Logit Slopes as Cumulative Odds Ratios

From a Cumulative-Logit Slope to an Odds Ratio

An ordinal response has several ordered categories, so its association with a predictor can be described at each division between lower and higher categories. For a chosen cutpoint \(j\), the relevant comparison is the event \(Y>j\) against its complement \(Y\le j\). A cumulative-logit slope has a direct interpretation for the odds of that comparison, provided its sign convention is kept explicit.

Use the cumulative-logit convention \(\operatorname{logit}P(Y\le j\mid x)=\alpha_j-\beta x\), where \(j\) indexes a cutpoint, \(x\) is a scalar predictor, and \(\alpha_j\) is the cutpoint parameter. Thus a positive \(\beta\) lowers the log-odds of being at or below a cutpoint as \(x\) increases; equivalently, it raises the odds of being above that cutpoint. The tutorial focuses on this upper-versus-lower cumulative comparison.

Definition: At cutpoint \(j\), the cumulative odds of being above the cutpoint are \(O_j(x)=P(Y>j\mid x)/P(Y\le j\mid x)\). The cumulative odds ratio comparing predictor values \(x+a\) and \(x\) is \(O_j(x+a)/O_j(x)\), where \(a\) is the size and direction of the predictor change.

For finite model parameters, the logistic cumulative probability is strictly between zero and one, so both events in this odds definition have positive probability. The central calculation is to invert the usual cumulative logit: the logit is written for \(Y\le j\), while the odds in the definition put \(Y>j\) in the numerator.

Deriving the Cumulative Odds Ratio

Theorem: Suppose the cumulative-logit model is \(\operatorname{logit}P(Y\le j\mid x)=\alpha_j-\beta x\). For a fixed cutpoint \(j\), the odds of \(Y>j\) rather than \(Y\le j\) are \(O_j(x)=\exp(\beta x-\alpha_j)\). Therefore, changing \(x\) from \(x\) to \(x+a\) multiplies those odds by \(\exp(\beta a)\).

Proof. Write \(p_j(x)=P(Y\le j\mid x)\). By the definition of the logit and the fact that \(P(Y>j\mid x)=1-p_j(x)\),

\operatorname{logit}p_j(x) =\log\!\left(\frac{p_j(x)}{1-p_j(x)}\right) =\alpha_j-\beta x.

Exponentiating gives the lower-versus-upper odds:

\frac{P(Y\le j\mid x)}{P(Y>j\mid x)} =\exp(\alpha_j-\beta x).

Taking the reciprocal gives the upper-versus-lower odds defined above. Evaluating those odds at \(x+a\) and \(x\) and dividing yields

\frac{O_j(x+a)}{O_j(x)} =\frac{\exp\{\beta(x+a)-\alpha_j\}}{\exp(\beta x-\alpha_j)} =\exp(\beta a).

The cutpoint parameter and starting predictor value cancel, which proves both formulas. \(\square\)

For a one-unit increase, \(a=1\), so the odds ratio is \(\exp(\beta)\). Its direction follows the sign of \(\beta\): a positive slope gives an odds ratio greater than one, a negative slope gives one below one, and a zero slope gives one. For a change of \(a\) units, the multiplier is \(\exp(\beta a)\), not simply \(\exp(\beta)\) unless \(a=1\).

Theorem: Under the shared-slope cumulative-logit model, the odds ratio for a given predictor change is the same at every cutpoint: for each \(j\), \(O_j(x+a)/O_j(x)=\exp(\beta a)\).

Proof. At any cutpoint \(j\), the first result gives \(O_j(x+a)/O_j(x)=\exp(\beta a)\). The right-hand side contains neither \(j\) nor \(\alpha_j\). Hence it has the same value at every cutpoint in the model. \(\square\)

This is the proportional-odds interpretation of a shared slope: the predictor change has a common multiplicative effect on the odds of being above each cutpoint. It does not say that the probabilities of every response category change by the same amount, or that every category has the same odds ratio when compared with another single category. The event being compared is cumulative: all categories above \(j\) versus all categories at or below \(j\).

Worked Interpretations

Worked Example: A Positive Slope

Suppose the fitted cumulative logit for a predictor \(x\) is \(\operatorname{logit}P(Y\le j\mid x)=0.4-0.7x\). Interpret a one-unit increase in \(x\) for the odds of \(Y>j\) versus \(Y\le j\).

Here \(\alpha_j=0.4\) and \(\beta=0.7\), so the one-unit odds ratio is \[ \exp(\beta)=\exp(0.7)\approx 2.0138. \] Thus, for this cutpoint, increasing \(x\) by one unit multiplies the odds of \(Y>j\) rather than \(Y\le j\) by about \(2.0138\), or raises those odds by about \(101.4\%\), since \((2.0137527-1)\times100\%\approx101.4\%\).

The formula can also be checked at two predictor values. At \(x=0\), the upper-versus-lower odds are \(\exp(-0.4)\approx0.6703\); at \(x=1\), they are \(\exp(0.3)\approx1.3499\). Using the unrounded values, their ratio is \(\exp(0.3)/\exp(-0.4)=\exp(0.7)\approx2.0138\). The displayed rounded odds need not divide to the same last decimal place.

