*(Q) How can I study interaction in factorial designs if the dependent variable is not normal? (Non parametric 2-way ANOVA) *(A) There are several ways: *References: *1) Shirley EAC. "Applications of Ranking Methods to Multiple Comparison Procedures and Factorial Experiments". Appl Statist 1987;36(2):205-213. *2) Conover WJ and Iman RL. "Rank Transformations as a Bridge between Parametric and Nonparametric Statistics". The American Statistician 1981;35:124-129. * Code written by Marta Garcia-Granero, sent to rlevesque by email on 2002/02/05. ******************** EXAMPLE ****************************. DATA LIST LIST /altit(f2.0) splenec(f2.0) fibrinog(f4.0). BEGIN DATA 1 1 528 1 1 444 1 1 338 1 1 342 1 1 338 1 2 434 1 2 331 1 2 312 1 2 575 1 2 472 2 1 294 2 1 254 2 1 352 2 1 241 2 1 291 2 2 272 2 2 275 2 2 350 2 2 350 2 2 466 END DATA. VARIABLE LABELS altit "Altitude". VALUE LABELS altit 1 "Sea level" 2 "15.000 ft." . VARIABLE LABELS splenec "Splenectomy". VALUE LABELS splenec 1 "No" 2 "Yes" . * SOLUTION #1: FOR BALANCED AND SQUARE DESIGNS ONLY * (AS THIS ONE IS: n=5; 2x2 design) * (in this situation, Chi square values are additive) *********************************************************. COMPUTE group = splenec+2*(altit=2) . * (for a 3x3 design: COMPUTE group = splenec+3*(altit=2)+6*(altit=3) * and so on...). EXECUTE . VARIABLE LABELS group "Combined group". VALUE LABELS group 1 "Sea lev/No" 2 "Sea lev/Yes" 3 "15.000 ft/No" 4 "15.000 ft/Yes" . NPAR TESTS /K-W=fibrinog BY group(1 4) /K-W=fibrinog BY altit(1 2) /K-W=fibrinog BY splenec(1 2). * Now, read the 3 Chi-square values + df from output and compute * Interaction Chi-square (and df) as difference: group - altit- splenec * The values have to be copied to the two first COMPUTE statements * (I still haven't found a solution for this). DO IF $casenum=1. COMPUTE chi=5.774-4.487-0.824. COMPUTE df=3-1-1. COMPUTE sig = 1-CDF.CHISQ(chi,df) . END IF. EXECUTE. FORMAT chi (F8.3) df (F2.0) sig (F8.4). REPORT FORMAT=LIST AUTOMATIC ALIGN(CENTER) /VARIABLES=chi df sig /MISSING=LIST /TITLE "Kruskal-Wallis test for interaction". * For unbalanced and/or non-square designs (2x3, 3x4...), the above * solution doesn't work (you can end up with a negative Chi-square!). * SOLUTION #2 (GENERAL): 2-WAY ANOVA (WITH TYPE 3 SS) WITH THE RANKS *****************************************************************************. DATA LIST LIST /altit(f2.0) splenec(f2.0) fibrinog(f4.0). BEGIN DATA 1 1 528 1 1 444 1 1 338 1 1 342 1 1 338 1 2 434 1 2 331 1 2 312 1 2 575 1 2 472 2 1 294 2 1 254 2 1 352 2 1 241 2 1 291 2 2 272 2 2 275 2 2 350 2 2 350 2 2 466 END DATA. VARIABLE LABELS altit "Altitude". VALUE LABELS altit 1 "Sea level" 2 "15.000 ft." . VARIABLE LABELS splenec "Splenectomy". VALUE LABELS splenec 1 "No" 2 "Yes" . * First, rank the dependent variable (as Kruskal-Wallis test does). RANK VARIABLES=fibrinog (A) /RANK INTO yranks /PRINT=NO /TIES=MEAN . * Now, a factorial ANOVA with the ranks. UNIANOVA yranks BY altit splenec. * According to Conover and Iman, the F values and significance are * approximately valid, but, if you want to have more appropriate * p-values you must calculate Kruskal-Wallis Chi squares with their * significance (according to Shirley). * The following MACRO does all the work (from ranking to Chi square * significance, + two profile plots with medians, to interpret the * interaction if significant). ******************** * MACRO DEFINITION * ********************. DEFINE rank2w (facta=!CHAREND ('/') /factb=!CHAREND ('/') /depvar=!CMDEND ). GRAPH /LINE(MULTIPLE)MED(!depvar) BY !facta BY !factb. GRAPH /LINE(MULTIPLE)MED(!depvar) BY !factb BY !facta. RANK VARIABLES=!depvar (A) /RANK INTO yranks /PRINT=NO /TIES=MEAN . UNIANOVA yranks BY !facta !factb /OUTFILE=EFFECT ('report_.sav'). GET FILE='report_.sav' /KEEP= source_ ss df. EXECUTE. SELECT IF((source_='CorModel') or (source_='CorTotal') or (source_='S2') or (source_='S3') or (source_='S4')). EXECUTE . COMPUTE pos=$casenum. execute. SORT CASES BY pos(D). compute varri=ss/df. if $casenum>1 varri=lag(varri,1). execute. DO IF $casenum>1. COMPUTE chi=ss/varri. COMPUTE sig = 1-CDF.CHISQ(chi,df) . END IF. execute. VARIABLE LABELS sig "Sig.". VARIABLE LABELS chi "Chi square". FORMAT ss (F8.3) df (F4.0) chi (F8.3) sig (F9.4). SORT CASES by pos(A). REPORT FORMAT=LIST AUTOMATIC ALIGN(CENTER) /VARIABLES=source_(label) ss df chi sig /TITLE "2-way Kruskal-Wallis". !ENDDEFINE. * MACRO call *. rank2w facta=altit /factb=splenec /depvar=fibrinog.