Deterministic vs. Stochastic Systems

The Problem with Planning for Averages

The most expensive analytical error in healthcare operations is not using the wrong model. It is using a deterministic model where a stochastic one is required — substituting a single expected value for a distribution of possible outcomes, and then acting surprised when reality deviates from plan.

A deterministic system is one in which the outputs are fully determined by the inputs and initial conditions. Given the same starting state and the same inputs, a deterministic system produces the same result every time. A factory that processes exactly one widget every four minutes, with parts arriving exactly every five minutes, is deterministic. You can calculate its throughput, its idle time, and its queue length with arithmetic.

A stochastic system is one in which at least some inputs, processing times, or state transitions are governed by probability distributions rather than fixed values. The outputs vary across runs even when the starting conditions are nominally identical. Nearly every healthcare delivery process is stochastic: patient arrivals are random, consultation durations vary, lab turnaround fluctuates, no-show rates shift by day of week and weather, discharge decisions depend on clinical judgment applied to variable patient trajectories.

The distinction matters because the behavior of stochastic systems cannot be correctly predicted by plugging average values into deterministic formulas. This is not a minor approximation error. It is a structural analytical failure that produces systematically wrong answers — answers that are always optimistic about capacity, always surprised by queues, and always behind on cost.

The Mathematical Mechanism: Why Averages Lie

The formal reason that average-based planning fails in stochastic systems is Jensen’s inequality: for any convex function f and random variable X, the expected value of f(X) is greater than or equal to f applied to the expected value of X. In notation: E[f(X)] >= f(E[X]).

Why does this matter operationally? Because the functions that govern healthcare system behavior — wait time as a function of utilization, cost as a function of demand, overtime as a function of patient volume — are convex. They curve upward. And when your input is variable (stochastic) rather than fixed (deterministic), the average outcome is always worse than the outcome you would predict from the average input.

Sam Savage named this the Flaw of Averages: plans based on average assumptions are wrong on average, and they are wrong in a specific direction — they underestimate cost, wait time, queue length, and resource consumption. This is not a statistical curiosity. It is the mechanism behind chronic understaffing, perpetual budget overruns, and the persistent gap between capacity plans and lived operational reality.

The Kingman Approximation Makes This Concrete

The Kingman formula (also called the VUT equation, popularized by Hopp and Spearman in Factory Physics) gives an approximation for expected waiting time in a single-server queue:

W_q ≈ (c_a² + c_s²) / 2 × (u / (1 - u)) × t_s

Where:

c_a² = squared coefficient of variation of inter-arrival times (arrival variability)
c_s² = squared coefficient of variation of service times (process variability)
u = server utilization (arrival rate / service rate)
t_s = mean service time

Three things are multiplied together: a variability factor (V), a utilization factor (U), and average processing time (T). The critical insight is the utilization factor: u/(1-u). At 50% utilization, this equals 1. At 80%, it equals 4. At 90%, it equals 9. At 95%, it equals 19. The relationship is a hyperbola — wait time does not degrade linearly as utilization rises. It explodes.

Now here is where the deterministic-stochastic distinction becomes consequential. In a deterministic system (c_a² = 0, c_s² = 0), the variability factor is zero. Wait time is zero regardless of utilization, up to 100%. A perfectly regular system with no variability can run at full capacity with no queue at all. This is the implicit mental model behind most healthcare staffing plans: if average demand is 40 patients per day and average capacity is 45 patients per day, we have enough. Deterministic reasoning says there is a 12.5% margin.

But real healthcare systems have variability. Typical coefficients of variation for ED arrivals run 0.8 to 1.2; for consultation times, 0.5 to 1.5 depending on case mix. Plug c_a² = 1.0 and c_s² = 1.0 into the Kingman formula at 89% utilization (40/45), and you get expected waiting time of approximately 8 times the mean service time. If average consultation is 20 minutes, expected wait is 160 minutes. The deterministic model predicted zero wait. The stochastic model predicts nearly three hours.

