1
$\begingroup$

I have a data set of ~700 studies with response variables. Currently to view a forest plot of the data looks like this.

If I use forest(dat$yi[1:100], dat$vi[1:100]) I get this.

I think the second output looks great, there is an opportunity to visualise the effect, but this is lost if I use the full dataset.

Any advice from the brains trust on alternative ways to display this information in a digestible way without sacrificing samples?

$\endgroup$
5
  • 1
    $\begingroup$ Could you plot these in a 2d plane where the x axis is the observed effect and the y axis is the width or half-width of the interval? $\endgroup$ Apr 14, 2020 at 3:51
  • 1
    $\begingroup$ How about a caterpillar plot? metafor-project.org/doku.php/plots:caterpillar_plot It's still essentially the same thing, but ordering the effects by magnitude already helps to make it look much neater. $\endgroup$
    – Wolfgang
    Apr 14, 2020 at 8:31
  • $\begingroup$ Further to this Wolfgang, is there a way to present average results for effect sizes based on a factor using forest? i.e. average yi, vi, grouped an additional column e.g. 'species'. This would be a neat way to add a summary-style forest plot. E.g. the plot here $\endgroup$
    – sleepy
    Apr 21, 2020 at 5:59
  • $\begingroup$ For anyone in future, a detailed answer to my question in the comment above can be found in this Cross Validated question $\endgroup$
    – sleepy
    Apr 21, 2020 at 6:02
  • $\begingroup$ Is there a way to re-overlay the dotted line at the 0 point? Once the effects are all plotted, it becomes quite obscured. $\endgroup$
    – sleepy
    Apr 21, 2020 at 6:31

1 Answer 1

2
$\begingroup$

This question raises the issue of why we use forest plots. What do they tell us? In general they display (a) the identity of the study, (b) the corresponding estimate, (c) the precision of that estimate, and (d) optionally, by using the order vertically the effect of a proposed moderator.

If we are willing to sacrifice the identity then we can plot the estimate against the precision with the choice of symbols representing the values of a categorical moderator if any. This is usually called a funnel plot and has the advantage that it will be familiar to the reader. Doing this for a continuous moderator is not viable so that would have to be sacrificed. For a large number of studies there would be over-printing so hollow symbols would be best or even using hexagonal binning.

If we have a continuous moderator we could plot the estimates against the value of the moderator with the size of the symbols proportional to the precision of the estimates. Again hollow symbols would work best here and hexagonal binning would not be an answer.

If the desire is to spot studies which might differ in some way from the others then calculating a suitable diagnostic, perhaps a leave-one-out statistic, and using it as the moderator would distinguish the influential or anomalous studies more readily than the standard forest plot.

$\endgroup$
2
  • $\begingroup$ Thanks mdewey. This has given me a lot to consider. I have a combination of categorical and continuous moderators - perhaps I will individually plot each effect size by each moderator to visualize individual effects. I will look in to how to size points by precision, this is a very good idea. Thanks for your insights. I will report back once I have tried these and accept your answer and explain what I ended up doing to solve my predicament. $\endgroup$
    – sleepy
    Apr 22, 2020 at 1:49
  • $\begingroup$ Your suggestion "If we have a continuous moderator we could plot the estimates against the value of the moderator with the size of the symbols proportional to the precision of the estimates." has worked very well for me. I am happy with this as a way to visualize the effects in relation to key moderators. $\endgroup$
    – sleepy
    Apr 22, 2020 at 5:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.