Sensitivity Analysis - Finite Mathematics

SECTION 1

Historical Context & Motivation

Every mathematical model rests on assumptions—estimated probabilities, projected payoffs, forecasted costs—and real-world data rarely match those assumptions exactly. Sensitivity analysis arose from the practical need to understand how robust a decision remains when the underlying parameters shift. Rather than treating a model's output as a single, definitive answer, sensitivity analysis asks: over what range of input values does the current optimal decision remain optimal? This question has driven developments across operations research, economics, and engineering since the mid-twentieth century.

1947 • Linear Programming & the Simplex Method

George Dantzig develops the simplex algorithm for linear programming. Researchers quickly realize that optimal solutions can change dramatically with small perturbations in constraint coefficients, motivating systematic parameter variation studies.

1954 • Post-Optimality Analysis Formalized

Researchers at RAND Corporation formalize post-optimality (sensitivity) analysis for LP models, establishing techniques to determine allowable ranges of objective-function coefficients and right-hand-side values without re-solving the entire problem.

1961 • Decision Analysis Emerges

Howard Raiffa and Robert Schlaifer publish foundational work on decision analysis under uncertainty. Sensitivity analysis becomes a standard component of evaluating decision trees and payoff matrices in business and policy contexts.

1980s • Spreadsheet Revolution

The proliferation of spreadsheet software (Lotus 1-2-3, later Excel) makes what-if analysis accessible to managers and analysts. Data tables and scenario managers embed sensitivity analysis into everyday business practice.

2000s–Present • Monte Carlo & Global Methods

Advances in computing power enable Monte Carlo simulation and global sensitivity methods (Sobol indices, Morris screening), extending sensitivity analysis beyond one-parameter-at-a-time approaches to high-dimensional, stochastic models.

In a finite mathematics course focused on decision analysis, sensitivity analysis addresses a central question: given a payoff table and a set of probabilities for states of nature, how much can those probabilities change before a different alternative becomes optimal? Answering this question transforms a static decision model into a dynamic tool for understanding risk and informing judgment under genuine uncertainty.

SECTION 2

Core Principles & Definitions

Before diving into computations, it is essential to ground the discussion in the vocabulary and conceptual framework of sensitivity analysis within decision theory. A standard decision problem involves a set of alternatives (actions the decision-maker can choose), a set of states of nature (uncertain outcomes beyond the decision-maker's control), payoffs associated with each alternative–state pair, and prior probabilities assigned to each state of nature. Sensitivity analysis systematically perturbs these probabilities to discover the boundaries of optimality.

Parameter of Interest

The specific probability (or set of probabilities) being varied. In a two-state problem the parameter is typically p, the probability of state 1, with the complementary probability 1 − p for state 2.

Expected Value (EV)

The probability-weighted average payoff for each alternative. Each alternative's EV is a linear function of p in a two-state framework, enabling straightforward graphical and algebraic analysis.

Crossover (Break-Even) Point

The value of p at which two alternatives yield the same expected value. At this threshold the decision-maker is indifferent between them; on either side of it, one alternative strictly dominates.

Range of Optimality

The interval of p values over which a particular alternative produces the highest expected value. Identifying this range is the primary deliverable of a sensitivity analysis.

Robustness

A decision is considered robust if it remains optimal across a wide range of plausible parameter values. A narrow range of optimality signals vulnerability and may warrant gathering additional information or hedging strategies.

✦

KEY TAKEAWAY

Think of sensitivity analysis like testing the structural integrity of a bridge by simulating different load scenarios. A single load test (one probability estimate) tells you the bridge holds under that specific weight, but systematically varying the load reveals the breaking point—the critical threshold where behavior changes. In decision analysis, the 'breaking point' is the crossover probability where your optimal choice switches to a different alternative.

SECTION 3

Visual Explanation — EV as a Function of p

The most powerful visual tool in sensitivity analysis for decision problems is the expected-value-versus-probability graph. Because each alternative's expected value is a linear function of p (in the two-state case), plotting all alternatives on the same axes produces a set of straight lines. The upper envelope of these lines—the highest line at each value of p—traces out the optimal decision as a function of the uncertain probability. Crossover points, where two lines intersect, mark the exact thresholds at which the optimal decision changes.

Three alternatives (A₁, A₂, A₃) plotted as linear EV functions of p. The yellow circles mark crossover points p*₁ ≈ 0.40 and p*₂ ≈ 0.67. For p < 0.40, A₃ is optimal; for 0.40 < p < 0.67, A₂ is optimal; for p > 0.67, A₁ is optimal.

