Method 4: Albatross plots
The albatross plot was first described by Harrison et al7 and is also discussed
in Higgins 22. This
method requires minimal data extraction or manipulation and allows data
to be synthesised even in circumstances when outcomes are reported in
multiple different formats or where no summary statistics are reported.
Reported results are split into two groups according to the direction of
effect; and then p-values are plotted against sample size. Where
necessary, 1-sided p-values need to be converted to 2-sided p-values (or
vice versa) to ensure consistency. An albatross plot allows us to
combine outcome data that was reported in a variety of different ways,
including from studies where only a p-value was provided. Under an
assumption of normality, you would expect results corresponding to the
same effect size to lie along a contour, with p-values generally getting
smaller as sample size increases. Contours can be added to the plot for
a range of different effect sizes based on standardised mean
differences, mean differences, odds ratios or other summary of choice.
Effect sizes can be estimated according to where the majority of points
lie. We have added contours to represent standardised mean differences
of 0.3, 0.6 and 0.9. Heterogeneity can also be explored visually by
looking at how closely trials tend to group together along a particular
contour.
Note that where p-values are obtained from studies that are clustered in
some way, adjustment of sample size is necessary. One method of doing
this is to calculate the effective sample size (E) using the sample size
(S), the reported intra-class correlation coefficient (ICC) and the
average cluster size (M) using the formula23
\begin{equation}
E=\frac{S}{1+ICC\ \times(M-1)}\nonumber \\
\end{equation}An alternative is to replace the sample size with the number of health
care professionals (or sites) as in method 3.
For illustration we produced a contour plot using the number of health
care professionals (or sites) as the sample size (Figure 1)