Did you know? You can have two mathematically valid regression lines for the same paired data, yet only one is appropriate for the prediction you were asked to make. One predicts Y from X; the other predicts X from Y. In general, reversing a fitted equation algebraically is not the same as fitting the other regression line.
This JC2 H2 Mathematics tutor guide for Bukit Timah readers answers a narrow but high-value A-Level 9758 question: Which regression line should I use to predict the unknown variable? The answer begins with reading the direction of the question, rather than picking the first equation the graphing calculator displays.
The 2026 H2 syllabus includes correlation, linear regression, interpolation and extrapolation, and use of the appropriate regression line for predictions. This page focuses on the prediction-direction mistake. For a broad overview of correlation and the full topic, read H2 Mathematics Correlation and Linear Regression.
Quick answer: predict the response, not the input
- Given X, estimate Y? Use the regression of Y on X; it predicts a Y value from an X value.
- Given Y, estimate X? Use the regression of X on Y; it predicts an X value from a Y value.
- Do not: take a Y-on-X equation, rearrange it and assume it has become the X-on-Y regression equation.
- After calculating: check the units, whether the prediction is inside the observed range, and whether a linear model is reasonable.
What a regression line is trying to make small
A Y-on-X least-squares fit measures the vertical differences between observed Y values and the fitted Y values. It chooses a line to reduce the sum of squared vertical residuals. An X-on-Y fit instead treats X as the variable being predicted and reduces the corresponding X residuals. Those are different optimisation questions whenever the data do not sit exactly on one straight line.
It is easy to miss this because an algebraic line can be written as either y = a + bx or x = c + dy. But the two least-squares fitting problems use different residual directions. Rewriting symbols does not rerun the fitting process.
Worked example: the two answers are not identical
Imagine a fictional paired-data exercise summarised as follows: the mean of X is 10, the mean of Y is 20, the standard deviations are 4 and 6 respectively, and the correlation is r = 0.75. The data show a positive association but are not perfectly linear. Use these as exercise summaries, not as measurements collected from Bukit Timah students.
For this exercise, the Y-on-X regression line is ŷ = 8.75 + 1.125x. The slope is positive and the line passes through the mean point (10, 20). If the task says estimate Y when X = 14, substitute into this line: 8.75 + 1.125(14) = 24.5. If the original Y unit were minutes, the prediction would be about 24.5 minutes.
The X-on-Y regression line is instead x̂ = 0.5y. This line also passes through (10, 20), but it solves a different least-squares problem. If the question says estimate X when Y = 26, calculate x̂ = 0.5(26) = 13. The fact that both equations pass through the means does not make them interchangeable.
Why simply reversing the first line fails
Suppose a student takes ŷ = 8.75 + 1.125x and rearranges it to x = (y − 8.75)/1.125. For y = 26 this gives approximately 15.3, not the correct X-on-Y fitted prediction of 13. This inversion asks which X value lies on the Y-on-X fitted line at that Y. It is not the regression fit that was designed to predict X.
When paired points are scattered around a line rather than perfectly on it, the fitted predictions are asymmetric. The useful habit is not to memorise an extra formula; it is to identify which quantity is known and which quantity must be estimated, then read the corresponding regression equation from the calculator or question.
Three checks that turn a number into a good answer
- Direction: Was the stated task Y from X or X from Y?
- Units: Does the reported estimate have the unit of the predicted variable?
- Range: Is the supplied predictor value in the region where data were observed? If not, label the result as extrapolation and explain its uncertainty.
The prediction is more credible when a scatter diagram suggests a roughly linear relationship and the requested predictor lies in the observed range. A correlation coefficient close to +1 or −1 supports a stronger linear relationship but does not prove causation; a high value of |r| is not permission to extrapolate indefinitely.
What a residual means — and what it does not
Suppose, in a separate fictional case, X = 14 and the actual observed Y is 22 while the Y-on-X fitted prediction is 24.5. The residual in Y units is 22 − 24.5 = −2.5. The negative sign means the observed Y falls 2.5 below the predicted Y. It does not mean that the equation must be shifted until every residual disappears.
Residuals can help a student recognise curvature or changing scatter that a straight-line model hides. Their meaning depends on which quantity has been predicted. A residual in Y units is not directly the same thing as a residual in X units.
