Professional Documents
Culture Documents
Meta Analysis 2019 2020
Meta Analysis 2019 2020
Meta Analysis 2019 2020
➢ The outcome, whether or not the patient died, was the same in all the studies
➢ The trials varied substantially in size: Md ≈ 100; 2 large scale trials (12,000
and 17,000 patients)
6 were statistically
significant (look at CI; they
do not include 1.0), 27 were
not (they had small sample
sizes and low statistical
power)
Conclusion of a narrative
or systematic review:
evidence against effect or
inconsistent results
Conclusion of MA: Effect
reasonably consistent
(most between 0.50 and
0.90) – appropriate
composite a summary
effect. The summary effect
is a RR of 0.79 (95% CI
0.72-0.87) – 21%
decrease in risk of death.
The p-value is 0.00000008
Those treated were more likely to survive
1. Integrating results in meta-analysis
(comparative) effectiveness
➢ Provide an updated view on a given topic
➢ Enhance establishing the generality of findings + identifying
SCIENCE
➢ Go broad!
2. Steps in meta-analysis
Step 2. Searching the literature
➢ Electronic searches:
➢ Type I error: broad searches normally retrieve a large number of studies
that no human could summarize
➢ Type II error: narrow searches routinely miss relevant studies
➢ Characteristics of studies
➢ Training coders
Information Details
Study level information Report information – e.g., ID, type of publication, year of
publication, funding, etc.
It needs to be coded only once Substantive issues– e.g., sample descriptors (age, sex,
for a given study and does not diagnostic, etc.), type of treatment, dose, theoretical
vary for different outcomes, orientation, etc.
follow-ups, sample breakdowns,
etc. Methods and procedures – e.g., sampling method, quality
of measures, follow-up measures, etc.
Effect size level information Dependent measure descriptors – e.g., type of ES,
outcome construct, etc.
Aspects that are specific to a ES (numerical) data – e.g., means and SD, ES based on,
particular quantitative relationship favoured group, calculated ES, page number where ES
or study finding was extracted, etc.
2. Steps in meta-analysis
Step 3. Coding
Manual Form
Manual
Form
“Specialist” coders
2. Steps in meta-analysis
Step 3. Coding
Paper MA software
Example of
FileMaker
Pro from
Lipsey and
Wilson (2001)
Spreadsheet
Example of
Excel from
Example of paper coding Borenstein et
from Cooper (2010) al. (2009)
2. Steps in meta-analysis
Step 3. Coding
Single ES
Hierarchy of ES
per study
2. Steps in meta-analysis
Step 3. Coding
➢ Strategies
Average
In a published MA…
In practice... The kind of data used in the primary studies will usually
lead to a pool of 2 or 3 ES that meet the criteria:
S pooled =
(n1 − 1)S12 + (n2 − 1)S 22 D = X1 − X 2 Correlation
Pre- and
post-scores
n1 + n2 − 2 2
S diff
between pre- and
VD =
post-scores are reported
2 2
S S n
VD = 1
+ 2 assuming σ1 ≠ σ2
n1 n2 S diff = S12 + S 22 − 2 r S1 S 2 assuming σ1 ≠ σ2
S diff = 2 S pooled
2
(1 − r ) assuming σ1 = σ2
2. Steps in meta-analysis
Step 4. Computing effect sizes
S within =
(n1 − 1)S12 + (n2 − 1)S 22 S diff
S within =
n1 + n2 − 2 2(1 − r ) Hedges’ g
n1 + n2 d2 1 d2
Vd = + 2(1 − r )
Vd = + 3
J = 1−
n1n2 2(n1 + n2 ) n 2n 4df − 1
n1 + n2 − 2 → Ind .Groups g = J d
n − 1 → MatchedGroups Vg = J 2 Vd
2. Steps in meta-analysis
Step 4. Computing effect sizes
Pairs recalled
+ ES
Events Non-Events N
Treated A B n1
Control C D n2
Dead Alive N
Treated 5 95 100
Control 10 90 100
2. Steps in meta-analysis
Step 4. Computing effect sizes
A n1
RR =
C n2
LogRR = ln(RR ) RR = exp(LogRR)
1 1 1 1
VLogRR = + + +
A n1 C n2
2. Steps in meta-analysis
Step 4. Computing effect sizes
A B AD
OR = =
C D BC
LogOR = ln(OR ) OR = exp(LogOR)
1 1 1 1
VLogOR = + + +
A B C D
2. Steps in meta-analysis
Step 4. Computing effect sizes
A C
RD = −
n1 n2 NO log transformation
AB CD
VRD = 3 + 3
n1 n2
2. Steps in meta-analysis
Step 4. Computing effect sizes
Dead Alive N
Treated 5 95 100
Control 10 90 100
1+ r
z = 0.5 ln
1− r e2 z −1
r = 2z
1 e +1
Vz =
n−3
Vr =
(1 − r )
2 2
n −1
2. Steps in meta-analysis
Step 4. Computing effect sizes
2. Steps in meta-analysis
Step 4. Computing effect sizes
Cell entries with a 0 can be problematic for the RR and OR. Adding
a constant to the cells of the 2 × 2 tables is a common solution.
Usually 0.5 is added to each cell of the 2 × 2 table only in those
tables with at least one cell equal to 0.
