SUBROUTINE SWILK (INIT, X, N, N1, N2, A, W, PW, IFAULT) C C ALGORITHM AS R94 APPL. STATIST. (1995) VOL.44, NO.4 C C Calculates the Shapiro-Wilk W test and its significance level C C IFAULT error code details from the R94 paper: C - 0 for no fault C - 1 if N1 < 3 C - 2 if N > 5000 (a non-fatal error) C - 3 if N2 < N/2, so insufficient storage for A C - 4 if N1 > N or (N1 < N and N < 20) C - 5 if the proportion censored (N-N1)/N > 0.8 C - 6 if the data have zero range (if sorted on input) C INTEGER N, N1, N2, IFAULT REAL X(*), A(*), PW, W REAL C1(6), C2(6), C3(4), C4(4), C5(4), C6(3), C7(2) REAL C8(2), C9(2), G(2) REAL Z90, Z95, Z99, ZM, ZSS, BF1, XX90, XX95, ZERO, ONE, TWO REAL THREE, SQRTH, QTR, TH, SMALL, PI6, STQR REAL SUMM2, SSUMM2, FAC, RSN, AN, AN25, A1, A2, DELTA, RANGE REAL SA, SX, SSX, SSA, SAX, ASA, XSX, SSASSX, W1, Y, XX, XI REAL GAMMA, M, S, LD, BF, Z90F, Z95F, Z99F, ZFM, ZSD, ZBAR C C Auxiliary routines C REAL PPND, POLY DOUBLE PRECISION ALNORM C INTEGER NCENS, NN2, I, I1, J LOGICAL INIT, UPPER C DATA C1 /0.0E0, 0.221157E0, -0.147981E0, -0.207119E1, * 0.4434685E1, -0.2706056E1/ DATA C2 /0.0E0, 0.42981E-1, -0.293762E0, -0.1752461E1, * 0.5682633E1, -0.3582633E1/ DATA C3 /0.5440E0, -0.39978E0, 0.25054E-1, -0.6714E-3/ DATA C4 /0.13822E1, -0.77857E0, 0.62767E-1, -0.20322E-2/ DATA C5 /-0.15861E1, -0.31082E0, -0.83751E-1, 0.38915E-2/ DATA C6 /-0.4803E0, -0.82676E-1, 0.30302E-2/ DATA C7 /0.164E0, 0.533E0/ DATA C8 /0.1736E0, 0.315E0/ DATA C9 /0.256E0, -0.635E-2/ DATA G /-0.2273E1, 0.459E0/ DATA Z90, Z95, Z99 /0.12816E1, 0.16449E1, 0.23263E1/ DATA ZM, ZSS /0.17509E1, 0.56268E0/ DATA BF1 /0.8378E0/, XX90, XX95 /0.556E0, 0.622E0/ DATA ZERO /0.0E0/, ONE/1.0E0/, TWO/2.0E0/, THREE/3.0E0/ DATA SQRTH /0.70711E0/, QTR/0.25E0/, TH/0.375E0/, SMALL/1E-19/ DATA PI6 /0.1909859E1/, STQR/0.1047198E1/, UPPER/.TRUE./ C PW = ONE IF (W .GE. ZERO) W = ONE AN = N IFAULT = 3 NN2 = N/2 IF (N2 .LT. NN2) RETURN IFAULT = 1 IF (N .LT. 3) RETURN C C If INIT is false, calculates coefficients for the test C IF (.NOT. INIT) THEN IF (N .EQ. 3) THEN A(1) = SQRTH ELSE AN25 = AN + QTR SUMM2 = ZERO DO 30 I = 1, N2 A(I) = PPND((REAL(I) - TH)/AN25,IFAULT) SUMM2 = SUMM2 + A(I) ** 2 30 CONTINUE SUMM2 = SUMM2 * TWO SSUMM2 = SQRT(SUMM2) RSN = ONE / SQRT(AN) A1 = POLY(C1, 6, RSN) - A(1) / SSUMM2 C C Normalize coefficients C IF (N .GT. 5) THEN I1 = 3 A2 = -A(2)/SSUMM2 + POLY(C2,6,RSN) FAC = SQRT((SUMM2 - TWO * A(1) ** 2 - TWO * * A(2) ** 2)/(ONE - TWO * A1 ** 2 - TWO * A2 ** 2)) A(1) = A1 A(2) = A2 ELSE I1 = 2 FAC = SQRT((SUMM2 - TWO * A(1) ** 2)/ * (ONE - TWO * A1 ** 2)) A(1) = A1 END IF DO 40 I = I1, NN2 A(I) = -A(I)/FAC 40 CONTINUE END IF INIT = .TRUE. END IF IF (N1 .LT. 3) RETURN NCENS = N - N1 IFAULT = 4 IF (NCENS .LT. 0 .OR. (NCENS .GT. 0 .AND. N .LT. 20)) RETURN IFAULT = 5 DELTA = FLOAT(NCENS)/AN IF (DELTA .GT. 