The diagram above illustrates the fundamental geometry of sensitivity analysis. Each alternative's expected value is a straight line because EV is computed as p × (payoff under state 1) + (1 − p) × (payoff under state 2), which is a linear equation in p. The upper envelope of these lines identifies the optimal decision for every possible value of p. Notice that A₃ (pink) has the highest EV when p is low, A₂ (violet) dominates the middle range, and A₁ (cyan) is best for high values of p. The crossover points are precisely the break-even probabilities where the decision-maker switches from one alternative to another.

SECTION 4

Mathematical Framework

The algebraic foundation of sensitivity analysis in a two-state decision problem is remarkably clean. Let the two states of nature be S₁ and S₂ with probabilities p and 1 − p respectively. For any alternative Aᵢ with payoffs aᵢ₁ (under S₁) and aᵢ₂ (under S₂), the expected value is a linear function of p.

EXPECTED VALUE (TWO-STATE)

EV(Aᵢ) = p × aᵢ₁ + (1 − p) × aᵢ₂ = aᵢ₂ + p × (aᵢ₁ − aᵢ₂)

Here p = P(S₁), aᵢ₁ is the payoff of alternative i under state S₁, and aᵢ₂ is the payoff under S₂. The expression has slope (aᵢ₁ − aᵢ₂) and intercept aᵢ₂.

Since each alternative's EV is a line in p, finding the crossover point between alternatives Aᵢ and Aⱼ reduces to solving a linear equation. Setting EV(Aᵢ) = EV(Aⱼ) and solving for p yields the break-even probability.

CROSSOVER PROBABILITY

p* = (aⱼ₂ − aᵢ₂) / [(aᵢ₁ − aᵢ₂) − (aⱼ₁ − aⱼ₂)]

This is valid when the denominator is nonzero, i.e., the two EV lines are not parallel. p* must lie in [0, 1] to be a meaningful crossover point within the decision problem.

GENERAL MULTI-STATE EXPECTED VALUE

EV(Aᵢ) = Σⱼ pⱼ × aᵢⱼ for j = 1, 2, …, n

For problems with more than two states, the expected value is a sum over all states. Sensitivity analysis in this case typically fixes all probabilities except one or two and varies them while adjusting the remaining probabilities to maintain the constraint Σpⱼ = 1.

💡

Why Linearity Matters

The linearity of EV in p guarantees that the optimal alternative can only change at crossover points—there are no hidden local optima or curvature effects. Between any two consecutive crossover points, one alternative dominates everywhere. This is what makes the graphical method both valid and complete for two-state problems.

SECTION 5

Types of Sensitivity Analysis in Decision Problems

Sensitivity analysis is not a single technique but a family of approaches, each suited to different aspects of the decision model. In finite mathematics and decision analysis, three primary types appear most frequently: one-way sensitivity analysis (varying a single parameter), two-way sensitivity analysis (varying two parameters simultaneously), and scenario analysis (evaluating discrete parameter combinations). Each provides a different lens on the robustness of the optimal decision.

Three primary types of sensitivity analysis compared. One-way analysis varies a single probability and produces crossover points. Two-way analysis varies two probabilities and yields optimality regions in a 2D plane. Scenario analysis evaluates discrete combinations of parameter values such as best-case, base-case, and worst-case.

Feature	One-Way	Two-Way	Scenario
Parameters Varied	1	2	All (discrete sets)
Output	Crossover probability, range of optimality	Optimality regions in 2D space	EV table for each scenario
Visualization	EV vs. p graph	2D region plot	Comparison table or tornado chart
Complexity	Low — algebra and graphing	Medium — simultaneous equations	Low to medium — enumeration

Comparison of sensitivity analysis types used in decision analysis

SECTION 6

Worked Example

A technology startup must choose among three product strategies: A₁ (aggressive launch), A₂ (moderate rollout), and A₃ (conservative pilot). Two states of nature exist: S₁ (favorable market, probability p) and S₂ (unfavorable market, probability 1 − p). The payoff table (in $thousands) is shown below.

