Regressions¶

This notebook demonstrates various regressions, and accompanies the videos in Week 8.

Setup¶

We're going to import our modules. Of particular note is statsmodels.formula.api, which gives us the regression APIs.

In [1]:

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import scipy.stats as sps
import statsmodels.api as sm
import statsmodels.formula.api as smf

We're going to use the Penguins data:

In [2]:

penguins = pd.read_csv('penguins.csv')

In [3]:

penguins.head()

Out[3]:

	species	island	bill_length_mm	bill_depth_mm	flipper_length_mm	body_mass_g	sex	year
0	Adelie	Torgersen	39.1	18.7	181.0	3750.0	male	2007
1	Adelie	Torgersen	39.5	17.4	186.0	3800.0	female	2007
2	Adelie	Torgersen	40.3	18.0	195.0	3250.0	female	2007
3	Adelie	Torgersen	NaN	NaN	NaN	NaN	NaN	2007
4	Adelie	Torgersen	36.7	19.3	193.0	3450.0	female	2007

I'm also going to create a utility function for plotting lines:

In [4]:

def plot_line(intercept, slope, xmin, xmax, **kwargs):
    xs = np.linspace(xmin, xmax, 100)
    ys = intercept + slope * xs
    plt.plot(xs, ys, **kwargs)

Explore Correlations¶

The Seaborn pairplot function shows a scatterplot matrix, allowing us to view possible correlations:

In [5]:

sns.pairplot(penguins[['species', 'sex', 'bill_length_mm', 'bill_depth_mm', 'flipper_length_mm', 'body_mass_g']], hue='species')

Out[5]:

<seaborn.axisgrid.PairGrid at 0x1da578ae8e0>

Let's also look at the distribution of sex and species, to see if we have sex imbalances between species:

In [6]:

sns.countplot(x='species', hue='sex', data=penguins)

Out[6]:

<matplotlib.axes._subplots.AxesSubplot at 0x1da58a9f490>

Nope!

Now a correlation matrix between our various numeric variables:

In [7]:

penguins.drop(columns=['year']).corr()

Out[7]:

	bill_length_mm	bill_depth_mm	flipper_length_mm	body_mass_g
bill_length_mm	1.000000	-0.235053	0.656181	0.595110
bill_depth_mm	-0.235053	1.000000	-0.583851	-0.471916
flipper_length_mm	0.656181	-0.583851	1.000000	0.871202
body_mass_g	0.595110	-0.471916	0.871202	1.000000

The .corr method on a data frame will compute correlations between all pairs of numeric variables, and show them in a table. The diagonal is 1, because a variable is always correlated with itself.

Dummy Variable Demo¶

Here's the code that demonstrates dummy-coding from the Categorical Predictors lecture:

In [8]:

pd.get_dummies(penguins['species'])

Out[8]:

	Adelie	Chinstrap	Gentoo
0	1	0	0
1	1	0	0
2	1	0	0
3	1	0	0
4	1	0	0
...	...	...	...
339	0	1	0
340	0	1	0
341	0	1	0
342	0	1	0
343	0	1	0

344 rows × 3 columns

And dropping the first level:

In [9]:

pd.get_dummies(penguins['species'], drop_first=True)

Out[9]:

	Chinstrap	Gentoo
0	0	0
1	0	0
2	0	0
3	0	0
4	0	0
...	...	...
339	1	0
340	1	0
341	1	0
342	1	0
343	1	0

344 rows × 2 columns

Flipper Regression¶

From the pair plot, the flipper length and body mass look like good candidates for regression. Let's show them with a regplot:

In [10]:

sns.regplot('flipper_length_mm', 'body_mass_g', data=penguins, line_kws={'color': 'orange'})

Out[10]:

<matplotlib.axes._subplots.AxesSubplot at 0x1da58d40d60>

Now we're going to fit a logistic regression.

