Package ubic.basecode.math.metaanalysis

Class CorrelationEffectMetaAnalysis

  • public class CorrelationEffectMetaAnalysis
extends MetaAnalysis
    Implementation of meta-analysis of correlations along the lines of chapter 18 of Cooper and Hedges, "Handbook of Research Synthesis". Both fixed and random effects models are supported, with z-transformed or untransformed correlations.
    Author:
    pavlidis

    • Constructor Detail

      • CorrelationEffectMetaAnalysis

        public CorrelationEffectMetaAnalysis()

      • CorrelationEffectMetaAnalysis

        public CorrelationEffectMetaAnalysis​(boolean fixed,
                                     boolean transform)

    • Method Detail

      • samplingVariance

        protected static double samplingVariance​(double r,
                                         double numsamples)
        Equation 18-10 from CH. For untransformed correlations. 

         v_i = ( 1 - r_i ˆ 2 ) ˆ 2 / ( n_i - 1 )

        I added a regularization to this, so that we don't get ridiculous variances when correlations are close to 1 (this happens). If the correlation is very close to 1 (or -1), we fudge it to be a value less close to 1 (e.g., 0.999)

        Parameters:
        r -
        n -
        Returns:

      • getBsv

        public double getBsv()

      • getE

        public double getE()

      • getN

        public double getN()

      • getP

        public double getP()

      • getQ

        public double getQ()

      • getV

        public double getV()

      • getZ

        public double getZ()

      • run

        public double run​(cern.colt.list.DoubleArrayList effects,
                  cern.colt.list.DoubleArrayList sampleSizes)
        Following CH section 2.2.

        There are four possible cases (for now):

        1. Fixed effects, Z-transformed. Weights are computed using CH eqns 18-8 and 18-3
        2. Fixed effects, untransformed. Weights are computed using CH eqns 18-10 and 18-3
        3. Random effects, Z-transformed. Weights are computed using CH eqns 18-20, 18-8 with 18-3
        4. Random effects, untransformed. Weights are computed using CH eqns 18-10, 18-20, 18-24.
        The default is untransformed, fixed effects.
        Parameters:
        correlations - - NOT fisher transformed. This routine takes care of that.
        sampleSizes -
        Returns:
        p-value. The p-value is also stored in the field p.

      • setFixed

        public void setFixed​(boolean fixed)

      • setTransform

        public void setTransform​(boolean transform)

      • fisherTransformedSamplingVariance

        protected double fisherTransformedSamplingVariance​(double sampleSize)
        Equation 18-8 from CH. For z-transformed correlations. 
         v_i = 1 / ( n_i - 3 )
        Parameters:
        n -
        Returns:

      • fisherTransformedSamplingVariances

        protected cern.colt.list.DoubleArrayList fisherTransformedSamplingVariances​(cern.colt.list.DoubleArrayList sampleSizes)
        Run equation CH 18-8 on a list of sample sizes.
        Parameters:
        sampleSizes -
        Returns:

      • samplingVariances

        protected cern.colt.list.DoubleArrayList samplingVariances​(cern.colt.list.DoubleArrayList effectSizes,
                                                           cern.colt.list.DoubleArrayList sampleSizes)
        Run equation CH 18-10 on a list of sample sizes and effects.
        Parameters:
        effectSizes -
        sampleSizes -
        Returns:
        See Also:
        samplingVariance(double,double)