From a Chance Question to a Trial
A question such as “How often are deliveries late?” sounds straightforward, but it leaves important details open. Does one trial mean one package, one delivery route, or one day of deliveries? How late must a delivery be to count as late? Is the question about a randomly chosen package or about whether a day includes any late package? Each choice defines a different probability question.
A simulation can only answer the question represented by its trial rules. Before assigning outcomes or repeating an experiment, specify exactly what is observed once, what results are possible, and which result or results count as the event of interest. The same discipline applies when describing a probability model, even if no simulation is being run.
For example, if one trial is the inspection of a single light bulb, its outcome might be “defective” or “not defective.” The event “the bulb is defective” contains one of those outcomes. If outcomes are more detailed—perhaps “cracked,” “does not light,” or “passes inspection”—an event may contain several outcomes. The outcomes must be defined so that one trial produces exactly one of them.
What a Trial Definition Needs
A useful definition should be specific enough that two people following it would identify the same trial and classify its result in the same way. In practice, state:
- The unit: one person, item, appointment, day, or other observational unit.
- The conditions: the setting or process under which that unit is observed.
- The observation window: when the trial starts and when it ends.
- The outcome rule: how to assign exactly one outcome to what happened.
- The event: the particular outcome or collection of outcomes relevant to the question.
A trial must also be consistent with the question’s target. “A randomly selected item from this lot is defective” and “a randomly selected box contains at least one defective item” use different units. A product can be a trial unit in the first question and a box in the second. Neither definition is automatically better; the appropriate one depends on what the probability is intended to describe.
Worked Example: Defining “Late” for One Workday
Suppose a manager asks for the chance that an employee arrives late. For a hypothetical model, define one trial as observing one scheduled employee on one specified workday. The shift begins at 9:00 a.m.; arrival at or before 9:00 is on time, and arrival after 9:00 is late. Assume the model assigns probability 0.12 to a late arrival on a randomly selected employee-workday.
Sample space. With the rule above, \(S=\{\text{on time},\text{late}\}\). These outcomes cannot both occur for the same trial, and the rule assigns every arrival time to one of them.
Event. Let \(L\) be the event that the employee arrives late. It contains the outcome “late,” so under the stated model \(P(L)=0.12\). The probability is for a randomly selected employee-workday—not necessarily for a randomly selected employee’s entire workweek.
Now compare a different question: what is the chance that the same employee is late on at least one of five scheduled workdays? That trial spans five days, and the event is “late on one or more days.” If we additionally assume each day’s lateness indicators are independent and each has probability 0.12, then
Thus the one-day probability is 0.12, while the five-day probability under the additional independence assumption is about 0.4723. The change is not a contradiction: the trial and event have changed. If the daily outcomes are not independent, the five-day probability cannot be calculated from 0.12 alone using this formula.
Outcomes Must Cover the Trial
An outcome list is complete when every possible result of the defined trial belongs to one of its categories. The categories should also be mutually exclusive: a single trial should not produce two different outcomes at once. These properties make the sample space a reliable basis for defining events.
Suppose appointment outcomes are classified as “on time” and “late,” but some patients cancel or do not arrive. Those two categories do not cover every booked appointment. One solution is to include cancellation and no-show as additional outcomes. Another is to restrict the trial to appointments that were attended. The choice depends on the question, but it must be made explicit because it affects which observations belong in the probability calculation.
Once the outcomes have been defined, an event is formed by selecting the outcomes that satisfy the question. For instance, in a sample space containing “on time,” “late,” “cancelled,” and “no-show,” the event “the appointment was not attended” contains the last two outcomes. The event is not a new kind of trial; it is a rule for grouping trial outcomes.
Worked Example: One Product or One Box?
Consider a hypothetical lot of 20 products, of which exactly 3 are defective and 17 are not. Assume a product is selected at random when a single product is the trial unit.
One-product trial. The outcomes are “defective” and “not defective,” and the event \(D\) is that the selected product is defective. Because each of the 20 products is equally likely to be selected, the model gives
This is a 15% probability that one randomly selected product from this lot is defective.