The first step is to create the OLS model. This has two inputs:

1. The formula specifying the regression model (in this case, predicting body mass with flipper length)
2. The data to tran the model on

Let's do it:

In [11]:

bm_mod = smf.ols('body_mass_g ~ flipper_length_mm', data=penguins)

Once we have created the model, we need to fit it. The fit method will return an object containing the results of fitting the model, such as its paramters; it can be summarized with summary:

In [12]:

bmf = bm_mod.fit()
bmf.summary()

Out[12]:

OLS Regression Results

Dep. Variable:	body_mass_g	R-squared:	0.759
Model:	OLS	Adj. R-squared:	0.758
Method:	Least Squares	F-statistic:	1071.
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	4.37e-107
Time:	19:13:14	Log-Likelihood:	-2528.4
No. Observations:	342	AIC:	5061.
Df Residuals:	340	BIC:	5069.
Df Model:	1
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	-5780.8314	305.815	-18.903	0.000	-6382.358	-5179.305
flipper_length_mm	49.6856	1.518	32.722	0.000	46.699	52.672

Omnibus:	5.634	Durbin-Watson:	2.190
Prob(Omnibus):	0.060	Jarque-Bera (JB):	5.585
Skew:	0.313	Prob(JB):	0.0613
Kurtosis:	3.019	Cond. No.	2.89e+03

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 2.89e+03. This might indicate that there are
strong multicollinearity or other numerical problems.

Now we need to check our residual assumptions. There are two plots to create:

• Residuals vs. Fitted
• Residual Q-Q

I'm going to create a function that plots them both:

In [13]:

def plot_lm_diag(fit):
    "Plot linear fit diagnostics"
    sns.regplot(x=fit.fittedvalues, y=fit.resid)
    plt.xlabel('Fitted')
    plt.ylabel('Residuals')
    plt.title('Residuals vs. Fitted')
    plt.show()

    sm.qqplot(fit.resid, fit=True, line='45')
    plt.title('Residuals')
    plt.show()

And now use it:

In [14]:

plot_lm_diag(bmf)

Standardizing Variables¶

We can also standardize our variables before fitting a model. This converts variable values into z-scores, such that

$$z_i = \frac{x_i - \bar{x}}{s}$$

Z-standardized (or z-normalized) variables have a mean of 0 and a standard deviation of 1 (and since $1^2 = 1$, the variance is also 1). The result is that model coefficients are in units of standard deviations.

Let's write a function to convert a series to its z-scores:

In [15]:

def zscore(xs):
    xbar = xs.mean()
    s = xs.std()
    return (xs - xbar) / s

And create standardized versions of our variables:

In [16]:

penguin_std = pd.DataFrame({
    'species': penguins['species'],
    'sex': penguins['sex'],
    'mass': zscore(penguins['body_mass_g']),
    'flipper': zscore(penguins['flipper_length_mm']),
    'bill_len': zscore(penguins['bill_length_mm']),
    'bill_depth': zscore(penguins['bill_depth_mm'])
})
penguin_std

Out[16]:

	species	sex	mass	flipper	bill_len	bill_depth
0	Adelie	male	-0.563317	-1.416272	-0.883205	0.784300
1	Adelie	female	-0.500969	-1.060696	-0.809939	0.126003
2	Adelie	female	-1.186793	-0.420660	-0.663408	0.429833
3	Adelie	NaN	NaN	NaN	NaN	NaN
4	Adelie	female	-0.937403	-0.562890	-1.322799	1.088129
...	...	...	...	...	...	...
339	Chinstrap	male	-0.251578	0.432721	2.175637	1.341320
340	Chinstrap	female	-0.999750	0.077145	-0.077282	0.480471
341	Chinstrap	male	-0.532143	-0.562890	1.040019	0.531109
342	Chinstrap	male	-0.126883	0.646066	1.259816	0.936215
343	Chinstrap	female	-0.532143	-0.207315	1.149917	0.784300

344 rows × 6 columns

And now we'll fit and describe a model using these:

In [17]:

bm_mod = smf.ols('mass ~ flipper', data=penguin_std)
bmf = bm_mod.fit()
bmf.summary()

Out[17]:

OLS Regression Results

Dep. Variable:	mass	R-squared:	0.759
Model:	OLS	Adj. R-squared:	0.758
Method:	Least Squares	F-statistic:	1071.
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	4.37e-107
Time:	19:13:14	Log-Likelihood:	-241.46
No. Observations:	342	AIC:	486.9
Df Residuals:	340	BIC:	494.6
Df Model:	1
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	6.939e-17	0.027	2.61e-15	1.000	-0.052	0.052
flipper	0.8712	0.027	32.722	0.000	0.819	0.924

Omnibus:	5.634	Durbin-Watson:	2.190
Prob(Omnibus):	0.060	Jarque-Bera (JB):	5.585
Skew:	0.313	Prob(JB):	0.0613
Kurtosis:	3.019	Cond. No.	1.00

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

The residual plots will be identical to the unstandardized model, except for the scales of the axes, because all we did was rescale and recenter the variables:

In [18]:

plot_lm_diag(bmf)

Mass and Bill Length¶

Let's do a regression we don't expect to work well: mass vs. bill length.

