Power and Sample Size Adapted from: Boulder 2004 Benjamin Neale Shaun Purcell I HAVE THE POWER!!!

41
Power Power and Sample Size and Sample Size Adapted from: Adapted from: Boulder 2004 Boulder 2004 Benjamin Neale Benjamin Neale Shaun Purcell Shaun Purcell I HAVE THE POWER!!!

Transcript of Power and Sample Size Adapted from: Boulder 2004 Benjamin Neale Shaun Purcell I HAVE THE POWER!!!

PowerPower and Sample Size and Sample Size

Adapted from:Adapted from:Boulder 2004Boulder 2004

Benjamin NealeBenjamin NealeShaun PurcellShaun Purcell

I HAVE THE

POWER!!!

OverviewOverview

Introduce Concept of Power via Introduce Concept of Power via Correlation Coefficient (ρ) ExampleCorrelation Coefficient (ρ) Example

Discuss Factors Contributing to Discuss Factors Contributing to PowerPower

Practical:Practical:• Simulating data as a means of Simulating data as a means of

computing powercomputing power• Using Mx for Power CalculationsUsing Mx for Power Calculations

3

Simple exampleSimple exampleInvestigate the linear relationship between Investigate the linear relationship between two random variables X and Y: two random variables X and Y: ρρ=0 vs. =0 vs. ρ≠ρ≠0 0

using the Pearson correlation coefficient.using the Pearson correlation coefficient.

Sample subjects at random from Sample subjects at random from populationpopulation Measure X andYMeasure X andY Calculate the measure of association Calculate the measure of association ρρ Test whether Test whether ρρ ≠≠ 0. 0.

4

How to Test How to Test ρρ ≠≠ 00

Assume data are normally Assume data are normally distributeddistributed

Define a null-hypothesis (Define a null-hypothesis (ρρ = 0) = 0) Choose an Choose an αα level (usually .05) level (usually .05) Use the (null) distribution of the Use the (null) distribution of the

test statistic associated with test statistic associated with ρρ=0=0 t=t=ρρ √√ [(N-2)/(1- [(N-2)/(1-ρρ22)] )]

5

How to Test How to Test ρρ ≠≠ 00

Sample N=40Sample N=40 r=.303, t=1.867, df=38, p=.06 r=.303, t=1.867, df=38, p=.06 αα

=.05=.05 Because observed p > Because observed p > αα, we fail , we fail

to reject to reject ρρ = 0 = 0

Have we drawn the correct Have we drawn the correct conclusion that p is genuinely conclusion that p is genuinely zero?zero?

6

= type I error rate = type I error rate probability of deciding probability of deciding ρ≠ρ≠ 0 0

(while in truth (while in truth ρρ=0)=0)

ααis often chosen to is often chosen to equal .05...why?equal .05...why?

DOGMA

7

N=40, r=0, nrep=1000, central N=40, r=0, nrep=1000, central t(38), t(38),

αα=0.05 (critical value 2.04)=0.05 (critical value 2.04)

8

Observed non-null Observed non-null distribution (distribution (ρρ=.2) =.2)

and null distributionand null distribution

9

In 23% of tests that In 23% of tests that ρρ=0, |t|=0, |t|>2.024 (>2.024 (αα=0.05), and thus =0.05), and thus correctly conclude that correctly conclude that ρρ 0. 0.

The probability of correctly The probability of correctly rejecting the null-hypothesis rejecting the null-hypothesis ((ρρ=0) is 1-=0) is 1-ββ, known as the power. , known as the power.

Hypothesis TestingHypothesis Testing

Correlation Coefficient hypotheses:Correlation Coefficient hypotheses: hhoo (null hypothesis) is ρ=0 (null hypothesis) is ρ=0

hha a (alternative hypothesis) is ρ ≠ 0(alternative hypothesis) is ρ ≠ 0 Two-sided test, where ρ > 0 or ρ < 0 are Two-sided test, where ρ > 0 or ρ < 0 are

one-sidedone-sided

Null hypothesis usually assumes no Null hypothesis usually assumes no effecteffect

Alternative hypothesis is the idea Alternative hypothesis is the idea being testedbeing tested

11

Summary of Possible Summary of Possible ResultsResults

H-0 trueH-0 true H-0 H-0 falsefalse

accept H-0accept H-0 1-1-ααββreject H-0reject H-0 αα 1-1-ββ

αα=type 1 error rate=type 1 error rate

ββ=type 2 error rate=type 2 error rate

1-1-ββ=statistical power=statistical power

Rejection of H0 Non-rejection of H0

H0 true

HA true

STATISTICS

R E

A L I T

Y

Nonsignificant result(1- α)

