## Basic Quantile Regression

By Gabriel Vasconcelos

## Introduction

Today we are going to talk about quantile regression. When we use the lm command in R we are fitting a linear regression using Ordinary Least Squares (OLS), which has the interpretation of a model for the conditional mean of $y$ on $x$. However, sometimes we may need to look at more than the conditional mean to understand our data and quantile regressions may be a good alternative. Instead of looking at the mean, quantile regressions will establish models for particular quantiles as chosen by the user. The most simple case when quantile regressions are good is when you have outliers in your data because the median is much less affected by extreme values than the mean (0.5 quantile). But there are other cases where quantile regression may be used, for example to identify some heterogeneous effects of some variable or even to give more robustness to your results.

## Outliers

The package we will be using for quantile regressions is the quantreg, which is very easy to use if you are already familiar with the lm function.

library(quantreg)


Our data is going to be very simple. The sample size will be 300 and we will have 10 variables. The $\beta$s will be either 1 or -1 depending on the index of the variable. This is our dgp:

$\displaystyle y_i = -x_{1,i}+x_{2,i}-x_{3,i}+\dots+x_{10,i} + \varepsilon_i$

where $x_i$ and $\varepsilon_i$ are all generated from a $N(0,1)$ distribution.

The first step is to generate the data and run a linear regression by OLS. As you can see in the results below, the estimates are very precise. The signals are always correct and the coefficients are close to 1 or -1.

N = 300 #sample size
P = 10 # number of variable

### Generate Data ###
betas = (-1)^(1:P)

set.seed(1) # Seed for replication
error = rnorm(N)
X = matrix(rnorm(N*P), N, P)
y = X%*%betas + error

summary(lm(y~X))

##
## Call:
## lm(formula = y ~ X)
##
## Residuals:
##      Min       1Q   Median       3Q      Max
## -2.94301 -0.58806 -0.03508  0.61445  2.55389
##
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)  0.03501    0.05739    0.61    0.542
## X1          -0.98513    0.05534  -17.80   <2e-16 ***
## X2           1.04319    0.05321   19.60   <2e-16 ***
## X3          -1.01520    0.05577  -18.20   <2e-16 ***
## X4           0.92930    0.05705   16.29   <2e-16 ***
## X5          -0.97971    0.05404  -18.13   <2e-16 ***
## X6           1.01255    0.05376   18.83   <2e-16 ***
## X7          -1.02357    0.05622  -18.21   <2e-16 ***
## X8           0.97456    0.05946   16.39   <2e-16 ***
## X9          -0.97097    0.05295  -18.34   <2e-16 ***
## X10          1.03386    0.05312   19.46   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.9747 on 289 degrees of freedom
## Multiple R-squared:  0.9274, Adjusted R-squared:  0.9249
## F-statistic:   369 on 10 and 289 DF,  p-value: < 2.2e-16


Now we are going to add some outliers to the data. We have 300 observations and only 5 are going to be outliers. The code below introduces the outliers and run OLS again. Estimates are now considerably worst than the first case with no outliers.

### Introducing outliers ###
set.seed(1)
outliers = rnorm(5, 0, 50)
error_o = error
error_o[1:5] = outliers

y_o = X%*%betas + error_o

## OLS ###
summary(lm(y_o~X))

##
## Call:
## lm(formula = y_o ~ X)
##
## Residuals:
##     Min      1Q  Median      3Q     Max
## -41.216  -1.105  -0.027   0.828  77.040
##
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)   0.1951     0.3367   0.579 0.562717
## X1           -1.3165     0.3247  -4.055 6.45e-05 ***
## X2            1.1355     0.3122   3.637 0.000326 ***
## X3           -1.0541     0.3272  -3.222 0.001421 **
## X4            0.9136     0.3347   2.729 0.006733 **
## X5           -1.2849     0.3171  -4.052 6.52e-05 ***
## X6            1.2834     0.3154   4.069 6.10e-05 ***
## X7           -0.7777     0.3298  -2.358 0.019055 *
## X8            1.4859     0.3488   4.260 2.77e-05 ***
## X9           -1.0915     0.3107  -3.513 0.000513 ***
## X10           0.5756     0.3117   1.847 0.065797 .
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 5.719 on 289 degrees of freedom
## Multiple R-squared:  0.3223, Adjusted R-squared:  0.2989
## F-statistic: 13.75 on 10 and 289 DF,  p-value: < 2.2e-16


If we run a quantile regression for the median like in the code below we can get good results once again. Note that in the OLS case the coefficients were not as close to 1 or -1 as in the quantile case below.

