Food Delivery Analytics: Testing Whether Late Orders Cause Customer Churn

You are a product manager at a food delivery app. The team believes that late deliveries cause customers to churn. What data would you look at to establish whether that hypothesis is true? Explain how you would define the key terms, which data you would pull, how you would tell a causal effect apart from a correlation, and what findings would convince you either way. ```hint Why else might they leave Customers who received a late order may differ from other customers in ways that have nothing to do with lateness. List those differences, then think about which comparison holds them fixed. ``` ### Clarifying Questions - The app has no subscription to cancel, so how should churn be defined: no order within a set number of days, or a lasting drop in how often someone orders? - What counts as late: arriving after the time promised at checkout, by how many minutes, and does an estimate that kept slipping during the order count? - What decision will the answer inform: investing in faster delivery, changing how delivery times are promised, or compensating customers after late orders? - Do we have order-level records of promised and actual delivery times, joined to customers, ratings, support contacts and refunds, over a long enough history? ### What a Strong Answer Covers - Precise definitions of churn, lateness and the time windows used - The specific datasets needed and how they join at the order and customer level - An analysis design that controls for confounders such as order frequency, delivery area, time of day, restaurant and outside events, ideally using comparisons where lateness is close to random - Dose-response and segment cuts: how late, how often, how recently, and for which customers - A clear rule for what evidence supports or refutes the hypothesis, and how the result feeds the decision ### Follow-up Questions - The first cut shows that customers with more late deliveries churn less than everyone else. How can that be, and what do you check next? - Could you run an experiment to measure the effect without deliberately making orders late? What would you randomize? - If lateness does drive churn, how would you choose between making deliveries faster and promising more conservative delivery times?