0.8) RETURN C C If W input as negative, calculate significance level of -W C IF (W .LT. ZERO) THEN W1 = ONE + W IFAULT = 0 GOTO 70 END IF C C Check for zero range C IFAULT = 6 RANGE = X(N1) - X(1) IF (RANGE .LT. SMALL) RETURN C C Check for correct sort order on range - scaled X C IFAULT = 7 XX = X(1)/RANGE SX = XX SA = -A(1) J = N - 1 DO 50 I = 2, N1 XI = X(I)/RANGE CCCCC IF (XX-XI .GT. SMALL) PRINT *,' ANYTHING' SX = SX + XI IF (I .NE. J) SA = SA + SIGN(1, I - J) * A(MIN(I, J)) XX = XI J = J - 1 50 CONTINUE IFAULT = 0 IF (N .GT. 5000) IFAULT = 2 C C Calculate W statistic as squared correlation C between data and coefficients C SA = SA/N1 SX = SX/N1 SSA = ZERO SSX = ZERO SAX = ZERO J = N DO 60 I = 1, N1 IF (I .NE. J) THEN ASA = SIGN(1, I - J) * A(MIN(I, J)) - SA ELSE ASA = -SA END IF XSX = X(I)/RANGE - SX SSA = SSA + ASA * ASA SSX = SSX + XSX * XSX SAX = SAX + ASA * XSX J = J - 1 60 CONTINUE C C W1 equals (1-W) calculated to avoid excessive rounding error C for W very near 1 (a potential problem in very large samples) C SSASSX = SQRT(SSA * SSX) W1 = (SSASSX - SAX) * (SSASSX + SAX)/(SSA * SSX) 70 W = ONE - W1 C C Calculate significance level for W (exact for N=3) C IF (N .EQ. 3) THEN PW = PI6 * (ASIN(SQRT(W)) - STQR) RETURN END IF Y = LOG(W1) XX = LOG(AN) M = ZERO S = ONE IF (N .LE. 11) THEN GAMMA = POLY(G, 2, AN) IF (Y .GE. GAMMA) THEN PW = SMALL RETURN END IF Y = -LOG(GAMMA - Y) M = POLY(C3, 4, AN) S = EXP(POLY(C4, 4, AN)) ELSE M = POLY(C5, 4, XX) S = EXP(POLY(C6, 3, XX)) END IF IF (NCENS .GT. 0) THEN C C Censoring by proportion NCENS/N. Calculate mean and sd C of normal equivalent deviate of W. C LD = -LOG(DELTA) BF = ONE + XX * BF1 Z90F = Z90 + BF * POLY(C7, 2, XX90 ** XX) ** LD Z95F = Z95 + BF * POLY(C8, 2, XX95 ** XX) ** LD Z99F = Z99 + BF * POLY(C9, 2, XX) ** LD C C Regress Z90F,...,Z99F on normal deviates Z90,...,Z99 to get C pseudo-mean and pseudo-sd of z as the slope and intercept C ZFM = (Z90F + Z95F + Z99F)/THREE ZSD = (Z90*(Z90F-ZFM)+Z95*(Z95F-ZFM)+Z99*(Z99F-ZFM))/ZSS ZBAR = ZFM - ZSD * ZM M = M + ZBAR * S S = S * ZSD END IF PW = REAL(ALNORM(DBLE((Y - M)/S), UPPER)) C RETURN END DOUBLE PRECISION FUNCTION ALNORM(X, UPPER) C C EVALUATES THE TAIL AREA OF THE STANDARDIZED NORMAL CURVE FROM C X TO INFINITY IF UPPER IS .TRUE. OR FROM MINUS INFINITY TO X C IF UPPER IS .FALSE. C C NOTE NOVEMBER 2001: MODIFY UTZERO. ALTHOUGH NOT NECESSARY C WHEN USING ALNORM FOR SIMPLY COMPUTING PERCENT POINTS, C EXTENDING RANGE IS HELPFUL FOR USE WITH FUNCTIONS THAT C USE ALNORM IN INTERMEDIATE COMPUTATIONS. C DOUBLE PRECISION LTONE,UTZERO,ZERO,HALF,ONE,CON, $ A1,A2,A3,A4,A5,A6,A7,B1,B2, $ B3,B4,B5,B6,B7,B8,B9,B10,B11,B12,X,Y,Z,ZEXP LOGICAL UPPER,UP C C LTONE AND UTZERO MUST BE SET TO SUIT THE PARTICULAR COMPUTER C CCCCC DATA LTONE, UTZERO /7.0D0, 18.66D0/ DATA LTONE, UTZERO /7.0D0, 38.00D0/ DATA