Alternative	S₁ (Favorable)	S₂ (Unfavorable)
A₁ (Aggressive)	120	−20
A₂ (Moderate)	70	30
A₃ (Conservative)	35	45

Payoff table ($thousands)

Sensitivity Analysis: Find All Crossover Points and Ranges of Optimality

1
Step 1 — Write EV Equations

Express each alternative's expected value as a linear function of p: EV(A₁) = 120p + (−20)(1 − p) = 140p − 20 EV(A₂) = 70p + 30(1 − p) = 40p + 30 EV(A₃) = 35p + 45(1 − p) = −10p + 45

EV(A₁) = 140p − 20, EV(A₂) = 40p + 30, EV(A₃) = −10p + 45
2
Step 2 — Find Crossover Points

A₁ vs. A₂: Set 140p − 20 = 40p + 30. Solving: 100p = 50, so p* = 0.50. A₂ vs. A₃: Set 40p + 30 = −10p + 45. Solving: 50p = 15, so p* = 0.30. A₁ vs. A₃: Set 140p − 20 = −10p + 45. Solving: 150p = 65, so p* = 13/30 ≈ 0.433. We check whether this crossover is relevant by verifying if either A₁ or A₃ is on the upper envelope at this point. At p = 0.433, EV(A₂) = 40(0.433) + 30 = 47.33, while EV(A₁) = EV(A₃) = 140(0.433) − 20 = 40.67. Since EV(A₂) > EV(A₁) = EV(A₃), this crossover lies below the upper envelope and is not an operational break-even point.

Relevant crossover points: p* = 0.30 (A₂ vs. A₃) and p* = 0.50 (A₁ vs. A₂)
3
Step 3 — Determine Ranges of Optimality

Evaluate which alternative has the highest EV in each interval defined by the crossover points: • For p < 0.30: Test p = 0. EV(A₁) = −20, EV(A₂) = 30, EV(A₃) = 45. A₃ is optimal. • For 0.30 < p < 0.50: Test p = 0.40. EV(A₁) = 36, EV(A₂) = 46, EV(A₃) = 41. A₂ is optimal. • For p > 0.50: Test p = 0.80. EV(A₁) = 92, EV(A₂) = 62, EV(A₃) = 37. A₁ is optimal.

A₃ optimal for 0 ≤ p ≤ 0.30 | A₂ optimal for 0.30 ≤ p ≤ 0.50 | A₁ optimal for 0.50 ≤ p ≤ 1.0
4
Step 4 — Interpret the Results

If the company's marketing team estimates a 60% chance of a favorable market (p = 0.60), then A₁ is optimal with EV = 140(0.60) − 20 = $64K. However, this decision is only robust as long as p stays above 0.50. If new data reduced the estimate to p = 0.45, the optimal decision would switch to A₂. The sensitivity analysis reveals that a mere 10-percentage-point downward revision in the probability estimate changes the recommended strategy entirely, signaling that additional market research could have significant value.

At p = 0.60, choose A₁ (EV = $64K). Decision switches to A₂ if p drops below 0.50.

SECTION 7

Strengths & Limitations

Sensitivity analysis is an indispensable complement to expected-value decision-making, but it is not a panacea. Understanding both its power and its boundaries is essential for using it wisely. The following table summarizes the principal strengths and limitations as encountered in finite mathematics and applied decision analysis.

Strengths	Limitations
Reveals how robust or fragile the optimal decision is against probability estimation errors.	One-way analysis varies only one parameter at a time, potentially missing interaction effects between multiple uncertain inputs.
Provides exact crossover probabilities through straightforward algebra—no simulation or iterative computation required.	Assumes payoff values are known with certainty; in practice, payoffs themselves may be uncertain and should also be tested.
Graphical representation (EV vs. p) offers powerful visual intuition that is accessible to non-technical stakeholders.	For problems with more than two or three states of nature, the graphical approach becomes impractical and algebraic methods grow complex.
Quantifies the value of additional information: if the decision is near a crossover point, gathering more data has high expected value.	Does not provide a probability distribution over outcomes—it tells you where the break-even is, not how likely it is that you are on one side or the other.

Strengths vs. limitations of sensitivity analysis in decision problems

✦

KEY TAKEAWAY

Sensitivity analysis is like stress-testing a financial portfolio: it does not predict which market conditions will occur, but it tells you precisely under which conditions your strategy fails. Knowing that your decision remains optimal as long as p stays within [0.30, 0.50] is far more informative than knowing only that 'A₂ is best at p = 0.40.' The width of the optimality range is itself a measure of decision quality.

SECTION 8

Connection to Advanced Decision Theory

The one-way sensitivity analysis studied here in finite mathematics is the entry point to a much richer set of techniques encountered in advanced operations research and data science. Understanding how the basic framework extends provides valuable context and motivation for further study. Two natural extensions are particularly important: Expected Value of Perfect Information (EVPI) and Bayesian updating with sensitivity.

