A Probability Model Is a Deliberate Simplification
A probability model describes the possible outcomes of a chance process and assigns a probability to each outcome. In Exam Practice on Normal Distribution Problems, a normal model helped us answer questions about a measured variable. Here, we step back and ask a more general question: why should we trust a model to represent the situation in the first place?
A model is not the chance process itself. It is a simplified description built for a particular purpose. It might represent one spin of a spinner, one draw from a container, or the response from one randomly selected user. A model can be useful when its outcomes match what can happen, its probabilities are reasonable for the process, and its assumptions are appropriate for the question. It need not describe every detail of reality; it does need to preserve the details that matter to the question being asked.
The full collection of outcomes a model considers is called its sample space. Outcomes may be categories, such as “clicked” and “did not click,” or values, such as the number of blue objects drawn. Naming the sample space helps make the model’s scope clear. For example, a model for the color of one draw does not automatically describe the number of blue objects in a series of draws.
To decide whether a model fits, ask what produces the randomness and where its probabilities come from. If a physical mechanism is designed to make outcomes equally likely, a theoretical model may assign equal probabilities, provided the relevant mechanism behaves as assumed. If probabilities come from collected data, the observed relative frequencies can inform a model for similar future cases—but only if the data and future cases are sufficiently comparable.
Make the Model’s Assumptions Visible
An assumption is a condition we treat as true in order to use a model. Some assumptions are about the mechanism: for example, that a spinner is balanced and every section is equally likely. Others are about how a process behaves over time: for example, that the chance of a response remains stable from one session to the next. A model that leaves its assumptions unstated can sound more certain or more general than it really is.
The question matters, too. A model might be adequate for describing one randomly selected event but not for predicting a long sequence of events. If the model assumes that events do not affect one another, then using it for a process where one result changes the next result could give misleading probabilities. The earlier tutorials Recognizing a Binomial Setting and Spotting Non-Binomial Situations discuss this issue for binomial models; the broader principle is that assumptions must match the process.
- What is one outcome, and what outcomes does the model include?
- What feature of the mechanism or data supports the assigned probabilities?
- Which assumptions are being made about fairness, stability, or dependence?
- Does the question concern one event or a process repeated over time?
- What features of the real situation are important enough that leaving them out could change the answer?
These questions are not a guarantee that a model is exactly true. They are a way to judge whether it is a reasonable tool for the specific question. A careful conclusion should therefore refer to what the model predicts, rather than presenting its result as a certainty about every real-world case.
Worked Example: A Spinner with Unequal Numbers of Sections
Worked Example: A Spinner with Unequal Numbers of Sections
A fictional classroom spinner has four equal-sized sections: two red, one blue, and one green. A student asks for the probability that one spin lands on red or green. Consider a model that treats the four sections as equally likely. Is it appropriate, and what probability does it give?
State. The chance process is one spin. The possible color outcomes are red, blue, and green. The model assumes that the pointer is equally likely to land on each of the four equal-sized sections and that the spin ends in one of those sections.
Plan. The model should reflect the physical layout, not treat the three colors as equally likely just because there are three color names. Since red occupies two of the four equal-sized sections, it has twice as many sections as either blue or green. The model is reasonable if the spinner is balanced and the pointer does not favor particular sections.
Do. Under those assumptions, the color probabilities are based on the number of sections of each color out of four:
The event “red or green” includes two red sections and one green section, or three of the four sections. Thus:
Conclude. If the spinner is balanced and each equal-sized section is equally likely, the model gives a 0.75 probability that one spin lands on red or green. The model would need reconsideration if the pointer were biased, the sections were not actually equal-sized, or the spinner could stop outside the colored sections.
Probabilities from a Mechanism or from Data
The spinner example used a theoretical probability model: the proposed probabilities came from the physical design of the spinner and an assumption about how it behaves. Other models use observed relative frequencies. Those two sources of probabilities have different limits.
An observed relative frequency describes what happened in a particular set of cases. It can be used as a reasonable estimate of a probability for similar cases, but it is not a promise that the same proportion will occur in the next group. The model’s usefulness depends on whether the cases that produced the data resemble the cases to which the model will be applied. A change in the users, setting, or time period may make the old probabilities less appropriate.
It is also important to distinguish a model for one randomly selected case from a model for a sequence. Repeating a single-case model requires additional assumptions. In particular, if the chance of an outcome changes after earlier events, or if the events are connected, the single-case probabilities alone may not describe the sequence. In a binomial setting, the earlier BINS checklist provides a structured way to assess the relevant requirements rather than assuming them automatically.
Worked Example: Using Past Website Sessions
Worked Example: Using Past Website Sessions
A fictional website team reviews 1,200 past sessions from a particular version of its site. In 216 sessions, the visitor clicked a new help button; in the other 984 sessions, the visitor did not. The team wants a model for the click outcome in one randomly selected session on the same version of the site. Is a model based on these data reasonable, and what assumptions should the team state?
State. Let the outcome for one selected session be either “clicked the help button” or “did not click the help button.” The proposed model uses the observed proportions from the 1,200 past sessions as probabilities.
Plan. The recorded sessions provide direct evidence about the click rate for those sessions. To use that rate as a model for another session, the team must judge whether the future session comes from a comparable group of visitors and whether the site design and circumstances remain similar. The model is for one session; using it to describe multiple sessions would require additional assumptions about how those sessions are selected and related.
