Preface to First Edition - lib

278 META-ANALYSISR> summary(BCG_DSL)Random effects ( DerSimonian-Laird ) meta-analysisCall: meta.DSL(ntrt = BCG[["BCGVacc"]], nctrl = BCG[["NoVacc"]],ptrt = BCG[["BCGTB"]], pctrl = BCG[["NoVaccTB"]],names = BCG$Study)------------------------------------OR (lower 95% upper)1 0.39 0.12 1.262 0.19 0.08 0.463 0.25 0.07 0.914 0.23 0.18 0.315 0.80 0.51 1.266 0.38 0.32 0.477 0.20 0.08 0.508 1.01 0.89 1.159 0.62 0.39 1.0010 0.25 0.14 0.4211 0.71 0.57 0.8912 1.56 0.37 6.5513 0.98 0.58 1.66------------------------------------SummaryOR= 0.47 95% CI ( 0.32,0.69 )Test for heterogeneity: X^2( 12 ) = 163.16 ( p-value 0 )Estimated random effects variance: 0.37Figure 15.4R output of the summary method for BCG_DSL.wald, 1975, Smith, 1980, for example). Research with statistically significantresults is potentially more likely to be submitted and published than workwith null or non-significant results (Easterbrook et al., 1991). The problemis made worse by the fact that many medical studies look at multiple outcomes,and there is a tendency for only those suggesting a significant effect tobe mentioned when the study is written up. Outcomes which show no cleartreatment effect are often ignored, and so will not be included in any later reviewof studies looking at those particular outcomes. Publication bias is likelyto lead to an over-representation of positive results.Clearly then it becomes of some importance to assess the likelihood of publicationbias in any meta-analysis. A well-known, informal method of assessingpublication bias is the so-called funnel plot. This assumes that the resultsfrom smaller studies will be more widely spread around the mean effect becauseof larger random error; a plot of a measure of the precision (such asinverse standard error or sample size) of the studies versus treatment effectfrom individual studies in a meta-analysis, should therefore be shaped like afunnel if there is no publication bias. If the chance of publication is greater© 2010 by Taylor and Francis Group, LLC