Year: 2013
Type: dissertation
Abstract: In clinical evaluation processes, meta-analysis is a statistical methodology to synthesise results of several trials for the purpose of quantitatively summarising the evidence expressed as a treatment effect (e.g, mean treatment difference for continuous outcomes).Especially in clinical perspectives, there is a growing interest in extending meta-analysis beyond estimat- ing an overall treatment effect, and. to produce results tailored to the individual patient or clinically relevant subgroups.Meta-regression (MR) is a technique for modeling relationships between the treatment effect and trial-level covariates, and also can be used for assessing how the patient characteristics affect the clinical efiectiveness in the context of meta-analysis.However, most meta-analysis methods build models on,aggregate data (AD) obtained by summarising individual patient data (IPD) which should have beerr measured originally in each trial, and thus a pooled treatment effect is estimated with ignoring scheme of sampling IPD.For this reason, the MR model have often been criticised.The MR model incorporates covariates as summary statistics on background factors of patients, such as a mean age and a proportion of male patients in each trial.This means that the patient characteristics are evaluated expediently with trial-level covariates in place of patient-level covariates.It causes a technical issue that is referred to as ecological bias, and leads to a Iimitation in interpre-tation.In pa,rticular, it is well known that, for a treatment-covariate interaction between a clinical treatment and a patient bharacteristic, a test using the MR model has seriously lower statistical power than that using the IPD-based model.Note that, for the inference of the treatment-covariate interaction, the MR model yields a result of just an 'across-trial interaction efiect' between the treatment efiect estimates and mean covariate values, not a 'within-trial interaction efiect' between individual outcomes and individual covariate values.As alternative solutions to this problems, some meta-analysis methods using IPD have been suggested.In these methods, once the original IPD including patient-level covariates are collected from all trials involved, any flexible statistical approaches, such as multilevel models and hierarchical random effect models, are applied to the IPD.Meta-analyses based on IPD allow one to achieve much more meaningful evaluation on the treatment-covariate interaction by separating it into the across-trial and the within-trial efiect; in particular, the IPD meta-analysis is an only way to assess the within-trial interaction efiect.However, use of the IPD may have a disadvantage related to their resources, such as substantial time and costs to obtain and pr6cess the IPDo And also,practitioners cannot alwars collect the IPD ttom all trials because the IPD Inight have beOn lost Or destroyed.For this reason,it has becolne ivreasingly important to consider situations where some trials pЮ 宙de IPD(IPD trials)and the others provide only AD(AD triab)。 Somё researchers have abё ady illvestigated how to combine IPD and AD in meta-analySis,especially when tお at品nt̲cOvariate interaction is of interest.■ om these backgrounds,we prOpose a metaanalysis method for estimating b6th the across-trial and the within-trial interaction.For the caSe that all trialS prOvide only AD, we■rst assume an IPD metttanalysis model including parameters of the across-t五 4 and the within-trial interaction elbct,and then llllarginalise the density of IPD with respect to imiSSing IPD.This proce,s produces a like五 hood for the AD available,and allowS one to get information on the withintrialinteraction by meta-analysing several AD t五 als.we en3phasise that the wtthintrial intё raction can be approximately estittated by using this likehhood even ifonly the Ap are available ttom eadh trialo Actu嵐 崎 ,SOme simulatiott studies suggested that the proposed IIlethod has potential benellts to the inference of the withintrial interaction in comparison with the existing MR approacho When S6me trials pЮ vide IPD and the others provide only AD,the proposed method is sil■ ■ ply extendOd to combillle IPD and ADo There,the likelihood for parameters to be estimtted is given by product of a hkehhood for the IPD trials and the marginを 山sed likelihood for the AD trials.This again allows one to get information on the within-tiialinteraction ibm the AD trials.ThЮ ugh simulation studieS,the pЮ posed nlethod provided sma■ er biases and smaller lnean-square errors ibr estillnator of the within- trial interaction in comparison with some e対 Sting lllleta-analysis IIIlethods,especially wh〕 n the proportiOn of available IPD was sman.And 4sO,Simulation studies investigated hOw the proportion of available IPE)arects the biases and the mean… squtte errors for estilnator of the within-triaHnteraction obtailled ttom the pFopOSed methol・ These resuns could。 • ra llseful guidance ifone considers how many IPD t五 a卜 shOuld be conected to presewe a deshed level of statistical power.Note that the prOposed method is applicable when parameters tO be estiIIlated can be asslllned as fⅨ ed elbcts;so that the treatllllent efFect and the treatmentcovariate inteFattion efFects are assllmed to be common across trials.As a breakthrough of the e対 sting meta-analysis methё ds,we pЮ pose a meta-alalysis method based on simulated IPD(SIPD),whiCh recOistructS the missing IPD for each trial and the,applies a standard IPD meta-analySiS ttOdelto each SIPDo We here discuss two types of samphng procedllres for generating the SIPD:frequentist and Bayesian prOcedllres.Since
Authors: ヤマグチ, ユウスケ, 山口, 祐介, Yamaguchi, Yusuke
Language: English
Subfield: Statistics, Probability and Uncertainty
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7016669599