When the graph becomes straight only after transformation
H2 Mathematics may involve a square, reciprocal or logarithmic transformation to achieve linearity. For example, if an exercise plots ln Y against X and finds ln Ŷ = a + bX, the original-scale prediction is Ŷ = e^(a+bX). Reporting a + bX as though it were measured in original Y units would be a scale error.
Always label the axes and transformed variable: y, ln y, y² or 1/y are different outputs. A transformed regression may describe the relationship better over the observed range, but it also requires correct back-transformation before an interpretation in the original units. Use only transformations and variables the particular question permits.
One-minute JC2 diagnostic
Without showing any statistics, ask: “I know a student’s practice time and want to estimate a task score. Which line?” The response should be score on practice time. Next ask: “I know a target task score and want to estimate the practice time associated with it. Which line?” The response should be practice time on score. The second question is not solved by casually reversing the first line.
Now give the two equations in the worked example but hide their labels. Ask which output each equation is designed to predict and whether an answer of 15.3 from algebraic inversion should be accepted. A learner who can justify the direction, not merely calculate quickly, has understood the structure.
Five preventable exam errors
- Using the first calculator equation: read both variable labels and the prompt before selecting.
- Algebraically inverting a fitted line: fit or retrieve the required regression direction instead.
- Predicting in the wrong unit: check the predicted variable and its scale.
- Ignoring the scatter diagram: a fitted line is not proof the data are close to linear.
- Treating extrapolation as an observed fact: a model prediction outside the data range has extra uncertainty.
Eight quick practice prompts and answers
- 1. Given height, estimate mass: use mass on height.
- 2. Given mass, estimate height: use height on mass.
- 3. With ŷ = 8.75 + 1.125x, x = 10 gives ŷ = 20.
- 4. With ŷ = 8.75 + 1.125x, x = 14 gives ŷ = 24.5.
- 5. With x̂ = 0.5y, y = 26 gives x̂ = 13.
- 6. Does rearranging the first line at y = 26 give the X-on-Y regression? No.
- 7. If observed Y = 22 and predicted Y = 24.5, residual Y − Ŷ = −2.5.
- 8. If the fitted response is ln Y = 1 + 0.2X, X = 0 gives original-scale prediction Y = e ≈ 2.718, not 1.
How a tutor makes the choice automatic
For a JC2 H2 Mathematics learner, begin with two arrows: given X → find Y and given Y → find X. Cover the equations. Ask the learner to explain what would count as the predictor and response. Only then reveal the fitted equations or graphing-calculator output. The next pass changes the variable requested, not the data: it is a more effective transfer exercise than ten repeated substitutions into one line.
A final explanation in natural language matters. A technically accurate estimate that confuses its units, pretends the line is causal or extrapolates beyond the data without warning is not a complete interpretation.
Frequently asked questions
Are the two regression lines ever the same?
They coincide as a single geometric line when the data have a perfect linear relationship, apart from degenerate cases where regression may be undefined. Otherwise, they generally differ.
Can a line of best fit prove one variable causes the other?
No. Linear association alone does not establish causation. Confounding variables and the study design matter.
Why does the graphing calculator show two line options?
The two directions answer different prediction questions. Calculator menus vary by approved model; identify the dependent quantity rather than memorising screen positions.
Should I always trust a prediction within the sample range?
No. Interpolation avoids the extra risk of extrapolation, but a poor linear fit, unusual observations or unexamined assumptions can still make a prediction unreliable.
What does the 2026 H2 Mathematics syllabus expect?
It includes interpreting correlation, obtaining and using appropriate regression lines, interpolation and extrapolation, and transformations used to achieve linearity. Check your examination year and the exact question wording.
Continue learning: four distinct JC2 H2 questions
- Why Standard Error Falls by √n
- One-Tailed or Two-Tailed Hypothesis Test?
- Which Regression Line Predicts Correctly? — this guide
- Line–Plane Angle vs Normal Vector
For the wider subject route, read the Bukit Timah JC H2 Mathematics guide and the Bukit Timah JC2 learning hub. These are subject and level overviews; this article owns only the specific question explained above.
Syllabus and scope: SEAB Singapore–Cambridge H2 Mathematics (9758), 2026 examination syllabus. The numbers in worked questions here are invented teaching examples, not past examination questions or actual student records. Refer to the syllabus for your examination year.