Two-Sided Delivery Platform: Rising Late Deliveries

You are the first analyst on a two‑sided delivery platform that handles both food and parcel orders. Late deliveries (arriving after the promised time) have risen. Answer the following, being explicit about datasets, metrics, identification, and design choices.

1. Quantify business impact

• Define the core outcome metrics and guardrails you will track (e.g., lateness rate, minutes‑late, cancel rate, refunds per order, reorder/retention, merchant/driver churn, driver idle time).
• Propose a simple causal model that translates an extra minute of lateness into customer reorder probability and refunds.
• What data fields do you need and from where (orders, merchant prep, courier app pings, maps/traffic, weather, support tickets)?
• Specify the exact calculation windows (e.g., compute 7‑day rolling lateness rate and 28‑day retention following an index order).

2. Find root causes and decide what to measure

• List at least five falsifiable causes of lateness (e.g., underestimated prep times, batching/stacking, driver supply shortage, routing/traffic spikes, long merchant handoff, cold/hot chain constraints).
• For each cause, specify the precise slice or feature you would inspect (e.g., prep_time_error = actual_prep − quoted_prep; supply_ratio = active_drivers/active_requests by 5‑min buckets; route_complexity; distance; weather_severity; merchant_hand_off).
• State the exact queryable fields needed and how you’d join tables to compute them.

3. Prioritize causes

• Propose a scoring framework (Impact × Effort × Confidence) and outline an analysis to rank causes with statistical control (e.g., fixed‑effects regression by merchant and hour‑of‑week controlling for distance and weather; compare partial effects on minutes‑late).
• How will you avoid confounding and simultaneity (e.g., instrument with exogenous weather/road closures; hold out holidays; difference‑in‑differences when supply shocks occur)?

4. Explain modality gap

• Data suggests food deliveries take longer than parcels, holding distance constant. Provide at least three competing hypotheses that could explain this (e.g., prep‑time variance, temperature‑sensitive packaging, merchant handoff queues, rejection rates).
• For each, describe a concrete test and the acceptance/rejection criterion (e.g., if handoff is the driver, the pickup_wait_time distribution for food first‑order stochastically dominates parcel after controlling for distance and hour‑of‑week; run quantile regression on pickup wait with merchant fixed effects).

5. Design an experiment to reduce lateness

• Choose one intervention (e.g., calibrating prep‑time estimates, earlier driver dispatch, disabling stacking for long‑distance food orders, hot‑bag priority).
• Specify: experimental unit (order vs merchant vs driver‑day), randomization scheme and stratification variables (city, hour‑of‑week, distance band, merchant SLA), power inputs and MDE for a baseline lateness rate of 12% (show the formula you’d use, not just words), primary metric and guardrails (e.g., minutes‑late, cancel rate, driver utilization, NPS), ramp plan, novelty and learning effects mitigation, and interference handling (e.g., cluster by merchant or zone to reduce spillover; geo‑randomization).
• Finally, define the exact decision rule to ship or roll back (e.g., 95% CI improvement ≥ 1.0 pp with no guardrail regressions beyond threshold).