Double your fun with double forest plots!

[#1]

The idiom “unable to see the forest for the trees” means that one fails to grasp the main issue(s) because of excessive attention to details. I love this expression as it fully captures why data visualization is so important to the modern quantitative scientist. Given the volume of data to review, the variety of analyses to perform, and the need to turn findings into actions quickly, researchers can easily get lost in the trees on a regular basis, especially if they do not make regular use of data visualization. I get it…useful data visualizations can be time consuming to produce, and there are some challenges for efficient review and quality control. But I believe this investment is worth the insight that is gained, coupled with the improved ability to communicate findings to others.

Another reason I love this expression is that it brings attention to my favorite plot, the forest plot (Figure 1)! There have been several improvements in Graph Builder over the years to make forest plots easier to produce, along with a steady stream of JMP add-ins to improve the efficiency of production. The most recent add-in made it easy to produce all manner of forest plots for JMP 17 and later releases, but it could only produce a single panel forest plot. With the release of the new Happy Little Forests add-in in the JMP Marketplace, analysts have a lot more options available.

Figure 1. Forest plot of primary and secondary endpoints for a clinical trial in chronic hepatitis B infection (Chang et al., 2006)

More than 10 years ago, some colleagues published one of my go-to references on the use of data visualization in medical research (Duke et al., 2015). What I especially love about this paper is the introduction (to me, at the very least) of the double forest plot (though the authors refer to it as a multi-panel dot plot). In their application, the authors present observed values of systolic blood pressure by treatment and sex in one panel, with treatment comparisons between active doses to placebo in a second panel. To this day, I reference this awesome graphic in presentations and short courses as an extremely useful and concise way to summarize a continuous endpoint (Figure 2).

Figure 2. Double forest plot of systolic blood pressure (Duke et al., 2005)

With my return to JMP, I explored how it would be possible to make double forest plots easier to produce. It led to some improvements in Graph Builder (thanks, Xan!) as well as some code for internal development to make it easier for the Life Sciences team to produce forest plots more easily within JMP Clinical. But everyone needs to benefit from this awesome visualization, so the Happy Little Forests add-in was born. For JMP 19 and later releases, this add-in makes it possible to produce two-panel dot-forest plots (Amit et al., 2008), which are commonly used in the analysis of adverse events (Figure 3), and double forest plots (Figure 4), as well as all of the functionality that was available in the Forest Plot add-in for single panel forest plots.

Figure 3. Dot-forest plot of HBV endpoints (Chang et al., 2006)

Figure 4. Double forest plot of diastolic blood pressure (simulated)

As shown in the images, double forest plots can:

1. Summarize observed values and treatment effects to a control group (e.g., Figure 2), producing a useful summary of individual treatment response along with comparisons to a standard of care.
2. Summarize observed values and change from baseline values over time (e.g., Figure 4), producing a more useful summary to the standard lengthy table of summary statistics that are often produced in clinical trials. For binary endpoints, one panel can summarize the proportion of individuals achieving a particular response, while the other panel can repeat this analysis subset to individuals not in response at baseline.
3. Summarize results within subgroups, and in the case of two-level subgroups, summarize the confidence interval of the treatment by subgroup interaction to communicate the strength of notable statistical differences between the subgroup levels.

Happy Little Forests is a first step for producing multi-panel forests. Some points for consideration:

1. Adding an animation button to switch subgroup and level columns (say, in Figure 2) to more easily make comparisons of subgroups within levels (i.e., treatment within subgroup) compared to levels within subgroups (i.e., levels within treatment, as presented). For a longitudinal presentation as presented in Figure 4, it has the effect of being able to more easily compare each visit across the treatments.
2. Adding an animation button to switch control arms for multi-arm studies, as in the Recurrence Report. As shown in Figure 2, confidence intervals are presented against a single control. Figure 3 displays entecavir minus lamivudine confidence intervals, though an animation button would allow users to produce lamivudine minus entecavir intervals if it was of interest. The ability to switch the comparator arm on the fly would make this interactive visualization even more useful in practice.
3. Allowing for additional panels. Continuous endpoints could potentially benefit from a three-panel plot with observed values in the left panel, change from baseline in the middle panel, and treatment effects in the right panel. It would necessitate eliminating annotated intervals for at least one of the panels (or moving to the left or right sides), but three panels may be too complicated for general use.

Finally, though I leveraged improvements in JMP 19 to make dot-forests and double forest plots easier to produce, it is absolutely possible to create these plots in earlier versions of JMP using a Multiple series stack in Data > Stack.

What new features would you like to see in Happy Little Forests? Comment in the JMP Marketplace.

References

Amit O, Heiberger RM & Lane PW. (2008). Graphical approaches to the analysis of safety data from clinical trials. Pharmaceutical Statistics 7: 20-35.

Chang TT, Gish RG, de Man R, et al. (2006). A comparison of entecavir and lamivudine for HBeAg-positive chronic hepatitis B. The New England Journal of Medicine 354: 1001-1010.

Duke SP, Bancken F, Crowe B, et al. (2015). Seeing is believing: Good graphic design principles for medical research. Statistics in Medicine 34: 3040-3059.