eval 6970: meta-analysis fixed-effect and random- effects models dr. chris l. s. coryn spring 2011

22
EVAL 6970: Meta-Analysis Fixed-Effect and Random- Effects Models Dr. Chris L. S. Coryn Spring 2011

Post on 19-Dec-2015

223 views

Category:

Documents


0 download

TRANSCRIPT

EVAL 6970: Meta-AnalysisFixed-Effect and Random-Effects Models

Dr. Chris L. S. CorynSpring 2011

Agenda

β€’ Fixed-effect models for meta-analysisβ€’ Random-effects models for meta-

analysisβ€’ Review questionsβ€’ In-class activity

Fixed-Effect Models

β€’ Under the fixed-effect model it is assumed that all studies share a common (true) effect size

β€’ All factors that could influence the effect size are the same across all studies (the true effect is the same, thus fixed)

β€’ The true (unknown) effect size is

Fixed-Effect Models

Fixed-effect model: True effect

Fixed-Effect Models

β€’ Given that all studies share the same true effect, the observed effect size varies from study to study only because of random (sampling) error

β€’ Although error is random, the sampling distribution of the errors can be estimated

Fixed-Effect Models

π‘Œ 𝑖=πœƒ+πœ€π‘–Fixed-effect model: True effects and sampling error

Fixed-Effect Models

β€’ To obtain an estimate of the population effect (to minimize variance) a weighted mean is computed using the inverse of each study’s variance as the study’s weight in computing the mean effect

Fixed-Effect Models

β€’ The weighted mean () is computed as

𝑀=βˆ‘π‘–=1

π‘˜

π‘Š π‘–π‘Œ 𝑖

βˆ‘π‘–=1

π‘˜

π‘Š 𝑖

Fixed-Effect Models

β€’ Where the weight assigned to each study is

β€’ Where is the within-study variance for study

π‘Š 𝑖=1𝑉 π‘Œ 𝑖

Fixed-Effect Models

β€’ With

β€’ And

β€’ And

𝑉𝑀=1

βˆ‘π‘–=1

π‘˜

π‘Š 𝑖

𝑆𝐸𝑀=βˆšπ‘‰π‘€

𝐿𝐿𝑀=π‘€βˆ’1.96×𝑆𝐸𝑀

π‘ˆπΏπ‘€=𝑀+1.96×𝑆𝐸𝑀

Random-Effects Models

β€’ Does not assume that the true effect is identical across studies

β€’ Because study characteristics vary (e.g., participant characteristics, treatment intensity, outcome measurement) there may be different effect sizes underlying different studies

Random-Effects Models

Random-effects model: True effects

Random-Effects Models

β€’ If the true effect size for a study is then the observed effect will be less than or greater than due to sampling error

β€’ The distance between the summary mean and the observed effect consists of true variation in effect sizes () and sampling error ()

Random-Effects Models

π‘Œ 𝑖=πœ‡+𝜁 𝑖+πœ€π‘–Random-effects model: True effect and observed effect

Random-Effects Models

β€’ The distance from to each depends on the standard deviation of the true effects across studies, which is represented as and for its variance

β€’ Each study’s variance is a function of both the within-study variance and and is the sum of these two values

Random-Effects Models

β€’ is estimated as

β€’ Where

𝑇 2=π‘„βˆ’π‘‘π‘“πΆ

𝑄=βˆ‘π‘–=1

π‘˜

π‘Š π‘–π‘Œ 𝑖2βˆ’

(βˆ‘π‘–=1

π‘˜

π‘Š π‘–π‘Œ 𝑖)2

βˆ‘π‘–=1

π‘˜

π‘Š 𝑖

Random-Effects Models

β€’ With

β€’ Where is the number of studies

β€’ And

𝑑𝑓=π‘˜βˆ’1

𝐢=βˆ‘π‘Š π‘–βˆ’βˆ‘π‘Š 𝑖

2

βˆ‘π‘Š 𝑖

Random-Effects Model

β€’ Under the random-effects model the weight assigned to each study is

β€’ Where is the within-study variance for study plus the between study variance

π‘Š π‘–βˆ—=

1

𝑉 π‘Œ 𝑖

βˆ—

Random-Effects Models

β€’ Where

β€’ With the weighted mean computed as

𝑉 π‘Œ 𝑖

βˆ— =𝑉 π‘Œ 𝑖+𝑇 2

π‘€βˆ—=βˆ‘π‘–=1

π‘˜

π‘Š π‘–βˆ—π‘Œ 𝑖

βˆ‘π‘–=1

π‘˜

π‘Š π‘–βˆ—

Random-Effects Models

β€’ With

β€’ And

β€’ And

𝑉𝑀 βˆ—=1

βˆ‘π‘–=1

π‘˜

π‘Š π‘–βˆ—

𝑆𝐸𝑀 βˆ—=βˆšπ‘‰ π‘€βˆ—

πΏπΏπ‘€βˆ—=π‘€βˆ—βˆ’1.96×𝑆𝐸𝑀 βˆ—

π‘ˆπΏπ‘€βˆ—=π‘€βˆ—+1.96×𝑆𝐸𝑀 βˆ—

Review Questions

1. When is it appropriate to use a fixed-effect model?

2. When is it appropriate to use a random-effects model?

3. How do the study weights differ for fixed-effect and random-effects models?

Today’s In-Class Activity

β€’ Individually, or in your working groups, download β€œData Sets 1-6 XLSX” from the course Website– Calculate the fixed-effect and random-effects

model weighted means for Data Sets 1, 2, 3, 4, 5, and 6

– Calculate the 95% confidence intervals (i.e., LL and UL) for the weighted means from Data Sets 1, 2, 3, 4, 5, and 6

– Use β€œMeta-Analysis with Means, Binary Data, and Correlations XLSX” to assist with the calculations