Consequences of Sequential Sampling for Meta-analysis

73
Consequences of Sequential Sampling for Meta-Analysis Lorenzo Braschi Diaferia, Juan Botella Ausina, Manuel Suero Su˜ ne Introduction What are sequential sampling rules? The CLAST rule The CLAST rule and meta-analysis Method Monte Carlo Simulations Simulation parameters Results Sample size Effect size estimation and bias Weighting Conclusions Consequences of Sequential Sampling for Meta-Analysis Lorenzo Braschi Diaferia 1 Juan Botella Ausina 2 Manuel Suero Su˜ ne 2 1 Facultad de Ciencias de la Salud Universidad Alfonso X el Sabio 2 Facultad de Psicolog´ ıa Universidad Aut´onoma de Madrid July 20th

Transcript of Consequences of Sequential Sampling for Meta-analysis

Page 1: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Consequences of Sequential Sampling forMeta-Analysis

Lorenzo Braschi Diaferia 1 Juan Botella Ausina 2

Manuel Suero Sune 2

1Facultad de Ciencias de la SaludUniversidad Alfonso X el Sabio

2Facultad de PsicologıaUniversidad Autonoma de Madrid

July 20th

Page 2: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Varying sample size

I Classical hypothesistesting requires a fixedsampling procedure to setα and β levels.

I But being able to adjustsample size as a result ofanalysis is cost-effective.

I Sequential sampling ruleshave been developed forseveral applications.

Page 3: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Varying sample size

I Classical hypothesistesting requires a fixedsampling procedure to setα and β levels.

I But being able to adjustsample size as a result ofanalysis is cost-effective.

I Sequential sampling ruleshave been developed forseveral applications.

Page 4: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Varying sample size

I Classical hypothesistesting requires a fixedsampling procedure to setα and β levels.

I But being able to adjustsample size as a result ofanalysis is cost-effective.

I Sequential sampling ruleshave been developed forseveral applications.

Page 5: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

What are sequential sampling rules?

A sequential sampling rule is a procedure for hypothesistesting that uses a smaller sample size relative to astandard or fixed sampling procedure.

The purpose of a sequential sampling rule is:

1. To offer a more flexible procedure for hypothesis testingregarding sample size.

2. To keep type I and type II error probabilities α and β

under control.

Page 6: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

What are sequential sampling rules?

A sequential sampling rule is a procedure for hypothesistesting that uses a smaller sample size relative to astandard or fixed sampling procedure.

The purpose of a sequential sampling rule is:

1. To offer a more flexible procedure for hypothesis testingregarding sample size.

2. To keep type I and type II error probabilities α and β

under control.

Page 7: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

What are sequential sampling rules?

A sequential sampling rule is a procedure for hypothesistesting that uses a smaller sample size relative to astandard or fixed sampling procedure.

The purpose of a sequential sampling rule is:

1. To offer a more flexible procedure for hypothesis testingregarding sample size.

2. To keep type I and type II error probabilities α and β

under control.

Page 8: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 9: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample size

I Divide region under the statistic’s distribution in threedifferent areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 10: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 11: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area

2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 12: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area

3. An uncertainty area that demands an increase in samplesize

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 13: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 14: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 15: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 16: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Composite Limited Adaptive Sequential Test (Botella,Ximenez, Revuelta, Suero, 2006)

I Set a lower and an upper boundary on sample sizeI Divide region under the statistic’s distribution in three

different areas:

1. A rejection area2. A manteinance area3. An uncertainty area that demands an increase in sample

size

I Calculate test statistic and associate p value.

I If p value falls within the uncertainty area, add moresubjects to the sample.

I Repeat until stopping conditions are met.

