Example 9.11 Regression Quantiles

The technique of estimating parameters in linear models by using the notion of regression quantiles is a generalization of the LAE or LAV least absolute value estimation technique. For a given quantile $q$, the estimate $\mb {b}^{*}$ of $\mb {\beta }$ in the model

\[  \mb {Y}=\mb {X}\mb {\beta } + \mb {\epsilon }  \]

is the value of $b$ that minimizes

\[  \sum _{t \in T} q|y_ t - x_ t b| - \sum _{t \in S} (1-q) |y_ t - x_ t b|  \]

where $T=\{ t|y_ t \geq x_ t b\} $ and $S=\{ t|y_ t \leq x_ t\} $. For $q=0.5$, the solution $\mb {b}^{*}$ is identical to the estimates produced by the LAE. The following routine finds this estimate by using linear programming.

   /* Routine to find regression quantiles                    */
   /* yname:    name of dependent variable                    */
   /* y:        dependent variable                            */
   /* xname:    names of independent variables                */
   /* X:        independent variables                         */
   /* b:        estimates                                     */
   /* predict:  predicted values                              */
   /* error:    difference of y and predicted.                */
   /* q:        quantile                                      */
   /*                                                         */
   /* notes:    This subroutine finds the estimates b         */
   /* that minimize                                           */
   /*                                                         */
   /*           q * (y - Xb) * e  + (1-q) * (y - Xb) * ^e     */
   /*                                                         */
   /* where e = ( Xb <= y ).                                  */
   /*                                                         */
   /* This subroutine follows the approach given in:          */
   /*                                                         */
   /* Koenker, R. and G. Bassett (1978). Regression           */
   /* quantiles. Econometrica. Vol. 46. No. 1. 33-50.         */
   /*                                                         */
   /* Basssett, G. and R. Koenker (1982). An empirical        */
   /* quantile function for linear models with iid errors.    */
   /* JASA. Vol. 77. No. 378. 407-415.                        */
   /*                                                         */
   /* When q = .5 this is equivalent to minimizing the sum    */
   /* of the absolute deviations, which is also known as      */
   /* L1 regression.  Note that for L1 regression, a faster   */
   /* and more accurate algorithm is available in the SAS/IML */
   /* routine LAV, which is based on the approach given in:   */
   /*                                                         */
   /* Madsen, K. and Nielsen, H. B. (1993). A finite          */
   /* smoothing algorithm for linear L1 estimation.           */
   /* SIAM J. Optimization, Vol. 3. 223-235.                  */
   /*---------------------------------------------------------*/
   start rq( yname, y, xname, X, b, predict, error, q);
      bound=1.0e10;
      coef = X`;
      m = nrow(coef);
      n = ncol(coef);
   /*-----------------build rhs and bounds--------------------*/
      e  = repeat(1,1,n)`;
      r  =  {0 0} || ((1-q)*coef*e)`;
      sign = repeat(1,1,m);

      do i=1 to m;
         if r[2+i] < 0 then do;
            sign[i]   = -1;
            r[2+i]    = -r[2+i];
           coef[i,]  = -coef[i,];
         end;
      end;

   l  = repeat(0,1,n) || repeat(0,1,m)  || -bound || -bound ;
   u  = repeat(1,1,n) || repeat(.,1,m)  ||   { . . }  ;

   /*-------------build coefficient matrix and basis----------*/
    a  = (        y`     || repeat(0,1,m)  ||   { -1 0 }   ) //
         ( repeat(0,1,n) || repeat(-1,1,m) ||   { 0 -1 }   ) //
         (      coef     ||     I(m)       || repeat(0,m,2)) ;
    basis = n+m+2 - (0:n+m+1);

   /*----------------find a feasible solution-----------------*/
   call lp(rc,p,d,a,r,,u,l,basis);