Type II error at rate β

Significant result(1-β)

Type I error at rate α

PowerPower The probability of rejecting the The probability of rejecting the

null-hypothesis depends on: null-hypothesis depends on: the significance criterion (the significance criterion (αα)) the sample size (N) the sample size (N) the effect size (NCP)the effect size (NCP)

“The probability of detecting a given effect size in a population from a sample of size N, using significance criterion α”

P(T)

T

alpha 0.05

Sampling distribution if HA were true

Sampling distribution if H0 were true

β α

POWER = 1 - β

Standard Case

Effect Size (NCP)

P(T)

T

alpha 0.1

Sampling distribution if HA were true

Sampling distribution if H0 were true

POWER = 1 - β ↑

Impact of less Impact of less conservative conservative α

β α

P(T)

T

alpha 0.01

Sampling distribution if HA were true

Sampling distribution if H0 were true

POWER = 1 - β↓

Impact of more Impact of more conservative conservative α

β α

P(T)

T

alpha 0.05

β α

Impact of increased sample size

Reduced variance of sampling distribution if HA is true

Sampling distribution if H0 is true

POWER = 1 - β↑

P(T)

T

alpha 0.05

Sampling distribution if HA were true

Sampling distribution if H0 were true

β α

POWER = 1 - β↑

Impact of increase in Effect Size

Effect Size (NCP)↑

Summary: Factors affecting Summary: Factors affecting powerpower

Effect SizeEffect Size Sample SizeSample Size Alpha LevelAlpha Level <Beware the False Positive!!!><Beware the False Positive!!!> Type of Data:Type of Data:

Binary, Ordinal, ContinuousBinary, Ordinal, Continuous Research DesignResearch Design

Uses of power Uses of power calculationscalculations

Planning a studyPlanning a study

Possibly to reflect on ns trend resultPossibly to reflect on ns trend result

No need if significance is achievedNo need if significance is achieved

To determine chances of study To determine chances of study successsuccess

Power Calculations via Power Calculations via SimulationSimulation

Simulate Data under theorized modelSimulate Data under theorized model

Calculate Statistics and Perform TestCalculate Statistics and Perform Test

Given α, how many tests p < αGiven α, how many tests p < α

Power = (#hits)/(#tests)Power = (#hits)/(#tests)

Practical: Empirical Practical: Empirical Power 1Power 1

Simulate Data under a model onlineSimulate Data under a model online

Fit an ACE model, and test for CFit an ACE model, and test for C

Collate fit statistics on boardCollate fit statistics on board

Practical: Empirical Practical: Empirical Power 2Power 2

First get First get http://www.vipbg.vcu.edu/neale/gen619/phttp://www.vipbg.vcu.edu/neale/gen619/power/power-raw.mx and put it into your ower/power-raw.mx and put it into your directorydirectory

Second, open this script in Mx, and note Second, open this script in Mx, and note both places where we must paste in the both places where we must paste in the datadata

Third, simulate data (see next slide)Third, simulate data (see next slide) Fourth, fit the ACE model and then fit the Fourth, fit the ACE model and then fit the

AE submodelAE submodel

Practical: Empirical Practical: Empirical Power 3Power 3

Simulation ConditionsSimulation Conditions 30% A30% A22 20% C20% C22 50% E 50% E22

Input:Input: A 0.5477 C of 0.4472 E of 0.7071A 0.5477 C of 0.4472 E of 0.7071 350 MZ 350 DZ350 MZ 350 DZ Simulate and use “Space Delimited” Simulate and use “Space Delimited”

option atoption at http://statgen.iop.kcl.ac.uk/workshop/http://statgen.iop.kcl.ac.uk/workshop/

unisim.html or click unisim.html or click herehere in slide show mode in slide show mode Click submit after filling in the fields and you Click submit after filling in the fields and you

will get a page of datawill get a page of data

Practical: Empirical Practical: Empirical Power 4Power 4

With the data page, use ctrl-a to select the With the data page, use ctrl-a to select the data, control-c to copy, switch to Mx (e.g. data, control-c to copy, switch to Mx (e.g. with alt-tab) and in Mx control-v to paste with alt-tab) and in Mx control-v to paste in both the MZ and DZ groups.in both the MZ and DZ groups.

Run the ace.mx script with the data Run the ace.mx script with the data pasted in and modify it to run the AE pasted in and modify it to run the AE model.model.