Concept	Basic (This Course)	Advanced Extension
Sensitivity Analysis	Vary p in a two-state payoff table; find crossover points algebraically	Global sensitivity (Sobol indices) over high-dimensional parameter spaces using Monte Carlo simulation
Value of Information	EVPI = EV with perfect info − EV of best alternative	Expected Value of Sample Information (EVSI) with Bayesian posterior probabilities and preposterior analysis
Decision Structure	Single-stage payoff matrix with discrete alternatives and states	Multi-stage decision trees, influence diagrams, and Markov decision processes
Uncertainty Modeling	Point probabilities varied over [0, 1]	Full probability distributions (Beta, Dirichlet) over uncertain parameters; robust optimization

Basic vs. advanced decision analysis concepts

The connection between sensitivity analysis and EVPI is worth emphasizing. If your sensitivity analysis reveals that the optimal decision is very sensitive to p—meaning the current estimate sits near a crossover point—then the EVPI is likely high, indicating that investing in better information (market research, expert consultation, pilot studies) could yield substantial returns. Conversely, if the optimal decision is robust over a wide range of p, the EVPI is low and additional research may not be cost-effective. In this way, sensitivity analysis directly informs resource allocation for uncertainty reduction.

ℹ

Looking Ahead

In courses on stochastic models and operations research, you will encounter parametric sensitivity analysis for linear programs (ranging objective coefficients and RHS values) and probabilistic sensitivity analysis using simulation. The algebraic intuition developed here—linearity, crossover points, ranges of optimality—transfers directly to those more complex settings.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

Explain why each alternative's expected value is a linear function of p in a two-state decision problem. What property of the expected value formula guarantees this linearity, and why does it matter for sensitivity analysis?

Show answer

The expected value formula EV(Aᵢ) = p × aᵢ₁ + (1 − p) × aᵢ₂ expands to aᵢ₂ + p(aᵢ₁ − aᵢ₂), which is of the form y = b + mp—a linear equation in p with slope m = (aᵢ₁ − aᵢ₂) and intercept b = aᵢ₂. This linearity arises because expected value is a weighted sum with weights that are linear in p. For sensitivity analysis, linearity guarantees that EV lines can cross at most once, meaning there is a single crossover point between any two alternatives. Consequently, the optimal alternative can only change at these intersection points, and between consecutive crossovers one alternative strictly dominates. Without linearity (e.g., if payoffs depended on p), the upper envelope could be curved, creating multiple crossings and local optima that would complicate the analysis significantly.

Question 2: BASIC CALCULATION

Consider two alternatives with the following payoffs: A₁ has payoffs 80 (S₁) and 20 (S₂); A₂ has payoffs 50 (S₁) and 40 (S₂). Let p = P(S₁). Find the crossover probability p* and state which alternative is optimal for p > p* and for p < p*.

Show answer

First, write the EV equations: EV(A₁) = 80p + 20(1 − p) = 60p + 20, and EV(A₂) = 50p + 40(1 − p) = 10p + 40. Setting them equal: 60p + 20 = 10p + 40, which gives 50p = 20, so p* = 0.40. For p > 0.40, A₁'s steeper slope causes its EV to exceed A₂'s, so A₁ is optimal. For p < 0.40, A₂ is optimal because its higher intercept dominates. At p = 0.40 exactly, the decision-maker is indifferent between the two alternatives, with EV(A₁) = EV(A₂) = 44.

Question 3: INTERMEDIATE

A manager faces three alternatives with the following payoff table (values in $thousands): A₁: S₁ = 100, S₂ = 10 A₂: S₁ = 60, S₂ = 50 A₃: S₁ = 40, S₂ = 55 Find all pairwise crossover points, determine which crossovers lie on the upper envelope, and state the complete ranges of optimality for each alternative.

Show answer

The EV equations are: EV(A₁) = 90p + 10, EV(A₂) = 10p + 50, EV(A₃) = −15p + 55. Pairwise crossovers: A₁ vs. A₂: 90p + 10 = 10p + 50 → 80p = 40 → p* = 0.50. A₂ vs. A₃: 10p + 50 = −15p + 55 → 25p = 5 → p* = 0.20. A₁ vs. A₃: 90p + 10 = −15p + 55 → 105p = 45 → p* ≈ 0.4286. At p = 0.4286, check A₂: EV(A₂) = 10(0.4286) + 50 = 54.29, while EV(A₁) = EV(A₃) ≈ 48.57. Since A₂ > A₁ = A₃ at this point, the A₁–A₃ crossover is below the envelope. The operative crossovers are p* = 0.20 (A₂ vs. A₃) and p* = 0.50 (A₁ vs. A₂). Ranges: A₃ optimal for 0 ≤ p ≤ 0.20, A₂ optimal for 0.20 ≤ p ≤ 0.50, A₁ optimal for 0.50 ≤ p ≤ 1.0.

