From a Prediction to a Decision
A regression line can estimate a response for a particular value of an explanatory variable. A company or agency might use that prediction to plan staffing, prepare supplies, or decide whether to investigate a possible problem. The prediction can inform a decision, but it does not make the decision automatically. Someone still needs to choose an action, decide what counts as important, and consider the limits of the model.
In “Units and Practical Significance,” we translated a slope into predicted response changes that matter in context. Here, we use a particular predicted response to consider an action. A useful decision explanation connects the prediction to the decision’s goal and threshold, then describes how much confidence the context warrants. It also avoids implying that a fitted line guarantees an outcome or proves that changing the explanatory variable causes a change in the response.
A prediction is not the same as an observed response. If a model predicts 136 visits, that does not mean exactly 136 visits will occur. Individual outcomes vary around the fitted line. As discussed in “Using \(s\) to Describe Prediction Accuracy,” the residual standard deviation \(s\) describes the typical size of prediction errors in the response variable’s units. It does not say that every error will be within \(s\), or establish a guaranteed range for a future response.
The action threshold comes from the situation, not from the regression calculation. For example, an agency might decide to add staff when a predicted number of visitors exceeds its current capacity. A company might investigate a production process when the predicted number of defective items reaches a level that makes investigation worthwhile. The model supplies a prediction; people responsible for the decision set the threshold and weigh practical consequences.
A Decision-Making Checklist
Before using a prediction, ask what decision it is meant to inform. Then check whether the prediction is relevant to that decision and whether the model is suitable for the cases at hand. The following steps provide a practical structure for explaining a model-based decision.
Name the action being considered and what response or outcome matters to it.
Confirm that the explanatory-variable value is within the observed range and that the cases and conditions are relevant to the decision.
Substitute the explanatory-variable value into the fitted line, and report the predicted response in context and with units.
Explain whether the prediction falls above, below, or near the threshold, and how that comparison relates to the proposed action.
Consider typical prediction error, residual patterns, how the data were collected, and whether the relationship or conditions might differ when the decision is made.
The checks in “Residuals and Outliers” and “Reading Residuals From Computer Output” can help identify observations or patterns that deserve attention. A strong \(r^2\) does not establish that a prediction is accurate enough for a particular decision; as covered in “\(r^2\) Is Not Percent Correct Predictions,” it is not a percentage of correct predictions. Use fit summaries for what they describe, and keep the decision’s actual consequences in view.
Worked Examples: Applying Predictions Carefully
Worked Example: Planning Staff at a Cooling Center
A fictional city agency uses a regression model to help plan staffing at a cooling center. Each case in the data is a day the center was open. The explanatory variable \(x\) is the afternoon temperature in degrees Celsius, and the response variable \(y\) is the number of visitors that day. The fitted line is \(\hat{y}=8+4x\). Temperatures in the data ranged from 25°C to 35°C, and the residual standard deviation was \(s=18\) visitors. For a particular staffing decision, the agency’s stated trigger for adding staff is a prediction of more than 130 visitors. A day’s temperature is forecast to be 32°C.
State. The agency must decide whether to add staff for the forecast day. Its stated rule is to add staff when the model predicts more than 130 visitors.
Plan. Use the fitted line to predict the number of visitors at 32°C, then compare the prediction with 130. The value 32°C is within the observed temperature range of 25°C to 35°C, so this is not extrapolation beyond the model’s recorded \(x\)-values. The prediction is relevant only to the extent that the forecast day and the center’s circumstances resemble the days represented in the data. The residual standard deviation is useful context for the typical size of prediction errors, but it is not a guaranteed error limit.
Do. Substitute 32 for \(x\):
The predicted number, 136 visitors, is 6 above the 130-visitor trigger. The typical residual size of 18 visitors is larger than that 6-visitor margin. This comparison does not calculate a probability that attendance will exceed 130; it does show that the prediction is relatively close to the trigger compared with the model’s typical prediction-error scale.