That is not a rounding error. That is the distance between a plan that looks adequate on a spreadsheet and an operational reality that produces patient complaints, staff burnout, and left-without-being-seen rates.

The Variability Buffering Law

Hopp and Spearman formalized a result they call the Variability Buffering Law: variability in a system will be buffered by some combination of capacity, inventory, and time. There is no fourth option. You pay for variability whether you plan for it or not — the only question is which currency.

In manufacturing, inventory is a buffer. In healthcare, there is no inventory of pre-treated patients. The buffers available are:

Capacity — staff, beds, rooms, equipment held in reserve beyond average demand
Time — patients wait (queues grow, access degrades, patients abandon)

When a healthcare system plans deterministically — staffing to average demand, budgeting to average cost — it has chosen to buffer variability with time. It has chosen, implicitly, to make patients wait. The Variability Buffering Law says this is not a management failure or a bad day. It is a mathematical certainty.

The operator’s choice is not whether to buffer variability but how. Capacity buffers cost money upfront (extra staff, unfilled appointment slots held in reserve). Time buffers cost money downstream (patient dissatisfaction, abandonment, downstream care delays, readmissions, staff overtime to catch up after surges). Deterministic planning hides this tradeoff by assuming variability away.

Healthcare Example: Behavioral Health Intake at a Rural FQHC

Consider a Federally Qualified Health Center in eastern Washington with a single behavioral health provider doing intake assessments. The center sees an average of 6 behavioral health intake requests per week. Each intake takes an average of 75 minutes. The provider has 8 hours per day, 5 days per week, yielding 32 available 75-minute intake slots per month — roughly 8 per week.

Deterministic analysis: Average demand is 6 per week. Average capacity is 8 per week. Utilization is 75%. The system appears to have a 25% capacity margin. No access problem.

Stochastic reality: Intake requests arrive according to a Poisson-like process (coefficient of variation approximately 1.0). Intake durations vary by patient acuity — some take 45 minutes, complex cases take 120 minutes (coefficient of variation approximately 0.6). Applying the Kingman approximation at 75% utilization with these variability parameters, expected wait for an available intake slot is roughly 1.5 times the mean service time — about 112 minutes of scheduling delay, translating to approximately 1-2 weeks of calendar wait given the weekly slot structure.

Now add a realistic complication: the behavioral health provider also covers crisis consultations, which are stochastic interruptions that further increase service time variability. During weeks when two or more crisis events occur, effective intake capacity drops to 5-6 slots. Utilization spikes above 90% for those weeks. The Kingman formula at 90% utilization predicts wait times approximately 6 times longer than at 75%. Calendar waits extend to 3-5 weeks — well past the window where a patient in behavioral health crisis remains engaged with care.

The FQHC’s grant report shows average utilization of 75% and average wait time of 10 days. Both numbers are accurate and both are misleading. The average obscures the bimodal reality: some weeks run smoothly, other weeks produce multi-week backlogs that drive patient abandonment. The patients who abandon — who never complete intake — never appear in the wait time statistics. This is survivorship bias compounding the flaw of averages: the reported average wait is calculated only from patients who successfully navigated the queue, not from all who entered it.

Warning Signs: You Are Planning Deterministically in a Stochastic System

Your staffing model uses average demand without a variability term. If the spreadsheet has one row for “expected patients per day” and derives staffing from that single number, it is deterministic.
Your capacity plan shows utilization targets above 85% and assumes this is sustainable. In any system with meaningful variability, sustained utilization above 85% produces nonlinear queue growth (per the Kingman formula). If your plan shows 85-90% target utilization and no explicit wait-time projection, it has assumed variability away.
Your budget uses point estimates for cost drivers. A grant budget that projects $4,200 average cost per patient served, multiplied by expected patient count, with no variance band or contingency derived from the cost distribution, is a deterministic plan in a stochastic environment.
Reported performance metrics are exclusively averages. Average wait time, average length of stay, average cost per encounter — without accompanying percentile data (P75, P90, P95) — mask the tail behavior where system failures concentrate.
“Good weeks” and “bad weeks” are attributed to luck or one-off events rather than to the expected variance of the system. If every bad week has a narrative explanation but no one has calculated whether the frequency of bad weeks is statistically expected given the system’s variability parameters, the organization is narrating around stochastic behavior rather than engineering for it.

