Value of Information

Paying to learn before you commit

A pharmaceutical division could license a compound today for $12 million certain or run a $2.5 million Phase II trial that might reveal success or failure before a larger manufacturing bet. The CEO asked a sharp question: "Are we paying $2.5 million for science, or are we paying to avoid a mistake?" That distinction is the value of information (VOI, how much a decision-maker should rationally pay to learn before acting).

Information is valuable only when it can change a decision or materially shift expected outcomes. A market study that confirms what you would do anyway is comfort spending. A study that steers you away from a $20 million loss is insurance. From Lesson 1, you rollback decision trees with EMV (expected monetary value, probability-weighted payoffs). VOI quantifies the EMV gain from learning, compared with acting on current beliefs. From Lesson 2, sensitivity shows which inputs matter; VOI asks whether buying data can move those inputs enough to alter the optimal branch.

Managers misuse VOI when they fund analytics without a decision fork tied to results. Dashboards that arrive after commitments are sunk storytelling. Pilots without predefined stop/go rules burn cash while pretending to learn. This lesson teaches expected value of perfect information (EVPI, upper bound on rational spend if learning removed all uncertainty relevant to the decision) and expected value of sample information (EVSI, value of a realistic study with noise), with rollback mechanics you can replicate in a spreadsheet.

Consider three organizational archetypes. Archetype A runs studies after the board vote to justify sunk spend: VOI is zero by construction. Archetype B runs pilots with stop/go rules but never computes EVSI: learning may help but spend is unbounded. Archetype C ties every study dollar to a tree fork and EVSI ceiling: learning is disciplined. This lesson equips you to move teams toward C without blocking legitimate non-EMV reasons (regulatory, reputational) when leadership documents overrides transparently.

Pharmaceutical Phase II trials, SaaS beta programs, and supplier audits before sole-source contracts share identical structure: pay now, observe signal, update beliefs, choose action. The math is rollback EMV; the culture is pre-registered rules. Finance partners with operations to ensure the study can actually change the capital path, not decorate it.

VOI sits at the intersection of Unit 6 tools. Decision trees (Lesson 1) supply the rollback structure. Sensitivity (Lesson 2) shows which inputs move outcomes; VOI asks whether paying to learn about a state (demand weak versus strong, borrower good versus bad) is cheaper than acting blind. Multi-criteria concerns (Lesson 4) sometimes override EMV VOI when learning serves regulatory or reputational goals. Communication (Lesson 5) must separate "EVSI does not justify cost" from "we still buy the study for policy reasons." This lesson focuses on EMV mechanics first, then names the override pattern explicitly.

Information only matters if decisions differ

Consider a simplified launch choice. Without study, EMV(launch) = $1.85M and EMV(wait) = $1.0M; you launch. Suppose a perfect oracle tells you demand will be weak before you launch. You would cancel and save losses. The value of that perfect signal is the difference between:

• Expected payoff with optimal decisions after knowing the truth, and
• Expected payoff when you choose optimally with current uncertainty.

If you would launch either way regardless of study outcome, EVPI = $0. Many corporate surveys fall here.

Formal setup:

1. Build a decision tree with imperfect or perfect information branch.
2. Roll back EMV for "with information" and "without information" policies.
3. VOI = EMV(with info) − EMV(without info) before subtracting study cost.
4. Buy information if VOI > study cost (for risk-neutral EMV).

Perfect information is a teaching upper bound: it assumes the signal always reveals the true state. Real studies misclassify favorable as unfavorable sometimes. EVSI ≤ EVPI always.

The managerial habit is to write decision rules before data: "If pilot conversion ≥ 8%, scale; else kill." Without rules, teams reinterpret ambiguous results to match prior hopes.

A practical test before funding any study: list the contingent actions for each possible signal. If the action column reads "launch" in every row, stop. You are buying confirmation, not information. If actions differ (launch / cancel / delay), proceed to EVPI and EVSI arithmetic.

Information structure must match the timeline. A study that returns after tooling is committed cannot change the launch decision; it can only refine marketing copy. Place the information branch before the irreversible spend in the tree. Real options logic (waiting, staging) often means the valuable study sits upstream of capital commitment, not parallel to it.

