Often we have additional data aside from the duration, and if applicable any censorships that occurred. In the regime dataset, we have the type of government the political leader was part of, the country they were head of, and the year they were elected. Can we use this data in survival analysis?
Yes, the technique is called survival regression -- the name implies we regress covariates (e.g., year elected, country, etc.) against a another variable -- in this case durations and lifetimes. Similar to the logic in the first part of this tutorial, we cannot use traditional methods like linear regression.
There are two popular competing techniques in survival regression: Cox's model and Aalen's additive model. Both models attempt to represent the hazard rate \lambda(t | x) as a function of t and some covariates x. In Cox's model, the relationship is defined:
\lambda(t | x) = b_0(t)\exp\left( b_1x_1 + ... + b_dx_d\right)
On the other hand, Aalen's additive model assumes the following form:
\lambda(t | x) = b_0(t) + b_1(t)x_1 + ... + b_d(t)x_d
Warning
This implementation is still experimental.
The estimator to fit unknown coefficients in Aalen's additive model is
located in estimators under AalenAdditiveFitter. For this
exercise, we will use the regime dataset and include the categorical
variables un_continent_name (eg: Asia, North America,...), the
regime type (e.g., monarchy, civilian,...) and the year the regime
started in, start_year.
Aalen's additive model typically does not estimate the individual
b_i(t) but instead estimates \int_0^t b_i(s) \; ds
(similar to the estimate of the hazard rate using NelsonAalenFitter
above). This is important to keep in mind when analyzing the output.
.. code:: python
from lifelines import AalenAdditiveFitter data.head()
| ctryname | cowcode2 | politycode | un_region_name | un_continent_name | ehead | leaderspellreg | democracy | regime | start_year | duration | observed |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Afghanistan | 700 | 700 | Southern Asia | Asia | Mohammad Zahir Shah | Mohammad Zahir Shah.Afghanistan.1946.1952.Mona... | Non-democracy | Monarchy | 1946 | 7 | 1 |
| Afghanistan | 700 | 700 | Southern Asia | Asia | Sardar Mohammad Daoud | Sardar Mohammad Daoud.Afghanistan.1953.1962.Ci... | Non-democracy | Civilian Dict | 1953 | 10 | 1 |
| Afghanistan | 700 | 700 | Southern Asia | Asia | Mohammad Zahir Shah | Mohammad Zahir Shah.Afghanistan.1963.1972.Mona... | Non-democracy | Monarchy | 1963 | 10 | 1 |
| Afghanistan | 700 | 700 | Southern Asia | Asia | Sardar Mohammad Daoud | Sardar Mohammad Daoud.Afghanistan.1973.1977.Ci... | Non-democracy | Civilian Dict | 1973 | 5 | 0 |
| Afghanistan | 700 | 700 | Southern Asia | Asia | Nur Mohammad Taraki | Nur Mohammad Taraki.Afghanistan.1978.1978.Civi... | Non-democracy | Civilian Dict | 1978 | 1 | 0 |
5 rows × 12 columns
I'm using the lovely library patsy here to create a covariance matrix from my original dataframe.
import patsy
# the '-1' term refers to not adding an intercept column (a column of all 1s).
X = patsy.dmatrix('un_continent_name + regime + start_year - 1', data, return_type='dataframe')X.columns['un_continent_name[Africa]', 'un_continent_name[Americas]', 'un_continent_name[Asia]', 'un_continent_name[Europe]', 'un_continent_name[Oceania]', 'regime[T.Military Dict]', 'regime[T.Mixed Dem]', 'regime[T.Monarchy]', 'regime[T.Parliamentary Dem]', 'regime[T.Presidential Dem]', 'start_year']
Below we create our fitter class. Since we did not supply an intercept
column in our matrix we have included the keyword fit_intercept=True
(True by default) which will append the column of ones to our
matrix. (Sidenote: the intercept term, b_0(t) in survival
regression is often known as the baseline hazard.)
