Statistical analysis
Statistical meta-analysis was performed with RStudio using themeta and metafor functions (RStudio Team (2015). RStudio:
Integrated Development for R. RStudio, Inc., Boston, MA URL
http://www.rstudio.com/). Statistical heterogeneity was not considered
during the evaluation of the appropriate model of statistical analysis
as the anticipated methodological heterogeneity of included studies did
not leave space for assumption of comparable effect sizes among studies
included in the meta-analysis [17].
Confidence intervals were set at 95%. We calculated pooled mean
differences (MD) and 95% confidence intervals (CI) with the
Hartung-Knapp-Sidik-Jonkman instead of the traditional Dersimonian-Laird
random effects model analysis (REM). The decision to proceed with this
type of analysis was taken after taking into consideration recent
reports that support its superiority compared to the Dersimonian-Laird
model when comparing studies of varying sample sizes and between-study
heterogeneity [18]. When variables
where expressed as median (range), median (interquartile range) or
interquartile range and sample size transformation where performed to
acquire the mean and standard deviation to include the studies in the
meta-analysis [19].
Subgroup analysis was planned to be conducted on the basis of pregnancy
trimester (1st, 2nd or 3rd), preeclampsia onset (early or late),
severity (mild or severe) and complications (eclampsia and HELLP
syndrome). Residual heterogeneity was planned to be explored by
conducting meta-regression analysis taking into account the following
parameters: year of publication, sample size (using a cut-off of 100
patients in at least one arm of the analysis), region (stratified in
North America, Europe and other countries), Newcastle-Ottawa Scale
score, study design, type of sample and definition of preeclampsia.
Meta-regression was not performed for covariate levels with <3
studies. Publication bias was assessed by examining the possibility of
small-study effects through the visual inspection of funnel plots. The
asymmetry of funnel plots was statistically evaluated using the Egger’s
regression and Begg-Mazumdar’s rank correlation tests. The Trim and Fill
function was also used to evaluate potential differences in summary
estimates after correction of asymmetry. Publication bias was evaluated
by examining the potential presence of small-study effects through the
visual inspection of contour enhanced and traditional funnel plots.
Contour enhanced funnel plots permit the assessment of statistical
significance of observed study estimates and may help differentiate if
asymmetry arises from publication bias or other variables such as study
quality.
Prediction intervals
Prediction intervals (PI) were also calculated, using the metafunction in RStudio, to evaluate the estimated effect that is expected
to be seen by future studies in the field. The estimation of prediction
intervals takes into account the inter-study variation of the results
and express the existing heterogeneity at the same scale as the examined
outcome.
Trial sequential analysis
To evaluate the information size, we performed trial sequential analysis
(TSA) which permits investigation of the type I error in the aggregated
result of meta-analyses performed for primary outcomes that were
predefined in the present meta-analysis. A minimum of 3 studies was
considered as appropriate to perform the analysis. Repeated significance
testing increases the risk of type I error in meta-analyses and TSA has
the ability to re-adjust the desired significance level by using the O‘
Brien-Flemming a-spending function. Therefore, during TSA sequential
interim analyses are performed that permit investigation of the impact
of each study in the overall findings of the meta-analysis. The risk for
type I errors was set at 5% and for type II errors at 20%. The TSA
analysis was performed using the TSA v. 0.9.5.10 Beta software
(http://www.ctu.dk/tsa/).