DoorDash · Statistics & Data Analysis

Diagnose and experiment to reduce late deliveries

TrueInterview

October 7, 2026 · 3 min read

You are the first analyst assigned to a two-sided delivery platform that carries both restaurant meals and parcels. The share of deliveries arriving after the promised time has increased. Respond to the prompts below, and be precise about the datasets, metrics, identification strategy, and design decisions you would use.

Quantify business impact: Specify the primary outcome metrics and guardrail metrics you would monitor, such as lateness rate, minutes late, cancellation rate, refunds per order, reorder or retention, merchant and driver churn, and driver idle time. Propose a simple causal model linking each additional minute of lateness to the probability that the customer reorders and to refunds. Which data fields do you require, and from which sources—orders, merchant preparation data, courier app pings, maps or traffic, weather, and support tickets? Define the exact calculation windows, for example a 7-day rolling lateness rate and 28-day retention measured after an index order.
Find root causes and decide what to measure: List at least five falsifiable causes of lateness, such as underestimated preparation times, batching or stacking, a shortage of driver supply, routing or traffic spikes, slow merchant handoff, and cold- or hot-chain constraints. For each cause, specify the exact slice or feature you would examine, for example $prep\_time\_error = actual\_prep - quoted\_prep$ , $supply\_ratio = active\_drivers / active\_requests$ computed in 5-minute buckets, route complexity, distance, weather severity, and merchant handoff. State the exact queryable fields required and how you would join tables to compute those features.
Prioritize causes: Propose a scoring framework such as $Impact \times Effort \times Confidence$ and outline an analysis that ranks causes with statistical controls, for example a fixed-effects regression by merchant and hour-of-week that controls for distance and weather and compares partial effects on minutes late. How would you avoid confounding and simultaneity, for instance by instrumenting with exogenous weather or road closures, excluding holidays, or using difference-in-differences when supply shocks occur?
Explain the modality gap: The data indicate that food deliveries take longer than parcel deliveries even when distance is held constant. Provide at least three competing hypotheses that could account for this, such as prep-time variance, temperature-sensitive packaging, merchant handoff queues, and rejection rates. For each hypothesis, describe a concrete test and the criterion for accepting or rejecting it; for example, if handoff is the driver, the $pickup\_wait\_time$ distribution for food should first-order stochastically dominate parcel after controlling for distance and hour-of-week, and you could run a quantile regression on pickup wait with merchant fixed effects.
Design an experiment to reduce lateness: Choose one intervention, such as calibrating prep-time estimates, dispatching drivers earlier, disabling stacking for long-distance food orders, or giving hot-bag orders priority. Specify the experimental unit—order, merchant, or driver-day—the randomization scheme and stratification variables such as city, hour-of-week, distance band, and merchant SLA; the power inputs and minimum detectable effect for a baseline lateness rate of 12%, showing the formula you would use rather than describing it in words; the primary metric and guardrails such as minutes late, cancellation rate, driver utilization, and NPS; the ramp plan; how you would mitigate novelty and learning effects; and how you would handle interference, for example by clustering by merchant or zone to reduce spillover or using geo-randomization. Finally, define the exact decision rule for shipping or rolling back, for example a $95\%$ confidence interval improvement of at least $1.0$ percentage points with no guardrail regressions beyond the threshold.

Overview: This question assesses a data scientist's skills in causal inference, metrics engineering, data instrumentation and table joins, root-cause analysis, and experiment design applied to operational delivery performance.

DoorDash · Statistics & Data Analysis

Diagnose and experiment to reduce late deliveries

TrueInterview

October 7, 2026 · 3 min read

Quantify business impact: Specify the primary outcome metrics and guardrail metrics you would monitor, such as lateness rate, minutes late, cancellation rate, refunds per order, reorder or retention, merchant and driver churn, and driver idle time. Propose a simple causal model linking each additional minute of lateness to the probability that the customer reorders and to refunds. Which data fields do you require, and from which sources—orders, merchant preparation data, courier app pings, maps or traffic, weather, and support tickets? Define the exact calculation windows, for example a 7-day rolling lateness rate and 28-day retention measured after an index order.
Find root causes and decide what to measure: List at least five falsifiable causes of lateness, such as underestimated preparation times, batching or stacking, a shortage of driver supply, routing or traffic spikes, slow merchant handoff, and cold- or hot-chain constraints. For each cause, specify the exact slice or feature you would examine, for example $prep\_time\_error = actual\_prep - quoted\_prep$ , $supply\_ratio = active\_drivers / active\_requests$ computed in 5-minute buckets, route complexity, distance, weather severity, and merchant handoff. State the exact queryable fields required and how you would join tables to compute those features.
Prioritize causes: Propose a scoring framework such as $Impact \times Effort \times Confidence$ and outline an analysis that ranks causes with statistical controls, for example a fixed-effects regression by merchant and hour-of-week that controls for distance and weather and compares partial effects on minutes late. How would you avoid confounding and simultaneity, for instance by instrumenting with exogenous weather or road closures, excluding holidays, or using difference-in-differences when supply shocks occur?
Explain the modality gap: The data indicate that food deliveries take longer than parcel deliveries even when distance is held constant. Provide at least three competing hypotheses that could account for this, such as prep-time variance, temperature-sensitive packaging, merchant handoff queues, and rejection rates. For each hypothesis, describe a concrete test and the criterion for accepting or rejecting it; for example, if handoff is the driver, the $pickup\_wait\_time$ distribution for food should first-order stochastically dominate parcel after controlling for distance and hour-of-week, and you could run a quantile regression on pickup wait with merchant fixed effects.
Design an experiment to reduce lateness: Choose one intervention, such as calibrating prep-time estimates, dispatching drivers earlier, disabling stacking for long-distance food orders, or giving hot-bag orders priority. Specify the experimental unit—order, merchant, or driver-day—the randomization scheme and stratification variables such as city, hour-of-week, distance band, and merchant SLA; the power inputs and minimum detectable effect for a baseline lateness rate of 12%, showing the formula you would use rather than describing it in words; the primary metric and guardrails such as minutes late, cancellation rate, driver utilization, and NPS; the ramp plan; how you would mitigate novelty and learning effects; and how you would handle interference, for example by clustering by merchant or zone to reduce spillover or using geo-randomization. Finally, define the exact decision rule for shipping or rolling back, for example a $95\%$ confidence interval improvement of at least $1.0$ percentage points with no guardrail regressions beyond the threshold.