Confidence Curves

Constructing the set of Confidence Curves from observed data

Overview

A function written in R to calculate the frequentist confidence distribution resulting from a point estimate and its associated error.

‘Posteriors without Priors’

"We are 99.9% confident the treatment is beneficial to the patient!"

Does this sound like a Bayesian statement about posterior probability of treatment benefit? In fact it is a frequentist confidence statement, fully compatiable with traditional frequentist analysis. Let’s explain:

What is a confidence distribution?

The confidence distribution is constructed by stacking every one-sided confidence interval (from 0-100) vertically to generate a cumulative distribution function. Figure 1 shows an example from real trial data where a point estimate of -0.22 (log odds ratio) was observed. The treatment effect in this example represents the odds of a patient in the treatment group doing poorly over the odds of a patient in the control group doing poorly. On this scale, smaller values are better. Values less than zero represented benefit of a treatment to the patient (lying in a Region of Benefit). We see the stack of one-sided intervals, including the 99.89% confidence interval which intersects with zero - zero being the boundary of the Region of Benefit. So, any values within the 99.89% confidence interval represent benefit to the patient.

confidence_distribution_INTERACT_annotated

Figure 1. Confidence distribution function from observed data that has a point estimate of -0.22 (a log odds ratio) made up of all one-sided confidence intervals, including the 5%, 25%, 50%, 75% 95% and 99.89% CI. The red dashed line represents the observed effect, and the blue line represents the line of no effect. Any treatment effect values less than zero represent benefit of treatment to the patient, and are in the region of benefit. The 99.89% CI covers the region of benefit.

Why would we do this?

By constructing this distribution, we have the flexibility to reason about a variety of treatment effects. For example, our treatment effect of interest in this example is “treatment benefit”. Given that the 99.98% confidence interval contains all values that represent treatment benefit, we may say: We are 99.89% confident that the treatment has benefit.

Flexible thresholds

We do not have to limit ourselves to statments about benefit. Say we are interested in an amount of benefit that we believe is actually meaningful: It is not enough that the magnitude of the effect is greater than zero, it had to be big enough to be clinically meaningful. That is possible! With confidenceCurves we can also specify a meaningful benefit. Figure 2 shows the same confidence distribution as in Figure 1, but this time we are interested in a different confidence interval: The interval covering the region of no meaningful benefit. In this case, the confidence interval covering the region of meaningful benefit is the 99.1% confidence interval, so the confidence interval for lack of meaningful benefit is \(1-0.991 = 0.009\). Therefore, we may say: We are 1% confidence that the treatment lacks meaningful benefit. If we were using the confidence in a lack of meaningful benefit or \(LMB\) to monitor for futility in an adaptive trial, we would not stop the trial early on the basis of this confidence value.

confidence_lob_interact_annotated_github

Figure 2. Confidence distribution from observed data that has a point estimate of -0.22 (a log odds ratio) made up of all one-sided confidence intervals, including the 5%, 25%, 50%, 75% 95% and 99.1% CIs. The red dashed line represents the observed effect, and the green dashed line represents the smallest clinically meaningful benefit. The 99.1% CI covers the regions of meaningful benefit. Any treatment effect greater than this represents no meaningful benefit to the patient, and are in the region of no meaningful benefit. The 99.1% CI covers the regions of meaningful benefit.

Confidence statments are valuable in clinical trials

The traditional frequentist p-value does not allow us to make such a useful statement, although we can still calculate it in this framework. Frequentist confidence statements are interpretable and intuitive and especially useful in the context of an adapative trial. During an adaptive trial, we can use frequentist confidence to express our confidence in a particular effect, and make interim decisions based on the strength of that evidence. We do not have to pay with control of Type I error, as we would if we used Bayesian posterior probability statements, and we do not have to construct a prior distribution either. Thats why these are “Posteriors without Priors”.

Reference

The confidence curves in this package are constructed using equations from Ian Marschner.

Marschner, I. “Confidence distributions for treatment effects in Clinical Trials: Posteriors without Priors”, Statistics in Medicine, 2024;43:1271-1289 https://doi.org/10.1002/sim.10000.