## Quantile Regression - Median ##
summary(rq(y_o ~ X, tau = 0.5), se = "boot", bsmethod= "xy")

##
## Call: rq(formula = y_o ~ X, tau = 0.5)
##
## tau: [1] 0.5
##
## Coefficients:
##             Value     Std. Error t value   Pr(>|t|)
## (Intercept)  -0.03826   0.08202   -0.46642   0.64127
## X1           -0.95730   0.07394  -12.94750   0.00000
## X2            0.99394   0.07612   13.05731   0.00000
## X3           -0.97981   0.07995  -12.25523   0.00000
## X4            0.92966   0.09319    9.97573   0.00000
## X5           -0.92781   0.08520  -10.88998   0.00000
## X6            0.98021   0.08114   12.08030   0.00000
## X7           -1.00993   0.06727  -15.01266   0.00000
## X8            1.03572   0.08043   12.87722   0.00000
## X9           -0.96732   0.07502  -12.89374   0.00000
## X10           1.02600   0.08775   11.69171   0.00000


## Treatment effects

Suppose now that we have the same model, without outliers, plus a treatment that is positive for lower values of $y$ and negative for bigger values of $y$. Only half of the sample will be treated and we want to see what we get when we try to estimate the effects of this treatment. The figure below plots the original $y$ against the treated $y$. The points in the 45 degrees line are the untreated observations.

Our treatment was simple adding 5 to the equation if the deterministic part of the equation is negative and subtracting 5 if it is positive. If we run OLS we can se below that the treatment effects are very poorly estimated and they are also not significant. That is because we are looking at the average effect of the treatment, which is 0 because half of the treated sample had an increase of 5 and the other half had a decrease of five, which is 0 on average.

#### More complicated data #####
set.seed(1)
treatment = sample(c(0,1),N,replace = TRUE)
aux = X%*%betas
treatment_error = rnorm(N)*treatment

y_tr = aux + ifelse(aux0,-5,0)*treatment + error + treatment_error
X_tr = cbind(treatment,X)

## Treated X Untreated
plot(y,y_tr)


Now lets try quantile regression for multiple quantiles (0.1 ,0.2 ,…,0.8, 0.9). The results are presented below. When we look at the middle quantiles like 0.5 and 0.6 we find that the treatment is not significant just like in the OLS case. However, as we move further to 0.1 or 0.9 we obtain significant results with estimates very close to the real treatment, which would be 5 or -5. The model is telling us that bigger values of $y$ are either untreated samples or treated samples that were previously small but became big because of the treatment of 5, that is why the treatment effect is close to 5 in the 0.9 quantile. Similarly, smaller values of $y$ are either untreated samples or treated samples that were previously big and became small because of the treatment of -5. We are modeling the quantiles of the treated $y$.

## OLS ##
summary(lm(y_tr ~ X_tr))

##
## Call:
## lm(formula = y_tr ~ X_tr)
##
## Residuals:
##     Min      1Q  Median      3Q     Max
## -7.5913 -2.0909  0.0168  1.8840  8.8915
##
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)
## (Intercept)   -0.08713    0.24512  -0.355 0.722502
## X_trtreatment -0.24555    0.35698  -0.688 0.492093
## X_tr          -0.40117    0.17283  -2.321 0.020973 *
## X_tr           0.55826    0.16594   3.364 0.000872 ***
## X_tr          -0.61606    0.17365  -3.548 0.000454 ***
## X_tr           0.56658    0.17766   3.189 0.001584 **
## X_tr          -0.37539    0.16829  -2.231 0.026477 *
## X_tr           0.43680    0.16788   2.602 0.009752 **
## X_tr          -0.29718    0.17523  -1.696 0.090972 .
## X_tr           0.39646    0.18585   2.133 0.033751 *
## X_tr          -0.34905    0.16499  -2.116 0.035233 *
## X_tr           0.83088    0.16593   5.007 9.64e-07 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 3.035 on 288 degrees of freedom
## Multiple R-squared:  0.2454, Adjusted R-squared:  0.2166
## F-statistic: 8.514 on 11 and 288 DF,  p-value: 5.243e-13