Intervention Levers

Reduce variability at the source. Smoothing demand (e.g., carve-out scheduling blocks for same-day behavioral health, pre-scheduled follow-ups to reduce walk-in variance) reduces the arrival variability coefficient. Standardizing intake protocols to reduce service time variance addresses the other term. Both directly shrink the V factor in the Kingman equation.

Build explicit capacity buffers. Staff to the 80th or 90th percentile of demand, not the mean. Hold appointment slots in reserve for stochastic surges rather than booking to 100% of template. This costs money — but the Variability Buffering Law guarantees that if you do not pay in capacity, you will pay in wait time.

Pool variable demand across servers. The square-root staffing law (Erlang, extended by Halfin and Whitt) shows that pooling variable demand across multiple servers produces sublinear growth in required capacity. Two behavioral health providers serving a combined panel need less total reserve capacity than two providers each managing separate panels. Pooling is the only free lunch in queueing theory — and it requires organizational redesign, not additional resources.

Measure and report distributional metrics. Replace average-only reporting with percentile reporting. The P90 wait time — the wait experienced by the worst-off 10% of patients — is a far more sensitive indicator of system stress than the mean.

What Software Should Surface

A capacity management tool operating in a stochastic environment should provide:

Real-time utilization with zone indicators — green/yellow/red mapped to the Kingman curve, not to arbitrary thresholds. Yellow should trigger at the utilization level where expected wait time exceeds the service-level target given the system’s measured variability.
Distributional demand forecasts — not a single demand number but a forecast distribution, showing the range of demand that is likely (P10 to P90) for the planning horizon. Every staffing recommendation should be stated as: “Staff X to meet average demand; staff Y to meet 90th percentile demand; the cost difference is Z.”
Wait-time percentile tracking — mean wait alongside P75 and P90, trended over time. The earliest degradation signal is P90 rising while the mean remains stable — the tail is growing before the average moves.
Abandonment rate as a system metric — tracked and trended, not as a patient behavior metric but as a system capacity metric. Rising abandonment at stable utilization means variability is increasing or buffers are eroding.
Scenario testing with stochastic inputs — “What happens to wait times if we lose one provider for three weeks?” should produce a probability distribution of outcomes, not a single revised average.

The Earliest Metric of Degradation

The coefficient of variation of wait time is the leading indicator. Before mean wait time rises — before utilization metrics breach thresholds — the spread of wait times increases. Some patients are seen promptly while others wait far longer than expected. This increasing dispersion is the first mathematical consequence of a system approaching the steep part of the utilization-delay curve. By the time the mean wait time rises noticeably, the system has been in stochastic stress for weeks.

Track the ratio of P90 wait time to P50 wait time. In a well-buffered system, this ratio is stable (typically 2:1 to 3:1). When it exceeds 4:1 and is trending upward, the system is entering the nonlinear zone — variability is outrunning capacity buffers. This is the signal to act, and it arrives before any average-based metric triggers an alarm.

Integration Hooks

Workforce: Staffing plans built on average patient volume are deterministic plans applied to a stochastic system. The gap between average-based and percentile-based staffing requirements is exactly the Variability Buffering Law operating on workforce capacity. Every understaffing event attributed to “an unusually busy day” is a stochastic event that a distributional staffing model would have anticipated. Workforce Module 1 (Workforce as Capacity Infrastructure) must incorporate variability parameters, not just headcount targets.

Public Finance: Grant budgets constructed from average cost-per-unit assumptions inherit the Flaw of Averages directly. If the cost per patient served has a right-skewed distribution (which it almost always does — some patients require far more resources than average), then budgeting at the mean guarantees overspend on the total program with high probability. Public Finance Module 6 (Financial Controls and Scenario Planning) should require Monte Carlo budget projections for any grant line item with coefficient of variation above 0.3.