Power and Sample Size Analysis - UCLA Integrated Substance

Report
Power and Sample Size Analysis
David Huang, Dr. PH.
Integrated Substance Abuse Treatment Programs,
University of California, Los Angeles
November 18, 2011
1
Outline of This Presentation


I. Introduction to Sample Size Analysis
Why estimation of required sample is important for a research
proposal


Rationales
What is sample size analysis

General framework: Rules of Thumb

Components:
(e.g., effect size, statistical power, correlations)

II. Strategies and Examples

Single group estimation

Two-group comparisons

Multivariate analysis

Nested (multi-level) sample

Repeated measured over time
2
How much data do we need
• How many subjects should be included in the
research.
Without considering the expenses, the more
data the better.
• It is not feasible to collect data on the entire
population of interest.
• Consider the collected data as a random
sample of the population of interest.
3
Rules of Thumb
Feasible in terms of budget and
research time frame
Sufficient data to ensure results to be
Accurate, Efficient, and Credible
4
What is Sample Size Analysis
Based upon the statistical test for the
main research question,
Sample size analysis is intended to
determine the minimal data (or sample
size) required for detecting a
significant research finding.
5
Example: An experimental Study
An experimental study will be conducted to
evaluate the treatment effectiveness between
the two treatment protocols (Suboxone vs
Methadone).
Outcome measures will include urine test
results, ASI score and satisfaction score at
three months after discharged.
6
Questions Related to Sample size
Calculation
How to measure treatment effectiveness? What
measures or indicators will be employed ?
The accuracy of the measures or indicators ?
How large a difference will be expected ? What is the
smallest effect size would be considered of importance?
Whether the new treatment is expected to better than
the old treatment (1-side or 2-side test).
Reliability of research findings will be (power or Alphalevel).
Whether subjects are nested within treatment clinics?
Whether there are follow-up data.
7
Components of Sample size
Calculation
What test statistic will be employed ?
Hypothesis Testing: The null hypothesis vs.
The alternative hypothesis
Alpha Level (or desired accuracy; width of
confidence interval)
Power
Effect size: expected differences and variation
of outcome measures
Sample size
8
Types of Statistics
Means (e.g., ASI score)
-- Compare 2 means (t-test)
-- Compare 3 or more means (ANOVA)
Proportions (e.g., abstinent rate)
-- Compare 2 proportions
Bivariate relationship – correlation (r)
Multiple regression – Multiple R2
Cluster sampling/multi-level
9
Hypothesis Testing
The null hypothesis: This hypothesis predicts
that there is no effect on the variable of
interest
The alternative hypothesis: This hypothesis
predicts that there is an effect on the variable
of interest (or a difference between groups).
Statistical tests look for evidence to reject the
null hypothesis and conclude the alternative
hypothesis (an effect is existing)
Sample size analysis: Determine the minimal
amount of data required.
10
Alpha level/Power from
Hypothesis Testing
True situation in the
population
H0
Ha
True
True
Research
findings
Do not
reject H0
Correct
(1- α )
Error
β
Reject H0
Error
α
Correct
(1- β)
“power”
11
Alpha Level and Power
Alpha level: Probability of incorrectly
concluding (from sample data) a significant
effect when it does not really exist in the
population (Type-I error).
-- Alpha level is usually set as .05
Power: Probability of correctly concluding
(from sample data) a significant effect when
it really exist in the population.
-- Power is usually set as .80
12
Effect size
Effect size – Standardized measure of the
magnitude of a difference or relationship.
-- How big a difference or relationship (in a
standardized metric) was detected in analysis
-- Effect sizes (for the same type of statistics) are
calculated on a common scale, which allows to
compare the effectiveness of different programs
In calculating sample size:
-- How big a difference or relationship do we want
to detect?
-- How big a difference is considered clinically
important?
13
Computing Effect Size
Various formulas depend on type of statistic
e.g., for difference in means (t-test)
d= (mean1 – mean2) /standard deviation
Various labels,
d for difference in two means
w (h) for difference in proportions
r for correlations
f for difference in many means (e.g., One-way ANOVA)
η2 for variance explained
R2 for multiple regression
14
Determining Effect Size
Based on substantive knowledge
Based on findings from prior research
Based on a pilot study
Use conventions
-- e.g., small, medium and large effect size
defined by Cohen
15
Magnitude of Effect size
From by Cohen, 1988
The bigger the effect size, the easier the detection.
small
medium
Large
d
0.20
0.50
0.80
w
f
η2
0.10
0.1
0.01
0.30
0.25
0.06
0.50
0.4
0.14
r
Partial R2
0.10
0.02
0.30
0.13
0.50
0.26
16
Steps for Sample Size Determination
Decide types of outcome statistics
