Logistic regression · OkCupid profiles, n = 59,826

One inch, three scales

Every OkCupid dating profile lists a height and a sex. In this set of 59,826 profiles, each person is listed as either male or female, and on average the men are several inches taller. Height alone is enough to guess which group a profile belongs to about 83% of the time.

We took that data and ran a logistic regression of sex on each person's stated height. The result is a single equation, but you can read its output in three ways: as log odds, as odds, or as a probability. Add one inch of height and each scale responds differently.

log(p1 − p) =−44.45+0.66· height
glm(male ~ height, family = binomial)

Here p = P(male) and height is in inches. That 0.66 is the whole story. Follow it through the three scales below.

5′7″67 in → 68 in

Step 1

From height to probability in three moves

Plug a height into the equation and you get log odds. Exponentiate to get odds. Divide odds by one plus odds to get a probability. These are three descriptions of the same prediction.

Log odds −0.10 −44.449 + 0.6619 × 67 Any number from −∞ to ∞. Zero means a coin flip.
Odds 0.90 e^(−0.10) Male profiles per female profile at this height.
P(male) 47.5% 0.90 / (1 + 0.90) Squeezed between 0 and 1.

Step 2

Now add one inch

Going from 5′7″ to 5′8″ …

Log odds: add 0.66, every time

constant · additive
+0.66 The same step at every height. The faint staircase has identical risers.

Odds: multiply by 1.94, every time

constant · multiplicative
×1.94

e0.66 = 1.94. The amount added keeps growing, but the multiplier never changes.

Probability: neither

not constant
+16.2 pts

Step 3

What each inch buys, from 60″ to 76″

Each bar is one inch of height. The two constant rows are flat, while the bottom two are not. Tap a bar to jump there.

Step 4

The full ledger

Read down the shaded columns: +0.66 on every row, ×1.94 on every row. The probability change rises to a peak near 67″ and falls off again.

HeightLog oddsChangeOddsChangeP(male)Change

Why

Why probability bends

Add

The model is linear in log odds. One inch always adds β₁ = 0.66, the way a slope works in ordinary regression.

Multiply

Odds = elog odds. Adding inside an exponent multiplies outside it: ea + 0.66 = ea × 1.94.

Bend

Probability has to stay between 0 and 1, so the same push in log odds moves it a lot near 50% and barely at all near 0% or 100%.

The size of a one-inch change in probability is close to the slope times p(1 − p):

Δp ≈ 0.66 × p × (1 − p)  →  largest at p = 0.5: 0.66 × 0.25 ≈ 0.165

That's the "divide by 4" rule. The steepest one-inch jump in P(male) is about 0.66 / 4 ≈ 16.5 points, and it happens near 67.2″, where the model is a coin flip. At 60″ or 76″ the same inch moves the probability by less than a point.

Values use the unrounded estimates (−44.449, 0.6619). At 69″ (5′9″) the model gives P(male) = 0.77.