FUNCTION ROBUST_LINEFIT,XIN,YIN,YFIT,SIG,SS, NUMIT=THIS_MANY, BISECT=TYPE, $ Bisquare_Limit=Bisquare_Limit, $ Close_Factor=Close_Factor ;+ ; NAME: ; ROBUST_LINEFIT ; ; PURPOSE: ; An outlier-resistant two-variable linear regression. ; EXPLANATION: ; Either Y on X or, for the case in which there is no true independent ; variable, the bisecting line of Y vs X and X vs Y is calculated. No ; knowledge of the errors of the input points is assumed. ; ; CALLING SEQUENCE: ; COEFF = ROBUST_LINEFIT( X, Y, YFIT, SIG, COEF_SIG, [ /BISECT, ; BiSquare_Limit = , Close_factor = , NumIT = ] ) ; ; INPUTS: ; X = Independent variable vector, floating-point or double-precision ; Y = Dependent variable vector ; ; OUTPUTS: ; Function result = coefficient vector. ; If = 0.0 (scalar), no fit was possible. ; If vector has more than 2 elements (the last=0) then the fit is dubious. ; ; OPTIONAL OUTPUT PARAMETERS: ; YFIT = Vector of calculated y's ; SIG = The "standard deviation" of the fit's residuals. If BISECTOR ; is set, this will be smaller by ~ sqrt(2). ; COEF_SIG = The estimated standard deviations of the coefficients. If ; BISECTOR is set, however, this becomes the vector of fit ; residuals measured orthogonal to the line. ; ; OPTIONAL INPUT KEYWORDS: ; NUMIT = the number of iterations allowed. Default = 25 ; BISECT if set, the bisector of the "Y vs X" and "X vs Y" fits is ; determined. The distance PERPENDICULAR to this line is used ; in calculating weights. This is better when the uncertainties ; in X and Y are comparable, so there is no true independent ; variable. Bisquare_Limit Limit used for calculation of ; bisquare weights. In units of outlier-resistant standard ; deviations. Default: 6. ; Smaller limit ==>more resistant, less efficient ; Close_Factor - Factor used to determine when the calculation has converged. ; Convergence if the computed standard deviation changes by less ; than Close_Factor * ( uncertainty of the std dev of a normal ; distribution ). Default: 0.03. ; SUBROUTINE CALLS: ; ROB_CHECKFIT ; ROBUST_SIGMA, to calculate a robust analog to the std. deviation ; ; PROCEDURE: ; For the initial estimate, the data is sorted by X and broken into 2 ; groups. A line is fitted to the x and y medians of each group. ; Bisquare ("Tukey's Biweight") weights are then calculated, using the ; a limit of 6 outlier-resistant standard deviations. ; This is done iteratively until the standard deviation changes by less ; than CLOSE_ENOUGH = CLOSE_FACTOR * {uncertainty of the standard ; deviation of a normal distribution} ; ; REVISION HISTORY: ; Written, H. Freudenreich, STX, 4/91. ; 4/13/93 to return more realistic SS's HF ; 2/94 --more error-checking, changed