def test_incidence():
# Check estimates in R:
# ftime = c(1, 1, 2, 4, 4, 4, 6, 6, 7, 8, 9, 9, 9, 1, 2, 2, 4, 4)
# fstat = c(1, 1, 1, 2, 2, 2, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0)
# cuminc(ftime, fstat)
#
# The standard errors agree with Stata, not with R (cmprisk
# package), which uses a different SE formula from Aalen (1978)
#
# To check with Stata:
# stset ftime failure(fstat==1)
# stcompet ci=ci, compet1(2)
ftime = np.r_[1, 1, 2, 4, 4, 4, 6, 6, 7, 8, 9, 9, 9, 1, 2, 2, 4, 4]
fstat = np.r_[1, 1, 1, 2, 2, 2, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0]
ci = CumIncidenceRight(ftime, fstat)
cinc = [
np.array([
0.11111111, 0.17037037, 0.17037037, 0.17037037, 0.17037037,
0.17037037, 0.17037037
]),
np.array([
0., 0., 0.20740741, 0.20740741, 0.20740741, 0.20740741, 0.20740741
]),
np.array([0., 0., 0., 0.17777778, 0.26666667, 0.26666667, 0.26666667])
]
assert_allclose(cinc[0], ci.cinc[0])
assert_allclose(cinc[1], ci.cinc[1])
assert_allclose(cinc[2], ci.cinc[2])
cinc_se = [
np.array([
0.07407407, 0.08976251, 0.08976251, 0.08976251, 0.08976251,
0.08976251, 0.08976251
]),
np.array([
0., 0., 0.10610391, 0.10610391, 0.10610391, 0.10610391, 0.10610391
]),
np.array([0., 0., 0., 0.11196147, 0.12787781, 0.12787781, 0.12787781])
]
assert_allclose(cinc_se[0], ci.cinc_se[0])
assert_allclose(cinc_se[1], ci.cinc_se[1])
assert_allclose(cinc_se[2], ci.cinc_se[2])
# Simple check for frequency weights
weights = np.ones(len(ftime))
ciw = CumIncidenceRight(ftime, fstat, freq_weights=weights)
assert_allclose(ci.cinc[0], ciw.cinc[0])
assert_allclose(ci.cinc[1], ciw.cinc[1])
assert_allclose(ci.cinc[2], ciw.cinc[2])
def test_kernel_cumincidence2():
# cases with tied times
n = 100
np.random.seed(3434)
x = np.random.normal(size=(n, 3))
time = np.random.randint(0, 10, size=n)
status = np.random.randint(0, 3, size=n)
CumIncidenceRight(time, status, exog=x, bw_factor=10000)
def test_kernel_cumincidence1():
# Check that when the bandwidth is very large, the kernel
# procedure agrees with standard cumulative incidence
# calculations. (Note: the results do not agree perfectly when
# there are tied times).
n = 100
np.random.seed(3434)
x = np.random.normal(size=(n, 3))
time = np.random.uniform(0, 10, size=n)
status = np.random.randint(0, 3, size=n)
result1 = CumIncidenceRight(time, status)
for dimred in False, True:
result2 = CumIncidenceRight(time, status, exog=x, bw_factor=10000,
dimred=dimred)
assert_allclose(result1.times, result2.times)
for k in 0, 1:
assert_allclose(result1.cinc[k], result2.cinc[k], rtol=1e-5)
def test_incidence2():
# Check that the cumulative incidence functions for all competing
# risks sum to the complementary survival function.
np.random.seed(2423)
n = 200
time = -np.log(np.random.uniform(size=n))
status = np.random.randint(0, 3, size=n)
ii = np.argsort(time)
time = time[ii]
status = status[ii]
ci = CumIncidenceRight(time, status)
statusa = 1 * (status >= 1)
sf = SurvfuncRight(time, statusa)
x = 1 - sf.surv_prob
y = (ci.cinc[0] + ci.cinc[1])[np.flatnonzero(statusa)]
assert_allclose(x, y)