Hi

Curious, it seems that a message I sent to the list didn't get through (?)

In may, someone asked for Cochran-Armitage test with SPSS. I replied, but, today, trying to recover a lost message of mine from the SPSS archives at the listserv, I realised my reply wasn't there. I'm sending it again:

SPSS gives a Linear-by-linear test, quite close, but not exactly the same.

*********************************** * COCHRAN-ARMITAGE TEST FOR TREND * ***********************************

* Reference: Alan Agresti (2002), Categorical Data Analysis (2nd Ed), pp 181-2, John Wiley & Sons *.

* Dataset #1: scoring is OK (example provided with SAS at:) (http://www.id.unizh.ch/software/unix/statmath/sas/sasdoc/stat/chap28/sect40.htm) * (SAS results): * Statistic (Z) 4.7918 * Asymptotic Test * One-sided Pr < Z <.0001 * Two-sided Pr < |Z| <.0001 * Sample Size = 161

DATA LIST LIST/dose adverse count (3 F8.0). BEGIN DATA 0 1 26 0 2 6 1 1 26 1 2 7 2 1 23 2 2 9 3 1 18 3 2 14 4 1 9 4 2 23 END DATA. VALUE LABEL adverse 1'No' 2'Yes'. WEIGHT BY count .

* Using SPSS linear by linear test *. CROSSTABS /TABLES=dose BY adverse /FORMAT= AVALUE TABLES /CELLS= COUNT ROW EXPECTED /STATISTIC=CHISQ.

* Cochran-Armitage test with MATRIX *.

* This code assumes the dependent variable is coded 0/1 (No/Yes) *. TEMPORARY. COMPUTE adverse=adverse-1. AGGREGATE /OUTFILE = 'c:\temp\aggdata.sav' /BREAK = dose /pi = MEAN(adverse) /ni = N.

MATRIX. PRINT/TITLE='Cochran-Armitage Test for Trend (Agresti 2002)'. GET pi /VAR=pi /FILE='c:\temp\aggdata.sav'. GET ni /VAR=ni /FILE='c:\temp\aggdata.sav'. * Scoring (change if required) *. COMPUTE xi={0; 1; 2; 3; 4}. * Reports *. PRINT {xi,ni&*pi,ni} /FORMAT='F8.1' /TITLE='Statistics' /CLABELS='Score i','Count i','N i'. PRINT MSUM(ni) /FORMAT='F8.0' /TITLE='Total sample size'. * Calculations *. COMPUTE p=MSUM(ni&*pi)/MSUM(ni). COMPUTE x=MSUM(xi&*ni)/MSUM(ni). COMPUTE num=(xi-x)&*ni&*pi. COMPUTE den=p*(1-p)*ni&*((xi-x)&**2). COMPUTE z=MSUM(num)/SQRT(MSUM(den)). COMPUTE zsig=2*(1-CDFNORM(ABS(z))). * Final report *. PRINT {z,z**2,zsig,zsig/2} /FORMAT='F8.3' /CLABELS='Z','Z²','2-sig','1-sig' /TITLE='Statistics & Significance'. COMPUTE observed={(ni-ni&*pi),ni&*pi}. COMPUTE expected=RSUM(observed)*CSUM(observed)/MSUM(ni). COMPUTE chisq=MSUM((observed-expected)&**2/expected). COMPUTE chisqp=1-CHICDF(chisq,NROW(observed)-1). COMPUTE reschi=chisq-z**2. COMPUTE reschip=1-CHICDF(reschi,NROW(observed)-2). PRINT {chisq,chisqp;reschi,reschip} /FORMAT='F8.3' /RLABELS='Pearson','Deviat.' /CLABELS='Chi²','Sig.' /TITLE='Pearson-Chi² & Deviation from trend (non-linearity)'. END MATRIX.

* As you can't specify scores for CROSSTABS, the key is recoding * the predictor to asign the scores you want. Suppose your data * don't reflect the scoring you want to give *.

* Dataset #2 (Agresti 2002) *. DATA LIST LIST /alcohol malform count (3 F8.0). BEGIN DATA 0 1 48 0 0 17066 1 1 38 1 0 14464 2 1 5 2 0 788 3 1 1 3 0 126 4 1 1 4 0 37 END DATA.

VALUE LABEL alcohol 0 ' 0' 1 '<1' 2 '1-2' 3 '3-5' 4 '>6'. VALUE LABEL malform 0'Absent' 1'Present'. WEIGHT BY count .

* Agresti gives the following scores: 0, 0.5, 1.5, 4, 7 *.

* The following will not give the results expected (Z²=6.57 p=0.010): *.

CROSSTABS /TABLES=alcohol BY malform /FORMAT= AVALUE TABLES /CELLS= COUNT ROW /STATISTIC=CHISQ. /* LxL test is 1.828 (p=0.176)*.

* Recoding and re-running *. TEMPORARY. RECODE alcohol (0=0) (1=0.5) (2=1.5) (3=4) (4=7) INTO alcscore . VARIABLE LABELS alcscore 'Scores for alcohol'. CROSSTABS /TABLES=alcscore BY malform /FORMAT= AVALUE TABLES /CELLS= COUNT ROW /STATISTIC=CHISQ. /* LxL is now 6.570 *.

* Cochran-Armitage test with MATRIX *.

* This time, MALFORM takes the correct values (0/1) *. AGGREGATE /OUTFILE='c:\temp\aggdata.sav' /BREAK=alcohol /pi = MEAN(malform) /ni=N.

MATRIX. PRINT/TITLE='Cochran-Armitage Test for Trend (Agresti 2002)'. GET pi /VAR=pi /FILE='c:\temp\aggdata.sav'. GET ni /VAR=ni /FILE='c:\temp\aggdata.sav'. * Scoring (change if required) *. COMPUTE xi={0; 0.5; 1.5; 4; 7}. * Reports *. PRINT {xi,ni&*pi,ni} /FORMAT='F8.1' /TITLE='Statistics' /CLABELS='Score i','Count i','N i'. PRINT MSUM(ni) /FORMAT='F8.0' /TITLE='Total sample size'. * Calculations *. COMPUTE p=MSUM(ni&*pi)/MSUM(ni). COMPUTE x=MSUM(xi&*ni)/MSUM(ni). COMPUTE num=(xi-x)&*ni&*pi. COMPUTE den=p*(1-p)*ni&*((xi-x)&**2). COMPUTE z=MSUM(num)/SQRT(MSUM(den)). COMPUTE zsig=2*(1-CDFNORM(ABS(z))). * Final report *. PRINT {z,z**2,zsig,zsig/2} /FORMAT='F8.3' /CLABELS='Z','Z²','2-sig','1-sig' /TITLE='Statistics & Significance'. COMPUTE observed={(ni-ni&*pi),ni&*pi}. COMPUTE expected=RSUM(observed)*CSUM(observed)/MSUM(ni). COMPUTE chisq=MSUM((observed-expected)&**2/expected). COMPUTE chisqp=1-CHICDF(chisq,NROW(observed)-1). COMPUTE reschi=chisq-z**2. COMPUTE reschip=1-CHICDF(reschi,NROW(observed)-2). PRINT {chisq,chisqp;reschi,reschip} /FORMAT='F8.3' /RLABELS='Pearson','Deviat.' /CLABELS='Chi²','Sig.' /TITLE='Pearson-Chi² & Deviation from trend (non-linearity)'. END MATRIX.

-- Regards, Marta mailto:biostatistics@terra.es

Marta Garcia-Granero, PhD Statistician