Converting statsmodels summary object to Pandas Dataframe

Question

I am doing multiple linear regression with statsmodels.formula.api (ver 0.9.0) on Windows 10. After fitting the model and getting the summary with following lines i get summary in summary object format.

X_opt  = X[:, [0,1,2,3]]
regressor_OLS = sm.OLS(endog= y, exog= X_opt).fit()
regressor_OLS.summary()


                          OLS Regression Results                            
==============================================================================
Dep. Variable:                      y   R-squared:                       0.951
Model:                            OLS   Adj. R-squared:                  0.948
Method:                 Least Squares   F-statistic:                     296.0
Date:                Wed, 08 Aug 2018   Prob (F-statistic):           4.53e-30
Time:                        00:46:48   Log-Likelihood:                -525.39
No. Observations:                  50   AIC:                             1059.
Df Residuals:                      46   BIC:                             1066.
Df Model:                           3                                         
Covariance Type:            nonrobust                                         
==============================================================================
                 coef    std err          t      P>|t|      [0.025      0.975]
------------------------------------------------------------------------------
const       5.012e+04   6572.353      7.626      0.000    3.69e+04    6.34e+04
x1             0.8057      0.045     17.846      0.000       0.715       0.897
x2            -0.0268      0.051     -0.526      0.602      -0.130       0.076
x3             0.0272      0.016      1.655      0.105      -0.006       0.060
==============================================================================
Omnibus:                       14.838   Durbin-Watson:                   1.282
Prob(Omnibus):                  0.001   Jarque-Bera (JB):               21.442
Skew:                          -0.949   Prob(JB):                     2.21e-05
Kurtosis:                       5.586   Cond. No.                     1.40e+06
==============================================================================

I want to do backward elimination for P values for significance level 0.05. For this i need to remove the predictor with highest P values and run the code again.

I wanted to know if there is a way to extract the P values from the summary object, so that i can run a loop with conditional statement and find the significant variables without repeating the steps manually.

Thank you.

Saturninasaturnine

The accepted answer shows how to convert the summary table to pandas DataFrame. However, for the use case of selection on p-values it is better to directly use the attribute results.pvalues, which is also used in the second answer. – Eductive

Answer 1

The answer from @Michael B works well, but requires "recreating" the table. The table itself is actually directly available from the summary().tables attribute. Each table in this attribute (which is a list of tables) is a SimpleTable, which has methods for outputting different formats. We can then read any of those formats back as a pd.DataFrame:

import statsmodels.api as sm

model = sm.OLS(y,x)
results = model.fit()
results_summary = results.summary()

# Note that tables is a list. The table at index 1 is the "core" table. Additionally, read_html puts dfs in a list, so we want index 0
results_as_html = results_summary.tables[1].as_html()
pd.read_html(results_as_html, header=0, index_col=0)[0]

Gelignite

This doesn't work for when using formula API. AttributeError: 'OLSResults' object has no attribute 'tables' – Keppel

What version are you on? I'm on python 3.6.5 and using the latest version of statsmodels, but didn't test older versions. – Gelignite

Python 3.6.5, statsmodels 0.9.0 – Keppel

Woops - forgot the summary method! Thanks for pointing that out. Answer is updated. – Gelignite

Why didn't I think of that? Borderline hacky but very neat. Here's an alternative using the csv methods, in case it comes in handy: pd.read_csv(pd.compat.StringIO(table.as_csv()), index_col=0) – Dalury

Since pandas>=0.25, pd.compat.StringIO has been removed. So instead use from io import StringIO. Worked for me on pandas==1.2.4! – Phylactery

Answer 2

An easy solution is just one line of code:

LRresult = (result.summary2().tables[1])

As ZaxR mentioned in the following comment, Summary2 is not yet considered stable, while it works well with Summary too. So this could be correct answer:

LRresult = (result.summary().tables[1])

This will give you a dataframe object:

type(LRresult)

pandas.core.frame.DataFrame

To get the significant variables and run the test again:

newlist = list(LRresult[LRresult['P>|z|']<=0.05].index)[1:]
myform1 = 'binary_Target' + ' ~ ' + ' + '.join(newlist)

M1_test2 = smf.logit(formula=myform1,data=myM1_1)

result2 = M1_test2.fit(maxiter=200)
LRresult2 = (result2.summary2().tables[1])
LRresult2

Shaeshaef

Summary2 is not yet considered stable, though looks close. See this discussion. – Gelignite

Also works for summary(). This should be the accepted answer – Janellajanelle

@user3357177, no does not. .summary2() returns a pandas.DataFrame, but .summary() returns statsmodels.SimpleTable. – Resentment

Answer 3

Store your model fit as a variable results, like so:

import statsmodels.api as sm
model = sm.OLS(y,x)
results = model.fit()

Then create a a function like below:

def results_summary_to_dataframe(results):
    '''take the result of an statsmodel results table and transforms it into a dataframe'''
    pvals = results.pvalues
    coeff = results.params
    conf_lower = results.conf_int()[0]
    conf_higher = results.conf_int()[1]

    results_df = pd.DataFrame({"pvals":pvals,
                               "coeff":coeff,
                               "conf_lower":conf_lower,
                               "conf_higher":conf_higher
                                })

    #Reordering...
    results_df = results_df[["coeff","pvals","conf_lower","conf_higher"]]
    return results_df

You can further explore all the attributes of the results object by using dir() to print, then add them to the function and df accordingly.