Now define a different trial: select four products from the lot without replacement, treating the four-product selection as one trial. Let \(B\) be the event that the selected group contains at least one defective product. The complement is that all four selected products are nondefective. Under random selection of a four-product group, there are \(\binom{20}{4}=4845\) possible groups, and \(\binom{17}{4}=2380\) groups containing only nondefective products. Therefore,
The probability is about 50.88% that a randomly selected group of four contains at least one defective product. The value differs from 0.15 because the trial is a group rather than one product, and the event asks whether the group contains any defect.
Specify the Denominator and the Event
A chance question often has an implicit denominator: which set of possible trials is being considered? The wording “among all booked appointments” includes cancellations and no-shows. The wording “among appointments that were attended” excludes them. Both can be legitimate, but they describe different probabilities. The trial definition should make the denominator clear before any count or probability is reported.
Worked Example: Late Arrivals Among All Appointments or Attended Appointments?
Suppose a hypothetical clinic records 100 booked appointments. Of these, 70 patients arrive on time, 12 arrive more than 10 minutes after the appointment time, 8 cancel, and 10 do not show up. Define the possible outcomes for a booked-appointment trial as “on time,” “late by more than 10 minutes,” “cancelled,” or “no-show.” These four categories are disjoint and account for all 100 appointments.
If the trial is one randomly selected booked appointment and the event is “the patient arrives more than 10 minutes late,” the event occurs 12 times out of 100:
The proportion in this hypothetical record is 0.12, or 12% of booked appointments. Now restrict the trial to an appointment that was attended. There are \(70+12=82\) attended appointments, and 12 were late by more than 10 minutes. The corresponding proportion is
That is about 14.63% of attended appointments. The two calculations use the same 12 late arrivals but different trial populations and denominators. The first describes booked appointments; the second describes attended appointments. Neither answers the other question.
A Short Justification: Why the Trial Changes the Probability
The five-day lateness calculation illustrates a general way to compute the probability of an event defined as “at least one.” The complement of “at least one late day” is “no days are late.” If each day has late-arrival probability \(q\), then its on-time probability is \(1-q\). Under independence, the probability of being on time on every one of \(k\) days is the product of the \(k\) daily probabilities, \((1-q)^k\). Since an event and its complement have probabilities that sum to 1,
This result depends on both the definition of the event and the independence assumption. It does not mean that “late” has one universal probability regardless of how the trial is defined. It shows why a question about one day and a question about several days can have different answers even when the daily chance is held fixed.
Common Mistakes
- Leaving the time window unstated. “Late” could mean late on one shift or late at least once over a week. State the window and count the relevant event over that window.
- Changing the unit midway through. A probability for one product is not the probability for a box containing at least one defective product. Name the trial unit before calculating.
- Using a denominator that does not match the question. If the trial is a booked appointment, cancellations and no-shows are part of the sample space unless the definition explicitly restricts the trial to attended appointments.
- Leaving possible outcomes out. Listing only “on time” and “late” fails to describe all booked appointments if cancellations and no-shows can occur. Add categories or clearly restrict which trials are included.
- Treating an event as an outcome. “At least one defect in a group of four” is an event containing many possible detailed group outcomes, not one individual product outcome.
Check Your Understanding
For each question, identify the trial unit, a suitable sample space, and the event of interest.
- A school asks for the chance that a bus is more than five minutes late on a randomly selected weekday. What details would make the trial definition precise?
- A factory asks for the chance that a randomly selected package contains at least one damaged item. How does the trial differ from selecting one item?
- A survey records “responded,” “declined,” and “could not be reached.” Why would a sample space with only “responded” and “declined” be incomplete for a trial defined as one call attempt?
- In a hypothetical set of 50 booked appointments, 36 are attended on time, 6 are attended more than 10 minutes late, 4 are cancelled, and 4 are no-shows. Calculate the observed proportion late among all booked appointments and among attended appointments.
- Explain why a probability for “late on one workday” need not equal the probability of “late at least once during a five-day workweek.”