Meta-analysis of controlled clinical trials /
Over the last twenty years there has been a dramatic upsurge in the application of meta-analysis to medical research. This has mainly been due to greater emphasis on evidence-based medicine and the need for reliable summaries of the vast and expanding volume of clinical research. At the same time th...
| Main Author: | |
|---|---|
| Format: | Book |
| Language: | English |
| Published: |
Chichester ; Hoboken, N.J. :
John Wiley & Sons,
2002
Chichester ; New York : 2002 Chichester ; Hoboken, NJ : 2002 |
| Series: | Statistics in practice (Chichester, England)
Statistics in practice Statistics in practice) |
| Subjects: |
Table of Contents:
- 1 Introduction
- 2. Protocol development
- 3. Estimating the treatment difference in an individual trial
- 4. Combining estimates of a treatment difference across trials
- 5. Meta-analysis using individual patient data
- 6. Dealing with heterogeneity
- 7. Presentation and interpretation of results
- 8. Selection bias
- 9. Dealing with non-standard data sets
- 10. Inclusion of trials with different study designs
- 11. A Bayesian approach to meta-analysis
- 12. Sequential methods for meta-analysis
- App. Methods of estimation and hypothesis testing.
- 1.1 The role of meta-analysis 1
- 1.2 Retrospective and prospective meta-analyses 3
- 1.3 Fixed effects versus random effects 5
- 1.4 Individual patient data versus summary statistics 6
- 1.5 Multicentre trials and meta-analysis 7
- 2 Protocol development 11
- 2.4 Outcome measures and baseline information 13
- 2.5 Sources of data 13
- 2.6 Study selection 14
- 2.7 Data extraction 15
- 2.8 Statistical analysis 16
- 2.8.1 Analysis population 16
- 2.8.2 Missing data at the subject level 17
- 2.8.3 Analysis of individual trials 18
- 2.8.4 Meta-analysis model 19
- 2.8.5 Estimation and hypothesis testing 19
- 2.8.6 Testing for heterogeneity 19
- 2.8.7 Exploration of heterogeneity 20
- 2.9 Sensitivity analyses 20
- 2.10 Presentation of results 21
- 3 Estimating the treatment difference in an individual trial 23
- 3.2 Binary data 25
- 3.2.1 Example: Stroke in hypertensive patients 25
- 3.2.2 Measurement of treatment difference 25
- 3.3 Survival data 32
- 3.3.1 Example: Mortality following myocardial infarction 32
- 3.3.2 Measurement of treatment difference 33
- 3.4 Interval-censored survival data 38
- 3.4.1 Example: Ulcer recurrence 38
- 3.4.2 Measurement of treatment difference 39
- 3.5 Ordinal data 42
- 3.5.1 Example: Global impression of change in Alzheimer's disease 42
- 3.5.2 Measurement of treatment difference 42
- 3.6 Normally distributed data 49
- 3.6.1 Example: Recovery time after anaesthesia 49
- 3.6.2 Measurement of treatment difference 50
- 4 Combining estimates of a treatment difference across trials 57
- 4.2 A general fixed effects parametric approach 58
- 4.2.1 A fixed effects meta-analysis model 58
- 4.2.2 Estimation and hypothesis testing of the treatment difference 58
- 4.2.3 Testing for heterogeneity across studies 60
- 4.2.4 Obtaining the statistics via weighted least-squares regression 61
- 4.2.5 Example: Stroke in hypertensive patients 61
- 4.2.6 Example: Mortality following myocardial infarction 69
- 4.2.7 Example: Ulcer recurrence 73
- 4.2.8 Example: Global impression of change in Alzheimer's disease 78
- 4.2.9 Example: Recovery time after anaesthesia 82
- 4.3 A general random effects parametric approach 88
- 4.3.1 A random effects meta-analysis model 88
- 4.3.2 Estimation and hypothesis testing of the treatment difference 88
- 4.3.3 Estimation of [tau superscript 2] using the method of moments 90
- 4.3.4 Obtaining the statistics via weighted least-squares regression 91
- 4.3.5 Example: Mortality following myocardial infarction 91
- 4.3.6 Example: Global impression of change in Alzheimer's disease 93
- 4.3.7 Example: Recovery time after anaesthesia 93
- 4.3.8 A likelihood approach to the estimation of [tau superscript 2] 94
- 4.3.9 Allowing for the estimation of [tau superscript 2] 97
- 5 Meta-analysis using individual patient data 99