EVPI: rollback with a perfect oracle

Take NovaWear from Lesson 1 without the $200k study cost clutter. Launch Now EMV = $1.85M. With perfect demand foresight, you launch only in strong demand ($5M) and avoid weak (−$2M). Suppose true strong probability is 0.55.

EMV(perfect info policy) = 0.55(5.0) + 0.45(0) = $2.75M (you never launch into weak)

EMV(without info, launch always) = $1.85M

EVPI = 2.75 − 1.85 = $0.90M

You should not pay more than $900,000 for a perfect crystal ball on demand before launch. Real studies cost $200k but are imperfect; compare EVSI to cost.

EVPI is not a budget to spend. It is a ceiling. If EVPI is $90k and a study costs $200k, do not buy on EMV grounds unless non-EMV benefits (regulatory, reputational) clear the gap.

EVPI also clarifies which uncertainty to resolve. Tornado charts (Lesson 2) rank inputs; EVPI ranks states of the world that change discrete choices.

General EVPI procedure on any tree:

1. Solve the tree with current uncertainty and record root EMV (no extra information).
2. For each chance node that information could resolve, allow the decision-maker to pick the best action per state after learning.
3. Weight state-optimal payoffs by prior probabilities.
4. Subtract step 1 EMV. That difference is EVPI for that uncertainty.

EVPI can be zero for one uncertainty (demand) while positive for another (competitive entry). Do not bundle "market research" without specifying which fork it informs.

EVSI and imperfect signals

Real pilots, surveys, and trials produce signals correlated with truth but not identical. A favorable study might still lead to weak demand (false comfort). Rollback structure:

1. Pay study cost.
2. Chance node for signal (favorable / unfavorable) with total probabilities summing to 1.
3. At each signal, decision node: act optimally given posterior beliefs.
4. Compare root EMV to no-study baseline.

Bayes' rule (updating prior probabilities after observing evidence) connects prior and posterior when you need rigor. For MBA trees, you may enter posterior probabilities directly with a footnote on calibration.

Example calibration check: if 40% of studies are unfavorable and you always cancel then, ensure unfavorable branch EMV reflects cancel payoff, not accidental launch.

EVSI = EMV(with sample study) − EMV(without study) before cost.

Buy if EVSI > cost. Compare EVSI to EVPI; gap shows noise tax.

When studies are cheap and EVPI large, even noisy EVSI can justify spend. When EVPI is tiny, do not over-engineer measurement.

Signal quality metrics help design studies. Sensitivity (in the statistical sense) is the chance a favorable signal appears when the state is truly favorable. Specificity is the chance an unfavorable signal appears when the state is truly unfavorable. Low specificity means false alarms that trigger unnecessary cancels; low sensitivity means false comfort that launches into weak demand. Improving specificity often raises EVSI more than enlarging sample size on a vague survey question.

Document likelihood tables when possible:

Unfavorable signal rates follow by complement. These numbers feed Bayes or direct posterior entry. A study with high false favorable rate on weak demand destroys EVSI even when EVPI is large.

Real options, timing, and organizational barriers

Real options (the right but not obligation to expand, defer, or abandon) overlap VOI. Waiting preserves option value when upside uncertainty is high and learning is cheap. Launching kills the wait option. Trees model defer as explicit branches; VOI quantifies learning branches.

Organizational failures that VOI exposes:

• Analysis after decision (study decorates approval already granted).
• Sunk cost fallacy continuing after negative signal to "save" prior spend.
• Uniform reporting that hides which metric triggers scale/kill.

Finance should ask for pre-mortem decision rules in the charter: signal thresholds, sample size, timeline, and owner.

VOI also applies outside capital projects: A/B tests in product, credit scoring before large exposure, supplier audits before sole-source contracts. The pattern is identical: fork, posterior, EMV compare.

Ethical note: if information arrives about harm (safety, fraud), EMV maximization may not govern; legal duty to act can override VOI framing.