First, we'll fit the regression:

In [19]:

bm_mod = smf.ols('mass ~ bill_len', data=penguin_std)
bmf = bm_mod.fit()
bmf.summary()

Out[19]:

OLS Regression Results

Dep. Variable:	mass	R-squared:	0.354
Model:	OLS	Adj. R-squared:	0.352
Method:	Least Squares	F-statistic:	186.4
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	3.81e-34
Time:	19:13:15	Log-Likelihood:	-410.02
No. Observations:	342	AIC:	824.0
Df Residuals:	340	BIC:	831.7
Df Model:	1
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	6.939e-17	0.044	1.59e-15	1.000	-0.086	0.086
bill_len	0.5951	0.044	13.654	0.000	0.509	0.681

Omnibus:	5.671	Durbin-Watson:	0.866
Prob(Omnibus):	0.059	Jarque-Bera (JB):	4.857
Skew:	-0.211	Prob(JB):	0.0882
Kurtosis:	2.597	Cond. No.	1.00

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

$R^2$ is much lower than the previous models, indicating this model explains much less of the variance.

Let's look at our diagnostic plots:

In [20]:

plot_lm_diag(bmf)

There's some noticable non-normality at the tails of the residuals, but it isn't bad. The residuals, however, exhibit notable heteroskedasticity,.

Multivariate Regression¶

This section goes with the Multiple Regression video.

Per-Species Intercepts¶

Let's add the species as a predictor, so we have per-species intercepts for the variable:

In [21]:

bm_mod = smf.ols('mass ~ flipper + species', data=penguin_std)
bmf = bm_mod.fit()
bmf.summary()

Out[21]:

OLS Regression Results

Dep. Variable:	mass	R-squared:	0.783
Model:	OLS	Adj. R-squared:	0.781
Method:	Least Squares	F-statistic:	405.7
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	1.25e-111
Time:	19:13:15	Log-Likelihood:	-223.79
No. Observations:	342	AIC:	455.6
Df Residuals:	338	BIC:	470.9
Df Model:	3
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	-0.0685	0.057	-1.207	0.228	-0.180	0.043
species[T.Chinstrap]	-0.2575	0.072	-3.577	0.000	-0.399	-0.116
species[T.Gentoo]	0.3327	0.119	2.801	0.005	0.099	0.566
flipper	0.7137	0.054	13.255	0.000	0.608	0.820

Omnibus:	3.390	Durbin-Watson:	2.556
Prob(Omnibus):	0.184	Jarque-Bera (JB):	3.397
Skew:	0.242	Prob(JB):	0.183
Kurtosis:	2.934	Cond. No.	6.47

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

Species is automatically dummy-coded with the first level dropped. We therefore get the following intercepts:

• Adelie: Intercept
• Chinstrap: Intercept + species[T.Chinstrap]
• Gentoo: Intercept + species[T.Gentoo]

Let's check fit:

In [22]:

plot_lm_diag(bmf)

And we'll plot the different lines on top of a scatterplot:

In [23]:

sns.scatterplot('flipper', 'mass', hue='species', data=penguin_std)
plot_line(bmf.params['Intercept'], bmf.params['flipper'], -2, 2, label='Adelie')
plot_line(bmf.params['Intercept'] + bmf.params['species[T.Gentoo]'],
          bmf.params['flipper'],
          -2, 2, label='Gentoo')
plot_line(bmf.params['Intercept'] + bmf.params['species[T.Chinstrap]'],
          bmf.params['flipper'],
          -2, 2, label='Gentoo')

Per-Species Slopes and Intercepts¶

Now let's add an interaction term between flipper and species.