- 5.2 Fixed effects models for normally distributed data 100
- 5.2.1 A fixed effects meta-analysis model 100
- 5.2.2 Estimation and hypothesis testing 101
- 5.2.3 Testing for heterogeneity in the absolute mean difference across studies 103
- 5.2.4 Example: Recovery time after anaesthesia 103
- 5.2.5 Modelling of individual patient data versus combining study estimates 105
- 5.2.6 Heterogeneity in the variance parameter across studies 105
- 5.3 Fixed effects models for binary data 107
- 5.3.1 A fixed effects meta-analysis model 107
- 5.3.2 Estimation and hypothesis testing 108
- 5.3.3 Testing for heterogeneity in the log-odds ratio across studies 109
- 5.3.4 Example: Stroke in hypertensive patients 110
- 5.3.5 Modelling of individual patient data versus combining study estimates 110
- 5.4 Fixed effects models for ordinal data 111
- 5.4.1 A fixed effects meta-analysis model 111
- 5.4.2 Estimation and hypothesis testing 113
- 5.4.3 Testing for heterogeneity in the log-odds ratio across studies 115
- 5.4.4 Example: Global impression of change in Alzheimer's disease 116
- 5.4.5 Modelling of individual patient data versus combining study estimates 117
- 5.4.6 Testing the assumption of proportional odds between treatments 117
- 5.4.7 A proportional odds model for studies and treatments 119
- 5.5 Fixed effects models for survival data 120
- 5.5.1 A fixed effects meta-analysis model 120
- 5.5.2 Estimation and hypothesis testing 121
- 5.5.3 Testing for heterogeneity in the log-hazard ratio across studies 122
- 5.5.4 Example: Mortality following myocardial infarction 122
- 5.5.5 Modelling of individual patient data versus combining study estimates 123
- 5.5.6 Testing the assumption of proportional hazards between treatments 124
- 5.5.7 A proportional hazards model for studies and treatments 124
- 5.6 Fixed effects models for interval-censored survival data 126
- 5.6.1 A fixed effects meta-analysis model 126
- 5.6.2 Estimation and hypothesis testing 127
- 5.6.3 Testing for heterogeneity in the log-hazard ratio across studies 127
- 5.6.4 Example: Ulcer recurrence 128
- 5.6.5 Modelling of individual patient data versus combining study estimates 128
- 5.6.6 Testing the assumption of proportional hazards between treatments across timepoints 129
- 5.6.7 A proportional hazards model for studies and treatments 130
- 5.7 The treatment difference as a random effect 131
- 5.8 Random effects models for normally distributed data 131
- 5.8.1 A random effects meta-analysis model 131
- 5.8.2 Estimation and hypothesis testing 132
- 5.8.3 Example: Recovery time after anaesthesia 133
- 5.8.4 The connection between the multilevel model and the traditional mixed effects linear model 134
- 5.9 Random effects models for binary data 136
- 5.9.1 A random effects meta-analysis model 136
- 5.9.2 Estimation and hypothesis testing 136
- 5.9.3 Example: Pre-eclampsia 139
- 5.10 Random effects models for other data types 142
- 5.10.1 A random effects meta-analysis model for ordinal data 142
- 5.10.2 Example: Global impression of change in Alzheimer's disease 143
- 5.11 Random study effects 144
- 5.11.1 Random study and study by treatment effects: normally distributed data 145
- 5.11.2 Example: Recovery time after anaesthesia 146
- 5.11.3 Random study and study by treatment effects: other data types 147
- 5.12 Comparisons between the various models 147
- 6 Dealing with heterogeneity 151
- 6.2 The use of a formal test for heterogeneity 152
- 6.3 The choice between a fixed effects and a random effects model 153
- 6.4 When not to present an overall estimate of treatment difference 154
- 6.5 The choice of an appropriate measure of treatment difference 156
- 6.6 Meta-regression using study estimates of treatment difference 157
- 6.6.1 Example: Global impression of change in Alzheimer's disease 160
- 6.6.2 Example: Recovery time after anaesthesia 161
- 6.6.3 Extension to study estimates of treatment difference from subgroups 163
- 6.7 Patient-level covariates 165
- 6.7.1 Adjustment for imbalance in prognostic factors 165