(e.g., mean, proportion, correlations,…)
Specify 1- or 2-tailed tests
Specify desired alpha level and power
Specify the desired effect size (from
literature, pilot study, or best guess)
17
General Rules: Required Sample Size
Detecting small effect sizes --> larger N
Smaller alpha or greater power --> larger N
2-tailed test --> larger N than 1-tailed test
Addition of covariates (e.g., ANCOVA) 
reduce error variance, then increase effect
size and decrease N
More parameters in model --> larger N
18
Effect Size vs. Number of Subjects per Group
for two-tailed t-test with α=.05, power=.80
Sample size per group
500
400
300
200
100
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Effect size
19
Functions of Power vs. Number of Subjects
per Group for two-tailed t-test with α=.05
General Rules: Required Sample Size
Cluster sampling/multi-level data structure:
-- Larger N as intra-class correlation increase
Follow-up with repeated measures:
-- More repeated measures, smaller N per
group
21
Power vs. Intra-class Correlation in the Cluster
Sample given Effect Size of 0.3, α of.05, and
20 Clusters
Power vs. Number of Subjects for testing linear
trend effect with repeated measures over time
given effect size of 0.3 and α=.05
800
time=3
Sample size
600
time=6
400
200
0
0.1
0.2
0.3
0.4
0.5
Power
0.6
0.7
0.8
0.9
Limitations of Sample Size
Analysis
The analyses do not generalize very well
Based on assumptions and educated guesses,
the analyses give a “best case scenario”
estimate of necessary sample size
It is a good strategy to compute the required
sample sizes by different levels of effect size
(or alpha, power), and then present the
required sample sizes in a range, instead of a
single number
24
Computational Software
•
•
•
Erdfelder, E., Paul, F., & Buchner, A. (1996).
GPOWER: A general power analysis program.
Behavior Research Methods, Instrument &
Computers, 28,1-11. (Free)
http://www.psycho.uniduesseldorf.de/abteilungen/aap/gpower3/
Elashoff, J. (2008) nQuery Advisor.
25
Examples
26
Example 1: A Cross-sectional Evaluation
Proposition 36 was implemented in California.
A study will be conducted to evaluate the
system impact of P36 on treatment
accessibility and outcomes among drugabused offenders. A total of 450 clients from
15 treatment programs will be screened and
recruited for the study.
Outcome measures will include urine test
results during treatment, ASI score and
satisfaction score at three months after
discharged.
27
Example 1: Research Aims
Estimate proportion of negative urine test, and
ASI score and satisfaction score at three months
after discharged.
Examine difference in treatment outcomes by
integrated and non-integrated programs.
Compare treatment retention by integrated,
women-only, other types of treatment programs.
Correlation of length of treatment retention with
ASI scores.
Examination of potential risk and protective
factors associated with treatment outcomes (e.g.,
ASI score, negative urine test).
28
Example 1: Corresponding Sample size Analysis
Description of means and rates (%)
For an estimation of mean of a continuous
measure (e.g., ASI score), the proposed
sample of 450 will be sufficient to estimate
ASI score within a range of ± 0.13 standard
deviation.
The proposed sample of N=450 will be
sufficient to estimate a rate (e.g., negative
rate) of 10% within a range of ± 4%, a rate of
30% within a range of ± 6%, and a rate of
50% within a range of ± 7%.
29
Example 1: Corresponding Sample size Analysis
Comparisons on outcome measures by
integrated and non-integrated programs.
Compare (means- independent t-test): The
required sample for detecting a medium
effect of 0.3 will be 176 per group. The
proposed sample of 225 per group will allow
detection of an effect size of 0.26.
Compare (%): Given the sample of N=225 per
group, the detectable difference on rates will
be 13%, 12%, and 9% when rates in the
study population are 50%, 30%, and 10 %,
respectively.
30
Example 1: Corresponding Sample size Analysis
Comparisons of means by three or more
groups (e.g., ANOVA).
For comparison of means among three groups
(e.g., integrated, women-only, and others),
the required sample for detecting a medium
effect of 0.25 (f) will be 159. The proposed
sample of 450 (e.g., 150 per group) will allow
detection of an effect size of 0.15.
31
Example 1: Corresponding Sample size Analysis
Correlation and Linear regression
In the simple regression (1 covariate), the
detectable correlation of 0.2 requires a
sample of 193. The sample of 450 will allow
detection of small-to-medium correlation of
0.13 or larger.
In a multiple regression with p covariates, the
required sample will increase, depending on
the partial correlation of p-1 covariates. the
sample of 450 will allow to detect medium
effect of R2=.02 for predicting ASI score from
5-10 covariates.
32
Example 1: Corresponding Sample size Analysis
Logistic regression for categorical measures
The sample of size of 450 should allow
detection of an odds ratio of 1.40 for a single
covariate model, given the rate of outcome in
the study population is 0.2.
The detectable effect size would increase
