# Survival Regression¶

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:

On the other hand, Aalen’s additive model assumes the following form:

## Cox’s Proportional Hazard 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:

Note a few facts about this model: the only time component is in the baseline hazard, \(b_0(t)\). In the above product, the partial hazard is a time-invariant scalar factor that only increases or decreases the baseline hazard. Thus a changes in covariates will only increase or decrease this baseline hazard.

### Lifelines implementation¶

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', show_progress=True)
cph.print_summary() # access the results using cph.summary
"""
duration col = week
event col = arrest
number of subjects = 432
number of events = 114
log-likelihood = -658.748
time fit was run = 2018-10-22 20:47:44 UTC
---
coef exp(coef) se(coef) z p lower 0.95 upper 0.95
fin -0.3794 0.6843 0.1914 -1.9826 0.0474 -0.7545 -0.0043 *
age -0.0574 0.9442 0.0220 -2.6109 0.0090 -0.1006 -0.0143 **
race 0.3139 1.3688 0.3080 1.0192 0.3081 -0.2898 0.9176
wexp -0.1498 0.8609 0.2122 -0.7058 0.4803 -0.5657 0.2662
mar -0.4337 0.6481 0.3819 -1.1358 0.2561 -1.1821 0.3147
paro -0.0849 0.9186 0.1958 -0.4336 0.6646 -0.4685 0.2988
prio 0.0915 1.0958 0.0286 3.1939 0.0014 0.0353 0.1476 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Concordance = 0.640
Likelihood ratio test = 33.266 on 7 df, p=0.00002
"""
```

To access the coefficients and the baseline hazard directly, you can use `cph.hazards_`

and `cph.baseline_hazard_`

respectively.

### Convergence¶

Fitting the Cox model to the data involves using gradient descent. Lifelines takes extra effort to help with convergence, so please be attentive to any warnings that appear. Fixing any warnings will generally help convergence. If you wish to see the fitting, there is a `show_progress`

parameter in `CoxPHFitter.fit`

function. For further help, see 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.

### Goodness of fit and prediction¶

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 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 additive model): `.predict_partial_hazard`

, `.predict_survival_function`

, etc.

```
X = rossi_dataset.drop(["week", "arrest"], axis=1)
cph.predict_partial_hazard(X)
cph.predict_survival_function(X)
cph.predict_survival_function(X, times=[5., 25., 50.])
```

### Plotting the coefficients¶

With a fitted model, an alternative 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', show_progress=True)
cph.plot()
```

### Plotting the effect of varying a covariate¶

After fitting, we can plot what the survival curves look like as we vary a single covarite while
holding everything else equal. This is useful to understand the impact of a covariate, *given the model*. To do this, we use the `plot_covariate_groups`

method and give it the covariate of interest, and the values to display.

```
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', show_progress=True)
cph.plot_covariate_groups('prio', [0, 5, 10, 15])
```

### Checking the proportional hazards assumption¶

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.

Non-proportional hazards is a case of *model misspecification*. Two suggestions are to look for ways to *stratify* a column (see below), or to go ahead with the current model but use `robust`

errors (in this case, the sandwhich error). In the latter case, you can specify this with with `CoxPHFitter.fit(..., robust=True)`

.

### Stratification¶

Sometimes one or more covariates may not obey the proportional hazard assumption. In this case, we can allow the covariate(s) to still be including in the model without estimating its effect. This is called stratification. At a high level, think of it as splitting the dataset into *N* datasets, defined by the covariate(s). Each dataset has its own baseline hazard (the non-parametric part ofthe model), but they all share the regression parameters (the parametric part of the model). Since covariates are the same within each dataset, there is no regression parameter for the covariates stratified on, hence they will not show up in the output. However there will be *N* baseline hazards under `baseline_cumulative_hazard_`

.