# Quantilie Regression for q = 0.1, 0.2, ... , 0.8, 0.9
summary(rq(y_tr ~ X_tr, tau = seq(0.1,0.9,0.1)), se = "boot", bsmethod= "xy")

##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.1
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)   -1.59418  0.22484   -7.09019  0.00000
## X_trtreatment -4.13074  0.47280   -8.73680  0.00000
## X_tr          -0.85152  0.17088   -4.98320  0.00000
## X_tr           1.10326  0.14030    7.86384  0.00000
## X_tr          -0.85055  0.18554   -4.58421  0.00001
## X_tr           0.68881  0.21612    3.18719  0.00159
## X_tr          -0.88047  0.16396   -5.36995  0.00000
## X_tr           0.68033  0.15287    4.45051  0.00001
## X_tr          -1.00234  0.15569   -6.43819  0.00000
## X_tr           0.99443  0.19484    5.10369  0.00000
## X_tr          -0.94634  0.17999   -5.25759  0.00000
## X_tr           0.92124  0.17348    5.31048  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.2
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)   -0.87790  0.18851   -4.65701  0.00000
## X_trtreatment -4.04359  0.43354   -9.32695  0.00000
## X_tr          -0.87484  0.17336   -5.04648  0.00000
## X_tr           0.90953  0.16734    5.43510  0.00000
## X_tr          -0.92386  0.15931   -5.79905  0.00000
## X_tr           0.84438  0.14318    5.89740  0.00000
## X_tr          -0.79882  0.16353   -4.88478  0.00000
## X_tr           0.86351  0.14477    5.96469  0.00000
## X_tr          -0.89046  0.14533   -6.12699  0.00000
## X_tr           0.80481  0.16133    4.98842  0.00000
## X_tr          -0.85986  0.14983   -5.73877  0.00000
## X_tr           1.05532  0.14247    7.40725  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.3
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)   -0.48650  0.15254   -3.18941  0.00158
## X_trtreatment -3.55483  0.39081   -9.09596  0.00000
## X_tr          -0.81732  0.15516   -5.26750  0.00000
## X_tr           0.97204  0.15176    6.40526  0.00000
## X_tr          -0.88749  0.12320   -7.20353  0.00000
## X_tr           0.80248  0.13892    5.77649  0.00000
## X_tr          -0.63208  0.15928   -3.96825  0.00009
## X_tr           0.85960  0.12409    6.92695  0.00000
## X_tr          -0.73083  0.13347   -5.47540  0.00000
## X_tr           0.80175  0.14581    5.49859  0.00000
## X_tr          -0.73295  0.14156   -5.17769  0.00000
## X_tr           0.96849  0.14074    6.88120  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.4
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)   -0.20518  0.17305   -1.18566  0.23673
## X_trtreatment -3.12524  0.82009   -3.81083  0.00017
## X_tr          -0.75969  0.17512   -4.33802  0.00002
## X_tr           0.75137  0.18705    4.01697  0.00008
## X_tr          -0.89673  0.15018   -5.97081  0.00000
## X_tr           0.81794  0.15558    5.25741  0.00000
## X_tr          -0.60392  0.18754   -3.22019  0.00143
## X_tr           0.76221  0.16089    4.73733  0.00000
## X_tr          -0.63207  0.24626   -2.56665  0.01077
## X_tr           0.70476  0.24297    2.90064  0.00401
## X_tr          -0.72050  0.21575   -3.33951  0.00095
## X_tr           0.97243  0.14277    6.81125  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.5
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)    0.01435  0.17259    0.08317  0.93377
## X_trtreatment -0.63174  0.97917   -0.64518  0.51933
## X_tr          -0.60768  0.19231   -3.15992  0.00175
## X_tr           0.54861  0.19838    2.76543  0.00605
## X_tr          -0.80869  0.16209   -4.98908  0.00000
## X_tr           0.73479  0.18775    3.91377  0.00011
## X_tr          -0.41427  0.21412   -1.93477  0.05400
## X_tr           0.64703  0.21863    2.95946  0.00334
## X_tr          -0.24361  0.22809   -1.06801  0.28641
## X_tr           0.37166  0.23587    1.57566  0.11620
## X_tr          -0.45650  0.19626   -2.32605  0.02071
## X_tr           0.92603  0.17611    5.25825  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.6
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)    0.25546  0.18342    1.39270  0.16478