We have also included the coef_penalizer option. During the estimation, a
linear regression is computed at each step. Often the regression can be
unstable (due to high
co-linearity
or small sample sizes) -- adding a penalizer term controls the stability. I recommend always starting with a small penalizer term -- if the estimates still appear to be too unstable, try increasing it.
aaf = AalenAdditiveFitter(coef_penalizer=1.0, fit_intercept=True)An instance of AalenAdditiveFitter
includes a fit method that performs the inference on the coefficients. This method accepts a pandas DataFrame: each row is an individual and columns are the covariates and
two individual columns: a duration column and a boolean event occurred column (where event occurred refers to the event of interest - expulsion from government in this case)
data = lifelines.datasets.load_dd()
X['T'] = data['duration']
X['E'] = data['observed']aaf.fit(X, 'T', event_col='E')After fitting, the instance exposes a cumulative_hazards_ DataFrame
containing the estimates of \int_0^t b_i(s) \; ds:
figsize(12.5,8)
aaf.cumulative_hazards_.head()| un_continent_name[Africa] | un_continent_name[Americas] | un_continent_name[Asia] | un_continent_name[Europe] | un_continent_name[Oceania] | regime[T.Military Dict] | regime[T.Mixed Dem] | regime[T.Monarchy] | regime[T.Parliamentary Dem] | regime[T.Presidential Dem] | start_year | baseline |
|---|---|---|---|---|---|---|---|---|---|---|---|
| -0.051595 | -0.082406 | 0.010666 | 0.154493 | -0.060438 | 0.075333 | 0.086274 | -0.133938 | 0.048077 | 0.127171 | 0.000116 | -0.029280 |
| -0.014713 | -0.039471 | 0.095668 | 0.194251 | -0.092696 | 0.115033 | 0.358702 | -0.226233 | 0.168783 | 0.121862 | 0.000053 | 0.143039 |
| 0.007389 | -0.064758 | 0.115121 | 0.170549 | 0.069371 | 0.161490 | 0.677347 | -0.271183 | 0.328483 | 0.146234 | 0.000004 | 0.297672 |
| -0.058418 | 0.011399 | 0.091784 | 0.205824 | 0.125722 | 0.220028 | 0.932674 | -0.294900 | 0.365604 | 0.422617 | 0.000002 | 0.376311 |
| -0.099282 | 0.106641 | 0.112083 | 0.150708 | 0.091900 | 0.241575 | 1.123860 | -0.391103 | 0.536185 | 0.743913 | 0.000057 | 0.362049 |
5 rows × 12 columns
AalenAdditiveFitter also has built in plotting:
aaf.plot(columns=['regime[T.Presidential Dem]', 'baseline', 'un_continent_name[Europe]'], iloc=slice(1,15))
Regression is most interesting if we use it on data we have not yet seen, i.e., prediction! We can use what we have learned to predict individual hazard rates, survival functions, and median survival time. The dataset we are using is available up until 2008, so let's use this data to predict the (already partly seen) possible duration of Canadian Prime Minister Stephen Harper.
ix = (data['ctryname'] == 'Canada') * (data['start_year'] == 2006)
harper = X.loc[ix]
print("Harper's unique data point", harper)Harper's unique data point
array([[ 0., 0., 1., 0., 0., 0., 0., 1.,
0., 0., 2003.]])
ax = plt.subplot(2,1,1)
aaf.predict_cumulative_hazard(harper).plot(ax=ax)
ax = plt.subplot(2,1,2)
aaf.predict_survival_function(harper).plot(ax=ax);
Warning
Because of the nature of the model, estimated survival functions of individuals can increase. This is an expected artifact of Aalen's additive model.
Lifelines has an implementation of the Cox propotional hazards regression model (implemented in
R under coxph). The idea behind the model is that the log-hazard of an individual is a linear function of their static covariates and a population-level baseline hazard that changes over time. Mathematically:
\lambda(t | X) = b_0(t) \exp{\left(\sum_{i=1}^d b_i x_i\right)}
Note a few facts about this model: the only time component is in the baseline hazard, b_0(t). In the above product, the second term is only a scalar factor that only increases or decreases the baseline hazard. Thus a change in a covariate will only increase or decrease this baseline hazard.
