| Vose Software

Industry: Government and Public Sector
Product: ModelRisk
Application: Social program analysis


A Housing-Assistance Program Budgeted at $3.6B Has a 44% Chance of Blowing Through Its Appropriation

A national housing-assistance program runs on a fixed annual appropriation: $3,600M, set by the budget committee and not easily reopened mid-year. The program office had built its request on point estimates — a 46% take-up rate, about 470,000 households served, an average case cost near $7,400 — which multiplied out to roughly $3.5B and looked comfortably inside the line. The committee approved it. What nobody had quantified was the chance the program would run out of money and need a supplemental appropriation.

That probability is the whole game in program budgeting, and it cannot be read off a point estimate. ModelRisk was used to forecast the distribution of total program cost and caseload, and to put a number on the over-run risk.

Why a point-estimate budget is a trap

A deterministic budget multiplies best-guess values together and reports one cost. But caseload is a count of people who show up, and how many show up depends on a take-up rate that varies, an eligible population that swells in a downturn, and a per-case cost that is right-skewed and itself climbs when the economy weakens. Multiply three central guesses and you get a single number that sits, by construction, near the middle of the real distribution — telling the committee nothing about the upside tail that triggers a supplemental.

The simulation, run over 100,000 trials, puts annual program cost at a mean of $3,688M, a P50 of $3,369M, a P90 of $5,832M, and a P95 of $6,799M. Against the $3,600M appropriation, the probability of an over-run is 43.8%. The point-estimate budget that looked safely inside the line is in fact a near coin-flip, with a P90 that exceeds the appropriation by more than 60%.

Distribution of total annual program cost against the appropriation

How caseload and cost are built

Total program cost is caseload multiplied by per-case cost, and each side is modelled from its own uncertain components — with one shared macroeconomic factor running through both.

  • Eligible population starts from a base pool of 1,000,000 households and is scaled by a macro-stress factor: a recession pushes more households below the means-test threshold, inflating the pool by up to roughly 45%.
  • Take-up rate — the fraction of eligible households that actually enroll — was modelled as a Beta distribution centred on 46%, the natural choice for a bounded participation rate, nudged slightly higher in bad years as more households apply. Caseload is eligible population × take-up, giving a mean of 468,770 served, P50 457,228, and P90 617,085.
  • Per-case annual cost was modelled as a LogNormal with a median near $7,400 and a long right tail (a minority of high-cost metropolitan cases dominate spending): mean $7,783, P90 $11,208. The same macro factor lifts the cost median in a downturn, as rents and arrears rise.

The decisive modelling choice is the shared macro factor. A recession does not politely raise eligibility while leaving costs alone — it raises both at once. Modelling caseload and per-case cost as independent would have let their swings partially cancel and produced a deceptively tight cost distribution. The common factor instead couples them, and that coupling is what fattens the upper tail where the supplemental-appropriation risk lives.

The budget breaks in two dimensions at once

Plotting caseload against total cost shows the program's budget does not live on a single axis — it lives in a plane, and the two axes move together.

Joint density of caseload and total program cost against the appropriation

The correlation between caseload and total cost is 0.63 — driven by the shared macro factor, not assumed away. The practical consequence is in the joint tail. The probability of landing in the danger quadrant — caseload above its 75th-percentile mark of 538,657 households and cost above the $3,600M appropriation — is 20.0%. If caseload and cost were treated as independent, that joint probability would be only 10.9%. Ignoring the coupling would understate the worst-case scenario — a surge in enrolment colliding with a surge in per-case cost — by nearly half. That is the exact scenario in which a program runs out of money and households are turned away mid-year.

What drives the cost spread

A rank-correlation tornado ranks the contributors to the spread in total program cost around its $3,369M median.

Tornado of drivers of program-cost spread

  • Per-case cost: ±$1,355M
  • Eligible population: ±$912M
  • Macro stress (shared factor): ±$912M
  • Take-up rate: ±$858M

Per-case cost is the single largest lever — its long LogNormal tail does most of the damage — but eligible population, the shared macro factor, and take-up are all close behind. There is no one dominant driver to manage; the over-run risk is broad-based, which argues for a contingency reserve rather than a single targeted control.

The inputs behind the forecast

The three core program uncertainties — a bounded take-up rate, a skewed per-case cost, and a macro-driven eligibility multiplier — are exactly the quantities a point-estimate budget replaces with single numbers.

Input distributions for take-up rate, per-case cost, and eligibility

The take-up Beta keeps participation inside [0, 1]; the per-case-cost LogNormal carries the right skew that a Normal would miss; the eligibility multiplier captures the macro swing that links the two. Collapsing any of them to its mean is what produced the comfortable-looking $3.5B that the simulation reveals to be a 44%-over-run gamble.

What the model changed

The program office had been ready to manage to a budget it believed was safe. The simulation reframed the appropriation as a risk position. It put the over-run probability at 43.8% — close enough to even odds that planning for a supplemental, rather than hoping to avoid one, became the responsible posture. It quantified the headroom needed for real safety: covering the P90 of $5,832M would require about $2.2B more than the appropriation, letting the committee decide explicitly how much over-run risk to fund versus accept. And by exposing the 0.63 coupling between caseload and cost, it showed that the dangerous case — a downturn driving enrolment and per-case cost up together, at 20% probability versus 11% under naive independence — is far more likely than an independence assumption would suggest.

ModelRisk functionality used

  • Monte Carlo simulation of total program cost over 100,000 trials, producing the cost distribution (mean $3,688M, P90 $5,832M) and the 43.8% over-run probability against the $3,600M appropriation.
  • Beta distribution for the bounded take-up rate (mean 46%) and LogNormal for the right-skewed per-case cost (median $7,400, P90 $11,208), each chosen to respect the variable's shape.
  • Shared macro common factor coupling eligible population and per-case cost, producing a 0.63 caseload–cost correlation instead of the artificially narrow spread independence would give.
  • Joint-density analysis quantifying the danger quadrant — high caseload and over budget — at 20.0% versus 10.9% under naive independence.
  • Tornado sensitivity ranking per-case cost (±$1,355M), eligibility, macro stress, and take-up as broadly comparable drivers, pointing to a contingency reserve over a single control.
  • Caseload forecasting (mean 468,770 households, P90 617,085) feeding both the cost output and the operational staffing question.

For a benefits program, Monte Carlo turns the budget from "it costs about $3.5B" into "there is a 44% chance it overruns $3.6B, and the worst case is a downturn driving caseload and cost up together" — the difference between funding a program and funding the risk of running it.