Report the -2log-likelihoods on the Report the -2log-likelihoods on the whiteboardwhiteboard

Optionally, keep a record of A, C, and E Optionally, keep a record of A, C, and E estimates of the first model, and the A and estimates of the first model, and the A and E estimates of the second model E estimates of the second model

Simulation of other Simulation of other types of datatypes of data

Use SAS/R/Matlab/MathematicaUse SAS/R/Matlab/Mathematica

Any decent random number Any decent random number generator will dogenerator will do

See See http://www.vipbg.vcu.edu/~neale/gehttp://www.vipbg.vcu.edu/~neale/gen619/power/sim1.sasn619/power/sim1.sas

27

RR

R is in your futureR is in your future Can do it manually with rnormCan do it manually with rnorm Easier to use mvrnormEasier to use mvrnorm

runmx at Matt Keller’s site:runmx at Matt Keller’s site: http://www.matthewckeller.com/html/mhttp://www.matthewckeller.com/html/m

x-r.htmlx-r.html27

library (MASS)mvrnorm(n=100,c(1,1),matrix(c(1,.5,.5,1),2,2),empirical=FALSE)

Mathematica ExampleMathematica ExampleIn[32]:=(mu={1,2,3,4}; sigma={{1,1/2,1/3,1/4},{1/2,1/3,1/4,1/5},{1/3,1/4,1/5,1/6},{1/4,1/5,1/6, 1/7}}; Timing[Table[Random[MultinormalDistribution[mu,sigma]],{1000}]][[1]])

Out[32]=1.1 Second

In[33]:=Timing[RandomArray[MultinormalDistribution[mu,sigma],1000]][[1]]

Out[33]=0.04 Second

In[37]:=TableForm[RandomArray[MultinormalDistribution[mu,sigma],10]]

Obtain mathematica from VCU http://www.ts.vcu.edu/faq/stats/mathematica.html

Theoretical Power Theoretical Power CalculationsCalculations

Based on Stats, rather than Based on Stats, rather than SimulationsSimulations

Can be calculated by hand Can be calculated by hand sometimes, but Mx does it for ussometimes, but Mx does it for us

Note that sample size and alpha-Note that sample size and alpha-level are the only things we can level are the only things we can change, but can assume different change, but can assume different effect sizeseffect sizes

Mx gives us the relative power levels Mx gives us the relative power levels at the alpha specified for different at the alpha specified for different sample sizessample sizes

Theoretical Power Theoretical Power CalculationsCalculations

We will use the power.mx script to We will use the power.mx script to look at the sample size necessary for look at the sample size necessary for different power levelsdifferent power levels

In Mx, power calculations can be In Mx, power calculations can be computed in 2 ways:computed in 2 ways: Using Covariance Matrices (We Do This Using Covariance Matrices (We Do This

One)One) Requiring an initial dataset to generate Requiring an initial dataset to generate

a likelihood so that we can use a chi-a likelihood so that we can use a chi-square testsquare test

Power.mx 1Power.mx 1! Simulate the data! Simulate the data

! ! 30% additive genetic30% additive genetic

! ! 20% common environment20% common environment

! ! 50% nonshared environment50% nonshared environment

#NGroups 3#NGroups 3

G1: model parametersG1: model parameters

CalculationCalculation

Begin Matrices;Begin Matrices;

X lower 1 1 fixedX lower 1 1 fixed

Y lower 1 1 fixedY lower 1 1 fixed

Z lower 1 1 fixedZ lower 1 1 fixed

End Matrices;End Matrices;

Matrix X 0.5477Matrix X 0.5477

Matrix Y 0.4472Matrix Y 0.4472

Matrix Z 0.7071Matrix Z 0.7071

Begin Algebra;Begin Algebra;

A = X*X' ; A = X*X' ;

C = Y*Y' ;C = Y*Y' ;

E = Z*Z' ;E = Z*Z' ;

End Algebra;End Algebra;

EndEnd

Power.mx 2Power.mx 2G2: MZ twin pairsG2: MZ twin pairs

CalculationCalculation

Matrices = Group 1Matrices = Group 1

Covariances A+C+ECovariances A+C+E || A+C _A+C _

A+CA+C || A+C+E /A+C+E /

Options MX%E=mzsim.covOptions MX%E=mzsim.cov

EndEnd

G3: DZ twin pairsG3: DZ twin pairs

CalculationCalculation

Matrices = Group 1Matrices = Group 1

H Full 1 1 H Full 1 1

Covariances A+C+ECovariances A+C+E || H@A+C _H@A+C _

H@A+CH@A+C || A+C+E /A+C+E /

Matrix H 0.5Matrix H 0.5

Options MX%E=dzsim.covOptions MX%E=dzsim.cov

EndEnd

Power.mx 3Power.mx 3! Second part of script! Second part of script

! Fit the wrong model to the simulated data ! Fit the wrong model to the simulated data

