2.4 Statistical analysis
Meta-analysis of this study included two aspects. In the primary aspects, pooled estimates of the OR were estimated for the associations of PD-L1 expression in clinical parameters, which were computed by natural logarithm OR, and its standard errors (SEs). In the second aspects, pooled estimates of HR of PD-L1 related to prognostic factors were performed by a random-effects model, in which natural logarithm HR and its SEs were calculated. The estimated HR of individual study was preferred to entered to this analysis if they was reported in the case of various adjusted factors. Several studies did not provided direct HR, but its information were available for synthesizing the estimated HR, and then methods described in Tierney et al.14 were performed to impute the estimated HR. We assumed clinicopathological and prognostic significance of PD-L1 expression was confounded by methods for PD-L1 immunostaining such as antibody, definition of PD-L1 positivity, IHC cutoffs. Studies reporting more than one type of them were separated as different data sets, correspondingly.
The heterogeneity across studies was investigated through Cochran’s Q test (p < 0.1) and Higgins I2, of which values 25%, 50% and 75% respectively indicated low, mild and high heterogeneity15. In order to identify the sources of heterogeneity, subgroup analyses were performed on the basis of grouping by assumed confounding factors. Pooled estimates synthesized by at least three studies were considered to be robust that no obvious fluctuation happened when removing one study at a time. Publication bias was determined by Begg’s and Egger’s tests16.