convergence criterion HF ; 5/94 --added BISECT option. HF. ; 8/94 --added Close_Factor and Bisquare_Limit options Jack Saba. ; 4/02 --V5.0 version, use MEDIAN(/EVEN) W. Landsman ;- ON_ERROR,2 IF N_ELEMENTS(THIS_MANY) GT 0 THEN ITMAX=THIS_MANY ELSE ITMAX=25 IF ( NOT KEYWORD_SET ( Close_Factor ) ) THEN Close_Factor = 0.03 DEL = 5.0E-07 EPS = 1.0E-20 N = N_ELEMENTS(XIN) ; First, shift X and Y to their centers of gravity: X0 = TOTAL(XIN)/N & Y0=TOTAL(YIN)/N X = XIN-X0 & Y = YIN-Y0 CC=FLTARR(2) SS=FLTARR(2) SIG=0. YFIT=YIN BADFIT=0 NGOOD=N ; Make sure the independent variables are not all the same. XRANGE=MAX(X)-MIN(X) AVEX= (TOTAL(ABS(X))/N) > EPS IF (XRANGE LT EPS) OR (XRANGE/AVEX LT DEL) THEN BEGIN message,'Independent variables the same. No fit possible.',/CON RETURN,0. ENDIF ; First guess: LSQ=0 YP=Y IF N GT 5 THEN BEGIN ; We divide the data into 2 groups and fit a line to their X and Y medians. S=SORT(X) & U=X[S] & V=Y[S] NHALF=N/2-1 X1=MEDIAN(U[0:NHALF],/EVEN) & X2=MEDIAN(U[NHALF+1:N-1],/EVEN) Y1=MEDIAN(V[0:NHALF],/EVEN) & Y2=MEDIAN(V[NHALF+1:N-1],/EVEN) IF ABS(X2-X1) LT EPS THEN BEGIN ; The X medians are too close. Select the end-points instead. X1=U[0] & X2=U[N-1] Y1=V[0] & Y2=V[N-1] ENDIF CC[1]=(Y2-Y1)/(X2-X1) & CC[0]=Y1-CC[1]*X1 YFIT = CC[0]+CC[1]*X ISTAT = ROB_CHECKFIT(YP,YFIT,EPS,DEL, SIG,FRACDEV,NGOOD,W,S) IF NGOOD LT 2 THEN LSQ=1 ENDIF IF (LSQ EQ 1) OR (N LT 6) THEN BEGIN ; Try a least-squares fit SX=TOTAL(X) & SY=TOTAL(Y) & SXY=TOTAL(X*Y) & SXX=TOTAL(X*X) D=SXX-SX*SX IF ABS(D) LT EPS THEN BEGIN PRINT,'ROBUST_LINEFIT: No fit possible.' RETURN,0. ENDIF YSLOP=(SXY-SX*SY)/D & YYINT=(SXX*SY-SX*SXY)/D IF KEYWORD_SET(TYPE) THEN BEGIN ; Get the X vs Y line. SYY=TOTAL(Y*Y) D=SYY-SY*SY IF ABS(D) LT EPS THEN BEGIN PRINT,'ROBUST_LINEFIT: No fit possible.' RETURN,0. ENDIF TSLOP=(SXY-SY*SX)/D & TYINT=(SYY*SX-SY*SXY)/D ; Now invert it to get the form Y=a+bX: IF ABS(TSLOP) LT EPS THEN BEGIN message,'No fit possible.',/CON RETURN,0. ENDIF XSLOP = 1./TSLOP & XYINT=-TYINT/TSLOP ; Now calculate the equation of the bisector of the 2 lines: IF YSLOP GT XSLOP THEN BEGIN A1=YYINT & B1=YSLOP & R1=SQRT(1.+YSLOP^2) A2=XYINT & B2=XSLOP & R2=SQRT(1.+XSLOP^2) ENDIF ELSE BEGIN A2=YYINT & B2=YSLOP & R2=SQRT(1.+YSLOP^2) A1=XYINT & B1=XSLOP & R1=SQRT(1.+XSLOP^2) ENDELSE YINT = (R1*A2+R2*A1)/(R1+R2) SLOP = (R1*B2+R2*B1)/(R1+R2) ; Now find the orthogonal distance to the line. Convert to normal ; coordinates. R = SQRT(1.