Problem: Delivery Cost per Order Increased — Design an Analysis Plan

Context

A restaurant reports a significant increase in delivery cost per order in the last month. Delivery cost is strictly defined as costs incurred from courier pickup at the restaurant to drop-off at the customer. It includes courier fees, platform fees, and delivery-related refunds/adjustments. It explicitly excludes food ingredients, kitchen labor, packaging materials, dine-in costs, and any non-delivery promotions.

You have the following datasets:

• orders(order_id, order_ts, customer_id, store_id, distance_km, promised_eta_min, actual_eta_min, platform, city_zone, courier_id, delivered_bool, cancel_reason)
• courier_payments(courier_id, order_id, pay_amount, surge_multiplier)
• fees(order_id, platform_fee, refund_amount, promo_amount)
• weather_by_zone(date, city_zone, precip_mm, temp_c, wind_kph)
• store_ops(store_id, open_hours_json, staffing_level, kitchen_prep_time_avg_min)

Assume “last month” refers to the most recent full calendar month in the store’s local timezone. The goal is to separate real operational changes from measurement artifacts and recommend data-driven interventions.

Tasks

1. Validate the metric definition and confirm the increase is not due to scope creep or denominator changes.
2. Decompose delivery cost per order by drivers:
   • Mix: platform, distance bands, zones, time-of-day.
   • Rate: pay per km/min, surge.
   • Execution: cancellations, reattempts, SLA breaches.
3. Identify plausible causal hypotheses (e.g., weather shocks, staffing shifts that increase pickup wait, platform policy changes) and quantify each using appropriate methods (e.g., difference-in-differences, fixed-effects regression, event studies).
4. Propose at least two actionable experiments to reduce cost (e.g., batching, zone remapping, pickup-wait time SLA) with success metrics, guardrails, power calculations, and expected bias sources.
5. List common pitfalls specific to this case (e.g., inadvertently including packaging/food costs, counting tips as costs, double-counting refunded orders, survivorship bias from excluding undelivered orders) and how you will avoid them.