   /*----------------find the optimal solution----------------*/
   l  = repeat(0,1,n) || repeat(0,1,m)  || -bound || {0} ;
   u  = repeat(1,1,n) || repeat(0,1,m)  ||   { . 0 }  ;
   call lp(rc,p,d,a,r,n+m+1,u,l,basis);

   /*---------------- report the solution-----------------------*/
   variable = xname`; b=d[3:m+2];
   do i=1 to m;
      b[i] =  b[i] * sign[i];
   end;
   predict = X*b;
   error = y - predict;
   wsum  = sum ( choose(error<0 , (q-1)*error , q*error) );

   print ,,'Regression Quantile Estimation' ,
           'Dependent Variable: ' yname ,
           'Regression Quantile: ' q ,
           'Number of Observations: ' n ,
           'Sum of Weighted Absolute Errors: ' wsum ,
            variable b,
            X y predict error;
   finish rq;

The following example uses data on the United States population from 1790 to 1970:

   z = { 3.929 1790 ,
         5.308 1800 ,
         7.239 1810 ,
         9.638 1820 ,
        12.866 1830 ,
        17.069 1840 ,
        23.191 1850 ,
        31.443 1860 ,
        39.818 1870 ,
        50.155 1880 ,
        62.947 1890 ,
        75.994 1900 ,
        91.972 1910 ,
       105.710 1920 ,
       122.775 1930 ,
       131.669 1940 ,
       151.325 1950 ,
       179.323 1960 ,
       203.211 1970 };

   y=z[,1];
   x=repeat(1,19,1)||z[,2]||z[,2]##2;
   run rq('pop',y,{'intercpt' 'year' 'yearsq'},x,b1,pred,resid,.5);

The results are shown in Output 9.11.1.

Output 9.11.1: Regression Quantiles: Results

Regression Quantile Estimation

  yname
Dependent Variable: pop

  q
Regression Quantile: 0.5

  n
Number of Observations: 19

  wsum
Sum of Weighted Absolute Errors: 14.826429

variable b
intercpt 21132.758
year -23.52574
yearsq 0.006549

X   y predict error
1 1790 3204100 3.929 5.4549176 -1.525918
1 1800 3240000 5.308 5.308 -4.54E-12
1 1810 3276100 7.239 6.4708902 0.7681098
1 1820 3312400 9.638 8.9435882 0.6944118
1 1830 3348900 12.866 12.726094 0.1399059
1 1840 3385600 17.069 17.818408 -0.749408
1 1850 3422500 23.191 24.220529 -1.029529
1 1860 3459600 31.443 31.932459 -0.489459
1 1870 3496900 39.818 40.954196 -1.136196
1 1880 3534400 50.155 51.285741 -1.130741
1 1890 3572100 62.947 62.927094 0.0199059
1 1900 3610000 75.994 75.878255 0.1157451
1 1910 3648100 91.972 90.139224 1.8327765
1 1920 3686400 105.71 105.71 8.669E-13
1 1930 3724900 122.775 122.59058 0.1844157
1 1940 3763600 131.669 140.78098 -9.111976
1 1950 3802500 151.325 160.28118 -8.956176
1 1960 3841600 179.323 181.09118 -1.768184
1 1970 3880900 203.211 203.211 -2.96E-12


The L1 norm (when $q=0.5$) tends to cause the fit to be better at more points at the expense of causing some points to fit worse, as shown by the following plot, which compares the L1 residuals with the least squares residuals.

Output 9.11.2: L1 Residuals vs. Least Squares Residuals

L1 Residuals vs. Least Squares Residuals


When $q=0.5$, the results of this module can be compared with the results of the LAV routine, as follows:

   b0 = {1 1 1};              /* initial value             */
   optn = j(4,1,.);           /* options vector            */

   optn[1]= .;                /* gamma default             */
   optn[2]= 5;                /* print all                 */
   optn[3]= 0;                /* McKean-Schradar variance  */
   optn[4]= 1;                /* convergence test          */

   call LAV(rc, xr, x, y, b0, optn);