+SLOP^2) & IF YINT GT 0. THEN R=-R U1 = SLOP/R & U2=-1./R & U3=YINT/R YP = U1*X+U2*Y+U3 ; = orthog. distance to line YFIT = FLTARR(N) ; to fool ROB_CHECKFIT SS=YP ENDIF ELSE BEGIN SLOP=YSLOP & YINT=YYINT YFIT = YINT+SLOP*X ENDELSE CC = [YINT,SLOP] ISTAT = ROB_CHECKFIT(YP,YFIT,EPS,DEL, SIG,FRACDEV,NGOOD,W,S) ENDIF IF ISTAT EQ 0 THEN GOTO,AFTERFIT IF NGOOD LT 2 THEN BEGIN message,'Data Dangerously Weird. Fit Questionable.',/CON BADFIT=1 GOTO,AFTERFIT ENDIF ; Now iterate until the solution converges: SIG_1= (100.*SIG) < 1.0E20 CLOSE_ENOUGH = Close_Factor * SQRT(.5/(N-1)) > DEL DIFF= 1.0E20 NIT = 0 WHILE( (DIFF GT CLOSE_ENOUGH) AND (NIT LT ITMAX) ) DO BEGIN NIT=NIT+1 SIG_2=SIG_1 SIG_1=SIG SX=TOTAL(W*X) & SY=TOTAL(W*Y) & SXY=TOTAL(W*X*Y) & SXX=TOTAL(W*X*X) D=SXX-SX*SX IF ABS(D) LT EPS THEN BEGIN message,'No fit possible.',/CON RETURN,0. ENDIF YSLOP = (SXY-SX*SY)/D & YYINT = (SXX*SY-SX*SXY)/D SLOP = YSLOP & YINT = YYINT IF KEYWORD_SET(TYPE) THEN BEGIN ; Get the X vs Y line. SYY=TOTAL(W*Y*Y) D=SYY-SY*SY IF ABS(D) LT EPS THEN BEGIN PRINT,'ROBUST_LINEFIT: No fit possible.' RETURN,0. ENDIF TSLOP=(SXY-SY*SX)/D & TYINT=(SYY*SX-SY*SXY)/D ; Now invert it to get the form Y=a+bX: IF ABS(TSLOP) LT EPS THEN BEGIN PRINT,'ROBUST_LINEFIT: No fit possible.' RETURN,0. ENDIF XSLOP=1./TSLOP & XYINT=-TYINT/TSLOP ; Now calculate the equation of the bisector of the 2 lines: IF YSLOP GT XSLOP THEN BEGIN A1=YYINT & B1=YSLOP & R1=SQRT(1.+YSLOP^2) A2=XYINT & B2=XSLOP & R2=SQRT(1.+XSLOP^2) ENDIF ELSE BEGIN A2=YYINT & B2=YSLOP & R2=SQRT(1.+YSLOP^2) A1=XYINT & B1=XSLOP & R1=SQRT(1.+XSLOP^2) ENDELSE YINT=(R1*A2+R2*A1)/(R1+R2) SLOP=(R1*B2+R2*B1)/(R1+R2) R=SQRT(1.+SLOP^2) & IF YINT GT 0. THEN R=-R U1=SLOP/R & U2=-1./R & U3=YINT/R YP=U1*X+U2*Y+U3 ; = orthog distance to line YFIT=FLTARR(N) & YFIT[*]=0. SS=YP ENDIF ELSE BEGIN YFIT = YINT+SLOP*X ENDELSE CC=[YINT,SLOP] ISTAT=ROB_CHECKFIT(YP,YFIT,EPS,DEL, SIG,FRACDEV,NGOOD,W,S, $ Bisquare_Limit=Bisquare_Limit ) IF ISTAT EQ 0 THEN GOTO,AFTERFIT IF NGOOD LT 2 THEN BEGIN PRINT,'ROBUST_LINEFIT: Data Dangerously Weird. Fit Questionable.' BADFIT=1 GOTO,AFTERFIT ENDIF DIFF = (ABS(SIG_1-SIG)/SIG) < (ABS(SIG_2-SIG)/SIG) ENDWHILE AFTERFIT: ; Untranslate the coefficients CC[0] = CC[0]+Y0-CC[1]*X0 IF N_PARAMS(0) GT 2 THEN YFIT = CC[0] + CC[1]*XIN IF KEYWORD_SET(BISECT) THEN RETURN,CC IF (N_PARAMS(0) GT 3) AND (SIG GT EPS) AND (NGOOD GT 2) THEN BEGIN ; Here we use an empirical formula to approximate the standard deviations ; of the coefficients. They are usually accurate to ~ 25%. SX2 = TOTAL(W*X*X) UU = S*S DEV = YIN-YFIT Y0 = TOTAL( W*DEV ) Q = WHERE(UU LE 1.0,COUNT) DEN1 = ABS(TOTAL( (1.-UU[Q])*(1.-5.*UU[Q]) )) SIG = ROBUST_SIGMA(DEV,/ZERO) ; Now empirically derived estimates of the uncertainties: SS[0] = SIG/SQRT(DEN1)/1.105 SS[1] = SS[0]/SQRT(SX2) ; Take the X shift into account: SS[0] = SQRT(SS[0]^2+X0*SS[1]^2) ENDIF IF BADFIT EQ 1 THEN CC=[CC,0.] RETURN,CC END