Title: Comparison of methods for the analysis of pairmatched cluster randomised trials: A simulation study
1Comparison of methods for the analysis of
pair-matched cluster randomised trials A
simulation study
Patty Chondros Prof John B Carlin Dr Obioha C
Ukoumunne Prof Jane M Gunn
2Overview
- Pair-matched cluster randomised trials (CRTs)
analysis issues - Methods for analysing continuous outcomes for
pair-matched CRTs - Simulation design and results
- Summary
3Why use pair-matched CRTs?
- When number of clusters is small ...
- Minimise risk of imbalance on known prognostic
factors - Increase precision of the estimated intervention
effect - Add credibility to the study
- Clusters are matched on risk factors associated
with the outcome - For each pair, one cluster is randomly allocated
to intervention arm one to the control arm - If matching is effective, pair-matched CRTs will
be more efficient than completely randomised CRTs - Matching correlation
- Correlation between cluster level outcome means
within pairs (strata)
4Analysis issues of pair-matched CRTs
- Analysis methods for pair-matched CRTs are more
limited than completely randomised CRTs - There is no replication of clusters within each
stratum and trial arm combination - Between-cluster variance cannot be estimated
- variation within each trial arm is confounded
with variation between-strata - variation within each stratum is confounded with
the intervention effect - Intra-cluster correlation (ICC) can not be
estimated - measures the positive correlation of individuals
within the same cluster - proportion of the total variance due to
between-cluster variance - For the analysis of pair-matched CRTs
- Standard error of the intervention effect is
derived from between-stratum information
5Cluster level methods
- Paired t-test on cluster means
- Two sample t-test on cluster means
- Ignore the matching
- Cluster level random effects regression (CL/RE)
- Random effects meta-analysis approach on cluster
means (Thompson et al, 1997) - Each stratum is treated as an individual study
- Effects of strata are treated as random effects
- Intervention effect is a weighted average of
differences between cluster means within strata
(pair of clusters) - Use t-based confidence intervals with dfk-1,
where k is the number of strata
6Individual level regression-based methods
- Marginal models using Generalised Estimating
Equations (GEEs) - Extension of GLM that allows for correlated
outcomes - Information sandwich (robust) standard errors
- gives consistent SEs even when correlation
structure is misspecified - Exchangeable working correlation
- Use t-based CIs
- Two approaches
- Adjust for the clustering effect, ignoring the
matching (df2k-2) - Treat strata (pairs of clusters) as clusters
(dfk-1) - where knumber of strata
- Random effects (RE) model
- Random effect of the cluster (or stratum) is
modelled explicitly
7For the analysis of pair-matched CRTs.
- Martin et al (1993) Diehr et al (1995) showed
that using cluster level two sample t-test
provides more efficient estimates than cluster
level paired t-test when number of stratalt10
matching is not effective - any gains in precision using matched analysis is
offset by loss in degrees of freedom - We extend this work to individual level
regression methods and meta-analysis approach for
the analysis of pair-matched CRTs
8Aim
- Evaluate the performance of the statistical
analysis methods that can be applied to CRTs with
a pair-matched design for continuous outcomes - Methods for analysing CRTs
- Cluster level analysis methods
- Individual level analysis methods
9Simulation design Parameter values
Continuous outcome Intervention effect 0
10Simulation design
- 2000 datasets generated for each combination of
design parameter values - Correlated data generated from random effects
model - Analysis methods applied to each data set
- For each data set saved the
- estimated intervention effect
- standard error of the estimate
- 95 confidence interval of the intervention
effect
11Model for generating correlated continuous outcome
Strata
Clusters
Individuals
12Model for generating correlated continuous outcome
Matching correlation
Intra-cluster correlation
13Simulation design Evaluation
- Coverage of 95 Confidence Intervals
- Percentage of confidence intervals that include
the true intervention effect - Coverage estimated with a margin of error of 1
- Mean width of confidence intervals
- Given two analysis methods with similar coverage,
method with narrower CIs is preferred as it
indicates greater precision
14Adjust for clustering effect Matching
correlation0.1Coverage
15Adjust for clustering effect Matching
correlation0.1 Mean width of 95 CIs
16Adjust for clustering effect Matching
correlation0.5Coverage
17Adjust for clustering effect Matching
correlation0.5 Mean width of 95 CIs
18Summary Analysis methods that ignore matching
- Cluster level 2 sample t-test provided good
coverage and had narrower CIs than cluster level
paired t-test when matching correlation was 0.1
number of strata was lt10 - GEEs adjusting for clustering effect
- generally provided good coverage when matching
correlation was 0.1 and number of strata lt10 - similar mean width of CI with cluster level 2
sample t-test - Generally, RE model adjusting for clustering
effect only, had poor coverage when number of
strata lt10 - All methods above provided good coverage when
number of strata was large and matching
correlation was 0.1 or less - mean width of CIs similar to CL paired t-test
(Results not shown)
19Methods that allow for matching Matching
correlation0.5Coverage
20Methods allowing for matching Matching
correlation0.5Coverage
21Summary Analysis methods that allow for matching
- Random effects meta-analysis approach performed
poorly when number of strata lt10 - GEEs treating strata as clusters had identical
results to cluster level paired t-test - Individual level random effects model with random
effects for strata gave very poor coverage as ICC
increased and/or cluster size increased - design effect increases
22Further work
- Unbalanced cluster size
- Binary outcomes
- Loss of cluster from a pair
- Adjustment of covariates (Donner, Taljaard, Klar,
2007)
23References
- Martin DC, Diehr P, Perrin EB, Koepsell TD
The Effect of Matching on the Power of Randomized
Community Intervention Studies. Statistics in
Medicine 1993, 12(3-4) - Diehr P, Martin DC, Koepsell T, Cheadle A
Breaking the matches in a paired t-test for
community interventions when the number of pairs
is small. Statistics in Medicine 1995, 14(13) -
- Thompson SG, Pyke SDM., Hardy RJ The design
and analysis of paired cluster randomized trials
an application of meta-analysis techniques.
Statistics in Medicine 1997, 16(18) - Klar N, Donner A The merits of matching in
community intervention trials A cautionary tale.
Statistics in Medicine 1997, 16(15) - Donner A, Taljaard M, Klar N The merits of
breaking the matches a cautionary tale.
Statistics in Medicine 2007, 26(13)
24Variance components