Predictions Using Statistical Models

Earlier prediction skills asked you to anticipate what a character would do next using foreshadowing and inference from a story. This lesson reframes prediction in a numerical setting: using a statistical model built from real data to estimate a value you have not directly measured. The reasoning is the same critical habit applied to a new kind of evidence, so you weigh a data-based claim with the same scrutiny you once gave a plot clue before you decide whether to believe it.

When a scatter plot of two variables shows a roughly straight-line pattern, that pattern can be summarized with a regression model, often called a line of best fit. Its equation is written \( \hat{y} = mx + b \), where \(x\) is the input value, \(\hat{y}\) is the predicted output, \(m\) is the slope describing the rate of change, and \(b\) is the value where the line crosses the vertical axis. Once this equation is known, you can substitute any value of \(x\) and calculate a prediction without needing a new measurement.

Before trusting a regression model, you check how well it actually fits the data using the correlation coefficient, \(r\). This value always falls between \( -1 \) and \(1\). A value close to \(1\) means a strong positive relationship, a value close to \( -1 \) means a strong negative relationship, and a value near \(0\) means the two variables show little to no linear relationship at all. A strong \(r\) tells you the line is a good summary of the pattern, but it never proves that changes in \(x\) cause the changes in \(y\).

A prediction made for an \(x\)-value that falls inside the range of the collected data is called interpolation, and it is generally the safer type of prediction because the model was built from evidence covering that exact interval. A prediction made for an \(x\)-value outside that range is called extrapolation. Extrapolated predictions can be far less reliable, because nothing guarantees that the same linear pattern continues once you leave the observed data. The graph below shows one regression line used both ways.

Statistical Model: An equation built from data that describes the relationship between two or more variables.

Line of Best Fit: A straight line that best summarizes the trend in a scatter plot of data points.

Slope and Y-Intercept: The values \(m\) and \(b\) in \( \hat{y} = mx + b \) that describe the rate of change and starting value of a linear model.

Correlation Coefficient: A number between \( -1 \) and \(1\) that measures the strength and direction of a linear relationship.

Interpolation: Using a model to predict a value that falls within the range of the original data.

Extrapolation: Using a model to predict a value that falls outside the range of the original data.

Residual: The difference between an actual data value and the value predicted by the model.

A researcher collects data on hours studied, \(x\), and test score, \(y\), for a group of students, ranging from \(1\) to \(9\) hours studied. The regression model is \( \hat{y} = 5.5x + 62 \), with a correlation coefficient of \(r = 0.91\), indicating a strong positive relationship.

To predict the score for a student who studied \(6\) hours, substitute \(x = 6\): \( \hat{y} = 5.5(6) + 62 = 95 \). Because \(6\) falls within the original data range, this is an interpolation, and the strong \(r\)-value supports treating it as a reasonably reliable estimate.

Now predict the score for a student who studied \(13\) hours: \( \hat{y} = 5.5(13) + 62 = 133.5 \). This result is impossible for most test scales and falls far outside the studied range of \(1\) to \(9\) hours, so it is an extrapolation. The example shows why a strong correlation inside the data does not guarantee an accurate prediction once you go beyond it.

Reading a statistical prediction critically means asking several questions at once: How strong is the reported correlation coefficient? Does the predicted \(x\)-value fall inside or outside the range of the collected data? Could another variable explain the pattern instead of the one being tested? These same evaluative habits carry over when predictions are expressed in other numerical forms, such as Predictions Using Probability, or when a trend is projected forward in time, as in Predictions from Data Trends. In every case, a prediction is only as trustworthy as the evidence and reasoning that produced it.

Related lessons

• Making Predictions