within a multivariate model, depending the
correlation of the main predictor with other
covariates. The detectable odds ratio will
range from 1.42 to 1.54 when partial R2 of
other covariates is 0.1 to 0.4.
33
Example 2: An Experimental Study
Under the main scope of CTN project, an
experimental study will be conducted to
evaluate the treatment effectiveness between
the two treatment protocols (Suboxone vs
Methadone). Subjects will be randomized to
one protocol.
Outcome measures will include treatment
negative urine test, ASI score and satisfaction
score at three months after discharged.
34
Example 2: An Experimental Study
In an experimental study, we need to think
about “statistically significant” versus
“clinically relevant” when considering effect
size.
A small effect size may not be clinically
meaningful. For example, a reduction of blood
pressure by two points.
If an incorrect conclusion has potential
adverse consequence on subjects, a lower
level of alpha and a higher power should be
selected.
35
Example 3: A Nested Study Design and
A Longitudinal Study
When individual subjects are recruited from
clusters (e.g., treatment programs, schools),
the correlation among subjects within each
cluster may need to be considered in sample
size analysis.
A Longitudinal study repeatedly collects data
across time. Each individual will have
repeated measures over time. The repeated
measures within each individual are
correlated.
36
Example 3: Approaches and Software
for Nested Data Structure
The specific statistical analysis for a nested
(cluster) data is Multi-level (Hierarchical)
Modeling or Generalized linear modeling.
Software for computing sample size for Multi-level or
Generalized linear models.
•
Spybrook, J., Raudenbush, S., Congdon, R., &
Martinez, A. (2009) Optimal Design for
Longitudinal and Multilevel Research:
Documentation for the “Optimal Design”
Software. Available at
http://www.wtgrantfoundation.org/resources/ove
rview/research_tools/research_tools
37
Example 3: Approaches to the Nested
Data Structure
Efficiency of proposed sample size will
decrease because of the intra-cluster
correlation. The required sample size could
be adjusted by an inflation factor:
1/[1+(m-1) ρ]
Here, m indicates average size in a treatment
program and ρ indicates intra-cluster
correlation.
38
Random effect regression for repeated
measures
The target sample size of 120 per group will allow
the detection of small-to-medium effects of about
d=.32 in detecting a difference in patterns over time
between the maintenance group and each of the
other two groups, with power=.80 and one-tailed
alpha=.05 assuming a moderate correlation of .50
over time and approximately 15% attrition (Hedeker
et al., 1999).
39
Software
•
•
•
•
Erdfelder, E., Paul, F., & Buchner, A. (1996). GPOWER: A
general power analysis program. Behavior Research
Methods, Instrument & Computers, 28,1-11. (Free)
http://www.psycho.uniduesseldorf.de/abteilungen/aap/gpower3/
Elashoff, J. (2008) nQuery Advisor.
Hedeker, D. RMASS2, Repeated Measures with Attrition:
Sample Size for 2 Groups. (free)
-- Old version available to download at
http://tigger.uic.edu/%7Ehedeker/ml.html
-- New on-line version available at
http://www.uic.edu/labs/biostat/projects.html
-- Software for computing sample size for general(ized) linear
models with repeated measures.
40
Software
•
Spybrook, J., Raudenbush, S., Congdon, R., & Martinez, A.
(2009) Optimal Design for Longitudinal and Multilevel
Research: Documentation for the “Optimal Design”
Software. Available at
http://www.wtgrantfoundation.org/resources/overview/rese
arch_tools/research_tools
•
SamplePower. SPSS
$$SPSS module for computing power/sample size.
•
Proc POWER and Proc GLMPOWER. SAS
$$ SAS procedures for computing power/sample size.
41
Software
•
PS-power/sample size (free) Available to download at
http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/Power
SampleSize
•
Dennis, M (1997) Power Analysis Worksheet. (free)
Available at
http://www.chestnut.org/LI/downloads/index.html [Click on
"Power Analysis Worksheet part way down web page]
Excel spreadsheet for calculating sample size for simple
designs/analyses
42
References
•
Cohen, J. (1988) Statistical Power Analysis for the Behavioral Sciences.
Hillsdale, NJ: Lawrence Erlbaum Assoc. Formulas and tables for
computing power and sample size--very comprehensive and somewhat
difficult--but very well-respected reference.
•
Fink, A. (2003) How to sample in surveys. Thousands Oaks, CA: Sage.
Simple step-by-step guide to sampling issues and procedures—with
table for looking up approximate sample size for proportions from 2category responses.
•
Kraemer, H. C. & Thiemann, S. (1987) How Many Subjects? Newbury
Park, CA: Sage. -- Formulas and tables for approximate sample size and
some good descriptions of issues.
•
Lipsey, M. W. (1990) Design Sensitivity. Newbury Park, CA: Sage. Nonmathematical discussion of many design/sample size issues.
•
Rudy, E. & Kerr, M. (1991) Unraveling the mystique of power analysis.
Heart & Lung, 20(5), 517-522. Description of power/sample size for
researchers.
43

similar documents