Worked Example: A Negative Slope

Now suppose \(\operatorname{logit}P(Y\le j\mid x)=0.2-(-0.5)x\), equivalently \(0.2+0.5x\). What happens to the odds of \(Y>j\) versus \(Y\le j\) when \(x\) increases by one unit?

In the specified convention the coefficient is \(\beta=-0.5\). The odds ratio for an increase of one is \[ \exp(-0.5)\approx0.6065307. \] The new odds are about \(0.6065\) times the old odds. The relative decrease, calculated from the unrounded multiplier, is \[ \{1-\exp(-0.5)\}\times100\% \approx39.3469\%, \] or about \(39.3\%\) to one decimal place.

For a check, at \(x=0\) the upper-versus-lower odds are \(\exp(-0.2)\approx0.8187308\). At \(x=1\) they are \(\exp(-0.7)\approx0.4965853\). Their exact ratio is \(\exp(-0.7)/\exp(-0.2)=\exp(-0.5)\). Comparing in the reverse direction, from \(x=1\) to \(x=0\), gives the reciprocal multiplier \(\exp(0.5)\approx1.6487\).

Worked Example: A Five-Unit Predictor Change

Consider a cutpoint with \(\alpha_j=0.5\) and a fitted slope \(\beta=0.12\). Compare \(x=2\) with \(x=7\), a five-unit increase.

The odds at \(x=2\) are \[ O_j(2)=\exp(0.12(2)-0.5)=\exp(-0.26)\approx0.7711. \] At \(x=7\), they are \[ O_j(7)=\exp(0.12(7)-0.5)=\exp(0.34)\approx1.4049. \] The exact odds ratio is therefore \[ \frac{O_j(7)}{O_j(2)} =\exp(0.34-(-0.26)) =\exp(0.60) \approx1.8221. \] Thus the five-unit increase multiplies the odds by about \(1.8221\), an increase of about \(82.2\%\). Applying the one-unit odds ratio five times gives the same result: \(\{\exp(0.12)\}^{5}=\exp(0.60)\). The comparison concerns odds, not a fixed percentage-point change in probability.

Odds Ratios Are Not Probability Ratios

An odds ratio describes how the odds change, not how many times larger the probability becomes. If the initial probability of \(Y>j\) is \(p\), its odds are \(p/(1-p)\). Multiplying those odds by an odds ratio \(R\) gives new odds \(R p/(1-p)\). Converting those odds back to a probability gives

p_{\mathrm{new}} =\frac{R p/(1-p)}{1+R p/(1-p)} =\frac{Rp}{1-p+Rp}.

Consequently, the probability change depends on the starting probability \(p\), even when the odds ratio \(R\) is held fixed. For example, if the initial probability is \(0.5\), the initial odds are \(1\). An odds ratio of \(\exp(0.7)\approx2.0138\) gives new odds of about \(2.0138\), corresponding to probability \(2.0138/(1+2.0138)\approx0.6682\). The probability has increased from \(0.5\) to about \(0.6682\), not multiplied by \(2.0138\).

Common Mistakes and a Practical Reading Rule

Reversing the event in the odds. The cumulative logit is written for \(Y\le j\), but the odds asked about here are \(Y>j\) divided by \(Y\le j\). Inverting the cumulative odds changes the sign in the exponent: the upper-versus-lower odds are \(\exp(\beta x-\alpha_j)\).
Using the wrong sign for the slope. In \(\operatorname{logit}P(Y\le j\mid x)=\alpha_j-\beta x\), a positive \(\beta\) raises the upper-versus-lower odds. Reading the minus sign as if the model were \(\alpha_j+\beta x\) reverses that interpretation.
Calling an odds ratio a probability multiplier. An odds ratio of two doubles odds, not probability. Convert the updated odds to a probability if a probability interpretation is required.
Ignoring the predictor scale. The one-unit multiplier depends on how \(x\) is measured. If \(x\) is measured in days, \(\exp(\beta)\) refers to one day; a ten-day change has multiplier \(\exp(10\beta)\).
Assuming the effect differs by cutpoint under a shared slope. In this model the cumulative odds ratio for a fixed predictor change is the same at every cutpoint. That equality concerns upper-versus-lower cumulative events, not separate category probabilities.

Key takeaway: With the convention \(\operatorname{logit}P(Y\le j\mid x)=\alpha_j-\beta x\), a change of \(a\) units in \(x\) multiplies the odds of \(Y>j\) versus \(Y\le j\) by \(\exp(\beta a)\). State the event, the direction and size of the predictor change, and that the result is about odds.

Check Your Understanding

For each question, keep the cumulative event and the stated sign convention explicit.

Under \(\operatorname{logit}P(Y\le j\mid x)=\alpha_j-\beta x\), derive the odds of \(Y>j\) versus \(Y\le j\).
If \(\beta=0.3\), what is the odds ratio for a one-unit increase in \(x\)? Interpret its direction.
If \(\beta=-0.4\), calculate the odds ratio for a three-unit increase. Does the odds rise or fall?
Why is the odds ratio for a fixed predictor change the same at every cutpoint under a shared-slope model?
An odds ratio is \(1.5\), and the starting probability of \(Y>j\) is \(0.2\). Find the starting odds, updated odds, and updated probability. Explain why the probability is not multiplied by \(1.5\).