Do. The observed proportion of clicks is \(216/1200=0.18\), and the observed proportion of non-clicks is \(984/1200=0.82\). The proposed model is therefore:
The counts agree with the total number of sessions: \(216+984=1200\). The model’s outcome categories fit the stated question because each session is classified as a click or no click. The numbers come from the observed sessions, not from a claim that the button has a fixed chance for every visitor in every circumstance.
Conclude. This is a reasonable starting model for one randomly selected session if the new session is comparable to the 1,200 recorded sessions and the site version and visitor mix have not changed substantially. It may be inappropriate for sessions after a redesign, for a different group of visitors, or for a question about a series of connected sessions without further assumptions. The model predicts a 0.18 click probability under its stated conditions; it does not guarantee that exactly 18% of any particular future group will click.
When the Process Changes the Next Outcome
A common model-fit problem occurs when a model treats repeated outcomes as if earlier outcomes have no effect, even though the real process changes after each outcome. Consider drawing objects from a container without replacing them. After the first draw, the container’s contents have changed. The probability for the next draw may therefore differ, and a model that assumes the same chance on each draw may not represent the process.
This does not mean every model for repeated draws is inappropriate. It means that its assumptions need to reflect whether objects are replaced and how many are available. The earlier tutorial Spotting Non-Binomial Situations explains why drawing without replacement can violate conditions for a binomial model. In the next example, the point is to compare a model that matches the actual process with one that ignores it.
Worked Example: Drawing Two Tokens Without Replacement
Worked Example: Drawing Two Tokens Without Replacement
A fictional bag contains five blue tokens and three yellow tokens. A student draws two tokens one after the other without replacing the first token. The question is the probability that both tokens are blue. Compare a model that follows the stated drawing process with a model that incorrectly assumes the blue probability stays at \(5/8\) on both draws.
State. The outcomes for the event of interest are “blue on both draws” and “not blue on both draws.” The actual process is drawing without replacement, so the contents of the bag change after the first token is drawn.
Plan. For a model that fits the process, the probability of a blue second token depends on the first draw. For the competing model, identify its assumption: it uses the original blue proportion, \(5/8\), as the probability on both draws. That assumption would be more plausible if the first token were replaced and the contents mixed before the second draw, but replacement is not part of the stated process.
Do. Under the actual without-replacement process, the probability of blue on the first draw is \(5/8\). Given a blue first draw, four blue tokens remain among seven total tokens, so the probability of blue on the second draw is \(4/7\). Therefore:
The model that incorrectly keeps the first-draw probability unchanged would calculate:
The second calculation is larger because it ignores the fact that a blue first draw leaves one fewer blue token in the bag. The difference between the two model results is approximately \(0.3906-0.3571=0.0335\), rounded to four decimal places.
Conclude. The without-replacement model is appropriate for the stated process and gives a probability of about 0.3571 that both tokens are blue. The unchanged-probability model does not fit the process as described because the first draw affects the contents available for the second draw. It could describe a different process, such as drawing with replacement, but not this one.
Common Mistakes and AP Exam Tips
Model selection is a reasoning task, not simply a choice of calculator command. Naming the model is not enough; a strong response explains why its description matches the situation and what assumptions support using it.
- Making outcomes equally likely without a reason. Outcomes are not automatically equally likely just because they are listed together. In the spinner example, colors were not equally likely because the colors occupied different numbers of equal-sized sections.
- Leaving out a possible outcome. If the model omits an outcome that can occur and matters to the question, it may not describe the process adequately. State what the model counts or classifies, and consider whether the categories cover the relevant possibilities.
- Treating past relative frequencies as guarantees. An observed proportion can inform a model for comparable future cases, but it does not ensure the same proportion in a new group. Explain why the future cases are similar enough for the model to be useful.
- Using one-event probabilities for a sequence without checking assumptions. A repeated-process model may require stable probabilities or independence. If the process changes after an outcome, say so and choose a model that reflects that change.
- Calling a model “correct” without naming its conditions. A more careful response says that a model is reasonable under stated assumptions, such as a balanced spinner or comparable website sessions. Models are useful representations, not proofs that reality follows them exactly.
- Ignoring the question’s scope. A model appropriate for one randomly selected session may not automatically answer a question about many sessions. Identify the event or process the question concerns before deciding which assumptions matter.
For full credit on a question asking whether a model is appropriate, identify the relevant outcomes, connect the probabilities to the mechanism or data, and state the assumptions that support applying the model. If an assumption does not fit, explain what part of the process conflicts with it. Avoid vague statements such as “the events are random”; describe how the process creates the randomness and whether the model’s conditions are plausible.
Key Takeaway
A probability model is useful when it represents the outcomes and chance mechanism relevant to the question. Its probabilities may come from a design or from data, but either way, the assumptions must be plausible for the situation. A good model statement makes those assumptions visible and keeps conclusions within the model’s intended scope.
Check Your Understanding
For each situation, identify the model-fit issue and state what assumption or evidence would matter.
- A balanced-looking spinner has three equal-sized sections labeled 1, 2, and 3. What assumption would support treating the three outcomes as equally likely?
- A fictional library finds that 30% of visitors in last month’s records borrowed a laptop. What would make that relative frequency a reasonable starting model for a randomly selected visitor this month?
- A container holds four orange and six purple tiles. A tile is drawn, kept out, and then a second tile is drawn. Why might using the original probability of orange on both draws fail to fit the process?
- A model includes only “yes” and “no” for a survey response. What should you consider if respondents can also skip the question?
- A model is designed for one randomly selected website session. What additional concern arises if it is used to describe the number of clicks in many sessions?