Title: Metaanalysis
1Meta-analysis
- What is it, is it useful?
2What is meta-analysis?
- An analysis of analyses trying to make
generalisations from a series of experiments in
an unbiased, quantitative way
3Making Generalisations I
- Traditional narrative review (X found this, on
the other hand Y found that ..) - Subjective
- Not quantitative
- Can be overwhelmed by large number of studies
(e.g. high CO2 research)
4Making Generalisations II
NB Cant take mean response Conclusions biased
because significance level of a study is a
function of the sample size i.e. some studies
are better than others
5Making Generalisations III
- Meta analysis
- weights each study according to sample size
General procedure - find a set of studies with
similar data - extract an effect size from each
study - calculate a weighted average effect size
across studies - construct a confidence
interval - divide data into groups (e.g.
terrestrial vs. marine) and test to see if groups
behave similarly
6Example
- Stomatal conductance (gs) of trees at elevated CO2
Open-top chamber, Antwerp
CO2 response of gs of Sitka spruce (Craig Barton)
71) Which studies?
- Should you exclude those of dubious quality?
FACE
pots
chambers
8Three Major Meta-analyses of gs
- Curtis Wang 1998 all studies published to
date (pots, chambers, branch bags etc) - Medlyn et al 2001 ECOCRAFT (long-term,
field-based experiments) but
cf. Curtis Wang showed studies gt 1 year were
consistent - Ainsworth Rogers 2007 FACE studies only (but
all growth forms)
9Potential Biases
- Non-significant results dont get published the
desk-drawer problem - Re-publication
- Overcoming bias
- Know the literature!
- Calculate how many studies would be needed to
give a barely non-significant result - Funnel plots
10Funnel Plots
no bias
Sample Size
Effect Size
bias
Hunter Schmidt 2004
data from Medlyn et al. 2001
112) Extracting data
- a) Obtain from publications
- (graphical data can be digitised e.g. DataThief)
- b) Ask nicely
- Whats needed?
- Mean, standard deviation, number of replicates
for control and experimental treatments - Remember to include these data in your own
articles!
123) Which data points?
- Data points are assumed to be independent.
- Different species OK (but beware Bazzaz
effect) - Different treatments OK (e.g. fertilized /
unfertilized) - What about repeated measures??
- Curtis Wang 1 data point from each
experiment mid-growing season, final year of
experiment - Ainsworth Rogers all measurements (98 data
points from 4 experiments!)
13Different leaves?
- What if data are presented separately for
different leaf categories, e.g. - age class
- aspect
- canopy height
- Try to make comparable across studies.
e.g. Barker et al. 2005 CO2 and photoinhibition
in snow gum
14A current problem
- Duke FACE experiment
- pine overstorey
- sweetgum understorey, some emergents
- pine gs measured on large number of trees each
year - sweetgum gs measured on 2 trees/ring, once
- BUT same number of replicates
- How to weight?
154) Calculating effect size
You can use a number of different statistics to
represent effect size e.g. correlation
coefficient r (ask Chris Lusk) Hedges d
standard deviation units
What does d mean? d effect size 0 none
0.2 small 0.5 medium 0.8 large
Xis are means of experimental (E) control (C)
groups si is the pooled standard deviation of
both groups
16Ln Response Ratio
- A metric that many ecologists are comfortable
with. - Response ratio r XiE / XiC
- Ln response ratio L ln (r)
- The variance of L is approximately equal to
Confidence interval is given by
Hedges et al. 1999
17Presenting Data
Rustad et al. 2001
185) Combining effect sizes
- Fixed effects model
- Studies have a common true effect size
- (differences just due to sampling error)
- Mixed effects model
- Studies have a common mean effect
- Random variation as well as sampling error
- (Most appropriate in ecological studies)
19Combining effect sizes Fixed Model
- The cumulated effect size across all expts is a
weighted average of the effect size estimates - Weights wi are the reciprocals of the sampling
variances, wi 1/vi. - Variance of d is
- Can then construct a confidence interval for d
20Combining effect sizes Mixed Model
- We need to modify the variance
- vi vi s2pooled
- where s2pooled is the pooled within-class
variance. - (Formula Gurevitch Hedges 1993)
- The weights wi are the reciprocals of the
sampling variances, wi 1/vi - Weights can be modified to avoid undue influence.
216) Division into Classes
- Effect of high CO2 might vary among studies e.g.
by - functional group
- nutrient availability
- exposure time
- Can test for differences among classes
- calculate between-class heterogeneity
- QB Si1 to m Sj 1 to kiwij (di - d)
22Presenting Comparisons
Effects of high CO2 on gs in FACE
Ainsworth Rogers 2007
23Beware Confounding
Medlyn et al. 2001
Number of studies Mature Young Coniferous 6
2 Broadleaf evergreen 3 0 Broadleaf
deciduous 1 17
24Philosophical Issues
- What does meta-analysis really tell us?
- Is meta-analysis a Good Thing??
Luo et al. 2006 Effects of high CO2 on carbon
accumulation in plant biomass high variability
ln
25References
- Gurevitch, J. and L.V. Hedges, 1993.
Meta-analysis Combining the results of
independent experiments. Pages 378-398 in S.M.
Scheiner and J. Gurevitch, editors. Design and
Analysis of Ecological Experiments. Chapman and
Hall, New York. - Rosenberg, M.S., D.C. Adams, and J. Gurevitch.
1997. MetaWin Statistical Software for
Meta-Analysis with Resampling Tests. Version 1.0.
Sinauer Associates, Sunderland, Massachusetts. - Hedges LV, Gurevitch JC, Curtis PS (1999) The
meta-analysis of response ratios in experimental
ecology. Ecology 80 1150-1156. - Hunter JE, Schmidt FL (2004) Methods of
Meta-Analysis, 2nd ed. HA 29.H847 - Cooper H, Hedges LV (1994) The Handbook of
Research Synthesis. Q180.55.M4.H35