Stage gates are VOI made operational. Phase 1 spend buys a signal; Phase 2 spend occurs only if signal clears a threshold. The gate converts EVSI into a process the board can audit. Without gates, teams treat pilot success as automatic permission to scale even when EVSI at pilot design was below cost.

Portfolio of studies: when EVPI is positive for several uncertainties, compare sequential versus parallel learning. Sequential learning (test demand, then test pricing) can be cheaper than one mega-study if the first signal eliminates branches. Trees with multiple information nodes get large quickly; keep the first model to one or two information forks for transparency.

Communicating VOI without overselling analytics

When recommending a pilot, state four numbers: EVPI (ceiling), EVSI (realistic gain), study cost, and net EMV benefit (EVSI − cost). Add one sentence on decision rule ("scale if conversion ≥ 8%"). If net benefit is negative, say so and list non-EMV reasons if leadership still approves (preview Lesson 5).

Avoid equating VOI with "budget for research department." EVPI is decision-specific and time-bound. A $900k EVPI for NovaWear demand does not justify unrelated customer satisfaction tracking.

Sensitivity on signal quality: if EVSI is $100k at current calibration, ask how high EVSI rises if unfavorable signals correctly cancel weak demand 90% of the time instead of 60%. That analysis guides study design more than enlarging scope.

Link VOI results to triggers in operating dashboards. If you skip the study because EVSI < cost, monitor the observable that would have been the signal (leading indicator churn, wait-list depth) and set a re-open threshold for reconsidering a study next quarter.

Worked EVPI on three actions (launch, delay, cancel)

A startup can Launch (EMV $800k), Delay one year (EMV $600k preserving option), or Cancel ($0). Perfect information on market size would lead to: Launch if Big (p=0.45, payoff $2M), Delay if Medium (p=0.35, payoff $700k after delay cost), Cancel if Small (p=0.20, $0).

EMV(perfect policy) = 0.45(2.0) + 0.35(0.7) + 0.20(0) = 0.90 + 0.245 = $1.145M

EMV(no info, best Launch) = $0.8M

EVPI = 1.145 − 0.8 = $0.345M

Pay up to $345k for perfect market-size resolution before acting. Imperfect Series A investor feedback costing $50k might justify if EVSI ≥ $50k after calibration.

Check: 0.45+0.35+0.20=1.00 ✓

Sample size and EVSI (introductory)

Larger samples usually improve signal quality but raise cost. Plot EVSI − cost versus sample size: often increasing at first, then flattening as EVSI approaches EVPI. Stop where marginal EVSI gain no longer exceeds marginal cost. This stops analytics teams from proposing $2M studies when $200k design captures most EVSI.

Pilot design choices (length, geography, metric) should target the state that changes actions. A pilot measuring wrong metric (social likes) when decision hinges on paid conversion has EVSI near zero regardless of sample size.

Expected value of perfect information with multiple chance nodes

When two uncertainties exist (demand and competitor entry), EVPI for joint perfect information is usually less than sum of individual EVPIs because learning one state may change actions before you need the second signal. Solve the tree with a single perfect-information node that resolves both uncertainties only if decisions require both; otherwise compute EVPI per decision-relevant state separately to avoid overpaying for redundant research.

Charter language for pilots

A pilot charter should include: decision owner, signal metric, threshold, max spend, stop date, and EMV comparison without pilot. Legal and finance sign before engineering builds. Charter absence is how organizations accumulate pilots that never connect to rollback.

Harbor litigation (Lesson 1) EVPI for perfect case outcome knowledge may be large, but insurance premium is a purchased imperfect signal reducing tail loss. Compare premium to reduction in covenant breach probability, not only EMV shift. Legal and treasury jointly own that VOI framing.

Practice problem 3 (full solution)

A retailer chooses Order Big (payoff $3M if Holiday hot, −$1M if cold) versus Order Small (payoff $1M either state). P(hot)=0.55 without research. Research costs $150k; after research, if signal Hot order Big, if Cold order Small. Signal accuracy: P(Hot signal|hot)=0.90, P(Cold signal|cold)=0.85.

Without research: EMV(Big)=0.55(3)+0.45(−1)=1.65−0.45=$1.2M; Small=$1M → Big wins.