The implementation of the Cox model in lifelines, called CoxPHFitter has a similar API to AalensAdditiveFitter. Like R, it has a print_summary function that prints a tabular view of coefficients and related stats.
This example data is from the paper here, avaible as load_rossi in lifelines.
from lifelines.datasets import load_rossi
from lifelines import CoxPHFitter
rossi_dataset = load_rossi()
cph = CoxPHFitter()
cph.fit(rossi_dataset, duration_col='week', event_col='arrest')
cph.print_summary() # access the results using cph.summary
"""
n=432, number of events=114
coef exp(coef) se(coef) z p lower 0.95 upper 0.95
fin -0.3790 0.6845 0.1914 -1.9806 0.0476 -0.7542 -0.0039 *
age -0.0572 0.9444 0.0220 -2.6042 0.0092 -0.1003 -0.0142 **
race 0.3141 1.3691 0.3080 1.0198 0.3078 -0.2897 0.9180
wexp -0.1511 0.8597 0.2121 -0.7124 0.4762 -0.5670 0.2647
mar -0.4328 0.6487 0.3818 -1.1335 0.2570 -1.1813 0.3157
paro -0.0850 0.9185 0.1957 -0.4341 0.6642 -0.4687 0.2988
prio 0.0911 1.0954 0.0286 3.1824 0.0015 0.0350 0.1472 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Concordance = 0.640
"""To access the coefficients and the baseline hazard directly, you can use cph.hazards_ and cph.baseline_hazard_ respectively.
Fitting the Cox model to the data involves using gradient descent. Lifelines takes extra effort to help with convergence. If you wish to see the fitting, there is a show_progress parameter in CoxPHFitter.fit function. For further help, see :ref:`Problems with convergence in the Cox Proportional Hazard Model`.
After fitting, the value of the maximum log-likelihood this available using cph._log_likelihood. Similarly, the score and Hessian matrix are available under _score_ and _hessian_ respectively. The _hessian_ can be used the find the covariance matrix of the coefficients.
After fitting, you may want to know how "good" of a fit your model was to the data. Aside from traditional approaches, two methods the author has found useful is to 1. look at the concordance-index (see below section on :ref:`Model Selection in Survival Regression`), available as cph.score_ or in the print_summary and 2. compare spread between the baseline survival function vs the Kaplan Meier survival function (Why? a small spread between these two curves means that the impact of the exponential in the Cox model does very little, whereas a large spread means most of the changes in individual hazard can be attributed to the exponential term). For example, the first figure below is a good fit, and the second figure is a much weaker fit.
After fitting, you can use use the suite of prediction methods (similar to Aalen's additve model above): .predict_partial_hazard, .predict_survival_function, etc.
cph.predict_partial_hazard(rossi_dataset.drop(["week", "arrest"], axis=1))With a fitted model, an altervative way to view the coefficients and their ranges is to use the plot method.
from lifelines.datasets import load_rossi
from lifelines import CoxPHFitter
rossi_dataset = load_rossi()
cph = CoxPHFitter()
cph.fit(rossi_dataset, duration_col='week', event_col='arrest')
cph.plot()
A quick and visual way to check the proportional hazards assumption of a variable is to plot the survival curves segmented by the values of the variable. If the survival curves are the same "shape" and differ only by a constant factor, then the assumption holds. A more clear way to see this is to plot what's called the logs curve: the loglogs (-log(survival curve)) vs log(time). If the curves are parallel (and hence do not cross each other), then it's likely the variable satisfies the assumption. If the curves do cross, likely you'll have to "stratify" the variable (see next section). In lifelines, the KaplanMeierFitter object has a .plot_loglogs function for this purpose.
The following is the loglogs curves of two variables in our regime dataset. The first is the democracy type, which does have (close to) parallel lines, hence satisfies our assumption:
from lifelines.datasets import load_dd
from lifelines import KaplanMeierFitter
data = load_dd()
democracy_0 = data.loc[data['democracy'] == 'Non-democracy']
democracy_1 = data.loc[data['democracy'] == 'Democracy']
kmf0 = KaplanMeierFitter()
kmf0.fit(democracy_0['duration'], event_observed=democracy_0['observed'])
kmf1 = KaplanMeierFitter()
kmf1.fit(democracy_1['duration'], event_observed=democracy_1['observed'])
fig, axes = plt.subplots()
kmf0.plot_loglogs(ax=axes)
kmf1.plot_loglogs(ax=axes)
axes.legend(['Non-democracy', 'Democracy'])
plt.show()
The second variable is the regime type, and this variable does not follow the proportional hazards assumption.
