topic Re: Question about standard reporting for plots of proportions and CIs over time in Statistical Procedures

Question about standard reporting for plots of proportions and CIs over time

mjkop56 — Wed, 20 Oct 2021 21:03:00 GMT

Below is a chart of proportions over time.

I’d like to add confidence intervals and a trend line through data such as this one. From articles I’ve seen, there seem to be two ways of doing this:

1. calculate the CIs for the observed proportions

2. run a model, and have the trend lines be predicted probabilities from the model and the 95% confidence intervals would go around the predicted probabilities. An example of this approach is in this article - "Columns indicate observed success rates. Points and error bars indicate model-predicted means and 95% confidence intervals, respectively." from https://www.biorxiv.org/content/10.1101/232868v2.full

My question: in a standard reports where there will be many of these charts shown, is one of these a "better" way? I ran a model because I also wanted to test if the trends were statistically significant, and the model also adjusted for clustering. Is this a good enough reason to take approach #2?

Any advice would be much appreciated.

Re: Question about standard reporting for plots of proportions and CIs over time

Rick_SAS — Thu, 21 Oct 2021 10:48:47 GMT

The second plot is sometimes called a "dynamite plot." Data visualization experts generally agree that dynamite plots are not the best way to summarize a data distribution.

You can read more about alternative visualizations in the article "Remaking a panel of dynamite plots."

I think it is a good idea to add CI's for the observed proportion. Be sure to specify on the graph what the intervals represent because there are several ways to visualize uncertainty in point estimates.

It is not clear to me that a trend line from a predictive model is the right way to get those CIs, but you know your data better than I do.

Re: Question about standard reporting for plots of proportions and CIs over time

mjkop56 — Thu, 21 Oct 2021 22:49:37 GMT

Thank you, Rick! This is very helpful.