Olist E-commerce Causal Analysis

Project Overview

This project investigates the causal effect of delayed delivery on customer satisfaction ratings in e-commerce, using data provided by Olist, the largest department store in Brazilian marketplaces.

Research Question

Our two research questions are:

1. What is the causal effect of delayed delivery on customer ratings?

2. What is the causal effect of delayed delivery on revenue?

Methodology

Our analysis employed two complementary causal inference approaches: Propensity Score Matching and Graphical Causal Models

1. Propensity Score Matching

Based on the DAG insights, we use propensity score matching to eliminate any variation that may come from our matching variables that may bias our causal effect. Once the variation has been removed and our matching variables are balanced, we are able to effectively estimate the causal effect of late deliveries on customer ratings.

We could define propensity score as the conditional probability of the observation being treated given the matching variables

Propensity Score = P(T=1| Matching Variables)

Approach:

The first step is to identify our matching variables—this is the identification step. Once these variables are determined, we match observations from the treatment group, where deliveries were late, to those in the control group, where deliveries were on time. By accurately estimating the Average Treatment Effect on the Treated (ATT), we can infer what customer ratings would have been if the deliveries had not been late for the treatment group.

Causal Estimand Identification

Based on the DAG, the causal effect is identified through the following estimand:

P(Rating|is_delivery_late,distance_km,season,Product_category_encoded,freight_value,Product_size)

realized estimand

Rating~is_delivery_late+distance_km+season+Product_category_encoded+freight_value+Product_size

Matching Variables:

• Distance (km)
• Season
• Product category
• Freight value
• Product size

Balance:

Achieved comparable covariate distributions between groups

Results (ATT):

-1.8 stars. On Average a late delivery causes customer ratings to drop by 1.8 stars.

Notebook:

📊 Propensity Score Matching

2. Graphical Causal Modeling

In Graphical Causal Models the goal is to model the data generating process based on what's observed from the DAG we designed. We achieve this by assigning a causal mechanism to each node in the graph.

Here is an example

the causal mechanisms are defined by functional causal models (FCMs) for non-root nodes and stochastic models for root nodes.

A causal mechanism describes how the values of a node are influenced by the values of its parent nodes.

• Notebook:

📊 Graphical Causal Model

We created a more in detailed report about the analysis that can be found here:

Final Report

Tools and Technologies

Python, DoWhy

Getting Started

Prerequisites

• Python 3.7 or higher
• pip (Python package installer)
• Git

Cloning the Repository

1. Open a terminal or command prompt.

2. Clone the repository using the following command:

   git clone https://github.com/yourusername/causal-attribution-analysis.git
   

   Replace yourusername with the actual username or organization name where the repository is hosted.

3. Navigate to the project directory:

   cd causal-attribution-analysis
   

Setting up the Virtual Environment

1. Create a virtual environment:

   • On Windows:

     py -3.9 -m venv venv
     

   • On macOS and Linux:

     python3 -m venv venv
     

2. Activate the virtual environment:

   • On Windows:

     venv\Scripts\activate
     

   • On macOS and Linux:

     source venv/bin/activate
     

Installing Dependencies

1. Ensure your virtual environment is activated (you should see (venv) at the beginning of your command prompt).

2. Install the required packages using the requirements.txt file:

   pip install -r requirements.txt
   

This will install all necessary dependencies for the project.

Deactivating the Virtual Environment

When you're done working on the project, you can deactivate the virtual environment by running:

deactivate

Contributors

• Juan Herrera
• Denisse Wohlstein
• Matteo Mennini
• Anna Natasha
• Avantika Gargya