1 Lecture 10: Meta-analysis of intervention studies • Introduction to meta-analysis • Selection of studies • Abstraction of information • Quality scores • Methods of analysis and presentation • Sources of bias
Jan 04, 2016
1
Lecture 10:Meta-analysis of intervention
studies
• Introduction to meta-analysis
• Selection of studies
• Abstraction of information
• Quality scores
• Methods of analysis and presentation
• Sources of bias
2
Definitions
• Traditional (narrative) review:– Selective, biassed
• Systematic review (overview):– Synthesis of studies of a research question– Explicit methods for study selection, data
abstraction, and analysis (repeatable)
• Meta-analysis:– Quantitative pooling of study results
3 Source: l’Abbé et al, Ann Intren Med 1987, 107: 224-233
4
Protocol preparation
• Research question
• Study “population”:– search strategy– inclusion/exclusion criteria
5
Protocol preparation
• Search strategy:– computerized databases (Medline, CINAHL,
Psychinfo, etc.):• test sensitivity and predictive value of search strategy
– hand-searches (reference list, relevant journals, colleagues)
– “grey” (unpublished) literature:• pro: publication bias• con: results less reliable
6
Identifying relevant studies for systematic reviews of RCTs in vision research (Dickerson, in Systematic Reviews, BMJ,1995)
• Sensitivity and precision” of Medline searching• Gold standard:
– registry of RCTs in vision research• extensive computer and hand searches
• contacts with investigators to clarify design
• Sensitivity:– proportion of known RCTs identified by the search
• “Precision”:– proportion of publications identified by search that were
RCTs
7Source: Chalmers + Altman, Systematic Reviews, BMJ Publishing Group, 1995
8Source: Chalmers + Altman, Systematic Reviews, BMJ Publishing Group, 1995
9Source: Chalmers + Altman, Systematic Reviews, BMJ Publishing Group, 1995
10
Protocol preparation
• Study “population”:– inclusion/exclusion criteria:
• language
• study design
• outcome of interest
• etc.
Source: Data abstraction form for meta-analysis project
11
Protocol preparation
• Data collection:– standardized abstraction form– number of abstractors– blinding of abstractors– rules for resolving discrepancies (consensus,
other)– use of quality scores
12Source: l’Abbé et al, Ann Intren Med 1987, 107:224-233
13
Analysis
• Measure of effect:– odds ratio, risk/rate ratio– risk/rate difference – relative risk reduction
• Graphical methods:– conventional (individual studies)– cumulative– exploring heterogeneity
14Source: Chalmers + Altman Systematic Reviews, BMJ Publishing Group, 1995
15Source: Chalmers + Altman Systematic Reviews, BMJ Publishing Group, 1995
16
Analyses
• Pooling results:– is it appropriate?
– equivalent to pooling results from multi-centre trials
– fixed (e.g., Mantel-Haenzel) methods• assume that all trials have same underlying treatment effect
– random effects methods (e.g., DerSimonian & Laird):• allow for heterogeneity of treatment effects
17Source: Chalmers + Altman Systematic Reviews, BMJ Publishing Group, 1995
18
19
20Source: l’Abbé et al, Ann Intren Med 1987, 107:224-233
21
Quality scores
• Rating scales and checklists to assess methodological quality of RCTs
• How should they be used?– Qualitative assessment– Exclusion of weaker studies– Weighting of estimates
22
Does quality of trials affect estimate of intervention efficacy? (Moher et al, 1998)
• Random sample of 11 meta-analyses of 127 RCTs • Replicated analysis • Used quality scales/measures• Results:
– masked abstraction provided higher quality score than unmasked
– low quality trials found stronger effects than high quality trials
– quality-weighted analysis resulted in lower statistical heterogeneity
23Source: Moher et al, Lancet 1998, 352: 609-13
24
Source: Moher et al, Lancet 1998, 352: 609-13
25Source: Moher et al, Lancet 1998, 352; 609-13
26
Unresolved questions about meta-analysis
• Apples and oranges?– Between-study differences in study population,
design, outcome measures, etc.
• Inclusion of weak studies?
• Publication bias– methods to evaluate impact – - particularly with small studies
• Is it better to do good original studies?
27
Large trials vs meta-analyses of smaller trials (Cappelleri et al, 1996)
• Selected meta-analyses from Medline and Cochrane pregnancy and childbirth database with at least 1 “large” study and 2 smaller studies:– sample size approach (n=1000+) - 79 meta-analyses– statistical power approach (adequate size to detect treatment
effect from pooled analysis - 61 meta-analyses
• Results:– agreement between larger trials and meta-analysis 82-90%
using random effects models– greater disagreement using fixed effects models
28
Large trials vs meta-analyses of smaller trials (Cappelleri et al, 1996)
• Results:– agreement between larger trials and meta-analysis 82-
90% using random effects models
– greater disagreement using fixed effects models
• Conclusion:– large and small trial results generally agree
– each type of trial has advantages and disadvantages:• large trials provide more stable estimates of effect
• small trials may better effect heterogeneity of clinical populations
29
Risk ratios from large studies vs pooled smaller studies (Cappeleri et al,1996)
(Left- sample size approach; right - statistical power approach)
Source: Cappeleri et al, JAMA 1996, 276: 1332-1338
30
Source: Cappeleri et al, JAMA 1996, 276: 1332-1338
31
Discrepancies between meta-analyses and subsequent large RCTs (LeLorier et al, 1997)
• Compared results of 12 large (n=1000+) RCTs with results of 19 prior meta-analyses (M-A)on same topics
• For total of 40 primary and secondary outcomes, agreement between large trial and M-A only fair (kappa = 0.35, 95% CI .06 to .64)
• Positive predictive value of M-A = 68%• Negative predictive value of M-A= 67%
32Source: Lelorier et al, NEJM 1997, 337: 536-42
33Source: Lelorier et al, NEJM 1997, 337: 536-42