Conclude. The prediction is above the agency’s stated trigger, so the model supports adding staff under that rule. Because the prediction is only 6 visitors above the trigger while typical prediction errors are about 18 visitors, the agency should not describe 136 as a certain count or treat the threshold comparison as decisive by itself. It could consider the consequences of understaffing versus adding staff, and use other relevant information if available.
This model describes an association between temperature and visits on the recorded days. It does not show that temperature alone causes a particular number of people to visit, and it does not guarantee the count for this future day. The agency should also consider whether the center’s hours, outreach, or other conditions are comparable to those in the data.
Worked Example: Deciding Whether to Inspect a Production Process
A fictional manufacturer models the number of rejected items in a batch, \(y\), using the machine’s operating time since its last service, \(x\), measured in hours. The fitted line is \(\hat{y}=1.5+0.8x\). The observed operating times ranged from 2 to 10 hours. The company uses a predicted 10 or more rejected items per batch as a trigger to inspect the process. A machine has been running for 8 hours.
The value 8 hours is inside the observed range. The model’s prediction is
The prediction is 2.1 items below the company’s trigger, because \(10-7.9=2.1\). If the company applies its stated rule strictly to the model prediction, it would not inspect the process based on this prediction alone. That is a conditional recommendation based on the company’s chosen trigger; it is not proof that the batch will have fewer than 10 rejected items.
The decision also depends on the consequences of errors. If missing a process problem is very costly, the company might choose to inspect at a lower predicted count. If an inspection is costly and the consequences of a short delay are small, it might keep the higher trigger. These are operational judgments, not facts determined by the regression line. The company should also review the residual scatter and residual plot, as discussed in earlier tutorials, before relying on the model.
Suppose someone instead uses the line at 11 hours:
Although 10.3 is above the trigger, 11 hours is beyond the observed range of 2 to 10 hours. This is extrapolation. The calculation is correct algebraically, but the fitted relationship may not continue in the same way beyond the data. The company should not treat that prediction as dependable just because it crosses the trigger; it may need data from that operating-time range.
Conclusion. At 8 hours, the model predicts 7.9 rejected items, below the stated inspection trigger of 10. That prediction can inform the decision, but it does not guarantee the batch’s outcome. A prediction at 11 hours is outside the observed range and needs an explicit extrapolation warning.
Worked Example: Preparing Trucks for River Cleanup
A fictional environmental agency models cleanup debris, \(y\), in tons, using rainfall \(x\), measured in millimeters during the previous day. Each case is one cleanup day. The fitted line is \(\hat{y}=0.6+0.35x\), and the recorded rainfall values ranged from 10 to 50 millimeters. The agency is considering dispatching an additional truck when predicted debris is at least 8 tons. The forecast rainfall is 20 millimeters.
The forecast value of 20 millimeters is within the observed range. The predicted debris amount is
The prediction is 0.4 ton below the 8-ton dispatch threshold. If the agency follows the threshold using the prediction alone, it would not dispatch the additional truck. But that comparison is close: the threshold is only 0.4 ton higher than the predicted amount. Before making the decision, the agency should consider the model’s typical prediction error in tons, if known, and whether the residuals show a pattern that could make predictions less dependable at this rainfall level.
The cases in the model are cleanup days, so the line predicts debris for a day under conditions represented by those cases. It does not predict the exact debris amount at every location along the river. Nor does it prove that rainfall alone caused the amount of debris: other factors, such as river flow or recent maintenance, could be related to both rainfall and debris. If a truck decision concerns a particular location rather than an overall daily amount, the agency should make sure the model’s response matches that decision.
Conclusion. At 20 millimeters of rainfall, the model predicts 7.6 tons of debris, slightly below the 8-ton threshold. That prediction is one useful input, but its small margin from the threshold and the fit of the model in the relevant setting matter. The agency should not present the model as a guarantee that debris will stay below 8 tons.
Limits to Mention When a Decision Is at Stake
A model-based recommendation is stronger when it names the conditions that could make the prediction less useful. The relevant limits depend on the situation, but several questions are worth asking consistently.
- Is the prediction within the data’s range? Predictions for explanatory-variable values beyond those used to fit the line are extrapolations. The fitted line may not continue to describe the relationship there.
