4
$\begingroup$

How can investigate publication bias for diagnostic tests in meta-analysis with R? I saw this topic "Publication bias in meta-analysis of diagnostic accuracy in R?" but there is no answer yet!

$\endgroup$
4
$\begingroup$

A lot of methods exist for investigating publication bias. The most general method, which can be used - to my knowledge - in any case, and therefore for diagnostic accuracy as well, is the funnel plot.

The metafor package can produce a funnel plot and can use the trim and fill technique for correction. However, the ideal form of a funnel plot depends on the particular statistic that is plotted. For most, such as cohens d it is pyramidal. For others like, for example, odds ratios it is not. In that case you have to transform the estimates first. Here is an example of an funnel plot using the trim and fill technique with log odds ratios: See the following example: http://www.metafor-project.org/doku.php/plots:funnel_plot_with_trim_and_fill

If you need other methods, let us know the method and your dependent variables.

Edit: If standard errors are not provided for your statistic, you can also use N instead. Or a transformation of it. Not really sure, because I have not used my knowledge on this topic for several years. However, I remember that the following book helped me bit:

Haidich, A. B. (2011). Meta-analysis in medical research. Hippokratia, 14(1), 29-37.

However, other users might be more experienced than I am on that particular topic. Just wanted to give some first ideas. ;)

$\endgroup$
  • 2
    $\begingroup$ There are of course also non-statistical methods. E.g. checking clinical trial registers for any existing trials that have not been published. Unlike for drug trials (and even for those some requirements would not have applied for older trials), I am not sure what the registration requirements are for diagnostic trials. $\endgroup$ – Björn Mar 1 '17 at 8:24
  • 2
    $\begingroup$ The question is specifically about diagnostic test studies. What statistic do you propose the OP should plot since there are several ways of summarising diagnostic studies which lead to two summary statistics not one? $\endgroup$ – mdewey Mar 1 '17 at 12:04
  • 2
    $\begingroup$ @mdewey: All? The main problem is to get the standard errors. Because I never saw them reported for gini, roc, sensitivity, specificity or overall correct classification, etc. (at least in psychological diagnostic) $\endgroup$ – StatisticsRat Mar 1 '17 at 13:39
  • 2
    $\begingroup$ Sensitivity and specificity are proportions, their standard errors are fairly obvious but how do you propose to plot them in a funnel plot since they are correlated? $\endgroup$ – mdewey Mar 1 '17 at 13:44
  • 2
    $\begingroup$ @mdewey sry, I was not very precise: I would plot them separately. However, the more I think about it the more I think less of my overall answer. - So, I hedged my answer a bit. ;) $\endgroup$ – StatisticsRat Mar 1 '17 at 13:54
3
$\begingroup$

One approach if using the sensitivity and specificity would be to plot them in the logit ROC space either as crosshairs (with the hairs having length proportional to the standard error in that direction) or as confidence ellipses. Small study effects would then be suggested if the larger ellipses (or the larger crosses) were concentrated in some part of the logit ROC space.

An alternative would be to summarise the studies using the diagnostic odds ratio and perform conventional summary plots of those (either forest or funnel).

The vignette for the R package mada has examples of these and other displays and their use.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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