Page 17: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule

Page 18: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

CLAST parameters and variables

Parameters:

1. Reference sample size for equivalent fixed sample size:NFSR .

2. Initial starting sample size: N1.

3. Upper and lower boundaries for the uncertainty areas:α1 and αu.

Variables:

1. Sample size at stopping: NSTOP

Page 19: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

CLAST parameters and variables

Parameters:

1. Reference sample size for equivalent fixed sample size:NFSR .

2. Initial starting sample size: N1.

3. Upper and lower boundaries for the uncertainty areas:α1 and αu.

Variables:

1. Sample size at stopping: NSTOP

Page 20: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

CLAST parameters and variables

Parameters:

1. Reference sample size for equivalent fixed sample size:NFSR .

2. Initial starting sample size: N1.

3. Upper and lower boundaries for the uncertainty areas:α1 and αu.

Variables:

1. Sample size at stopping: NSTOP

Page 21: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

CLAST parameters and variables

Parameters:

1. Reference sample size for equivalent fixed sample size:NFSR .

2. Initial starting sample size: N1.

3. Upper and lower boundaries for the uncertainty areas:α1 and αu.

Variables:

1. Sample size at stopping: NSTOP

Page 22: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

CLAST parameters and variables

Parameters:

1. Reference sample size for equivalent fixed sample size:NFSR .

2. Initial starting sample size: N1.

3. Upper and lower boundaries for the uncertainty areas:α1 and αu.

Variables:

1. Sample size at stopping: NSTOP

Page 23: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule for Student’s paired t test

Botella et al. (2006) found that the CLAST rule uses asample size 20-40% smaller than a fixed rule of equivalentpower.

Page 24: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule and meta-analysis

Main problem: How to incorporate primary CLAST studiesinto a meta-analysis?

We need to know:

1. Bias of estimations of effect size under CLAST.

2. Appropiate weights for primary CLAST studies.

Page 25: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule and meta-analysis

Main problem: How to incorporate primary CLAST studiesinto a meta-analysis?

We need to know:

1. Bias of estimations of effect size under CLAST.

2. Appropiate weights for primary CLAST studies.

Page 26: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule and meta-analysis

Main problem: How to incorporate primary CLAST studiesinto a meta-analysis?

We need to know:

1. Bias of estimations of effect size under CLAST.

2. Appropiate weights for primary CLAST studies.

Page 27: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 28: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 29: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 30: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 31: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 32: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 33: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Monte Carlo Simulations

I Monte Carlo simulations of random samples using R.

I Student’s paired t test.

I Population parameters: µ and σ (so that the status ofthe null H0 is TRUE/FALSE).

I For all studies, µ = δ.

I Sample of initial size N1 drawn from theoreticalpopulation of differences D ∼ N(δ,1).

I New data added drawn from the same population.

Page 34: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 35: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}

I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 36: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 37: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUE

I If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 38: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 39: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 40: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 41: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Simulation parameters

I N fixed= {16, 20, 24, 30, 40}I Theoretical effect size δ = {0, 0.1, 0.2, 0.3, ... , 1}

I If δ = 0, then H0 = TRUEI If δ > 0, then H0 = FALSE

I 105 iterations for each condition (combination of Nfixed and δ, 55 total scenarios)

I α1 = 0.01

I αu = 0.25

I α1 and αu values have been selected so that the overallα level converges to α = 0.05 in the long run (afterBotella et al., 2006)

Page 42: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Results

Page 43: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

CLAST rule mean sample size at stopping fordifferent starting effect sizes

Page 44: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Effect size estimation and bias under CLAST

We used the d as an estimator of δ for the paired t test:

d =D

SD· c(m) (1)

Bias was measured as the difference between the estimatorand the true value of the effect size:

Bias = d−δ (2)

Page 45: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Estimation of δ under the CLAST and fixedsampling rules

Page 46: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Effect size estimation and bias under CLAST

No significant bias for effect size estimation.

Primary CLAST studies can be incorporated into ameta-analysis without adding significant bias.

Page 47: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to weight primary CLAST studies?

Common procedure: weight studies by inverse of variance ofestimator d .

We need to know var(d)

var(d) =∑wi ·di

∑wi(3)

Where wi = N

Page 48: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to calculate w?

To calculate var(d), we need to weight studies by N, that is,sample size.

Which sample size?

In CLAST, we have three possible candidates for sample sizeweighting:

1. Equivalent sample size under the fixed sampling rule,NFSR .

2. Starting sample size, N1.

3. Sample size at stopping, NSTOP

Page 49: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to calculate w?

To calculate var(d), we need to weight studies by N, that is,sample size.

Which sample size?

In CLAST, we have three possible candidates for sample sizeweighting:

1. Equivalent sample size under the fixed sampling rule,NFSR .

2. Starting sample size, N1.

3. Sample size at stopping, NSTOP

Page 50: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to calculate w?

