# Disagreeing significance tests for an Ordinal Logistic Regression

I calculated both confidence interval and p-value on the same model using R, and the two tests of significance disagreed with one another.

I am comparing how 3 groups differ in terms of an ordinal variable, once a separate numeric predictor was taken into account. The model indicated that group was a significant predictor, and when I graphed the probabilities, I noticed that two of the groups were very similar, so I created a subset of the data only containing these groups to see if group was still a statistically significant predictor.

The p-value indicated that it was not (p=0.64), but the confidence interval indicated that it was a significant predictor (CI 2.5% = 0.68, CI 95% = 1.28).

I'm not sure where to go from here. I did a couple of other tests (which make a few extra assumptions and so aren't ideal), and all of them indicated non-significance. I've been using confidence intervals for this entire project, so suddenly switching to p-values is a bit jarring and I would prefer not to turn to tests & models that are less ideal.

I wish to know why this happened, what it means, and, if possible, whether this is due to a mistake in my code.

here's the code for the model & statistical tests I used:

library(MASS)
#model
DvS<-polr(ordinal~numeric+group, data=DvSata, Hess = TRUE)
summary(DvS)

#confidence intervals & odds ratios
ci<-confint(DvS)
exp(coef(DvS))
exp(cbind((OR = coef(DvS)), ci))

#p-value
st <- coef(summary(DvS))
pval <- pnorm(abs(st[, "t value"]),lower.tail = FALSE)* 2
st <- cbind(st, "p value" = round(pval,3))
st
$$$$
`

Recall that the null hypothesis for an odds ratio (OR) is that $$\text{OR} = 1$$, which occurs when $$\text{log(OR)} = 0$$. The confidence interval you produced is around the OR; it contains 1, and therefore is nonsignificant, consistent with the p-value. If you produce a confidence interval around the $$\text{log(OR)}$$ (i.e., the coefficient itself), you will see that it will contain 0.