Ordinal logistic regression is a statistical method that predicts an ordered category outcome — like "low," "medium," or "high" — from one or more predictor variables, telling you how much each predictor pushes someone toward a higher category. If your survey asked people to rate something on a scale (strongly disagree to strongly agree, or a pain rating of mild/moderate/severe) and you want to know what predicts that rating, this is the test you're looking for — and it's the one most people accidentally get wrong by running plain linear regression instead.
Key Takeaways
- Ordinal logistic regression is used when your outcome variable has ordered categories (e.g., rankings, Likert-style ratings) rather than continuous numbers or unordered groups.
- It works by modeling the odds of being at or below each threshold of the outcome, producing odds ratios you interpret one predictor at a time.
- The most common version assumes proportional odds — that each predictor's effect is the same across every threshold of the outcome.
- You report it in APA 7 with the odds ratio (OR), 95% confidence interval, Wald statistic or z, and p value for each predictor.
- Tools like StatRyx automatically check whether ordinal logistic regression fits your data, run it, and generate the APA write-up — no coding or SPSS menus required.
When should I use ordinal logistic regression?
Use ordinal logistic regression when your dependent variable is a set of categories that have a natural order but no fixed numeric spacing between them. Classic examples: education level (high school, bachelor's, master's, doctorate), a satisfaction rating (dissatisfied, neutral, satisfied), or disease severity (mild, moderate, severe).
The key word is ordered. The categories rank from low to high, but you can't assume the "distance" from mild to moderate equals the distance from moderate to severe. That's exactly why linear regression is the wrong choice — it treats your outcome as if those gaps were equal and measured in real numbers.
Here's a quick way to decide:
| Your outcome variable | Correct test |
|---|---|
| Continuous number (age, income, test score) | Linear regression |
| Two unordered categories (yes/no, pass/fail) | Binary logistic regression |
| 3+ unordered categories (car brand, blood type) | Multinomial logistic regression |
| 3+ ordered categories (Likert scale, severity level) | Ordinal logistic regression |
If your outcome has just two levels, you want binary logistic regression instead — and if you're stuck choosing, StatRyx picks the right model from your data automatically.
What is the intuition behind it?
Ordinal logistic regression estimates the odds of scoring in a higher category versus all lower categories, then assumes a predictor shifts those odds by the same amount at every cut-point in your scale.
Imagine a satisfaction scale with three levels: dissatisfied (1), neutral (2), satisfied (3). Instead of predicting the exact category, the model quietly splits your scale into a series of yes/no questions:
- Is the person above "dissatisfied" (i.e., neutral or satisfied)?
- Is the person above "neutral" (i.e., satisfied)?
Each split is like a mini logistic regression. The proportional odds assumption says a predictor — say, hours of training — has the same effect on crossing every one of those thresholds. That's what lets the model summarize everything into a single odds ratio per predictor, which keeps your results simple to interpret.
How does it actually work? The mechanics
The model produces a log-odds equation for the cumulative probability of being at or below each category. When you translate the coefficients out of log-odds, you get odds ratios:
- An OR greater than 1 means higher values of that predictor increase the odds of landing in a higher outcome category.
- An OR less than 1 means higher values push toward a lower category.
- An OR of 1 means the predictor has no effect.
You also get threshold (cut-point) estimates, which mark where one category ends and the next begins. Most researchers ignore these and focus on the predictor odds ratios, which carry the interpretation.
A worked example with real numbers
Suppose we survey 200 employees and ask them to rate job satisfaction as dissatisfied, neutral, or satisfied. We want to know whether hours of weekly training predicts higher satisfaction, controlling for years of tenure.
After running ordinal logistic regression, we get:
- Training hours: OR = 1.42, 95% CI [1.18, 1.71], z = 3.71, p < .001
- Tenure (years): OR = 1.05, 95% CI [0.97, 1.14], z = 1.21, p = .226
Here's what each number means in plain terms:
Training hours (OR = 1.42): For each additional hour of weekly training, the odds of being in a higher satisfaction category are 1.42 times greater — a 42% increase — holding tenure constant. The 95% CI [1.18, 1.71] doesn't include 1, and p < .001, so this effect is statistically significant.
Tenure (OR = 1.05): Each extra year of tenure raises the odds of higher satisfaction by only 5%, but the confidence interval [0.97, 1.14] includes 1 and p = .226, so we can't conclude tenure has a real effect here.
The headline: training matters, tenure doesn't. That single, defensible sentence is exactly what a dissertation committee or reviewer wants.
What counts as significant?
A predictor is statistically significant when its p value is below your alpha level (usually .05) and its 95% confidence interval for the odds ratio does not contain 1. In the example above, training hours (p < .001, CI [1.18, 1.71]) clears both bars; tenure (p = .226, CI [0.97, 1.14]) fails both. The confidence interval and the p value should always agree — if the CI crosses 1, the p value will be above .05.
Checking the proportional odds assumption
Before trusting the results, you must confirm the proportional odds assumption holds — that each predictor's effect is consistent across every threshold of your outcome. This is commonly tested with a Brant test or a likelihood-ratio test. If the assumption is violated (a significant test result), your single odds ratio per predictor is misleading, and you should switch to a partial proportional odds model or multinomial logistic regression instead.
This assumption check trips up a huge number of researchers because SPSS doesn't run the Brant test out of the box — you often need R or a custom syntax script. StatRyx runs the proportional odds check automatically and warns you if your model violates it, so you don't report a result that a reviewer can quietly dismantle.
How do I report ordinal logistic regression in APA 7?
In APA 7 style, report the odds ratio, its 95% confidence interval, the Wald or z statistic, and the p value for each predictor, with test statistics and p italicized and no leading zero on p. A model-level line and a clear predictor sentence look like this:
An ordinal logistic regression predicted job satisfaction from weekly training hours and tenure (N = 200). Training hours significantly predicted higher satisfaction, OR = 1.42, 95% CI [1.18, 1.71], z = 3.71, p < .001, whereas tenure did not, OR = 1.05, 95% CI [0.97, 1.14], p = .226.
Report your effect at the odds-ratio scale (it's more interpretable than raw log-odds), and always include the confidence interval — APA 7