- 6.7.2 Investigation of potential sources of heterogeneity 166
- 6.7.3 Example: Global impression of change in Alzheimer's disease 167
- 6.7.4 Meta-regression using individual patient data 168
- 6.7.5 Example: Recovery time after anaesthesia 168
- 6.8 An investigation of heterogeneity: Aspirin in coronary heart disease 170
- 6.9 A strategy for dealing with heterogeneity 174
- 7 Presentation and interpretation of results 175
- 7.2 Structure of a report 176
- 7.2.2 Methods 176
- 7.3 Graphical presentation 182
- 7.3.1 A confidence interval plot 183
- 7.3.2 A radial plot 186
- 7.4 Clinically useful measures of treatment difference 189
- 7.4.1 Simple transformations of the treatment difference parameter 190
- 7.4.2 Probability of doing better on treatment than on control 192
- 7.4.3 The number needed to treat 194
- 8 Selection bias 197
- 8.2 An investigation of publication bias: Intravenous magnesium following acute myocardial infarction 199
- 8.3 A funnel plot 199
- 8.4 Statistical methods for the detection and correction of publication bias 205
- -- 8.4.1 A test of funnel plot asymmetry 205
- 8.4.2 Rosenthal's file-drawer method 208
- 8.4.3 Models for the probability of selection 210
- 8.5 Bias due to selective reporting within studies 213
- 9 Dealing with non-standard data sets 215
- 9.2 No events in treatment arms of individual trials 216
- 9.3 Different rating scales or methods of assessment across trials 220
- 9.4 Different times of assessment across trials 225
- 9.5 Combining trials which report different summary statistics 228
- 9.5.1 Continuous outcomes 228
- 9.5.2 Ordinal data 231
- 9.5.3 Survival data 233
- 9.6 Imputation of the treatment difference and its variance 233
- 9.6.1 Absolute mean difference for continuous outcomes 233
- 9.6.2 The log-hazard ratio for survival data 235
- 9.7 Combining summary statistics and individual patient data 236
- 9.8 Combining p-values 237
- 10 Inclusion of trials with different study designs 241
- 10.2 More than two treatment groups 242
- 10.2.1 A fixed effects meta-analysis model 242
- 10.2.2 A random effects meta-analysis model 243
- 10.2.3 Random study effects 244
- 10.2.4 Example: First bleeding in cirrhosis 245
- 10.3 Dose--response relationships 249
- 10.4 Multicentre trials 253
- 10.5 Cross-over trails 254
- 10.6 Sequential trials 255
- 11 A Bayesian approach to meta-analysis 259
- 11.2 A Bayesian approach to the random effects model for study estimates 261
- 11.3 Choice of the prior distribution 263
- 11.4 Implementation using the BUGS software 265
- 11.4.1 Example: Recovery time after anaesthesia 267
- 11.5 Bayesian meta-regression 268
- 11.6 A Bayesian random effects model based on individual patient data 270
- 11.6.1 Normally distributed data 271
- 11.6.2 Binary data 273
- 11.6.3 Ordinal data 274
- 11.6.4 Study-level and patient-level covariates 276
- 11.6.5 Random study effects 276
- 11.7 Incorporating data from other treatment comparisons 279
- 11.8 An empirical prior distribution for the heterogeneity parameter 282
- 12 Sequential methods for meta-analysis 285
- 12.2 A proactive cumulative meta-analysis 286
- 12.2.1 Choice of a sequential design 287
- 12.2.2 A fixed effects model 291
- 12.2.3 A random effects model 292
- 12.2.4 Example: The triangular test for a primary efficacy outcome 293
- 12.2.5 Estimation of the heterogeneity parameter 296
- 12.3 A reactive cumulative meta-analysis 296
- 12.3.1 Example: Endoscopic haemostasis for bleeding peptic ulcers 297
- 12.3.2 Alternative approaches to a formal stopping rule 303
- Appendix Methods of estimation and hypothesis testing 305
- A.2 The method of least squares 306
- A.3 The method of weighted least squares 308
- A.4 Iterative maximum likelihood estimation 308
- A.5 Likelihood, efficient score and Fisher's information 311
- A.6 Iteratively weighted least squares 312
- A.7 Maximum likelihood methods for general linear mixed models 314
- A.8 Iterative generalized least squares for normally distributed data 316
- A.9 Marginal quasi-likelihood and penalized quasi-likelihood methods for discrete data 317