2. Steps in meta-analysis
Step 4. Computing effect sizes
➢ Example
➢ dat.bcg – 13 studies of the effectiveness of the BCG vaccine
against tuberculosis (Colditz et al., 1994)
number of treated
(vaccinated) subjects that were
tuberculosis positive
Let us move to R…
2. Steps in meta-analysis
Step 4. Computing effect sizes
➢ Example
➢ Example_means – 15 studies of the differences between men
and women in maths anxiety
1 Carroll 94 22 60 92 20 60
2 Grant 98 21 65 92 22 65
3 Peck 98 28 40 88 26 40
4 Donat 94 19 200 82 17 200
5 Stewart 98 21 50 88 22 45
6 Young 96 21 85 92 22 85
Let us move to R…
2. Steps in meta-analysis
Step 4. Computing effect sizes
Example...
n1 + n2
Example SMD = t
n1n2
Situation 2.
Some articles report an ES measure and others use another ES
measure
In a published MA…
Individual studies
Effect size (ES)
Precision
Study weights
p-values
Summary effect
Effect size (ES)
Precision
p-value
Heterogeneity of ES
weighted mean M= i =1
k
W
i =1
i
1
Wi =
1 VYi
VM = k
weighted variance
W
i =1
i within-study variance
W Y
2
k
*
WiYi i i
M =* i =1
W Y 2 − i =1
k
Q= k
W i*
1
W i* =
i i k
i =1
W i =1
i i =1 VY*i
1
df = k − 1 VM * = k
within-study variance +
i
W 2
W *
between-study variance (τ2)
C = Wi −
i
i =1
W i
SEM, 95% CI, Z-value to test H0: θ = 0
(for a difference) or θ = 1 (for a ratio)
2. Steps in meta-analysis
Step 5. Combining effect sizes
➢ FEM gives much more weight to studies with higher sample size
than REM – the weights under the REM are more balanced
➢ CI are narrower with FEM – effects that are significant under a
FEM may no longer be significant under a REM – REM more
conservative
➢ REM become FEM when distributions are homogeneous (tau-
squared = 0)
➢ Assumptions of FEM rarely plausible
➢ Historically, most MA in Psychological Bulletin have used FEM
➢ General advise: use REM a priori
➢ Area of active debate among statisticians
2. Steps in meta-analysis
Step 5. Combining effect sizes
➢ Example
➢ dat.bcg – 13 studies of the effectiveness of the BCG vaccine
against tuberculosis (Colditz et al., 1994)
number of treated
(vaccinated) subjects that were
tuberculosis positive
Let us move to R…
2. Steps in meta-analysis
Step 5. Combining effect sizes
The risk of tuberculosis
infection in vaccinated
individuals is on average
Fitting a REM (method = REML by default) half as large as the infection
risk without the vaccination
Transformed to RR = exp
(-0.7145) = 0.49 (95% CI:
0.34 – 0.70)
2. Steps in meta-analysis
Step 5. Combining effect sizes
i =1
➢ Q has low power when the number of studies is low and when
sample size within studies is low
Q
➢ Larger values of I2, the more heterogeneity
The amount of
heterogeneity in the true
Fitting a REM (method = REML by default) log RR is estimated to be
0.3132
Considerable
heterogeneity
v The distribution of ES is
heterogeneous
2. Steps in meta-analysis
Step 5. Combining effect sizes
analogous to one-way
weighted multiple regression –
ANOVA
Meta-regression
2. Steps in meta-analysis
Step 6. Moderators analyses
2. Steps in meta-analysis
Step 6. Moderators analyses
Q = Qbetween + Qwithin
2. Steps in meta-analysis
Step 6. Moderators analyses
Analogous
XXX to one-way ANOVA
dat.bcg – 13 studies of the effectiveness of the BCG vaccine against
tuberculosis (Colditz et al., 1994)
Moderator: type of allocation (alternate, random, systematic)
H 0 : β1 = β 2 = 0
The type of allocation method
v does not influence the average
effectiveness of the vaccine
β1= 0.4478 and β2 = 0.5369 estimate how much larger the average
Average log RR for random log RR values are when using alternate and systematic allocation,
allocation respectively.
Β1 + β1 = -0.9658 + 0.4478 = -0.5180 is the average log RR for
studies using alternate allocation
Β1 + β1 = -0.9658 + 0.5369 = -0.4289 is the average log RR for
studies using systematic allocation
2. Steps in meta-analysis
Step 6. Moderators analyses
◼ Q for the model (QM) indicates whether the regression model explains a
significant portion of the variability across ES
◼ Q for the residual (QE) indicates whether the remaining variability across ES is
homogeneous
Q = QModel + QError
2. Steps in meta-analysis
Step 6. Moderators analyses
Meta-regression
dat.bcg – 13 studies of the effectiveness of the BCG vaccine against tuberculosis
(Colditz et al., 1994)
Moderators: year of publication (year), absolute latitude (ablat)
Tau-squared without moderators was
0.3132. So, (0.3132-0.1108)/0.3132 = 65%
v
of the total amount of heterogeneity can
be accounted for by including the two
moderators in the model
H0: β1= β2 = 0
We can reject H0 but only absolute latitude appears to have a significant
influence on the effectiveness of the vaccine
▪ Steps:
1. Eliminate iteratively the most extreme small-studies’ ES until
symmetry is reached.
2. Estimate new ES.
3. Put eliminated back + impute until symmetry: this step corrects
against underestimating the average ES variance, but does not affect
the estimate of the mean.
2. Steps in meta-analysis
Step 7. Exploring biases
The estimated effect of the vaccine is smaller with the missing studies filled
in, but results still indicate that the effect is statistically significant
2. Steps in meta-analysis
Step 7. Exploring biases
3. Individual-Participant-Data meta-analysis
▪If you are going to use R, read the book by Chen and Peace (2013)
Applied meta-analysis with R (Chapman & Hall/CRC Biostatistics
Series)
Web pages of interest: