25013 - Power and sample size for independence tests in RxC tables

Sample 25013: Power and sample size for independence tests in RxC tables

Power and sample size for independence tests in RxC tables

Contents:

Purpose / Requirements / Usage / Details / Limitations / See Also / References

NOTE: Beginning in SAS 9.4 TS1M4, use the CUSTOM statement in SAS/STAT PROC POWER to do these power computations.

PURPOSE:

Computes approximate power for Pearson and Likelihood Ratio chi-square tests of independence in PROC FREQ.

REQUIREMENTS:

Only Base SAS Software is required.

USAGE:

Follow the instructions in the Downloads tab of this sample to save the POWERRxC macro definition. Replace the text within quotes in the following statement with the location of the POWERRxC macro definition file on your system. In your SAS program or in the SAS editor window, specify this statement to define the POWERRxC macro and make it available for use:

   %inc "<location of your file containing the POWERRxC macro>";

Following this statement, you may call the POWERRxC macro. See the Results tab for an example.

The following parameters are required:

row=variable: Specifies the variable defining the rows of the table.
col=variable: Specifies the variable defining the columns of the table.

These parameters are optional:

data=data-set-name

Specifies the SAS data set to be analyzed. If the DATA= option is not supplied, the most recently created SAS data set is used.

count=variable

Specifies a variable containing the cell counts of the table. Omit this option if each observation in the DATA= data set represents only a single entry in the table. This variable, if specified, is used in the WEIGHT statement in PROC FREQ.

level=value

Specifies the significance level or size of the test. This is the limit you place on the probability of a type 1 error. The specified value should be between 0 and 1. The default value is 0.05 which sets the probability of a type 1 error at 0.05.

nrange=value | value-list

Specifies the sample size or list of sample sizes for which approximate power is to be computed. If omitted, the actual sample size is used. You may specify a list of values separated by commas, a range of the form x TO y or x TO y BY z, or a combination of these. However, you must surround the NRANGE= value with %str( ) if any commas appear in it. For example,

nrange=20 to 200 by 20

nrange=%str(20,50,100,140)

nrange=%str(10, 20, 50 to 100 by 10)

DETAILS:

Agresti (1990 and 2002) discusses the details of computing approximate power for Pearson and deviance chi-square tests in contingency tables using the noncentral chi-square distribution.

LIMITATIONS:

Limited error checking is done. If the DATA= option is specified, be sure the named data set exists. If DATA= is not specified, a data set must have been created previously in the current SAS session. Be sure that the variables specified in the ROW=, COL= and COUNT= options exist in the data set. You can run PROC CONTENTS on the data set prior to using this macro to verify the data set name and the names of variables.

SEE ALSO:

Discussion of power analysis for 2x2 tables
POWER2x2 macro to compute power of a test comparing two independent proportions.
POWER macro to calculate power-related measures for retrospective and prospective analyses for linear models fit to normal responses.

REFERENCES:

Agresti, A. (1990 and 2002), Categorical Data Analysis, New York: John Wiley & Sons, Inc.

These sample files and code examples are provided by SAS Institute Inc. "as is" without warranty of any kind, either express or implied, including but not limited to the implied warranties of merchantability and fitness for a particular purpose. Recipients acknowledge and agree that SAS Institute shall not be liable for any damages whatsoever arising out of their use of this material. In addition, SAS Institute will provide no support for the materials contained herein.

Example 1: Power for 2x2 table with cell count data

In this example, each observation is a cell count instead of a single response, so a count variable (FREQ) is created to hold the cell counts. The observed sample size is used for computing approximate power.

        
     data a;
        do row=1 to 2; 
          do col=0,1;
            input freq @@; output;
          end;
        end; 
        datalines;
        3 11
        6 2
        ;
     
     /* Define the POWER2x2 macro */
     %inc "<location of your file containing the POWERRxC macro>";
        
     %powerRxC(row=row, col=col, count=freq)

Results:

                  Approximate Power of Chi-square Tests for Independence
                                      Test Level=.05

                                     Power of        Power
                                      Pearson       of L.R.
                               n    Chi-square    Chi-square

                              22      0.69093       0.70345

Example 2: Power for 2x2 table with individual data

This example is identical to the previous one except that raw, individual data are used instead of summarized, cell-count data. That is, each observation in data set AA contributes a count of one in one cell of the table. With individual data, a count variable is not needed.

     data aa;
        do row=1 to 2;
          do col=0,1;
            input freq @@; 
            do i=1 to freq; 
              drop i freq;
              output; 
            end;
          end;
        end; 
        datalines;
        3 11
        6 2
        ;

        %powerRxC(row=row, col=col)

Results: The results are identical to Example 1.

Example 3: Power for RxC table

Suppose subjects are to be given one of four treatments and a binary response will be recorded for each subject resulting in a 4x2 table. Determine the sample size needed to detect a 0.1 difference in response probabilities with adequate power. This is done by entering cell values that exhibit the worst case scenario -- three rows with response probability .2 (=.05/(.2+.05)) and one row with response probability .1 (=.025/(.225+.025)). Treatment sample sizes are to be equal, so the marginal row (treatment) probabilities are all 0.25. Here, column 2 is regarded as the response category, though the columns could be reversed without effect.

     data bb;
        do trt=1 to 4; 
          do response=0,1;
            input prob @@; output;
          end;
        end;
        datalines;
        .2   .05
        .2   .05
        .2   .05
        .225 .025
        ;

        %powerRxC(row=trt, col=response, count=prob, 
                  nrange=%str(20,50,100 to 1000 by 100))

Results:

                  Approximate Power of Chi-square Tests for Independence
                                      Test Level=.05

                                      Power of        Power
                                       Pearson       of L.R.
                                n    Chi-square    Chi-square

                               20      0.06570       0.06735
                               50      0.09111       0.09562
                              100      0.13763       0.14762
                              200      0.24106       0.26305
                              300      0.34986       0.38282
                              400      0.45603       0.49715
                              500      0.55420       0.59999
                              600      0.64131       0.68842
                              700      0.71608       0.76178
                              800      0.77855       0.82086
                              900      0.82955       0.86728
                             1000      0.87038       0.90299

Beginning in SAS 9.4 TS1M4, the same results can be obtained using the CUSTOM statement in PROC POWER. As described in the documentation of the CUSTOM statement, an exemplary data set is created and the chi-square statistic and degrees of freedom are obtained for this data set. The chi-square statistic is used to provide the noncentrality value (PRIMNC=) in PROC POWER. In this example, the treatments are to have equal sample sizes so a single instance of each is created in the exemplary data set, BB (one instance involves one observation for each response level). Note that this is 4 times the proportion (0.25) of each treatment in the data. Hence, the chi-square statistic is divided by four for use as the noncentrality value.

The following statements produce the power analysis and yields the same power values as the PowerRxC macro above.

      data BB;
        do trt=1 to 4; 
        input p0 p1;
          y=0; py=p0/(p0+p1); output;
          y=1; py=p1/(p0+p1); output;
        end;
        datalines;
        .2   .05
        .2   .05
        .2   .05
        .225 .025
        ;
      proc freq data=bb;
        weight py;
        table trt*y/chisq;
        ods output chisq=cs(where=(statistic="Chi-Square")) ;
        run;
      data _null_;
        set cs;
        call symput('cs',value/4);
        call symput('df',df);
        run;
      proc power;
        custom
          dist   = chisquare
          primnc = &cs
          testdf = &df
          ntotal = 20, 50, 100 to 1000 by 100
          power  = .;
        run;

Example 4: Power for range of sample sizes in a 2x2 table

Example from Agresti (1990) pp 241-243. Column 1 probabilities for each row are hypothesized to be .63 and .57. Row sample sizes are to be equal, so the marginal row probabilities are both 0.5.

     
     data c; 
        do row=1 to 2;
          do col=0,1;
            input freq @@; 
            output; 
          end;
        end; 
        datalines; 
        .315 .185 
        .285 .215
        ;

        %powerRxC(data=c, row=row, col=col, count=freq, 
                  nrange=%str(20,50 to 200 by 50))

Results:

                  Approximate Power of Chi-square Tests for Independence
                                      Test Level=.05

                                     Power of        Power
                                      Pearson       of L.R.
                              n     Chi-square    Chi-square

                              20      0.05864       0.05864
                              50      0.07174       0.07176
                             100      0.09395       0.09398
                             150      0.11651       0.11656
                             200      0.13935       0.13941

Example 5: Power for observed 3x2 table

Get power for a .10 level test performed on previously-collected data.

     data d; 
        input type $ sex $ ;
        datalines;
        a m
        b m
        c f
        b m
        c f
        c f
        b f
        a m
        a f
        a m
        b m
        c f
        b m
        c f
        c f
        b f
        a m
        a f
        ;

        %powerRxC(row=type, col=sex, level=.10)

Results:

                  Approximate Power of Chi-square Tests for Independence
                                      Test Level=.10

                                     Power of        Power
                                      Pearson       of L.R.
                               n    Chi-square    Chi-square

                              18      0.77402       0.87034

Type:	Sample
Topic:	SAS Reference ==> Procedures ==> FREQ Analytics ==> Nonparametric Analysis Analytics ==> Longitudinal Analysis Analytics ==> Descriptive Statistics Analytics ==> Exact Methods Analytics ==> Categorical Data Analysis SAS Reference ==> Procedures ==> POWER

Date Modified:	2020-01-21 11:16:29
Date Created:	2005-01-13 15:03:36

Product Family	Product	Host	SAS Release
			Starting	Ending
SAS System	SAS/STAT	All	n/a	n/a

Support

Sample 25013: Power and sample size for independence tests in RxC tables

Power and sample size for independence tests in RxC tables

Operating System and Release Information