Causal Impact Analysis: Competitor Entry Effects on Rossmann Store Sales

Project Overview

This project uses Propensity Score Matching (PSM), a causal inference technique, to isolate the true impact of competitor entry on Rossmann drugstore sales. Instead of simple before/after comparisons, we rigorously control for confounding factors to answer: "Did competitor entry actually hurt sales, or would sales have changed anyway?"

---

Problem Statement

Organizations struggle to understand the true impact of market changes (competitor entry, policy changes, etc.) versus natural market trends. A store's sales might drop, but was it the competitor's fault, or just seasonal decline? This project tackles that challenge.

Business Question: How much did competitor entry impact Rossmann store sales?

---

Methodology

Technique: Propensity Score Matching (PSM)

Why PSM?

• Handles selection bias: Stores that got competitors might be naturally different from stores without competitors
• Works with observational data: No randomized experiment needed
• Creates comparable groups: Matches similar stores (one with competitor, one without)
• Isolates causal effect: Compares matched pairs to measure true impact

How It Works

1. Calculate Propensity Scores

   • Build a logistic regression model to estimate: "Given store characteristics, what's the probability it gets a competitor?"
   • Output: Score 0-1 for each store

2. Match Treated & Control Stores

   • Match each treated store (got competitor) with a control store (no competitor) that has a similar propensity score
   • Create 357 matched pairs (limited by number of control stores)

3. Calculate Treatment Effect

   • Compare sales: Treated Store Sales - Control Store Sales
   • Average Treatment Effect (ATE) = mean difference across all pairs
   • Confidence interval tells us if effect is statistically significant

4. Validate & Check Robustness

   • Verify matching quality (are matched stores actually similar?)
   • Calculate 95% confidence interval
   • Assess statistical significance

---

Key Findings

Metric	Value
Average Treatment Effect (ATE)	+60.92 sales
Standard Deviation	2,162.02
95% Confidence Interval	[-164.11, +285.96]
Statistical Significance	NOT SIGNIFICANT

Interpretation

Competitor entry had NO statistically significant effect on Rossmann sales.

• ATE = +60.92 (small, positive)
• CI includes zero → effect could be negative, positive, or zero
• We cannot confidently say competitor hurt or helped
• With 95% confidence, true effect is between -164 and +286 sales per store

---

Data & Features

Dataset

• Rossmann Store Sales (1,017,209 transactions, 1,115 stores)
• Time Period: 2013-2015
• Key Variables:
  • Sales: Daily store revenue
  • CompetitionOpenSinceYear/Month: When competitor entered
  • HasCompetition: Treatment indicator (1 = competitor, 0 = no competitor)
  • Promo, StateHoliday, SchoolHoliday: Confounders to control

Confounders Controlled

• Promotions (boost sales independently)
• State/School Holidays (affect shopping behavior)
• Day of Week (weekends differ from weekdays)
• Store Type (different formats sell differently)

---

Visualizations

Plot 1: Propensity Score Distribution

Shows treated vs control store propensity scores. Good overlap validates matching approach.

Plot 2: Matched Pairs Scatter

Compares treated vs control sales for each pair. Points near diagonal line = matching worked; spread shows variability in effects.

Plot 3: Treatment Effect Distribution

Histogram of individual effects with confidence interval. Centered near zero, includes zero → no significant effect.

---

Limitations & Caveats

1. Limited Time Window (2013-2015)

   • Data only covers 3 years; competitors entered gradually
   • No clear "pre-competitor" vs "post-competitor" periods
   • Limits causal identification

2. Unconfoundedness Assumption

   • Assumes all confounders measured (promotion, holidays, store type)
   • Unmeasured confounders (store reputation, manager quality) could bias results
   • Cannot be tested; only sensitivity analysis available

3. High Variability

   • Standard deviation (2,162) much larger than ATE (60.92)
   • Large differences between individual stores
   • Effect is imprecisely estimated

4. Selection Bias

   • Stores that got competitors might be in growing markets (confounding)
   • Even with matching, baseline differences may remain

---

Interpretation & Business Implications

What This Means:

• Competitor entry did NOT significantly hurt Rossmann sales in this analysis
• Natural variation and confounders explain most of the change observed
• Any impact is too small or uncertain to detect with this data

Why This Matters:

• Suggests competitors less threatening than feared, OR
• Data quality/timeframe insufficient to detect true effect, OR
• Other factors (price, location, loyalty) matter more than competitor presence

---

Code Structure

Impact-Analysis/
├── data/
│   ├── train.csv           
│   └── store.csv           
├── analysis.ipynb         
├── README.md           
└── plots/
    ├── psm_distribution.html
    ├── matched_pairs.html
    └── effect_distribution.html

---

How to Run

1. Install dependencies:

   pip install pandas numpy scikit-learn scipy plotly

2. Load data:

   import pandas as pd
   df = pd.read_csv('train.csv').merge(pd.read_csv('store.csv'), on='Store')

3. Run analysis:

   • Open analysis.ipynb in Jupyter
   • Execute cells sequentially
   • View interactive plots

---

Key Learnings: Causal Inference Concepts

Confounding: Other factors change alongside treatment. We controlled for promotions, holidays, store type to isolate competitor effect.

Propensity Score: Probability of treatment based on pre-treatment characteristics. Used to match comparable stores.

Matching: Paired treated (competitor) and control (no competitor) stores with similar propensity scores, making comparison fair.

Statistical Significance: 95% confidence interval includes zero → effect not significantly different from no effect.