To calculate var(d), we need to weight studies by N, that is,sample size.

Which sample size?

In CLAST, we have three possible candidates for sample sizeweighting:

1. Equivalent sample size under the fixed sampling rule,NFSR .

2. Starting sample size, N1.

3. Sample size at stopping, NSTOP

Page 51: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to calculate w?

To calculate var(d), we need to weight studies by N, that is,sample size.

Which sample size?

In CLAST, we have three possible candidates for sample sizeweighting:

1. Equivalent sample size under the fixed sampling rule,NFSR .

2. Starting sample size, N1.

3. Sample size at stopping, NSTOP

Page 52: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to calculate w?

To calculate var(d), we need to weight studies by N, that is,sample size.

Which sample size?

In CLAST, we have three possible candidates for sample sizeweighting:

1. Equivalent sample size under the fixed sampling rule,NFSR .

2. Starting sample size, N1.

3. Sample size at stopping, NSTOP

Page 53: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

How to calculate w?

To calculate var(d), we need to weight studies by N, that is,sample size.

Which sample size?

In CLAST, we have three possible candidates for sample sizeweighting:

1. Equivalent sample size under the fixed sampling rule,NFSR .

2. Starting sample size, N1.

3. Sample size at stopping, NSTOP

Page 54: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Mean CLAST weights for NFSR = 24

Page 55: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Mean CLAST weights for NFSR = 24

Page 56: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Mean CLAST weights for NFSR = 24

Page 57: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Mean CLAST weights for NFSR = 24

Page 58: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Conclusions

I No previous research on how to incorporate sequentialsampling rules into meta-analysis.

I Small to insignificant bias in effect size estimation underCLAST.

I Problem of weighting only when we mix CLAST andstandard primary studies.

I Use mean of NFSR and initial N.

Page 59: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Conclusions

I No previous research on how to incorporate sequentialsampling rules into meta-analysis.

I Small to insignificant bias in effect size estimation underCLAST.

I Problem of weighting only when we mix CLAST andstandard primary studies.

I Use mean of NFSR and initial N.

Page 60: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Conclusions

I No previous research on how to incorporate sequentialsampling rules into meta-analysis.

I Small to insignificant bias in effect size estimation underCLAST.

I Problem of weighting only when we mix CLAST andstandard primary studies.

I Use mean of NFSR and initial N.

Page 61: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Conclusions

I No previous research on how to incorporate sequentialsampling rules into meta-analysis.

I Small to insignificant bias in effect size estimation underCLAST.

I Problem of weighting only when we mix CLAST andstandard primary studies.

I Use mean of NFSR and initial N.

Page 62: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Conclusions

I No previous research on how to incorporate sequentialsampling rules into meta-analysis.

I Small to insignificant bias in effect size estimation underCLAST.

I Problem of weighting only when we mix CLAST andstandard primary studies.

I Use mean of NFSR and initial N.

Page 63: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

Thank you for yourattention!

Page 64: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 65: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 66: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 67: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 68: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 69: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 70: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 71: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 72: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.

Page 73: Consequences of Sequential Sampling for Meta-analysis

Consequences ofSequential

Sampling forMeta-Analysis

Lorenzo BraschiDiaferia, Juan

Botella Ausina,Manuel Suero

Sune

Introduction

What are sequentialsampling rules?

The CLAST rule

The CLAST rule andmeta-analysis

Method

Monte Carlo Simulations

Simulation parameters

Results

Sample size

Effect size estimation andbias

Weighting

Conclusions

The CLAST rule algorithm for the paired t test

1. Determine the necessary sample under the fixed stopping rule(NFSR) as a function of desired power and the assumed effectsize.

2. Determine the initial sample size, N1 = 0.5NFSR , and themaximum sample size, Nmax = 1.5NFSR .

3. Execute the experiment with a sample of size N1 and analyzedata.

3.1 If p > αU , stop the experiment and maintain H0.3.2 If p ≤ α1, stop the experiment and reject H0.3.3 If α1 < p ≤ αU , add N2 = 1 to the sample.

4. If N1 +N2 = Nmax stop the experiment and decide:

4.1 If p ≤ 0.05 reject H0.4.2 If p > 0.05 maintain H0.

5. If N1 +N2 < Nmax reanalyze data with N1 +N2 and return tostep 3.