## X_trtreatment  1.36457  1.24271    1.09806  0.27310
## X_tr          -0.60693  0.21221   -2.86008  0.00455
## X_tr           0.64917  0.21300    3.04782  0.00252
## X_tr          -0.83220  0.17072   -4.87459  0.00000
## X_tr           0.73617  0.18348    4.01219  0.00008
## X_tr          -0.45447  0.19288   -2.35628  0.01913
## X_tr           0.76270  0.21047    3.62385  0.00034
## X_tr          -0.45186  0.25197   -1.79332  0.07397
## X_tr           0.46886  0.22511    2.08276  0.03816
## X_tr          -0.56117  0.22729   -2.46900  0.01413
## X_tr           0.94549  0.19311    4.89600  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.7
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)    0.52520  0.14537    3.61291  0.00036
## X_trtreatment  3.30124  0.61742    5.34682  0.00000
## X_tr          -0.67172  0.15957   -4.20960  0.00003
## X_tr           0.74401  0.15205    4.89331  0.00000
## X_tr          -0.93090  0.15068   -6.17788  0.00000
## X_tr           0.78792  0.15054    5.23389  0.00000
## X_tr          -0.72815  0.13243   -5.49819  0.00000
## X_tr           0.84332  0.14158    5.95641  0.00000
## X_tr          -0.71329  0.16296   -4.37707  0.00002
## X_tr           0.59229  0.17063    3.47111  0.00060
## X_tr          -0.76839  0.16055   -4.78593  0.00000
## X_tr           0.98971  0.15318    6.46091  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.8
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)    0.79126  0.14003    5.65083  0.00000
## X_trtreatment  3.70886  0.36971   10.03188  0.00000
## X_tr          -0.73646  0.14104   -5.22148  0.00000
## X_tr           0.71236  0.12350    5.76817  0.00000
## X_tr          -0.86668  0.13143   -6.59424  0.00000
## X_tr           0.75392  0.13379    5.63521  0.00000
## X_tr          -0.76165  0.10571   -7.20496  0.00000
## X_tr           0.86711  0.11852    7.31598  0.00000
## X_tr          -0.73505  0.13053   -5.63128  0.00000
## X_tr           0.80531  0.13801    5.83513  0.00000
## X_tr          -0.79344  0.12981   -6.11250  0.00000
## X_tr           1.00810  0.13979    7.21169  0.00000
##
## Call: rq(formula = y_tr ~ X_tr, tau = seq(0.1, 0.9, 0.1))
##
## tau: [1] 0.9
##
## Coefficients:
##               Value    Std. Error t value  Pr(>|t|)
## (Intercept)    1.16385  0.16462    7.07006  0.00000
## X_trtreatment  4.08872  0.37080   11.02661  0.00000
## X_tr          -0.93217  0.15900   -5.86256  0.00000
## X_tr           0.82984  0.12705    6.53140  0.00000
## X_tr          -0.93147  0.13285   -7.01143  0.00000
## X_tr           0.70112  0.12903    5.43359  0.00000
## X_tr          -0.90992  0.13283   -6.85018  0.00000
## X_tr           0.84127  0.14036    5.99349  0.00000
## X_tr          -0.66466  0.15369   -4.32462  0.00002
## X_tr           0.84578  0.17838    4.74134  0.00000
## X_tr          -0.84474  0.13833   -6.10682  0.00000
## X_tr           0.98760  0.16193    6.09902  0.00000

This entry was posted in R and tagged , , , , , . Bookmark the permalink.

### 4 Responses to Basic Quantile Regression

1. Cody Custis says:

“If we run OLS we can se below that the treatment effects are very poorly estimated and they are also not significant. That is because we are looking at the average effect of the treatment, which is 0 because half of the treated sample had an increase of 5 and the other half had a decrease of five, which is 0 on average.”

The treatment effects shouldn’t be significant in linear regression, since the average effect of the treatment is, as you state, zero. You’ve created a non-linear effect, linear models shouldn’t be able to estimate it well. The quantile methods estimate an effect because you’ve changed the shape of the distribution and find heavy effects in the tails.

Like

2. Pingback: Distilled News | AnalytiXon