# How to compare two nested linear models (in Python, using statsmodels)

See all solutions.

Model $A$ is said to be “nested” in model $B$ if the predictors included in $A$ are a subset of those included in $B$. In such a situation, how can we determine if the larger model (in this case $B$) is significantly better than the smaller (reduced) model? We can use an Extra Sums of Squares test, also called a partial $F$-test, to compare two nested linear.

This technique will also help us with another question. If we have a multivarate linear model,

$\hat{y}=\beta_0 + \beta_1x_1 + \beta_2x_2 + \cdots + \beta_kx_k,$

how can we test the influence of only some of the coefficients? If we remove some of the coefficients, we have a smaller model nested in the larger one, so the question is the same.

## Solution

The solution below uses an example dataset about car design and fuel consumption from a 1974 Motor Trend magazine. (See how to quickly load some sample data.)

We will create two models, one nested inside the other, in a natural way in this example. But this is not the only way to create nested models; it is just an example.

1
2
from rdatasets import data
df = data('mtcars')


Consider a model using number of cylinders (cyl) and weight of car (wt) to predict its fuel efficiency (mpg). We create this model and perform an ANOVA to see if the predictors are significant. We use the Ordinary Least Squares module from statsmodels.

1
2
3
4
5
from statsmodels.formula.api import ols
add_model = ols('mpg ~ cyl + wt', data = df).fit()

import statsmodels.api as sm

df sum_sq mean_sq F PR(>F)
cyl 1.0 817.712952 817.712952 124.043687 5.424327e-12
wt 1.0 117.162269 117.162269 17.773034 2.220200e-04
Residual 29.0 191.171966 6.592137 NaN NaN

In the final column of output we see that all numbers are below $0.05$, which suggests that both predictors are significant. A natural question to ask is whether the two predictors have an interaction effect. Let’s create a model containing the interaction term.

1
2
int_model = ols('mpg ~ cyl*wt', data = df).fit()
sm.stats.anova_lm(int_model, typ= 1)

df sum_sq mean_sq F PR(>F)
cyl 1.0 817.712952 817.712952 145.856269 1.280635e-12
wt 1.0 117.162269 117.162269 20.898350 8.942713e-05
cyl:wt 1.0 34.195767 34.195767 6.099533 1.988242e-02
Residual 28.0 156.976199 5.606293 NaN NaN

As seen in the final column of output, there is a significant interaction between the two predictors (bottom number being below $0.05$).

We now have one model (add_model) nested inside a larger model (int_model). To check which model is better, we can conduct an ANOVA comparing the two models. We use the anova_lm function from statsmodels.

1
2
from statsmodels.stats.anova import anova_lm

df_resid ssr df_diff ss_diff F Pr(>F)
0 29.0 191.171966 0.0 NaN NaN NaN
1 28.0 156.976199 1.0 34.195767 6.099533 0.019882

We have just performed this hypothesis test:

$H_0 =$ the two models are equally useful for predicting the outcome

$H_a =$ the larger model is significantly better than the smaller model

In the final column of the output, called Pr(>F), the only number in that column is our test statistic, $0.019882$. Since is below our chosen threshold of $0.05$, we reject the null hypothesis, and prefer to use the second model.

This method can be used to check if covariates should be included in the model, or if additional variables should be added as well.