Question 4: APPLIED

A hospital administrator must select a staffing plan for flu season. Plan A (full staffing) costs $500K but handles surges efficiently, yielding a net benefit of $200K if flu season is severe (S₁) and $−100K if mild (S₂). Plan B (partial staffing) yields $80K if severe and $60K if mild. Plan C (contract nurses on-call) yields $50K if severe and $90K if mild. The CDC estimates a 55% chance of a severe season. (a) Which plan is optimal at p = 0.55? (b) How far would the probability of a severe season need to drop before the administrator should switch plans? (c) What does this imply about the value of improved flu forecasting?

Show answer

(a) EV(A) = 200p + (−100)(1 − p) = 300p − 100. EV(B) = 80p + 60(1 − p) = 20p + 60. EV(C) = 50p + 90(1 − p) = −40p + 90. At p = 0.55: EV(A) = 65, EV(B) = 71, EV(C) = 68. Plan B is optimal at p = 0.55. (b) The relevant crossovers: B vs. C: 20p + 60 = −40p + 90 → 60p = 30 → p* = 0.50. B vs. A: 20p + 60 = 300p − 100 → 160 = 280p → p* ≈ 0.571. Plan B is optimal for 0.50 ≤ p ≤ 0.571. If p drops below 0.50, Plan C becomes optimal. Since p = 0.55 is only 5 percentage points above the B-to-C crossover, the decision is moderately sensitive. (c) Because the current estimate (0.55) sits only 0.05 above the lower crossover, improved forecasting that could narrow the confidence interval on p would have significant decision value. If the estimate were revised downward by even a modest amount, the optimal plan would change, potentially saving tens of thousands of dollars in misallocated resources.

Question 5: CRITICAL THINKING

Prove that in a two-state decision problem with n alternatives, the maximum number of segments on the upper envelope of EV lines (and hence the maximum number of alternatives that can be optimal for some value of p ∈ [0, 1]) is n, but the maximum number of crossover points on the upper envelope is n − 1. Then explain why, in practice, many alternatives are typically dominated and never appear on the upper envelope at all.

Show answer

Each alternative's EV is a line in p, and the upper envelope is the pointwise maximum of n lines. A classical result from computational geometry establishes that the upper envelope of n lines in the plane has at most n segments and n − 1 breakpoints (crossover points). To see this, note that two distinct non-parallel lines intersect at most once, and each line can appear on the upper envelope as at most one contiguous segment (because two lines can cross only once, a line that drops below another cannot re-emerge above it later). Therefore the segments are ordered left to right, and n segments are separated by at most n − 1 transition points. In practice, many alternatives are dominated: an alternative Aᵢ is dominated if there exists another alternative Aⱼ such that EV(Aⱼ) ≥ EV(Aᵢ) for all p ∈ [0, 1], or more subtly, if the crossover between Aᵢ and other lines falls outside [0, 1] or below a third line (as in the worked example where A₁ vs. A₃ crossed below A₂). Domination is common because real-world decision problems often include alternatives that are uniformly mediocre—neither the best under optimistic nor pessimistic scenarios—and these never appear on the upper envelope regardless of the probability estimate.

SUMMARY

Lesson Summary

Summary

Sensitivity analysis in decision analysis systematically varies the probability of each state of nature to determine the range of optimality for each alternative. In a two-state problem, each alternative's expected value is a linear function of the probability parameter p, yielding straight lines that can be plotted on an EV-versus-p graph. The crossover (break-even) points—where two EV lines intersect—mark the exact thresholds at which the optimal decision changes, and the upper envelope of all lines traces out the best alternative across the full range of p.

The crossover probability formula p* = (aⱼ₂ − aᵢ₂) / [(aᵢ₁ − aᵢ₂) − (aⱼ₁ − aⱼ₂)] provides the algebraic solution for any pair of alternatives. Sensitivity analysis enhances decision-making by revealing robustness: a wide range of optimality signals a safe decision, while a narrow range near the current probability estimate suggests the decision is fragile and that investing in better information may be worthwhile. Extensions include two-way sensitivity analysis, scenario analysis, and connections to the Expected Value of Perfect Information (EVPI).