Sometimes a covariate may not obey the proportional hazard assumption. In this case, we can allow a factor without estimating its effect to be adjusted. To specify categorical variables to be used in stratification, we define them in the call to fit:
cph.fit(rossi_dataset, 'week', event_col='arrest', strata=['race'])
cph.print_summary() # access the results using cph.summary
"""
n=432, number of events=114
coef exp(coef) se(coef) z p lower 0.95 upper 0.95
fin -0.3775 0.6856 0.1913 -1.9731 0.0485 -0.7525 -0.0024 *
age -0.0573 0.9443 0.0220 -2.6081 0.0091 -0.1004 -0.0142 **
wexp -0.1435 0.8664 0.2127 -0.6746 0.4999 -0.5603 0.2734
mar -0.4419 0.6428 0.3820 -1.1570 0.2473 -1.1907 0.3068
paro -0.0839 0.9196 0.1958 -0.4283 0.6684 -0.4677 0.3000
prio 0.0919 1.0962 0.0287 3.1985 0.0014 0.0356 0.1482 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Concordance = 0.638
"""Warning
This implementation is still experimental.
Often an individual will have a covariate change over time. An example of this is hospital patients who enter the study and, at some future time, may recieve a heart transplant. We would like to know the effect of the transplant, but we cannot condition on whether they recieved the transplant naively. Consider that if patients needed to wait at least 1 year before getting a transplant, then everyone who dies before that year is considered as a non-transplant patient, and hence this would overestimate the hazard of not recieving a transplant.
We can incorporate changes over time into our survival analysis by using a modification of the Cox model above. The general mathematical description is:
\lambda(t | X) = b_0(t) \exp{\left(\sum_{i=1}^d b_i x_i(t)\right)}
Note the time-varying x_i(t) to denote that covariates can change over time. This model is implemented in lifelines as CoxTimeVaryingFitter. The dataset schema required is different than previous models, so we will spend some time describing this.
Lifelines requires that the dataset be in what is called the long format. This looks like one row per state change, including an ID, the left (exclusive) time point, and right (inclusive) time point. For example, the following dataset tracks three unique subjects.
| id | start | stop | group | z | event |
|---|---|---|---|---|---|
| 1 | 0 | 8 | 1 | 0 | False |
| 2 | 0 | 5 | 0 | 0 | False |
| 2 | 5 | 8 | 0 | 1 | True |
| 3 | 0 | 3 | 1 | 0 | False |
| 3 | 3 | 12 | 1 | 1 | True |
5 rows × 6 columns
In the above dataset, start and stop denote the boundaries, id is the unique identifier per subject, and event denotes if the subject died at the end of that period. For example, subject ID 2 had variable z=0 up to and including the end of time period 5 (we can think that measurements happen at end of the time period.), after which it was set to 1.
So if this is the desired dataset, it can be built up first from smaller datasets. To do this we can use some helper functions provided in lifelines.
Typically, data will be in a format that looks like it comes out of a relational database. You may have a "base" table with ids, durations, and a censorsed flag, and possibly static covariates. Ex:
| id | duration | event | var1 |
|---|---|---|---|
| 1 | 10 | True | 0.1 |
| 2 | 12 | False | 0.5 |
2 rows × 4 columns
You'll also have secondary dataset that reference taking future measurements. Example:
| id | time | var2 |
|---|---|---|
| 1 | 0 | 1.4 |
| 1 | 4 | 1.2 |
| 1 | 8 | 1.5 |
| 2 | 0 | 1.6 |
4 rows × 3 columns
where time is the duration from the entry event. Here we see subject 1 had a change in their var2 covariate at the end of time 4 and at the end of time 8. We can use to_long_format to transform the base dataset into a long format and add_covariate_to_timeline to fold the covariate dataset into the original dataset.
