Meta-Analysis, a practical story by Darko Medin - part I
In this article (first of the series), i will discuss some important aspects of Meta-Analysis in Systematic Reviews. Let me start with the obvious. Main Statistical and even plain logic is that the larger sample, the better the evidence derived from it. Moving from smaller samples to large phase IV clinical trials shows a trend of improving evidence on all aspects.
A logical step to get even more credible evidence is to actually try to compare these parameters across multitude of studies and derive a pooled result. Most credible results are from larger trials being pooled, but smaller studies might also be pooled or even a combination of different sized studies (sometimes even parts of the same study might be pooled to improve evidence output).
Meta-analysis is considered as one of the most trustworthy evidence based frameworks in the literature and is considered as one of the main components of clinical validation of study results.
This does comply with many Statistical theory and plain logic, but when trying to derive such results, Statistical theory also plays a pivotal role to address potential large scale issues and problems too. Very often such result are based on a study level, compared to having each subject data present in the dataset and as such two different approaches exist in the Meta-analysis framework. Methodologies between studies might vary, populations might vary, treatments might vary, follow up periods might vary, actual period of the study might vary, even the statistical methods themselves might vary… and of course the results might vary too. Addressing all this potential variation is one of the keys of a good Meta-analysis. So i will place this is number one of the key Meta-Analysis aspects, which is HETEROGENEITY and how to address and interpret the Heterogeneity in results.
Principles of contribution of single studies to pooled effects are another key to determine the best Meta-analysis model. This is number two of the key Meta-Analysis principles called STUDY WEIGHTING. Heterogeneity and many other aspects in the Meta-analysis are drawn from single studies effects, but also their deviations, confidence intervals, credible intervals, errors etc. Different Meta-analysis methods are most often varying in this principle.
This brings the next question. Creating a METHODOLOGICAL FRAMEWORK for Meta-analysis. There are many methods of which most will be discussed in this series, from inverse variance Bayesian frameworks and Network-Meta-analysis .
Also these frameworks might have different assumptions like Fixed/Random effects models, distribution assumptions, methodological assumptions, clinical assumptions, dependence assumptions and many others, which brings next section which is ASSUMPTIONS of a Meta-Analysis.
Bias is generally a key thing needs to be addressed. Before it can be addressed, it must be analyzed. Bias reduction can sometimes be limited but it can be reported and included in interpretation. This marks the section of BIAS analysis. There are many types of bias in Meta-analyses like Selection Bias, Publication bias, Citation Bias or Study design biases, reporting biases and many others. In fact Meta-analysis is one of the best frameworks to identify such biases.
Last but not least, Meta-analysis also requires some Data Science skills, especially in the validation of results, which brings me to the one of the most important sections, called VALLIDATION.
In the next article I will be discussing some of the software where Meta-analysis can be efficiently performed.
By Darko Medin
Msc Faculty of Mathematics and Natural Sciences
Data Science/AI/Biostatistics Expert
