The kettle had just clicked off when my wife came into the kitchen, holding her phone at arm's length. Rain ticked against the window, and the number on her screen was large enough to read before she said a word.
"Two hundred dollars," she said. "For riding the train without paying."
I thought of the young man I often saw on the 7:42, the one who slowed beside the card reader, glanced at it, and kept walking.
Then I pulled a napkin from beneath the sugar bowl.
"What are you doing?"
"Looking for the one probability that changes the answer."
She lowered the phone. "You do not know how often inspectors check."
"I do not need to know it yet."
The number nobody can give you
"Suppose I ask a student what score they think they will probably get tomorrow," I said. "Any student, the night before the exam."
"They cannot tell me. Seventy? Ninety? It depends which questions come up, how they slept, whether the one topic they skipped is on the paper. The honest answer is that they do not know, and no amount of pushing produces a number."
"Now I ask that same student something else. Will you score at least twenty?"
"Yes," she said. "Obviously yes. After a year of classes, of course they clear twenty."
"Instantly, and with complete confidence. Same student, same exam, the same missing knowledge. Only the shape of the question changed. We are poor at pulling a number out of the air, and very good at saying which side of a line we are on."
"And this helps with a train fine how?"
"Because this decision has a line too. Find it, and I stop needing the number I was never going to get. That is what I am looking for on this napkin."
The bet hidden inside one ride
I flattened the napkin between the sugar bowl and her phone, then drew a fork in the road.
Above one branch I wrote Pay $10. That branch was boring in the best possible way. You knew the cost before the train moved.
Above the other I wrote Do not pay. Most of the time, that branch appeared to cost nothing. But if an inspector arrived, it cost $200.
I labelled the chance of that happening p.
Expected monetary cost of not paying = p × $200
The expected monetary cost of paying was simply $10. No probability was needed on that branch because the fare was certain.
"So now you need p," she said.
"Eventually. First I want the value of p that makes these two branches cost exactly the same."
The five-percent line
I wrote the equality in the middle of the napkin:
$10 = p × $200
Then I divided both sides by 200.
p = $10 / $200 = 0.05 = 5%
In general, p* = fare / fine. With these numbers, p* was 5 percent. There it was.
"One ride in twenty?" my wife asked.
"At exactly one in twenty, the two choices tie. A little above it, paying has the lower expected monetary cost. A little below it, not paying does, but only inside this deliberately narrow one-ride model."
I circled the five. "This is the threshold probability."
She did not look convinced. "But five percent is not how often they actually check."
"No. It is not a prediction."
"Then what did you learn?"
"Where the answer flips. I may not know where reality stands, but I know the border it has to cross."
The border, not the prediction
That distinction is the whole lesson.
When a decision depends on a probability nobody can give you, the instinct is to hunt for the missing number. You search reports, ask experts, argue over estimates, and sometimes discover that the probability is changing, confidential, or simply unknowable.
Threshold analysis starts from the other direction. It asks: At what probability would I change my choice?
The square is the choice. The circle is the uncertain event. Once the payoffs are on the branches, the crossing point falls out of the model even while the true probability remains unknown.
A doctor can calculate the disease probability at which treatment becomes worthwhile. A product team can calculate the adoption probability that would justify development. An investor can calculate the failure probability that would erase a return. The subject changes; the logic does not.
The threshold does not remove uncertainty. It tells you which uncertainty can reverse the decision.
My wife touched the circle on the napkin. "So the unknown stays unknown, but it stops being shapeless."
"Exactly."
The lesson in three lines
Paying costs $10, every time, with no surprises. Skipping costs nothing on most days and $200 on the day an inspector reaches your seat, so on average it costs p × $200, where p is your chance of being checked.
The two choices cost the same when $10 = p × $200. Solve that and p is 5 percent. In general the break-even point is p* = fare / fine.
Five percent is not a guess at how often inspectors appear. It is the line where the answer flips. If the real chance is higher than 5 percent, paying costs you less in the long run. If it is lower, skipping does. If you cannot tell which side you are on, that is the one thing worth finding out.
The question you can actually answer
My wife was not finished. "A border is useful only if you know which side you are standing on."
She was right. The threshold alone does not settle the decision. You still need evidence about the inspection probability, or at least an honest range of values you consider plausible.
The green line never moves because paying always costs $10. The red line rises with p because a $200 fine becomes more expensive on average as inspections become more likely. Their crossing is the five-percent threshold.
If every plausible value of p lies below 5 percent, not paying has the lower expected monetary cost in this narrow model. If every plausible value lies above 5 percent, paying does. If the plausible range crosses 5 percent, the decision is sensitive.
That gives you a question you can actually answer: Could any believable inspection probability cross the five-percent line?
The threshold is not permission to replace evidence with a hunch. It is a map showing exactly where better evidence can change the answer.
What the average leaves out
"So below five percent, your model tells him not to pay?" my wife asked.
"No. It tells us only what a deliberately narrow, money-only, one-ride model says."
The calculation leaves out the law, honesty, social cost, repeat penalties, time, stress, and every consequence that does not fit into the two dollar amounts on the napkin. A model can explain why a temptation exists without endorsing the person who follows it.
Even at exactly 5 percent, where the two branches have the same expected monetary cost, they do not create the same experience. Paying is a small, predictable loss. Not paying is usually uneventful until, on one particular morning, it is not.
My wife looked at the napkin and said, "An average never has to take a morning off work to sit in a courtroom. You do."
I drew a box around her sentence. It was not part of the equation, but it belonged in the decision.
The lesson you take off the train
An average never has to take a morning off work to sit in a courtroom. You do.
By then the tea was cold, and the rain had softened to a whisper. The train had almost disappeared from the conversation, which is how I knew the example had done its job.
Expected value compares uncertain choices by multiplying each outcome by its probability and adding the results. It describes the long-run average inside the model, not a promise about the next ride.
Threshold probability asks a different question: at what probability do the preferred choices switch?
For a simple decision, the method can fit into three questions:
- What does the certain option cost?
- What can happen on the uncertain option, and with what probability?
- At what probability do their expected values become equal?
If every plausible probability stays on one side of the threshold, the decision is robust. If the plausible range crosses it, the decision is sensitive, and you have found the uncertainty worth investigating.
This is an expected value real-life example precisely because the arithmetic is simple while the human decision is not.
The mathematics does not know Toronto. Replace the fare, the fine, and even the subject of the choice. The structure survives: model the branches, find where they tie, and ask whether reality could plausibly cross that line.
Seeing the crossing in a real decision tree
A napkin is enough for one choice, one uncertain event, and two clean outcomes. Real decisions are less polite. They split again, carry several costs at once, and hide the probability you care about deep inside the tree.
In Decision Tree Pro, you can build the same structure, then use the Sensitivity Analyzer to vary p across a range. The software draws the competing expected values and makes the crossing point visible instead of leaving it buried in algebra. In a larger model, the sensitivity index helps separate inputs that can truly change the decision from inputs that merely look important.
The sensitivity analysis guide shows the same logic with a job-offer decision. When observations later arrive, the Bayesian inference tools can update the uncertain probability without pretending that a small sample is certainty.
The next morning, the platform was still wet. The young man from the 7:42 slowed beside the card reader, just as he always did. My wife tapped her own card, and the reader gave its ordinary green chirp. The mathematics had found the border. It had not made the choice for him.
A fictional kitchen-table story built around a real probability calculation and an announced $200 first-offence fine.