from lifelines.utils import to_long_format
from lifelines.utils import add_covariate_to_timeline
base_df = to_long_format(base_df, duration_col="T")
df = add_covariate_to_timeline(base_df, cv, duration_col="time", id_col="id", event_col="event")| id | start | stop | var1 | var2 | event |
|---|---|---|---|---|---|
| 1 | 0 | 4 | 0.1 | 1.4 | False |
| 1 | 4 | 8 | 0.1 | 1.2 | False |
| 1 | 8 | 10 | 0.1 | 1.5 | True |
| 2 | 0 | 12 | 0.5 | 1.6 | False |
4 rows × 6 columns
From the above output, we can see that subject 1 changed state twice over the observation period, finally expiring at the end of time 10. Subject 2 was a censored case, and we lost them after time 2.
You may have multiple covariates you wish to add, so the above could be streamlined like so:
from lifelines.utils import to_long_format
from lifelines.utils import add_covariate_to_timeline
base_df = to_long_format(base_df, duration_col="T")
df = base_df.pipe(add_covariate_to_timeline, cv1, duration_col="time", id_col="id", event_col="event")\
.pipe(add_covariate_to_timeline, cv2, duration_col="time", id_col="id", event_col="event")\
.pipe(add_covariate_to_timeline, cv3, duration_col="time", id_col="id", event_col="event")One additional flag on add_covariate_to_timeline that is of interest is the cumulative_sum flag. By default it is False, but turning it to True will perform a cumulative sum on the covariate before joining. This is useful if the covariates describe an incremental change, instead of a state update. For example, we may have measurements of drugs administered to a patient, and we want to the covariate to reflect how much we have administered since the start. In contrast, a covariate the measure the temperature of the patient is a state update. See :ref:`Example cumulative total using ``add_covariate_to_timeline``` to see an example of this.
For an example of pulling datasets like this from a SQL-store, see :ref:`Example SQL queries and transformations to get time varying data`.
Once your dataset is in the correct orientation, we can use CoxTimeVaryingFitter to fit the model to your data.
from lifelines import CoxTimeVaryingFitter
ctv = CoxTimeVaryingFitter()
ctv.fit(df, id_col="id", event_col="event", start_col="start", stop_col="stop")
ctv.print_summary()
ctv.plot()Unlike the other regression models, prediction in a time-varying setting is not possible normally. To predict, we would need to know the covariates values beyond the current time, but if we knew that, we would also know if the subject was still alive or not. For this reason, there are no prediction methods attached to CoxTimeVaryingFitter.
If censorship is present, it's not appropriate to use a loss function like mean-squared-error or mean-absolute-loss. Instead, one measure is the concordance-index, also known as the c-index. This measure evaluates the accuracy of the ordering of predicted time. It is infact a generalization of AUC, another common loss function, and is interpreted similarly:
- 0.5 is the expected result from random predictions,
- 1.0 is perfect concordance and,
- 0.0 is perfect anti-concordance (multiply predictions with -1 to get 1.0)
The measure is implemented in lifelines under lifelines.utils.concordance_index and accepts the actual times (along with any censorships) and the predicted times.
Lifelines has an implementation of k-fold cross validation under lifelines.utils.k_fold_cross_validation. This function accepts an instance of a regression fitter (either CoxPHFitter of AalenAdditiveFitter), a dataset, plus k (the number of folds to perform, default 5). On each fold, it splits the data
into a training set and a testing set fits itself on the training set and evaluates itself on the testing set (using the concordance measure).
from lifelines import CoxPHFitter
from lifelines.datasets import load_regression_dataset
from lifelines.utils import k_fold_cross_validation
regression_dataset = load_regression_dataset()
cph = CoxPHFitter()
scores = k_fold_cross_validation(cph, regression_dataset, 'T', event_col='E', k=3)
print(scores)
print(np.mean(scores))
print(np.std(scores))
#[ 0.5896 0.5358 0.5028]
# 0.542
# 0.035