We do this with flipper * species, which expands to flipper + species + flipper:species:

In [24]:

bm_mod = smf.ols('mass ~ flipper * species', data=penguin_std)
bmf = bm_mod.fit()
bmf.summary()

Out[24]:

OLS Regression Results

Dep. Variable:	mass	R-squared:	0.790
Model:	OLS	Adj. R-squared:	0.786
Method:	Least Squares	F-statistic:	252.2
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	2.22e-111
Time:	19:13:16	Log-Likelihood:	-218.25
No. Observations:	342	AIC:	448.5
Df Residuals:	336	BIC:	471.5
Df Model:	5
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	-0.1761	0.074	-2.393	0.017	-0.321	-0.031
species[T.Chinstrap]	-0.1888	0.101	-1.871	0.062	-0.387	0.010
species[T.Gentoo]	0.1579	0.135	1.172	0.242	-0.107	0.423
flipper	0.5757	0.081	7.095	0.000	0.416	0.735
flipper:species[T.Chinstrap]	0.0305	0.138	0.222	0.825	-0.240	0.301
flipper:species[T.Gentoo]	0.3821	0.122	3.139	0.002	0.143	0.621

Omnibus:	5.822	Durbin-Watson:	2.534
Prob(Omnibus):	0.054	Jarque-Bera (JB):	5.758
Skew:	0.317	Prob(JB):	0.0562
Kurtosis:	3.034	Cond. No.	10.8

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

AIC is improving slightly, not a lot.

Look at assumption checks:

In [25]:

plot_lm_diag(bmf)

Getting better, but still some curve on the Q-Q and doesn't quite look homoskedastic. (Watch the video for more explanation.)

Let's plot these different lines:

In [26]:

sns.scatterplot('flipper', 'mass', hue='species', data=penguin_std)
plot_line(bmf.params['Intercept'], bmf.params['flipper'], -2, 2, label='Adelie')
plot_line(bmf.params['Intercept'] + bmf.params['species[T.Gentoo]'],
          bmf.params['flipper'] + bmf.params['flipper:species[T.Gentoo]'],
          -2, 2, label='Gentoo')
plot_line(bmf.params['Intercept'] + bmf.params['species[T.Chinstrap]'],
          bmf.params['flipper'] + bmf.params['flipper:species[T.Chinstrap]'],
          -2, 2, label='Gentoo')

Sexual Dimorphism¶

Now let's look at penguin sex as a predictor, and we'll keep the flipper:species interaction:

In [27]:

bm_mod = smf.ols('mass ~ flipper + species * sex + flipper:species', data=penguin_std)
bmf = bm_mod.fit()
bmf.summary()

Out[27]:

OLS Regression Results

Dep. Variable:	mass	R-squared:	0.876
Model:	OLS	Adj. R-squared:	0.873
Method:	Least Squares	F-statistic:	285.9
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	7.40e-142
Time:	19:13:17	Log-Likelihood:	-125.89
No. Observations:	333	AIC:	269.8
Df Residuals:	324	BIC:	304.0
Df Model:	8
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	-0.7717	0.076	-10.090	0.000	-0.922	-0.621
species[T.Chinstrap]	0.2359	0.120	1.968	0.050	3.11e-05	0.472
species[T.Gentoo]	1.0222	0.120	8.509	0.000	0.786	1.259
sex[T.male]	0.7474	0.063	11.790	0.000	0.623	0.872
species[T.Chinstrap]:sex[T.male]	-0.5059	0.124	-4.086	0.000	-0.749	-0.262
species[T.Gentoo]:sex[T.male]	-0.0023	0.109	-0.021	0.983	-0.217	0.212
flipper	0.2861	0.069	4.172	0.000	0.151	0.421
flipper:species[T.Chinstrap]	0.1817	0.126	1.442	0.150	-0.066	0.429
flipper:species[T.Gentoo]	0.1260	0.117	1.074	0.284	-0.105	0.357

Omnibus:	0.719	Durbin-Watson:	2.050
Prob(Omnibus):	0.698	Jarque-Bera (JB):	0.836
Skew:	0.083	Prob(JB):	0.658
Kurtosis:	2.819	Cond. No.	16.2

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

This is good, but some non-significant relationships. In particular, the flipper/species interaction disappeared.

Assumption checks:

In [28]:

plot_lm_diag(bmf)

That is looking pretty good.

