Patsy: Contrast Coding Systems for categorical variables

Note

This document is based heavily on this excellent resource from UCLA.

A categorical variable of K categories, or levels, usually enters a regression as a sequence of K-1 dummy variables. This amounts to a linear hypothesis on the level means. That is, each test statistic for these variables amounts to testing whether the mean for that level is statistically significantly different from the mean of the base category. This dummy coding is called Treatment coding in R parlance, and we will follow this convention. There are, however, different coding methods that amount to different sets of linear hypotheses.

In fact, the dummy coding is not technically a contrast coding. This is because the dummy variables add to one and are not functionally independent of the model’s intercept. On the other hand, a set of contrasts for a categorical variable with k levels is a set of k-1 functionally independent linear combinations of the factor level means that are also independent of the sum of the dummy variables. The dummy coding isn’t wrong per se. It captures all of the coefficients, but it complicates matters when the model assumes independence of the coefficients such as in ANOVA. Linear regression models do not assume independence of the coefficients and thus dummy coding is often the only coding that is taught in this context.

To have a look at the contrast matrices in Patsy, we will use data from UCLA ATS. First let’s load the data.

Example Data

In [1]: import pandas as pd

In [2]: url = 'http://www.ats.ucla.edu/stat/data/hsb2.csv'

In [3]: hsb2 = pd.read_table(url, delimiter=",")

It will be instructive to look at the mean of the dependent variable, write, for each level of race ((1 = Hispanic, 2 = Asian, 3 = African American and 4 = Caucasian)).

Treatment (Dummy) Coding

Dummy coding is likely the most well known coding scheme. It compares each level of the categorical variable to a base reference level. The base reference level is the value of the intercept. It is the default contrast in Patsy for unordered categorical factors. The Treatment contrast matrix for race would be

In [4]: from patsy.contrasts import Treatment

In [5]: levels = [1,2,3,4]

In [6]: contrast = Treatment(reference=0).code_without_intercept(levels)

In [7]: print(contrast.matrix)
[[ 0.  0.  0.]
 [ 1.  0.  0.]
 [ 0.  1.  0.]
 [ 0.  0.  1.]]

Here we used reference=0, which implies that the first level, Hispanic, is the reference category against which the other level effects are measured. As mentioned above, the columns do not sum to zero and are thus not independent of the intercept. To be explicit, let’s look at how this would encode the race variable.

In [8]: contrast.matrix[hsb2.race-1, :][:20]
Out[8]: 
array([[ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  1.,  0.],
       [ 0.,  0.,  0.],
       [ 0.,  0.,  1.],
       [ 0.,  1.,  0.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  1.,  0.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.],
       [ 0.,  0.,  1.]])

This is a bit of a trick, as the race category conveniently maps to zero-based indices. If it does not, this conversion happens under the hood, so this won’t work in general but nonetheless is a useful exercise to fix ideas. The below illustrates the output using the three contrasts above

In [9]: from statsmodels.formula.api import ols

In [10]: mod = ols("write ~ C(race, Treatment)", data=hsb2)

In [11]: res = mod.fit()

In [12]: print(res.summary())
                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  write   R-squared:                       0.107
Model:                            OLS   Adj. R-squared:                  0.093
Method:                 Least Squares   F-statistic:                     7.833
Date:                Mon, 20 Jul 2015   Prob (F-statistic):           5.78e-05
Time:                        17:45:25   Log-Likelihood:                -721.77
No. Observations:                 200   AIC:                             1452.
Df Residuals:                     196   BIC:                             1465.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
===========================================================================================
                              coef    std err          t      P>|t|      [95.0% Conf. Int.]
-------------------------------------------------------------------------------------------
Intercept                  46.4583      1.842     25.218      0.000        42.825    50.091
C(race, Treatment)[T.2]    11.5417      3.286      3.512      0.001         5.061    18.022
C(race, Treatment)[T.3]     1.7417      2.732      0.637      0.525        -3.647     7.131
C(race, Treatment)[T.4]     7.5968      1.989      3.820      0.000         3.675    11.519
==============================================================================
Omnibus:                       10.487   Durbin-Watson:                   1.779
Prob(Omnibus):                  0.005   Jarque-Bera (JB):               11.031
Skew:                          -0.551   Prob(JB):                      0.00402
Kurtosis:                       2.670   Cond. No.                         8.25
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

We explicitly gave the contrast for race; however, since Treatment is the default, we could have omitted this.

Simple Coding

Like Treatment Coding, Simple Coding compares each level to a fixed reference level. However, with simple coding, the intercept is the grand mean of all the levels of the factors. See User-Defined Coding for how to implement the Simple contrast.

In [13]: contrast = Simple().code_without_intercept(levels)

In [14]: print(contrast.matrix)
[[-0.25 -0.25 -0.25]
 [ 0.75 -0.25 -0.25]
 [-0.25  0.75 -0.25]
 [-0.25 -0.25  0.75]]

In [15]: mod = ols("write ~ C(race, Simple)", data=hsb2)

In [16]: res = mod.fit()

In [17]: print(res.summary())
                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  write   R-squared:                       0.107
Model:                            OLS   Adj. R-squared:                  0.093
Method:                 Least Squares   F-statistic:                     7.833
Date:                Mon, 20 Jul 2015   Prob (F-statistic):           5.78e-05
Time:                        17:45:25   Log-Likelihood:                -721.77
No. Observations:                 200   AIC:                             1452.
Df Residuals:                     196   BIC:                             1465.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
===========================================================================================
                              coef    std err          t      P>|t|      [95.0% Conf. Int.]
-------------------------------------------------------------------------------------------
Intercept                  51.6784      0.982     52.619      0.000        49.741    53.615
C(race, Simple)[Simp.1]    11.5417      3.286      3.512      0.001         5.061    18.022
C(race, Simple)[Simp.2]     1.7417      2.732      0.637      0.525        -3.647     7.131
C(race, Simple)[Simp.3]     7.5968      1.989      3.820      0.000         3.675    11.519
==============================================================================
Omnibus:                       10.487   Durbin-Watson:                   1.779
Prob(Omnibus):                  0.005   Jarque-Bera (JB):               11.031
Skew:                          -0.551   Prob(JB):                      0.00402
Kurtosis:                       2.670   Cond. No.                         7.03
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

Sum (Deviation) Coding

Sum coding compares the mean of the dependent variable for a given level to the overall mean of the dependent variable over all the levels. That is, it uses contrasts between each of the first k-1 levels and level k In this example, level 1 is compared to all the others, level 2 to all the others, and level 3 to all the others.

In [18]: from patsy.contrasts import Sum

In [19]: contrast = Sum().code_without_intercept(levels)

In [20]: print(contrast.matrix)
[[ 1.  0.  0.]
 [ 0.  1.  0.]
 [ 0.  0.  1.]
 [-1. -1. -1.]]

In [21]: mod = ols("write ~ C(race, Sum)", data=hsb2)

In [22]: res = mod.fit()

In [23]: print(res.summary())
                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  write   R-squared:                       0.107
Model:                            OLS   Adj. R-squared:                  0.093
Method:                 Least Squares   F-statistic:                     7.833
Date:                Mon, 20 Jul 2015   Prob (F-statistic):           5.78e-05
Time:                        17:45:25   Log-Likelihood:                -721.77
No. Observations:                 200   AIC:                             1452.
Df Residuals:                     196   BIC:                             1465.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
=====================================================================================
                        coef    std err          t      P>|t|      [95.0% Conf. Int.]
-------------------------------------------------------------------------------------
Intercept            51.6784      0.982     52.619      0.000        49.741    53.615
C(race, Sum)[S.1]    -5.2200      1.631     -3.200      0.002        -8.437    -2.003
C(race, Sum)[S.2]     6.3216      2.160      2.926      0.004         2.061    10.582
C(race, Sum)[S.3]    -3.4784      1.732     -2.008      0.046        -6.895    -0.062
==============================================================================
Omnibus:                       10.487   Durbin-Watson:                   1.779
Prob(Omnibus):                  0.005   Jarque-Bera (JB):               11.031
Skew:                          -0.551   Prob(JB):                      0.00402
Kurtosis:                       2.670   Cond. No.                         6.72
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

This correspons to a parameterization that forces all the coefficients to sum to zero. Notice that the intercept here is the grand mean where the grand mean is the mean of means of the dependent variable by each level.

In [24]: hsb2.groupby('race')['write'].mean().mean()
Out[24]: 51.67837643678162

Backward Difference Coding

In backward difference coding, the mean of the dependent variable for a level is compared with the mean of the dependent variable for the prior level. This type of coding may be useful for a nominal or an ordinal variable.

In [25]: from patsy.contrasts import Diff

In [26]: contrast = Diff().code_without_intercept(levels)

In [27]: print(contrast.matrix)
[[-0.75 -0.5  -0.25]
 [ 0.25 -0.5  -0.25]
 [ 0.25  0.5  -0.25]
 [ 0.25  0.5   0.75]]

In [28]: mod = ols("write ~ C(race, Diff)", data=hsb2)

In [29]: res = mod.fit()

In [30]: print(res.summary())
                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  write   R-squared:                       0.107
Model:                            OLS   Adj. R-squared:                  0.093
Method:                 Least Squares   F-statistic:                     7.833
Date:                Mon, 20 Jul 2015   Prob (F-statistic):           5.78e-05
Time:                        17:45:25   Log-Likelihood:                -721.77
No. Observations:                 200   AIC:                             1452.
Df Residuals:                     196   BIC:                             1465.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
======================================================================================
                         coef    std err          t      P>|t|      [95.0% Conf. Int.]
--------------------------------------------------------------------------------------
Intercept             51.6784      0.982     52.619      0.000        49.741    53.615
C(race, Diff)[D.1]    11.5417      3.286      3.512      0.001         5.061    18.022
C(race, Diff)[D.2]    -9.8000      3.388     -2.893      0.004       -16.481    -3.119
C(race, Diff)[D.3]     5.8552      2.153      2.720      0.007         1.610    10.101
==============================================================================
Omnibus:                       10.487   Durbin-Watson:                   1.779
Prob(Omnibus):                  0.005   Jarque-Bera (JB):               11.031
Skew:                          -0.551   Prob(JB):                      0.00402
Kurtosis:                       2.670   Cond. No.                         8.30
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

For example, here the coefficient on level 1 is the mean of write at level 2 compared with the mean at level 1. Ie.,

In [31]: res.params["C(race, Diff)[D.1]"]
Out[31]: 11.541666666666577

In [32]: hsb2.groupby('race').mean()["write"][2] - \
   ....:      hsb2.groupby('race').mean()["write"][1]
   ....: 
Out[32]: 11.541666666666664

Helmert Coding

Our version of Helmert coding is sometimes referred to as Reverse Helmert Coding. The mean of the dependent variable for a level is compared to the mean of the dependent variable over all previous levels. Hence, the name ‘reverse’ being sometimes applied to differentiate from forward Helmert coding. This comparison does not make much sense for a nominal variable such as race, but we would use the Helmert contrast like so:

In [33]: from patsy.contrasts import Helmert

In [34]: contrast = Helmert().code_without_intercept(levels)

In [35]: print(contrast.matrix)
[[-1. -1. -1.]
 [ 1. -1. -1.]
 [ 0.  2. -1.]
 [ 0.  0.  3.]]

In [36]: mod = ols("write ~ C(race, Helmert)", data=hsb2)

In [37]: res = mod.fit()

In [38]: print(res.summary())
                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  write   R-squared:                       0.107
Model:                            OLS   Adj. R-squared:                  0.093
Method:                 Least Squares   F-statistic:                     7.833
Date:                Mon, 20 Jul 2015   Prob (F-statistic):           5.78e-05
Time:                        17:45:25   Log-Likelihood:                -721.77
No. Observations:                 200   AIC:                             1452.
Df Residuals:                     196   BIC:                             1465.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
=========================================================================================
                            coef    std err          t      P>|t|      [95.0% Conf. Int.]
-----------------------------------------------------------------------------------------
Intercept                51.6784      0.982     52.619      0.000        49.741    53.615
C(race, Helmert)[H.2]     5.7708      1.643      3.512      0.001         2.530     9.011
C(race, Helmert)[H.3]    -1.3431      0.867     -1.548      0.123        -3.054     0.368
C(race, Helmert)[H.4]     0.7923      0.372      2.130      0.034         0.059     1.526
==============================================================================
Omnibus:                       10.487   Durbin-Watson:                   1.779
Prob(Omnibus):                  0.005   Jarque-Bera (JB):               11.031
Skew:                          -0.551   Prob(JB):                      0.00402
Kurtosis:                       2.670   Cond. No.                         7.26
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

To illustrate, the comparison on level 4 is the mean of the dependent variable at the previous three levels taken from the mean at level 4

In [39]: grouped = hsb2.groupby('race')

In [40]: grouped.mean()["write"][4] - grouped.mean()["write"][:3].mean()
Out[40]: 3.1690613026819818

As you can see, these are only equal up to a constant. Other versions of the Helmert contrast give the actual difference in means. Regardless, the hypothesis tests are the same.

In [41]: k = 4

In [42]: 1./k * (grouped.mean()["write"][k] - grouped.mean()["write"][:k-1].mean())
Out[42]: 0.79226532567049546

In [43]: k = 3

In [44]: 1./k * (grouped.mean()["write"][k] - grouped.mean()["write"][:k-1].mean())
Out[44]: -1.3430555555555561

Orthogonal Polynomial Coding

The coefficients taken on by polynomial coding for k=4 levels are the linear, quadratic, and cubic trends in the categorical variable. The categorical variable here is assumed to be represented by an underlying, equally spaced numeric variable. Therefore, this type of encoding is used only for ordered categorical variables with equal spacing. In general, the polynomial contrast produces polynomials of order k-1. Since race is not an ordered factor variable let’s use read as an example. First we need to create an ordered categorical from read.

In [45]: _, bins = np.histogram(hsb2.read, 3)

In [46]: try: # requires numpy master
   ....:     readcat = np.digitize(hsb2.read, bins, True)
   ....: except:
   ....:     readcat = np.digitize(hsb2.read, bins)
   ....: 

In [47]: hsb2['readcat'] = readcat

In [48]: hsb2.groupby('readcat').mean()['write']
Out[48]: 
readcat
0    46.000000
1    44.980392
2    53.356436
3    60.127660
Name: write, dtype: float64
In [49]: from patsy.contrasts import Poly

In [50]: levels = hsb2.readcat.unique().tolist()

In [51]: contrast = Poly().code_without_intercept(levels)

In [52]: print(contrast.matrix)
[[-0.6708  0.5    -0.2236]
 [-0.2236 -0.5     0.6708]
 [ 0.2236 -0.5    -0.6708]
 [ 0.6708  0.5     0.2236]]

In [53]: mod = ols("write ~ C(readcat, Poly)", data=hsb2)

In [54]: res = mod.fit()

In [55]: print(res.summary())
                            OLS Regression Results                            
==============================================================================
Dep. Variable:                  write   R-squared:                       0.320
Model:                            OLS   Adj. R-squared:                  0.309
Method:                 Least Squares   F-statistic:                     30.73
Date:                Mon, 20 Jul 2015   Prob (F-statistic):           2.51e-16
Time:                        17:45:25   Log-Likelihood:                -694.54
No. Observations:                 200   AIC:                             1397.
Df Residuals:                     196   BIC:                             1410.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
==============================================================================================
                                 coef    std err          t      P>|t|      [95.0% Conf. Int.]
----------------------------------------------------------------------------------------------
Intercept                     51.1161      2.018     25.324      0.000        47.135    55.097
C(readcat, Poly).Linear       11.3501      5.348      2.122      0.035         0.803    21.897
C(readcat, Poly).Quadratic     3.8954      4.037      0.965      0.336        -4.066    11.857
C(readcat, Poly).Cubic        -2.4598      1.998     -1.231      0.220        -6.400     1.480
==============================================================================
Omnibus:                        9.741   Durbin-Watson:                   1.699
Prob(Omnibus):                  0.008   Jarque-Bera (JB):               10.263
Skew:                          -0.535   Prob(JB):                      0.00591
Kurtosis:                       2.703   Cond. No.                         13.7
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

As you can see, readcat has a significant linear effect on the dependent variable write but not a significant quadratic or cubic effect.

User-Defined Coding

If you want to use your own coding, you must do so by writing a coding class that contains a code_with_intercept and a code_without_intercept method that return a patsy.contrast.ContrastMatrix instance.

In [56]: from patsy.contrasts import ContrastMatrix

In [57]: def _name_levels(prefix, levels):
   ....:     return ["[%s%s]" % (prefix, level) for level in levels]
   ....: 

In [58]: class Simple(object):
   ....:     def _simple_contrast(self, levels):
   ....:         nlevels = len(levels)
   ....:         contr = -1./nlevels * np.ones((nlevels, nlevels-1))
   ....:         contr[1:][np.diag_indices(nlevels-1)] = (nlevels-1.)/nlevels
   ....:         return contr
   ....: 

In [59]: def code_with_intercept(self, levels):
   ....:         contrast = np.column_stack((np.ones(len(levels)),
   ....:                                     self._simple_contrast(levels)))
   ....:         return ContrastMatrix(contrast, _name_levels("Simp.", levels))
   ....: