The COO Wants a No-Show Model by Quarter End
A 12-site orthopedic group schedules about 210,000 appointments per year. Overall no-show rate is 18.4% against a specialty benchmark of roughly 11%. The COO has asked your team for 'a predictive model that flags patients likely to no-show so the front desk can double-book them.' Constraints: the central scheduling center has 2 FTE; your team has 0.5 FTE of analyst capacity and a $60,000 budget; you have three years of scheduling history. A reason code is recorded for only 41% of no-shows, and scheduling templates differ across the 12 sites. One site already runs at 9.7% no-shows and uses 7-day reminder calls plus a self-pay deposit policy. Physicians at two sites have publicly objected to overbooking because of past wait-time complaints.
1. Rewrite this request as a business problem statement with a measurable objective and an explicit decision rule.
Read the worked answer
The request as stated presumes both the solution (a model) and the intervention (double-booking). Neither is established. Reframe in three parts:
1. What decision changes? The schedulers' real decision is how to allocate a fixed number of appointment slots across patients and how much reminder effort to spend. Overbooking is only one of at least four levers: reminder intensity, lead-time limits, deposit requirements, and overbooking. The model must attach to one of these.
2. What is the outcome metric? No-show rate alone is the wrong target - you can drive it to zero by overbooking until patients wait 90 minutes. Use slot utilization (completed visits ÷ available slots) as the primary metric, with door-to-provider wait time and provider idle minutes as balancing measures, and revenue per clinic day as the financial readout.
3. Decision rule. Example: 'For appointments with predicted no-show probability above X, the site may overbook up to 1 slot per provider-half-day.' Specify X and the cap up front, because that is where physician trust is won or lost.
Baseline first: the 9.7% site is a natural experiment. Before any modeling, produce no-show rates by site, lead time, day of week, payer, and visit type. If lead time alone explains most of the variation, a scheduling rule beats a model on cost and maintainability.
2. How do you decide whether to build the predictive model or recommend process changes first, given the budget and staffing constraints?
Read the worked answer
Treat it as a portfolio decision with an explicit ordering test.
Step 1 - quantify the size of the prize. 18.4% of 210,000 appointments is roughly 38,600 lost slots per year. Multiply by contribution margin per visit to get the annualized opportunity. If the number is, say, $4M, then even a modest, cheap intervention is worth more than an expensive one with a longer runway.
Step 2 - check the cost curve of each lever. Reminder calls cost labor (2 FTE can reach perhaps 40–60 calls per day - a fraction of volume); lead-time caps and deposits cost policy friction; overbooking costs patient wait time. A model feeding a double-booking rule only pays off after you know no-show probability is highly variable and predictable within your own data.
Step 3 - the data sufficiency test. With 41% of reasons unrecorded and templates varying by site, a model trained across sites will mostly learn site and template effects. Before committing, run a lead-time × prior-no-show × payer baseline. If a simple 5-variable logistic model captures nearly all the lift of a gradient-boosted model in cross-validation, deploy the simple one - you have 0.5 FTE to maintain it.
Recommended sequencing: a 90-day rules-and-reminders pilot at 2–3 sites (including the physician-objecting sites, with a wait-time guarantee) as the data collection window, then build the model only for the residual, genuinely unpredictable portion.
3. The two objecting physician groups must be brought on board. How do you structure the pilot and the success criteria?
Read the worked answer
Objection to overbooking is rational - it externalizes cost onto the waiting room. Address it with design, not persuasion.
Randomize at the provider-half-day level, not by site, so both arms experience the same staff, season, and case mix. This also gives a defensible effect size rather than a before/after comparison.
Cap the intervention. Limit overbooking to one slot per half-day, only for appointments created within 14 days of the visit (short lead time reduces the wait-time risk), and only when predicted risk exceeds the threshold. Publish the rule so clinicians can see it is not discretionary.
Pre-register the success criteria: primary = slot utilization; non-inferiority guardrail = door-to-provider wait time no worse than +5 minutes and provider idle time reduced; secondary = patient satisfaction item on access. Define in advance what result would stop the pilot.
Report back to the physicians in their own terms - idle minutes recovered, end-of-day overruns avoided, and any wait-time change shown with confidence intervals rather than a raw average. Give the two sites a mid-pilot review at 6 weeks with authority to halt. Front-line ownership is the difference between a pilot that produces evidence and one that produces grievances.