With research (calibrated): P(Hot signal)=0.55(0.9)+0.45(0.15)=0.495+0.0675=0.5625 (approx). Work decision subtree:

After Hot signal: order Big, EMV≈0.9(3)+0.1(−1)=2.7−0.1=$2.6M conditional... simplified aggregate EMV(with policy) ≈ $1.35M before cost (illustrative rollup).

EVSI ≈ 1.35−1.2 = $0.15M = $150k. Cost $150k → borderline. Check signal probs documented.

Managerial read: negotiate research to $120k or improve cold-signal specificity before holiday commit.

Bayesian update primer (two-state)

Prior P(hot)=0.55. Observe Hot signal with likelihood P(signal Hot|hot)=0.90 and P(signal Hot|cold)=0.15.

Posterior P(hot|Hot) = 0.55(0.90) / [0.55(0.90)+0.45(0.15)] = 0.495 / (0.495+0.0675) = 0.495/0.5625 ≈ 0.88.

Use posterior in subtree after Hot signal. Cold signal symmetric. Bayes keeps probability bookkeeping coherent when calibration data exist.

EVPI zero examples (teaching)

• Study confirms launch either way → EVPI=0.
• Perfect info arrives after irreversible invest → EVPI=0 for that decision (too late).
• Two states, same optimal action → EVPI=0.

Listing zero cases prevents wasted analytics budget.

Reuse NovaWear parameters from Lesson 1.

Part A: Without study

Launch Now EMV = $1.85M (optimal without study path).

Study First rolled back EMV = $1.75M after $0.2M cost.

Part B: EVSI of proposed study

EMV(with study policy as modeled) before subtracting cost = $1.95M (from Lesson 1 intermediate).

EVSI = 1.95 − 1.85 = $0.10M = $100,000

Subtract cost $200,000 → net −$100,000 vs launching without study.

Part C: Check

VOI before cost ($100k) < cost ($200k) → do not buy study on EMV.

Part D: Managerial read

Marketing may still want study for narrative risk; finance quantifies EMV gap. Negotiate cheaper study or design that raises EVSI (better discrimination). Sensitivity from Lesson 2: if favorable signal pushes strong demand to 85% probability, EVSI rises enough to justify purchase.

Ridge Manufacturing from Lesson 1 considers a $40k demand test before choosing Buy versus Lease.

Part A: Without test

Lease net $390k certain dominates Buy EMV $50k.

Part B: Test structure

Test costs $40k. Signal High (p=0.55) matches true High demand; signal Low (p=0.45) matches true Low. After High signal, EMV(Buy)=$400k net on high branch only... simplified: if High signal, Ridge buys with EMV $400k on buy path versus lease $390k → Buy. After Low signal, EMV(Buy)=−$300k → Lease.

EMV(with test, optimal) = 0.55(max(400,390)) + 0.45(max(−300,390)) = 0.55(400) + 0.45(390) = 220 + 175.5 = $395.5k before test cost.

Subtract test cost: $395.5k − $40k = $355.5k.

Part C: EVSI check

EMV without test (optimal Lease) = $390k.

EVSI = 395.5 − 390 = $5.5k before cost; after cost net −$34.5k.

Do not buy test on EMV. Check: signal probs 0.55+0.45=1.00 ✓

Part D: Managerial read

Perfect discrimination would raise EVSI toward EVPI on avoiding −$300k low branch, but imperfect test still rarely flips to Buy on Low signal. Ridge should Lease unless test quality improves or Buy floor policy changes.

A lender chooses Extend $10M credit or Decline. If extend, types are Good (p=0.70, NPV +$1.2M) or Bad (p=0.30, NPV −$2.0M).

Without info, EMV(extend) = 0.7(1.2)+0.3(−2.0)=0.84−0.60=+$0.24M; decline = 0 → extend wins.

With perfect type knowledge: extend only Good → EMV = 0.7(1.2)+0.3(0)=$0.84M

EVPI = 0.84 − 0.24 = $0.60M

Pay up to $600k for perfect underwriting signal; realistic bureau pull costing $50k is easily justified if EVSI is near EVPI.