To specify categorical variables to be used in stratification, we define them in the call to `fit`

:

```
from lifelines.datasets import load_rossi
from lifelines import CoxPHFitter
rossi_dataset = load_rossi()
cph = CoxPHFitter()
cph.fit(rossi_dataset, 'week', event_col='arrest', strata=['race'], show_progress=True)
cph.print_summary() # access the results using cph.summary
"""
duration col = week
event col = arrest
strata = ['race']
number of subjects = 432
number of events = 114
log-likelihood = -620.564
time fit was run = 2018-10-23 02:45:52 UTC
---
coef exp(coef) se(coef) z p lower 0.95 upper 0.95
fin -0.3788 0.6847 0.1913 -1.9799 0.0477 -0.7537 -0.0038 *
age -0.0576 0.9440 0.0220 -2.6198 0.0088 -0.1008 -0.0145 **
wexp -0.1428 0.8670 0.2128 -0.6708 0.5023 -0.5598 0.2743
mar -0.4388 0.6448 0.3821 -1.1484 0.2508 -1.1878 0.3101
paro -0.0858 0.9178 0.1958 -0.4380 0.6614 -0.4695 0.2980
prio 0.0922 1.0966 0.0287 3.2102 0.0013 0.0359 0.1485 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Concordance = 0.638
Likelihood ratio test = 109.634 on 6 df, p=0.00000
"""
cph.baseline_cumulative_hazard_.shape
# (49, 2)
```

### Weights & Robust Errors¶

Observations can come with weights, as well. These weights may be integer values representing some commonly occuring observation, or they may be float values representing some sampling weights or inverse probability weights. In the `CoxPHFitter.fit`

method, an option is present for specifying which column in the dataframe should be used as weights. See example below.

Generally, unless your weights are integers should

## Aalen’s Additive model¶

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.

```
from lifelines import AalenAdditiveFitter
from lifelines.datasets import load_dd
data = load_dd()
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
X = patsy.dmatrix('un_continent_name + regime + start_year', data, return_type='dataframe')
X = X.rename(columns={'Intercept': 'baseline'})
```

```
X.columns.tolist()
```

```
['baseline',
'un_continent_name[T.Americas]',
'un_continent_name[T.Asia]',
'un_continent_name[T.Europe]',
'un_continent_name[T.Oceania]',
'regime[T.Military Dict]',
'regime[T.Mixed Dem]',
'regime[T.Monarchy]',
'regime[T.Parliamentary Dem]',
'regime[T.Presidential Dem]',
'start_year']
```

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)
```

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)

```
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 \(\int_0^t b_i(s) \; ds\):

```
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[T.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 duration of former Canadian Prime Minister Stephen Harper.

```
ix = (data['ctryname'] == 'Canada') & (data['start_year'] == 2006)
harper = X.loc[ix]
print("Harper's unique data point:")
print(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.

## Cox’s Time Varying Proportional Hazard model¶

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:

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.

### Dataset creation for time-varying regression¶

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. Since `event`

is 1 in that row, we conclude that the subject died at time 8,

This desired dataset can be built up 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 alive, 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

We will perform a light transform to this dataset to modify it into the “long” format.

```
from lifelines.utils import to_long_format
base_df = to_long_format(base_df, duration_col="duration")
```

The new dataset looks like:

id | start | stop | var1 | event |
---|---|---|---|---|

1 | 0 | 10 | 0.1 | True |

2 | 0 | 12 | 0.5 | False |

2 rows × 5 columns

You’ll also have secondary dataset that references future measurements. This could come in two “types”. The first is when you have a variable that changes over time (ex: administering varying medication over time, or taking a tempature over time). The second types is an event-based dataset: an event happens at some time in the future (ex: an organ transplant occurs, or an intervention). We will address this second type later. The first type of dataset may look something like:

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 `add_covariate_to_timeline`

to fold the covariate dataset into the original dataset.

```
from lifelines.utils import add_covariate_to_timeline
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 track of them after time 12.

You may have multiple covariates you wish to add, so the above could be streamlined like so:

```
from lifelines.utils import add_covariate_to_timeline
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")
```

If your dataset is of the second type, that is, event-based, your dataset may look something like the following, where values in the matrix denote times since the subject’s birth, and `None`

or `NaN`

represent the event not happening (subjects can be excluded if the event never occurred as well) :

```
print(event_df)
id E1
0 1 1.0
1 2 NaN
2 3 3.0
...
```

Initially, this can’t be added to our baseline dataframe. However, using `utils.covariates_from_event_matrix`

we can convert a dataframe like this into one that can be easily added.

```
from lifelines.utils import covariates_from_event_matrix
cv = covariates_from_event_matrix(event_df, id_col="id")
print(cv)
event id duration E1
0 1 1.0 1
1 3 3.0 1
...
base_df = add_covariate_to_timeline(base_df, cv, duration_col="time", id_col="id", event_col="E")
```

For an example of pulling datasets like this from a SQL-store, and other helper functions, see Example SQL queries and transformations to get time varying data.

### Cumulative sums¶

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 the covariate to reflect how much we have administered since the start. Event columns do make sense to cumulative sum as well. In contrast, a covariate to measure the temperature of the patient is a state update, and should not be summed. See Example cumulative total using and time-varying covariates to see an example of this.

### Delaying time-varying covariates¶

`add_covariate_to_timeline`

also has an option for delaying, or shifting, a covariate so it changes later than originally observed. One may ask, why should one delay a time-varying covariate? Here’s an example. Consider investigating the impact of smoking on mortality and available to us are time-varying observations of how many cigarettes are consumed each month. Unbeknownst to us, when a subject reaches critical illness levels, they are admitted to the hospital and their cigarette consumption drops to zero. Some expire while in hospital. If we used this dataset naively, we would see that not smoking leads to sudden death, and conversely, smoking helps your health! This is a case of reverse causation: the upcoming death event actually influences the covariates.

To handle this, you can delay the observations by time periods:

```
from lifelines.utils import covariates_from_event_matrix
base_df = add_covariate_to_timeline(base_df, cv, duration_col="time", id_col="id", event_col="E", delay=14)
```

### Fitting the model¶

Once your dataset is in the correct orientation, we can use `CoxTimeVaryingFitter`

to fit the model to your data. The method is similar to `CoxPHFitter`

, expect we need to tell the `fit`

about the additional time columns.

Fitting the Cox model to the data involves using gradient descent. Lifelines takes extra effort to help with convergence, so please be attentive to any warnings that appear. Fixing any warnings will generally help convergence. For further help, see Problems with convergence in the Cox Proportional Hazard Model.

```
from lifelines import CoxTimeVaryingFitter
ctv = CoxTimeVaryingFitter()
ctv.fit(df, id_col="id", event_col="event", start_col="start", stop_col="stop", show_progress=True)
ctv.print_summary()
ctv.plot()
```

### Short note on prediction¶

Unlike the other regression models, prediction in a time-varying setting is not trivial. To predict, we would need to know the covariates values beyond the observed times, but if we knew that, we would also know if the subject was still alive or not! However, it is still possible to compute the hazard values of subjects at known observations, the baseline cumulative hazard rate, and baseline survival function. So while `CoxTimeVaryingFitter`

exposes prediction methods, there are logicial limitations to what these predictions mean.

## Model Selection in Survival Regression¶

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)

A fitted model’s concordance-index is present in the print_summary(), but also available under the score_ property. Generally, the measure is implemented in lifelines under lifelines.utils.concordance_index and accepts the actual times (along with any censorships) and the predicted times.

```
from lifelines import CoxPHFitter
from lifelines.datasets import load_rossi
rossi = load_rossi()
cph = CoxPHFitter()
cph.fit(rossi, duration_col="week", event_col="arrest")
# method one
cph.print_summary()
# method two
print(cph.score_)
# method three
from lifelines.utils import concordance_index
print(concordance_index(rossi['week'], -cph.predict_partial_hazard(rossi).values, rossi['arrest']))
```

However, there are other, arguably better, methods to measure the fit of a model. Included in print_summary is the log-likelihood, which can be used in an AIC calculation <https://en.wikipedia.org/wiki/Akaike_information_criterion>, and the log-likelihood ratio statistic <https://en.wikipedia.org/wiki/Likelihood-ratio_test>. Generally, I personally loved this article by Frank Harrell, “Statistically Efficient Ways to Quantify Added Predictive Value of New Measurements” <http://www.fharrell.com/post/addvalue/>.

### Cross Validation¶

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
```