Pantograph

Thank you Michael B for the help. – Saturninasaturnine

No problem, if it worked please mark the answer as correct! Happy coding/data sci-ing!! – Pantograph

Super useful function! – Electrify

Answer 4

You may write as below.It will be a easy fix and work almost appropriate every time.

lr.summary2()

Emissive

Answer 5

I still don't think there is a clean answer that captures the query in its totality. Here is one way to capture everything in two dataframes (one for the middle table, one for the metrics on the top and bottom).

def reform_df(dft):
    # quick and dirty stacking of cols 2,3 on 0,1
    dfl = dft[[0,1]]
    dfr = dft[[2,3]]
    dfr.columns = 0,1
    dfout = pd.concat([dfl,dfr])
    dfout.columns=['Parameter','Value']
    return dfout

def model_summary_to_dataframe(model):
    # first the middle table      
    results_df = pd.DataFrame(model.summary().tables[1])
    results_df = results_df.set_index(0)
    results_df.columns = results_df.iloc[0]
    results_df = results_df.iloc[1:]
    results_df.index.name='Parameter'

    # now for the surrounding information
    metrics_top = reform_df(pd.DataFrame(model.summary().tables[0]))
    metrics_bot = reform_df(pd.DataFrame(model.summary().tables[2]))
    metrics_df = pd.concat([metrics_top,metrics_bot])

    return pd.DataFrame(results_df),metrics_df

Biauriculate

Answer 6

The code below puts all the metrics into a dictionary accessible by key. The intermediate result is actually a DataFrame you can use, I did not make the coefficients into a dictionary, but you can apply a similar method but then two levels deep dict[var][metric].

In order to make the keys easy to type, I converted some of the metric names into more easily typed versions. E.g. "Prob(Omnibus):" becomes prob_omnibus such that you can access the value by res_dict['prob_omnibus'].

import pandas as pd

res = sm.OLS(y, X).fit()
model_results_df = []
coefficient_df = None
for i, tab in enumerate(res.summary().tables):
    header, index_col = None, None
    if i == 1:
        coefficient_df = pd.read_html(tab.as_html(), header=0, index_col=0)[0]
    else:
        df = pd.read_html(tab.as_html())[0]
        model_results_df += [df.iloc[:,0:2], df.iloc[:,2:4]]

model_results_df = pd.DataFrame(np.concatenate(model_results_df), columns=['metric', 'value'])
model_results_df.dropna(inplace=True, axis=0)
model_results_df.metric = model_results_df.metric.apply(lambda x : x.lower().replace(' (', '_')
                                                        .replace('.', '').replace('(', '_')
                                                        .replace(')', '').replace('-', '_')
                                                       .replace(':', '').replace(' ', '_'))

res_dict = dict(zip(model_results_df.metric.values, model_results_df.value.values))
res_dict['f_statistic']

Regiment

Answer 7

If you want the surrounding information, try the following:

import pandas as pd
dfs = {}
fs = fa_model.summary()
for item in fs.tables[0].data:
    dfs[item[0].strip()] = item[1].strip()
    dfs[item[2].strip()] = item[3].strip()
for item in fs.tables[2].data:
    dfs[item[0].strip()] = item[1].strip()
    dfs[item[2].strip()] = item[3].strip()
dfs = pd.Series(dfs)

Biauriculate

Answer 8

It works but I found a small bug in item[3] This is the fix:

import pandas as pd
dfs = {}
fs = stepwise_fit.summary()
for item in fs.tables[0].data:
    #print("item " + str(item))
    dfs[item[0].strip()] = item[1].strip()
    dfs[item[2].strip()] = str(item[3]).strip()
for item in fs.tables[2].data:
    dfs[item[0].strip()] = item[1].strip()
    dfs[item[2].strip()] = str(item[3]).strip()
dfs = pd.Series(dfs)
print(type(dfs))

Antlion

Answer 9

This worked for me:

st, data, ss2 = summary_table(result, alpha=0.05)

df = pd.DataFrame( data=data, columns=ss2 )

Treasonous