual elements

Each participant made 6 binary choices across three macro-scenarios (TV purchase / vacation / unexpected medical bill). Their choices in this wave (wave 2) are compared to an earlier wave (wave 1) where the same offers were presented as plain financial tables — providing a clean within-subject consistency measure.

---

Dataset

The analysis Rmd reads from data/sample.csv — a fully synthetic sample of ~265 rows (≈132 binary choices, 22 simulated participants) generated to:

• Respect the original schema column-by-column
• Match the marginal distributions of condicion_experimental and macroescenario
• Reproduce the basic mechanic that participants tend to pick the alternative with lower total cost (CTC)

The original 8,000-row dataset is not redistributed because it was provided under the terms of an academic course. See data/README.md for the full schema.

---

Methodology

1. EDA

• Demographics (gender, financial literacy, monthly liquidity)
• Choice rates across the four offer attributes
• Consistency vs. wave 1 by experimental group
• Comprehension errors (CTC / VC / APR recall) by experimental group

2. Discrete Choice Models (mlogit)

Model	Type	Notes
modelp1	Conditional logit	Baseline: only CAE + CantCuotas as alternative-specific attributes
modelp2	Conditional logit + interactions	Adds participant-level interactions: mujer × T3, lit_fin × T3, estima_bien × T3, scenario fixed effects
modelp3	Mixed logit	Same spec + random coefficients with correlation, capturing unobserved heterogeneity across participants

3. ML Comparison (caret)

Predictive comparison on a held-out test split:

• CART (rpart)
• SVM linear (svmLinear)
• KNN (knn)
• Random Forest (rf)

Models compared via resamples() on accuracy and Kappa.

---

Key Findings

1. Offer attributes matter most. Across all models the strongest predictors of choice are the APR (CAE) and the number of installments — visual treatment is a second-order effect compared to the underlying financials.
2. Conditional logits (modelp1, modelp2) found no significant T3 effect on choice. Standard discrete-choice analysis suggests that emphasizing financial information does not meaningfully change which offer people pick.
3. The mixed logit (modelp3) tells a different story. Once we allow for unobserved heterogeneity across participants, T3 does show a significant effect on the probability of choosing the left alternative, suggesting that simpler logits were masking real but heterogeneous responses.
4. No interaction with demographics is significant. Being a woman, having higher financial literacy, or correctly estimating costs does not significantly modulate the T3 effect — at least not at this sample size.
5. Policy implication: the evidence is inconclusive but suggests a T3-style minimum standard is defensible — it guarantees all users see the relevant information at the same visual weight, even if the average behavioral effect is small.
6. Prediction: Random Forest gave the best test-set accuracy among the ML models, but logit / mixed-logit remain preferable when the goal is to interpret the role of each variable.

---

Tech Stack

R mlogit caret rpart kernlab randomForest ggplot2 dplyr kableExtra

---

Project Structure

credit-choice-experiment/
├── README.md
├── analysis.Rmd                    # Full analysis (EDA + mlogit + ML)
├── data/
│   ├── README.md                   # Schema reference
│   └── sample.csv                  # Synthetic 265-row sample
└── .github/workflows/
    └── render.yml                  # CI: render Rmd → HTML → Pages

---

How to Reproduce

Option A — Read the rendered report (no install)

Just open https://jsanchez-ds.github.io/credit-choice-experiment/. The CI rebuilds the report on every push to main.

Option B — Run locally

install.packages(c(
  "rmarkdown", "knitr", "tidyverse", "dplyr", "readr", "ggplot2",
  "gridExtra", "ggcorrplot", "psych", "mlogit", "fixest", "glmnet",
  "caret", "kernlab", "earth", "randomForest", "kableExtra",
  "stargazer", "modelsummary", "gt"
))

rmarkdown::render("analysis.Rmd", output_dir = "docs", output_file = "index.html")

The Rmd reads data/sample.csv from the repo, so no data download is needed.

---

Context

Originally developed for IN5162 — Marketing Engineering, taught by Prof. Marcel Goic at the Universidad de Chile (Industrial Engineering, 2023). Reframed and rewritten in English for portfolio presentation.

---

Author

Jonathan Sánchez

• GitHub: @jsanchez-ds
• Universidad de Chile — Industrial Engineering