Drop interaction, due to non-significant coefficients:

In [29]:

bm_mod = smf.ols('mass ~ flipper + species * sex', data=penguin_std)
bmf = bm_mod.fit()
bmf.summary()

Out[29]:

OLS Regression Results

Dep. Variable:	mass	R-squared:	0.875
Model:	OLS	Adj. R-squared:	0.873
Method:	Least Squares	F-statistic:	380.2
Date:	Sat, 10 Oct 2020	Prob (F-statistic):	6.74e-144
Time:	19:13:17	Log-Likelihood:	-127.17
No. Observations:	333	AIC:	268.3
Df Residuals:	326	BIC:	295.0
Df Model:	6
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
Intercept	-0.7034	0.062	-11.305	0.000	-0.826	-0.581
species[T.Chinstrap]	0.0968	0.076	1.280	0.202	-0.052	0.246
species[T.Gentoo]	0.9982	0.108	9.273	0.000	0.786	1.210
sex[T.male]	0.7233	0.061	11.767	0.000	0.602	0.844
species[T.Chinstrap]:sex[T.male]	-0.4188	0.106	-3.953	0.000	-0.627	-0.210
species[T.Gentoo]:sex[T.male]	0.0549	0.090	0.612	0.541	-0.122	0.231
flipper	0.3592	0.049	7.291	0.000	0.262	0.456

Omnibus:	0.476	Durbin-Watson:	2.046
Prob(Omnibus):	0.788	Jarque-Bera (JB):	0.596
Skew:	0.069	Prob(JB):	0.742
Kurtosis:	2.845	Cond. No.	9.89

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

This is definitely the best AIC we've seen. Let's look at those assumption check plots:

In [30]:

plot_lm_diag(bmf)

That looks quite good.

Bootstrap¶

Now let's bootstrap some of these model parameters. We'll start with a function that will compute bootstrap samples of the rows of a data frame, and pass them to a function to get a statistic:

In [31]:

def boot_rows(df, stat, nboot=10000):
    n = len(df)
    obs = stat(df)
    boots = [stat(df.sample(n=n, replace=True)) for i in range(nboot)]
    return obs, np.quantile(boots, [0.025, 0.975])

Now we'll write a function to train our full OLS model, and return the flipper length coefficient:

In [32]:

def flipper_coefficient(df):
    bm_mod = smf.ols('mass ~ flipper + species * sex', data=df)
    bmf = bm_mod.fit()
    return bmf.params['flipper']

And we'll bootstrap the flipper length coefficient:

In [33]:

boot_rows(penguin_std, flipper_coefficient, nboot=1000)

Out[33]:

(0.35920897258861856, array([0.2687583 , 0.45152572]))

That matches the CIs from the model itself pretty closely!

Simultaneous Bootstrap of Multiple Parameters¶

In the video, I said you could do this. Here's the code.

First, we need a boot_rows function that can deal with the stat function returning a series of multiple statistics:

In [34]:

def boot_rows(df, stat, nboot=10000):
    n = len(df)
    obs = stat(df)
    boots = [stat(df.sample(n=n, replace=True)) for i in range(nboot)]
    if isinstance(obs, pd.Series):
        # we have multiple statistics - stack them into a frame
        boots = pd.DataFrame.from_records(boots)
        # compute confidence intervals - columns will be params, and rows CI bounds
        conf = boots.quantile([0.025, 0.975])
        # transpose will put the params on rows
        # join with frame-ified version of the original observations
        return obs.to_frame('value').join(conf.transpose())
    else:
        return obs, np.quantile(boots, [0.025, 0.975])

And a function to return the parameters from a trained model:

In [35]:

def flipper_params(df):
    bm_mod = smf.ols('mass ~ flipper + species * sex', data=df)
    bmf = bm_mod.fit()
    return bmf.params

And let's bootstrap all of this:

In [36]:

boot_rows(penguin_std, flipper_params, nboot=1000)

Out[36]:

	value	0.025	0.975
Intercept	-0.703441	-0.810038	-0.596531
species[T.Chinstrap]	0.096813	-0.037675	0.235350
species[T.Gentoo]	0.998247	0.803651	1.184524
sex[T.male]	0.723339	0.603583	0.837445
species[T.Chinstrap]:sex[T.male]	-0.418757	-0.631728	-0.198684
species[T.Gentoo]:sex[T.male]	0.054908	-0.118899	0.228228
flipper	0.359209	0.272918	0.453373

In [ ]: