Meta-Analysis
Chloe Mirzayi
May 14th, 2019
Meta-Analysis
Systematic Review Meta-Analysis
Define Identify Summarize Statistical
topic studies findings Analysis
• Clarify issue • Define • Synthesize • Calculate
of interest inclusion/excl narrative of effect sizes
• Create usion criteria results • Determine
standardized • Database • Present and heterogeneity
search terms search interpret • Assess bias
• Assess study findings • Secondary
quality analyses
Extract effect
sizes
• From each article included in
your systematic review, extract
the main finding (the effect
size) and variance
• This could be an odds
ratio, regression
coefficient, or a difference
in means
Example: B=0.42, SE=0.13,
OR 1.52 from Gurung et al.
2017
Calculate standardized
effect size
• Convert this main finding and
variance into a standardized form
(typically Cohen’s d or Hedge’s g)
• Recall from the quantitative
section of this course that
standardization is just a simple
formula!
Example: B=0.42, SE=0.13
• Do all the studies and the observed
effect sizes stem from the same
homogenous population?
• If yes, we can use a “fixed-effects”
Determine model
• If no, we can use a “random-effects”
heterogeneity model
• But how do we know?
• And what is a “homogenous
population” in this context anyway?
Cochran’s Q
• The larger the Q, the more heterogeneous
the population
• Sounds a bit like a chi-squared test
statistic (because it is!)
Measures of • Unfortunately it doesn’t work well with
too many or too few studies
heterogeneity
I2
• The percentage of variation across studies
due to study variation rather than chance
• Doesn’t that sound like an R2?
• Higher I2 means more heterogeneity
Create the
forest plot
Assess publication
bias (funnel plot)
• X-axis: effect size
• Y-axis: standard error
• Look for symmetry
around the estimate
(line in the middle)
• Asymmetry is
evidence of
publication bias
Low evidence
of bias
• Most symmetrical
• Slight evidence of
bias based on
missing studies in
lower left part of
funnel
Evidence of
bias
• No clear symmetry
• Studies outside
the funnel
Evidence of
bias
• Clear asymmetry
• Evidence of bias
against contrary
results
See if there is a difference
between certain subpopulations,
treatments, or study designs
within our larger meta-analysis
Subgroup
analyses
Repeat the meta-analysis
process comparing the subgroup
to a reference group
Discussion of Thakkar et al. 2016
• What treatment were they interested in?
• Did they find that the treatment worked?
• How do you know?
• Did they use a random-effects or fixed-effects model?
• Did they perform any subgroup analyses?
Discussion of Thakkar et al. 2016
• Here is the funnel plot
from the supplement
from the article. Do you
think there’s evidence of
publication bias?
• Based on the article and
our discussion, would you
recommend text
messaging to improve
medication adherence?