Concept map / Linking

Share

Linking scores across assessments

Compare a regression prediction with a linear concordance between two related tests, and see why only one of them runs the same way in both directions.

Prerequisites RegressionNext topics Score reporting

1One Test X score, two Test Y answers

The same 120 students take Test X (mean 500, SD 100) and Test Y (mean 21, SD 5). A college that accepts either wants a table that turns an X score into a Y score. The two tests follow different blueprints, so their scores cannot be equated. They can be linked.

The green line is the regression of Y on X: the average Y score of students with a given X score. The gold line is a concordance: the Y score at the same percentile. Move the correlation and the student. The shaded bar shows where a new student’s Y score would likely fall.

20030040050060070080010152025303540Test Y scoreTest X scorer = 0.60
Regression of Y on XConcordance (same z score)95% prediction interval
Regression prediction25.5
Concordance28.5
95% prediction interval17.4 to 33.6

A student at 650 on Test X is 1.50 SD above the mean. Students with that score average 25.5 on Test Y, and the Y score at the same percentile is 28.5. They differ by 3.0 points.

2Predict where a round trip lands

Tests X and Y correlate 0.60. A student scores 700 on Test X. You predict their Test Y score with the regression of Y on X, then predict Test X from that Y score with the regression of X on Y. Where do they land on Test X?

3The two lines differ by a factor of r

The only difference between the formulas is the correlation r. Regression pulls each prediction toward the mean by that factor. That gives the smallest prediction error, but predicting X from Y needs a different line. Concordance matches z scores, so one line serves both directions, at the cost of larger prediction errors.

y^=yˉ+r sYsX (x−xˉ)y∗=yˉ+sYsX (x−xˉ)\hat y = \bar y + \htmlClass{term-r}{r}\,\frac{s_Y}{s_X}\,(x - \bar x) \qquad y^{*} = \bar y + \frac{s_Y}{s_X}\,(x - \bar x)

Now: ŷ = 21 + 0.60 × (5 / 100) × (650 − 500) = 25.50 and y* = 21 + (5 / 100) × (650 − 500) = 28.50.

How the prediction interval is computed

The 95% interval for one new student is ŷ ± t × s × √(1 + 1/n + (x − x̄)² / ((n − 1)s²ₓ)), with residual standard deviation s = 4.02 and t = 1.98 for 118 degrees of freedom: ± 8.06 points.

4Check yourself

The tasks use the sliders in section 1 and tick themselves when you meet them.

  • Move the student to at least 700 on Test X with a correlation of 0.60 or lower, so the concordance value sits at least 4 points above the regression prediction.Now: X = 650, r = 0.60, concordance minus regression = 3.0
  • Keep the correlation below 0.70 and find the student for whom regression and concordance agree within 0.25 points.Now: r = 0.60, difference 3.00 at X = 650
  • Raise the correlation until the 95% prediction interval for one student is narrower than 8 points on Test Y.Now: interval width 16.1 points at r = 0.60

5Regression and concordance in base R

The snippet builds 120 pairs with a sample correlation of exactly 0.6, as the lab does. It fits the regression with lm() and writes the concordance as one line of base R. At X = 650 it prints a prediction of 25.5 with an interval from 17.44 to 33.56, and a concordance of 28.5. The round trip from 700 returns 572.

Regression and concordance between two testsShow the code
set.seed(11)

# 120 students take Test X (mean 500, SD 100) and Test Y
# (mean 21, SD 5). Make the correlation exactly 0.6.
n <- 120
r <- 0.6
z1 <- scale(rnorm(n))[, 1]
z2 <- scale(resid(lm(rnorm(n) ~ z1)))[, 1]
x <- 500 + 100 * z1
y <- 21 + 5 * (r * z1 + sqrt(1 - r^2) * z2)

# Regression prediction with a 95% individual interval
fit <- lm(y ~ x)
predict(fit, data.frame(x = 650), interval = "prediction")

# Linear concordance: match z scores instead
concord <- function(s) mean(y) + sd(y) * (s - mean(x)) / sd(x)
concord(650)

# Round trip X to Y and back with the two regressions
back <- lm(x ~ y)
y_hat <- predict(fit, data.frame(x = 700))
predict(back, data.frame(y = y_hat))

Uses base R only.

What a linked score means