Check: 0.70+0.30=1.00 ✓

Launch vs Cancel without test: Launch EMV $500k. Perfect test would let you launch only when Success (p=0.40, payoff $2M) and cancel otherwise (Failure p=0.60 → $0).

1. Compute EMV with perfect information policy.
2. Compute EVPI.
3. Test costs $300k and is imperfect: EVSI = $350k. Buy or not?

Solution

1. EMV(perfect) = 0.40(2.0)+0.60(0)=$800k

2. EVPI = 800 − 500 = $300k

3. EVSI $350k > cost $300k → buy (barely). Note EVSI > EVPI example impossible with true perfection; here EVSI is stated as model output. Check EVSI ≤ EVPI in real calibrated models.

Explain in a paragraph why a customer survey costing $80k with EVSI $25k might still be approved.

Solution

EMV alone rejects the spend (25 < 80). Approval may rely on non-EMV benefits: regulatory evidence, investor narrative, risk committee comfort, or option value on a follow-on decision. Document the override; do not pretend EMV justified it.

Organizations run dozens of parallel studies. VOI prioritizes spend. Rank pending pilots by EVSI minus cost per dollar of capital at risk. A $50k survey with EVSI $120k on a $5M branch outranks a $200k study with EVSI $80k on a $500k branch even though the second study costs more in absolute terms.

VOI also ends vanity metrics. If no decision fork uses the metric, EVPI for that metric is zero. Dashboards should map each tile to a fork or be demoted to background monitoring.

Pair with Lesson 2 sensitivity: if tornado shows decision hinges on churn, EVPI for perfect churn knowledge is high; if tornado shows rent is negligible, do not fund occupancy study.

Communication template: "EVPI $900k; proposed study EVSI $100k at cost $200k; recommend decline on EMV; approve only if reputational benefit ≥ $100k documented."

Ethics override: safety signals may require action with negative EMV VOI; legal documents override.

Extended practice: biotech Phase II (full walk)

Problem: License $12M now versus Phase II $4M with success p=0.35, NPV $20M, failure NPV $0.

EMV(Phase II) = 0.35(20) − 4 = $3M beats license $1.5M.

EVPI with perfect trial: success → continue, failure → license at $12M? If failure leads to license instead of $0, perfect policy EMV rises. If failure → $0, EVPI for trial outcome equals gain from avoiding failed Phase II spend. Compute per tree.

Managerial read: Phase II wins on EMV; VOI of perfect efficacy signal caps spend on supplementary biomarker studies.

Check: probabilities sum 1.0 ✓

Deep dive: documenting non-EMV overrides

When board approves study despite EVSI < cost, memo must state override category: regulatory, reputational, investor, option value. Quantify override minimum dollars so future audits understand trade.

Deep dive: sequential learning

Run cheap pilot first; if signal positive, run expensive trial. Tree with two information nodes; EVSI of second study conditional on first reduces wasted Phase III spend.

Closing integration with Unit 6

Lesson 1 trees supply forks. Lesson 2 sensitivity ranks which state matters. Lesson 3 VOI prices learning. Lesson 4 MCDM handles non-dollar goals. Lesson 5 communicates all four. VOI without communication becomes spreadsheet trivia executives ignore.

Value of information closes the loop on uncertainty: pay to learn only when learning moves optimal actions enough to justify cost under EMV, unless leadership documents a non-EMV override with eyes open.

Key takeaways

• Information has decision value only when it can change optimal actions or materially shift EMV.
• EVPI is the EMV gain from perfect foresight; it caps rational spend on learning.
• EVSI values realistic imperfect studies; buy when EVSI exceeds cost for risk-neutral EMV.
• Predefine signal-based decision rules before collecting data.
• Pair VOI with sensitivity to focus measurement on uncertainties that move choices.

1. For a pending pilot at your firm, write stop/go rules and estimate whether EVSI could exceed pilot cost.
2. What is EVPI in words for that decision (not only dollars)?
3. Continue to Lesson 4: Multi-Criteria Decision Making.