- Do the cases match the decision? As discussed in “Group Averages Versus Individual Predictions,” a model fitted to group averages predicts group averages, not individual outcomes. More broadly, a model about one kind of case does not automatically apply to a different kind of case.
- How large is typical prediction error? Compare \(s\), in response units, with the prediction’s distance from the decision threshold. This comparison describes scale; it does not give a guaranteed error range or a probability of crossing the threshold.
- Do residuals suggest a problem? A curved pattern, changing scatter, or an unusual observation can signal that a simple linear model may not describe all parts of the data equally well. A single \(r^2\) value does not show these details.
- How were the cases obtained? “Population Scope and Generalizing Results” and “Sampling Method and Reliability of a Model” explain why a model that fits a convenience sample may not generalize well to a broader group.
- Have conditions changed? The relationship in earlier data may not apply after a change in equipment, policy, customer behavior, weather patterns, or measurement methods. A model may need to be checked against newer data.
- Would acting on the predictor change the response? As explained in “Association Versus Causation in Regression,” an observational association does not by itself establish that changing \(x\) causes a change in \(y\). Use “the model predicts” rather than claiming the explanatory variable will cause the response to change.
Do not make every answer a long list of generic cautions. Identify the limits that matter to this decision and explain why they matter. If the prediction is well within the observed range and far from the action threshold, one concern may be less urgent than when the prediction is near the threshold. A close comparison deserves particular care when typical errors are large relative to the margin.
Common Mistakes and AP Exam Tips
A complete response shows the connection between the model and the proposed action, while staying precise about what the model supports.
- Calling the prediction an observed fact. “There will be 136 visitors” overstates the model. Say “the model predicts 136 visitors” and acknowledge that the actual count may differ.
- Treating a threshold as a statistical law. A trigger such as 130 visitors is a practical rule chosen for a purpose. The line does not prove that one extra visitor makes an action necessary.
- Ignoring the margin and typical error. Saying only that a prediction is above a threshold can hide that it is barely above it. Compare the margin with \(s\) when \(s\) is provided, without turning that comparison into a probability claim.
- Using \(r^2\) as a prediction guarantee. \(r^2\) describes the fraction of response variation accounted for by the linear relationship in the data. It is not the percentage of future predictions that will be correct.
- Forgetting the cases and units. State what one case represents and name the predicted response in its units. Do not turn a prediction for days, sites, or groups into a claim about individual people or locations without justification.
- Making a causal recommendation from an association. A model may help forecast a response without showing that deliberately changing the explanatory variable will produce the predicted change.
- Leaving out extrapolation or changing conditions. Check the observed \(x\)-range and consider whether the future setting resembles the data used to fit the line.
For full credit, calculate the predicted response correctly, interpret it in context with units, compare it with the stated decision criterion, and give a recommendation that follows from that comparison. Then qualify the recommendation with the most relevant limits—for example, the prediction’s margin from the threshold, typical residual size, observed range, or population scope. Avoid claiming certainty or causation unless the study design and evidence support those claims.
Check Your Understanding
For each question, connect the fitted-line prediction to a decision and state an appropriate limit.
- A fictional library models daily visitors \(y\) from the maximum temperature \(x\), in degrees Celsius, using \(\hat{y}=20+5x\). The library considers adding staff above a prediction of 160 visitors. Find and interpret the prediction when \(x=28\), then compare it with the threshold.
- In question 1, the observed temperatures ranged from 18°C to 30°C. Is a prediction at 32°C an extrapolation? Explain what that means for using the prediction.
- A company’s model predicts 52 orders, compared with an action threshold of 50 orders. The residual standard deviation is \(s=9\) orders. What does comparing the margin with \(s\) tell you, and what does it not tell you?
- A model fitted to average daily visitor counts for several stores is used to predict a particular customer’s behavior. What is the mismatch between the model’s cases and the proposed prediction?
- An observational regression predicts that higher advertising spending is associated with higher sales. Why would it be too strong to conclude from the line alone that increasing advertising will cause the predicted sales increase?