! to calculate power! to calculate power

#NGroups 3#NGroups 3

G1 : model parametersG1 : model parameters

CalculationCalculation

Begin Matrices;Begin Matrices;

X lower 1 1 freeX lower 1 1 free

Y lower 1 1 fixedY lower 1 1 fixed

Z lower 1 1 freeZ lower 1 1 free

End Matrices;End Matrices;

Begin Algebra;Begin Algebra;

A = X*X' ;A = X*X' ;

C = Y*Y' ; C = Y*Y' ;

E = Z*Z' ; E = Z*Z' ;

End Algebra;End Algebra;

EndEnd

Power.mx 4Power.mx 4G2 : MZ twinsG2 : MZ twins

Data NInput_vars=2 NObservations=350Data NInput_vars=2 NObservations=350

CMatrix Full File=mzsim.covCMatrix Full File=mzsim.cov

Matrices= Group 1Matrices= Group 1

Covariances A+C+ECovariances A+C+E || A+C _A+C _

A+C A+C | | A+C+E /A+C+E /

Option RSidualsOption RSiduals

EndEnd

G3 : DZ twinsG3 : DZ twins

Data NInput_vars=2 NObservations=350Data NInput_vars=2 NObservations=350

CMatrix Full File=dzsim.covCMatrix Full File=dzsim.cov

Matrices= Group 1Matrices= Group 1

H Full 1 1H Full 1 1

Covariances A+C+ECovariances A+C+E || H@A+C _H@A+C _

H@A+C H@A+C | | A+C+E /A+C+E /

Matix H 0.5Matix H 0.5

Option RSidualsOption RSiduals

! Power for alpha = 0.05 and 1 df! Power for alpha = 0.05 and 1 df

Option Power= 0.05,1 Option Power= 0.05,1

EndEnd

35

Model IdentificationModel Identification

Necessary ConditionsNecessary Conditions

Sufficient ConditionsSufficient Conditions

Algebraic TestsAlgebraic Tests

Empirical TestsEmpirical Tests

35

36

Necessary ConditionsNecessary Conditions

Number of Parameters < or = Number of Parameters < or = Number of StatisticsNumber of Statistics

Structural Equation Model usually Structural Equation Model usually count variances & covariances to count variances & covariances to identify variance componentsidentify variance components

What is the number of What is the number of statistics/parameters in a univariate statistics/parameters in a univariate ACE model? Bivariate?ACE model? Bivariate? 3

6

37

Sufficient ConditionsSufficient Conditions

No general sufficient conditions for No general sufficient conditions for SEMSEM

Special case: ACE modelSpecial case: ACE model Distinct Statistics (i.e. have different Distinct Statistics (i.e. have different

predicted valuespredicted values VP = a2 + c2 + e2VP = a2 + c2 + e2 CMZ = a2 + c2CMZ = a2 + c2 CDZ = .5 a2 + c2CDZ = .5 a2 + c2

37

38

Sufficient Conditions 2Sufficient Conditions 2 Arrange in matrix formArrange in matrix form

1 1 1 a2 VP1 1 1 a2 VP 1 1 0 c2 = CMZ1 1 0 c2 = CMZ .5 1 0 e2 CDZ.5 1 0 e2 CDZ

A x = bA x = b

If A can be inverted then can find AIf A can be inverted then can find A--

11bb 38

39

Sufficient Conditions 3Sufficient Conditions 3

39

Solve ACE modelCalc ng=1Begin Matrices; A full 3 3 b full 3 1End Matrices;Matrix A1 1 11 1 0.5 1 0Labels Col A A C ELabels Row A VP CMZ CDZMatrix b ! Data, essentially1.8.5Labels Col B StatisticLabels Row B VP CMZ CDZBegin Algebra; C = A~; x = A~*b;End Algebra;Labels Row x A C EEnd

40

Sufficient Conditions 4Sufficient Conditions 4

What if not soluble by inversion?What if not soluble by inversion? Empirical:Empirical:

1 Pick set of parameter values T1 Pick set of parameter values T11

2 Simulate data2 Simulate data 3 Fit model to data starting at T3 Fit model to data starting at T22 (not T (not T11)) 4 Repeat and look for solutions to step 3 4 Repeat and look for solutions to step 3

that are perfect but have estimates not that are perfect but have estimates not equal to Tequal to T11

If equally good solution but different If equally good solution but different values, reject identified model values, reject identified model hypothesis hypothesis

40

ConclusionConclusion

Power calculations relatively simple Power calculations relatively simple to doto do

Curse of dimensionalityCurse of dimensionality Different for raw vs summary Different for raw vs summary

statisticsstatistics Simulation can be done many waysSimulation can be done many ways No substitute for research designNo substitute for research design