ZERO,HALF,ONE,CON /0.0D0,0.5D0,1.0D0,1.28D0/ DATA A1, A2, A3, $ A4, A5, A6, $ A7 $ /0.398942280444D0, 0.399903438504D0, 5.75885480458D0, $ 29.8213557808D0, 2.62433121679D0, 48.6959930692D0, $ 5.92885724438D0/ DATA B1, B2, B3, $ B4, B5, B6, $ B7, B8, B9, $ B10, B11, B12 $ /0.398942280385D0, 3.8052D-8, 1.00000615302D0, $ 3.98064794D-4, 1.98615381364D0, 0.151679116635D0, $ 5.29330324926D0, 4.8385912808D0, 15.1508972451D0, $ 0.742380924027D0, 30.789933034D0, 3.99019417011D0/ C ZEXP(Z) = DEXP(Z) C UP = UPPER Z = X IF (Z .GE. ZERO) GOTO 10 UP = .NOT. UP Z = -Z 10 IF (Z .LE. LTONE .OR. UP .AND. Z .LE. UTZERO) GOTO 20 ALNORM = ZERO GOTO 40 20 Y = HALF * Z * Z IF (Z .GT. CON) GOTO 30 C ALNORM = HALF - Z * (A1- A2 * Y / (Y + A3- A4 / (Y + A5 + A6 / $ (Y + A7)))) GOTO 40 C 30 ALNORM = B1* ZEXP(-Y)/(Z - B2 + B3/ (Z +B4 +B5/(Z -B6 +B7/ $ (Z +B8 -B9/ (Z +B10 +B11/ (Z + B12)))))) C 40 IF (.NOT. UP) ALNORM = ONE - ALNORM RETURN END REAL FUNCTION PPND(P, IFAULT) C C ALGORITHM AS 111 APPL. STATIST. (1977), VOL.26, NO.1 C C PRODUCES NORMAL DEVIATE CORRESPONDING TO LOWER TAIL AREA OF P C REAL VERSION FOR EPS = 2 **(-31) C THE HASH SUMS ARE THE SUMS OF THE MODULI OF THE COEFFICIENTS C THEY HAVE NO INHERENT MEANINGS BUT ARE INCLUDED FOR USE IN C CHECKING TRANSCRIPTIONS C STANDARD FUNCTIONS ABS, ALOG AND SQRT ARE USED C C NOTE: WE COULD USE DATAPLOT NORPPF, BUT VARIOUS APPLIED C STATISTICS ALGORITHMS USE THIS. SO WE PROVIDE IT TO C MAKE USE OF APPLIED STATISTICS ALGORITHMS EASIER. C REAL ZERO, SPLIT, HALF, ONE REAL A0, A1, A2, A3, B1, B2, B3, B4, C0, C1, C2, C3, D1, D2 REAL P, Q, R INTEGER IFAULT DATA ZERO /0.0E0/, HALF/0.5E0/, ONE/1.0E0/ DATA SPLIT /0.42E0/ DATA A0 / 2.50662823884E0/ DATA A1 / -18.61500062529E0/ DATA A2 / 41.39119773534E0/ DATA A3 / -25.44106049637E0/ DATA B1 / -8.47351093090E0/ DATA B2 / 23.08336743743E0/ DATA B3 / -21.06224101826E0/ DATA B4 / 3.13082909833E0/ DATA C0 / -2.78718931138E0/ DATA C1 / -2.29796479134E0/ DATA C2 / 4.85014127135E0/ DATA C3 / 2.32121276858E0/ DATA D1 / 3.54388924762E0/ DATA D2 / 1.63706781897E0/ C IFAULT = 0 Q = P - HALF IF (ABS(Q) .GT. SPLIT) GOTO 1 R = Q*Q PPND = Q * (((A3*R + A2)*R + A1) * R + A0) / * ((((B4*R + B3)*R + B2) * R + B1) * R + ONE) RETURN 1 R = P IF (Q .GT. ZERO)R = ONE - P IF (R .LE. ZERO) GOTO 2 R = SQRT(-ALOG(R)) PPND = (((C3 * R + C2) * R + C1) * R + C0)/ * ((D2*R + D1) * R + ONE) IF (Q .LT. ZERO) PPND = -PPND RETURN 2 IFAULT = 1 PPND = ZERO RETURN END REAL FUNCTION POLY(C, NORD, X) C C C ALGORITHM AS 181.2 APPL. STATIST. (1982) VOL. 31, NO. 2 C C CALCULATES THE ALGEBRAIC POLYNOMIAL OF ORDER NORED-1 WITH C ARRAY OF COEFFICIENTS C. ZERO ORDER COEFFICIENT IS C(1) C REAL C(NORD) POLY = C(1) IF(NORD.EQ.1) RETURN P = X*C(NORD) IF(NORD.EQ.2) GOTO 20 N2 = NORD-2 J = N2+1 DO 10 I = 1,N2 P = (P+C(J))*X J = J-1 10 CONTINUE 20 POLY = POLY+P RETURN END