How to treat ordinal predictors in the context of multiple linear regression

Question

Hi all, I have a question regarding an analysis I’m trying to do right now concerning data of 100 patients. I have a transformed normally distrubuted continuous outcome Y. My predictor X is 12-scale ordinal predictor (disease severity score using multiple subdomains, minimum total score is 0 and maximum is 12). One thing to note is that the scores 0,1 and 12 do not occur in these patients. I want to do multiple linear regression analyses to analyse the association between Y and X (and some covariates such as sex, age and medication use etc), but the literature on how to handle ordinal predictors is a bit too overwhelming for me. Ordinal logistic regression (swithing X and Y) is not an option, since the research question and perspective changes too much in that way. A few questions regarding this topic:

1. Can I choose to treat this ordinal predictor as a continuous predictor? If so, what are some arguments generally in favor of doing so (quite a few categories for example)?
2. If I were to treat it as a continous predictor, how can I statistically test beforehand whether this is an‘’okay’’ thing to do (I work with Rstudio)? I’m reading about comparing AIC levels and such..
3. If that is not possible, which of the methods (of handeling ordinal predictors) is most used and accepted in clinical research?

Thank you in advance for your help and feedback!

With kind regards

Answer 1

1. Can I treat the ordinal predictor as continuous?

Short answer: Yes, you can - and it's often done in practice, especially if:

1. The ordinal variable has many levels (generally ≥5 is a common heuristic),
2. The distances between levels are believed to be approximately equal (i.e., the "interval" assumption isn't wildly implausible),
3. You're primarily interested in testing linear trends (e.g., "does increasing disease severity relate to a change in Y?").

Arguments in favor:

1. Simplicity and interpretability: You get one coefficient that reflects the linear trend across levels.1.
2. Efficiency: Fewer parameters to estimate compared to modeling each level (e.g., as dummies).
3. Power: Less loss of degrees of freedom = more statistical power.

Common in practice: Especially when the variable is derived from summing multiple subdomains (as your severity score is), many researchers treat such scores as continuous.

2. How can I test whether it’s “okay” to treat it as continuous?

Best practices in R:

a. Compare the linear model vs categorical model:

# Treating X as continuous

model_cont <- lm(Y ~ X + sex + age + medication, data = yourdata)

# Treating X as factor

model_cat <- lm(Y ~ factor(X) + sex + age + medication, data = yourdata)

# Compare models

anova(model_cont, model_cat) # Likelihood ratio test

AIC(model_cont, model_cat) # Compare AIC values

If the ANOVA p-value is not significant, the linear term is sufficient - treating X as continuous is defensible.

AIC: Lower = better. A small AIC difference (<2) suggests both models are comparable.

b. Test linearity of effect:

You can use a restricted cubic spline or generalized additive model (GAM) to check for non-linearity:

library(splines)

model_spline <- lm(Y ~ ns(X, df = 3) + sex + age + medication, data = yourdata)

anova(model_cont, model_spline)

or

library(mgcv)

model_gam <- gam(Y ~ s(X) + sex + age + medication, data = yourdata)

plot(model_gam, se = TRUE)

If spline or GAM suggests a nearly straight line, that supports using a linear term.

3. If not, what’s the most accepted method in clinical research?

If treating as continuous is not defensible:

Treat as categorical (factor): This is the most conservative and most commonly used method in clinical research.

Pros: No assumptions about linearity or equal spacing.

Cons: More parameters; less power.

Polynomial contrasts or splines:

Sometimes used to model non-linear but smooth effects.

Splines (e.g., natural splines) allow flexibility without estimating 10 dummy variables.

Ordinal regression: As you mentioned, not applicable since Y is continuous.