Choosing Explanatory and Response Variables

Start With the Question

A regression line is used to describe or predict one variable from another. Before fitting or using one, decide which variable is the explanatory variable and which is the response variable. The choice comes from the question being asked: what information will be used to predict, explain, or summarize what outcome?

For example, a researcher might ask how a student's study time is related to a quiz score, or might ask whether quiz scores can help predict study time. The same paired measurements can address both questions, but the prediction target changes. Consequently, the regression line changes too.

The roles are not determined just by which measurement was collected first, or by which variable is easier to measure. They should reflect the study question and the intended use of the line. In standard simple linear regression, both variables are quantitative. Their pairing matters: each \(x\) value must belong to the same individual or unit as its corresponding \(y\) value.

Roles Depend on the Prediction Target

Suppose a data set records daily outdoor temperature and a building's daily electricity use. If the question is “How does electricity use tend to vary with temperature?”, temperature is explanatory and electricity use is the response. If the question is instead “Can electricity use help predict the outdoor temperature?”, electricity use is explanatory and temperature is the response. The observations have not changed; the direction of prediction has.

Choosing an explanatory variable does not establish that it causes the response. A variable can be useful for prediction without being a cause. Even when a study question is phrased in terms of an effect, the design and evidence—not the labels \(x\) and \(y\)—determine whether a causal conclusion is justified.

Why Reversing the Roles Changes the Line

For paired quantitative observations \((x_i,y_i)\), the least-squares line predicting \(y\) from \(x\) has the form \(\hat{y}=a+bx\). Its slope is chosen to fit the response values using the explanatory values. If the roles are reversed, the fitted line instead predicts \(x\) from \(y\), written \(\hat{x}=c+dy\). In general, this is a separately fitted line, not the algebraic inverse of the first.

The distinction can be shown using the sample correlation \(r\), the sample standard deviations \(s_x\) and \(s_y\), and the standard slope formulas for simple least-squares regression. The slope predicting \(y\) from \(x\) is \(b=r(s_y/s_x)\); the slope predicting \(x\) from \(y\) is \(d=r(s_x/s_y)\). These formulas give a concise justification for why the two fitted directions are related but not generally interchangeable.

If the first line were simply the exact inverse of the second, their slopes would have product 1. But the product of the actual regression slopes is \(r^2\). Unless the observations lie exactly on a straight line, \(|r|<1\), so \(r^2<1\). Thus the fitted lines generally are not inverses. When the relationship is perfectly linear, \(|r|=1\), the product is 1 and the two fitted lines are exact inverses (provided the slope is nonzero).

Both fitted lines pass through the point \((\bar{x},\bar{y})\), but sharing that point does not make them the same line. Reversing the roles changes which variable's deviations from its mean are used to fit the predictions. That difference is why predicting scores from study hours is not equivalent to solving the scores-on-hours equation for hours.

Worked Fits in Both Directions

The following examples calculate both directions from paired data. To calculate the slope of \(y\) predicted from \(x\), use the sum of products of deviations divided by the sum of squared deviations in \(x\). For the reverse line, divide that same sum of products by the sum of squared deviations in \(y\). The intercept in each direction is chosen so its line passes through the sample means.

Choosing Roles Carefully

A practical way to decide the roles is to state the question as a prediction sentence: “Use ___ to predict ___.” The first blank is the explanatory variable and the second is the response. Then check that the paired data, units, and fitted line correspond to that sentence. If the question does not identify a prediction target, a regression direction has not yet been chosen.

Common Mistake: Treating the Reverse Line as an Inverse

A common error is to fit a line predicting \(y\) from \(x\), rearrange its equation to solve for \(x\), and call the result the regression line predicting \(x\) from \(y\). This assumes the fitted line predicts every observed \(y\) value exactly. Usually it does not; the reverse regression is fitted separately to minimize prediction errors in \(x\), rather than errors in \(y\).

For example, the quiz-score line above rearranges to \(\text{study hours}=(\text{score}-42.5)/7\), with slope \(1/7\approx0.1429\) hours per point. But the separately fitted reverse slope is \(0.14\) hours per point. They differ because the first equation is an algebraic rearrangement, not a new least-squares fit. In this particular small example the numbers are close, but closeness does not make the procedures equivalent.