Journal of Industrial Engineering Research, 1(10) Special 2015, Pages: 1-7
IWNEST PUBLISHER
Journal of Industrial Engineering Research
(ISSN: 2077-4559)
Journal home page: http://www.iwnest.com/AACE/
Corresponding Author: Nurul Afiqah Misran, Department of Computational and Theoretical Sciences, Faculty of Science,
International Islamic University Malaysia,Kuantan Campus,25200 Kuantan, Pahang, Malaysia
Tel: +6013 3634550; E-mail: fiqamis12@gmail.com
An Empirical Assessment of Meta-Analysis Estimates from Multi- level Studies
Nurul Afiqah Misran and Nik Ruzni Nik Idris
Department of Computational and Theoretical Sciences, Faculty of Science, International Islamic University Malaysia,Kuantan
Campus,25200 Kuantan, Pahang, Malaysia
ARTICLE INFO ABSTRACT
Article history:
Received 3 August 2015
Accepted 28 October 2015
Available online 31 October 2015
Keywords:
Meta-analysis;Combine data;
Individual patient data; Aggregate
data; Simulation; Bias
Background: A conventional meta-analysis may be performed using studies which are
available at individual patient level (IPD) or aggregate level (AD). Presently however,
meta-analysis that combine the two levels of studies is increasingly common. The
implications of utilising different levels of data on the overall estimates have not been
fully explored. Objective: This study examined the efficacy of the estimates of overall
treatment effect from AD, IPD and the mixed AD: IPD studies, and investigated how
they differ from the true treatment effect. Additionally, this study investigated the
influence of the ratio of AD: IPD on the precision of the overall treatment effects
estimates. The bias, root mean-square-error (RMSE) and coverage probability were
used to assess the efficiency of the overall estimates. Results: The results showed that
the IPD meta-analysis produced better estimates in terms of RMSE compared to AD
meta-analysis and the mixed AD:IPD meta-analysis. For the cases where both the AD
and IPD studies were available, our findings showed that the combined AD : IPD data
produced better estimates, in terms of precision, than utilising the AD alone.
Conclusion: It is therefore recommended that available IPD should always be included
in a conventional meta-analysis using summary level data as significant statistical
benefit is gained by pooling the two levels of data.
© 2015 IWNEST Publisher All rights reserved.
To Cite This Article: Nurul Afiqah Misran and Nik Ruzni Nik Idris., An Empirical Assessment of Meta-Analysis Estimates from Multi-
level Studies. J. Ind. Eng. Res., 1(10): 1-7, 2015
INTRODUCTION
Meta-analysis is a statistical technique for integrating quantitative results from several sources. The main
aim is to provide conclusions based on the whole body of research. A traditional meta-analysis involves
integration of aggregate data (AD) which is extracted from the individual study publications [1]. Typical AD
includes a mean difference for continuous outcomes or the number of events and participants for binary
outcomes. The overall treatment effect is computed by taking the weighted average of the effects across the
trials using methods such as the inverse variance method [2] or the Mantel-Haenszel method [3] for the binary
data. Alternatively, meta-analysis may be performed based on individual patient data (IPD), where raw data
from individual study is obtained and synthesized directly. Although it has numerous advantages compared to
the traditional meta-analysis, particularly in terms of type of analyses that can be done, IPD meta-analysis is
usually relatively costly and time consuming [4][5]. Another potential problem for IPD analysis is that IPD are
seldom or may not be available from all the individual studies.
Combining available IPD with the AD has been advocated [6][7] in order to maximized available
information and allows larger number of patients and greater part of the evidence-based to be included.
Currently however, combined-study level meta-analysis is not very common. A review of 199 meta-analyses [8]
which has both the IPD and AD available, found 33 combined the data in their analysis and 166 did not. The
review noted that the articles that combined the IPD and the AD in their studies had, on average, IPD available
in 64% of the studies, while articles which did not combine the studies had an average of 90% IPD available.
The implications of these ratios on the overall meta-analysis estimates are yet to be explored. A study [9]
examined the models used for combining the IPD and AD in meta-analysis of continuous outcomes. In part of
their study, estimates extracted from IPD-only (i.e. only IPD were utilized when both type are available) were
compared with those from the combinations of both IPD and AD. The study was based on real data from a
Hypertension study [10]. The results showed that the benefit of combining the IPD with AD increases as the
proportion of IPD studies decreases. The comparison was made against the estimates from all IPD studies.