<p> Small-study effects, which are factors resulting in dependencies between treatment effect size and precision, are an important source of bias in meta-analyses of randomized controlled trials. However, established nonparametric tests for detection of small-study effects that are based on rank correlation lack statistical power, while established parametric tests that are based on linear regression are not robust in the presence of between-study heterogeneity. </p><p> A novel method for detection of small-study effects is proposed that is designed to overcome these limitations. The method uses repeated one-sample Wald-Wolfowitz runs tests to evaluate the null hypothesis of serial independence among trial treatment effect size estimates that are ranked by precision. This dissertation describes lower-tailed, upper-tailed, and two-tailed versions of the proposed method for detection of small-study effects and compares the proposed method to established tests using simulation. The novel method is implemented in Stata using various procedures for control of type 1 error, including the Bonferroni and Sidak corrections, Hochberg’s step-up procedure, and the Benjamini-Hochberg procedure for control of the false discovery rate. The type 1 error rate and power of the novel method are then compared to those of existing tests, including the nonparametric rank correlation test of Begg and Mazumdar and the commonly-used regression-based tests of Egger, Harbord, and Peters. Factors known to affect the performance of established tests, including effect size, number of trials in each meta-analysis, degree of between-study heterogeneity, and degree and type of publication bias (a specific cause of small-study effects) are simulated to reflect characteristics of meta-analyses in the biomedical literature. </p><p> The simulation demonstrated that all of the procedures evaluated for control of type 1 error in the novel method maintained an error rate below the nominal rate under all scenarios, suggesting that any of these procedures may be used to implement the novel method. In contrast, error rates for the established tests of Begg and Mazumdar, Egger, Harbord, and Peters were at or above the nominal rate under most scenarios. The lower-tailed, upper-tailed, and two-tailed novel tests showed little power in excess of the type 1 error rate under all conditions. In contrast, established tests demonstrated variable power depending on the conditions. Specifically, the power of established tests increased with an increase in effect size, an increase in the number of trials in each meta-analysis, an increase in the severity of publication bias, and publication bias that operated by effect size rather than by <i> p</i>-value. In contrast, the power of established tests decreased with an increase in heterogeneity. Overall, Egger’s test demonstrated the highest power. Despite the low power of the novel method, selected circumstances under which it may be useful are described.</p>
| Identifer | oai:union.ndltd.org:PROQUEST/oai:pqdtoai.proquest.com:10137622 |
| Date | 14 July 2016 |
| Creators | Bucci, Jay Robert |
| Publisher | The University of Texas Health Science Center at San Antonio |
| Source Sets | ProQuest.com |
| Language | English |
| Detected Language | English |
| Type | thesis |
Page generated in 0.0022 seconds