EMPIRICAL LIKELIHOOD INFERENCE FOR THE ACCELERATED FAILURE TIME MODEL USING KENDALL ESTIMATING EQUATUION By Yinghua Lu June 29 th 2009 Georgia State University
EMPIRICAL LIKELIHOOD INFERENCE FOR THE
ACCELERATED FAILURE TIME MODEL USING KENDALL ESTIMATING EQUATUION
By Yinghua Lu
June 29th 2009
Georgia State University
Contents
Introduction
Main Procedure
Simulation Study
Real Application
Conclusion
Introduction – AFT Model
Accelerated Failure Time (AFT) Model: Very popular. Similar to the classic linear regression:
where Y=ln(T).
Different methods are developed OLS Non-monotone estimating equations Monotone estimating equations with normal approximation.
Introduction – Kendall’s Tau
Let {X1,Y1} and {X2, Y2} be two observations of two variables.
Kendall’s tau coefficient is defined as:
where nc is the number of [sign(X1-X2) = sign(Y1-Y2)], nd is the number of [sign(X1-X2) = -sign(Y1-Y2)].
Sen (1968) proposed
ε(b)=Y-bX
U(b) is non-increasing in b.
Introduction – Empirical Likelihood
A nonparametric method
Based on a data-driven likelihood ratio function
Without specifying a parametric family of distributions for the data.
The shape of confidence regions
Joins the reliability of the nonparametric methods and the efficiency of the likelihood methods.
Introduction – Empirical Likelihood
For X1,X2,…,Xn, the likelihood function is defined by
Let X1,X2,…,Xn be n independent samples, the empirical cumulative distribution (ECDF) at x is
The nonparametric likelihood of the CDF can be defined as
Introduction – Empirical Likelihood
Likelihood ratio:
Owen (2001) proved
Introduction – Brief History
Traced back to Thomas and Grunkemeier (1975)
Summarized and discussed in Owen (1988, 1990, 1991, 2001)
Qin and Jing (2001) and Li and Wang (2003): the limiting distribution EL ratio is a weighted chi-square distribution.
Zhou (2005) and Zhou and Li (2008): Logrank and Gehan estimators, and Buckley-James estimator.
Main Procedure – Preliminaries
Let T1,…,Tn be a sequence of random variables and Ti > 0. Let Z1,…,Zn be their corresponding covariates sequence.
Z and β are px1 vectors.
We observe and
Define
We employee the estimating equation as follow:
Main Procedure – Preliminaries
We can rewrite it as a U-statistic with symmetric kernel,
Similar to Fygenson and Ritov (1994),
where R and J are defined similarly in Fygenson and Ritov (1994).
Main Procedure – Preliminaries
The asymptotic variance of generalized estimate of β is
The numerator can be estimated by
The denominator can be estimated by
Then we can construct the confidence interval as
Main Procedure – Empirical Likelihood
Let and
Apply the idea of Sen (1960), we define
where W’s are independently distributed.
Main Procedure – Empirical Likelihood
Let be a probability vector. Then the empirical likelihood function at the value β is given by
For this function, reaches its maximum when
Thus, the empirical likelihood ratio at β is defined by
Main Procedure – Empirical Likelihood
By Lagrange Multiplier method for logarithm transformation of above equation, we write
Setting the partial derivative of G with respect to p to 0, we have
then
Main Procedure – Empirical Likelihood
Plug into the previous equation, we obtain
So, for all the p’s
We have
Main Procedure – Empirical Likelihood
Theorem 1 Under the above conditions, converges in distribution to , where is a chi-square random variable with p degrees of freedom.
Confidence region for β is given by
EL confidence region for the q sub-vector
Of
Theorem 2 Under the above conditions, converges in distribution to , where is a chi-square random variable with q degrees of freedom.
confidence region for is given by
Simulation Study – EL vs. NA
Consider the AFT model:
Model 1: (skewed error distribution)
Z ~ Uniform distribution in [-1, 1].
The censoring time C ~ Uniform distribution in [0, c], where c controls the censoring rate.
The error term has the standard extreme value distribution, which is skewed to the right.
Simulation Study – EL vs. NA
Model 2: (symmetric error distribution ).
Z ~ Uniform distribution in [0.5, 1.5].
The censoring time C is defined as 2exp(1)+c.
The error term has the standard Normal distribution N(0,1), which is symmetric.
Setting:
Repetition: 10000
Censoring Rate 15% 30% 45% 60%
Sample Size 30 50 75 100
Simulation Study – EL vs. NA
Results for model 1:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
15%
30CP* 0.8686 0.8986 0.9221 0.9427
AL** 1.4354 1.5751 1.7102 1.9030
50CP 0.8856 0.9084 0.9338 0.9516
AL 1.0857 1.1840 1.2936 1.4138
75CP 0.8880 0.9123 0.9405 0.9589
AL 0.8730 0.9500 1.0402 1.1412
100CP 0.8937 0.9152 0.9425 0.9607
AL 0.7520 0.8065 0.8960 0.9793
Simulation Study – EL vs. NA
Results for model 1:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
30%
30CP* 0.8669 0.8946 0.9163 0.9366
AL** 1.6870 1.8175 2.0101 2.1739
50CP 0.8768 0.8984 0.9272 0.9452
AL 1.2694 1.3635 1.5124 1.6253
75CP 0.8828 0.9057 0.9372 0.9515
AL 1.0218 1.1044 1.2174 1.3113
100CP 0.8911 0.9108 0.9418 0.9594
AL 0.8810 0.9479 1.0497 1.1352
Simulation Study – EL vs. NA
Results for model 1:
1-α=0.90 1-α=0.95CR n Wald EL Wald EL
45%
30CP*
0.8494 0.8720 0.9081 0.9188AL**
2.0324 2.1770 2.4216 2.5976
50CP
0.8699 0.8846 0.9233 0.9333AL
1.5241 1.5961 1.8160 1.9005
75CP
0.8798 0.8999 0.9336 0.9437AL
1.2293 1.2953 1.4647 1.5278
100CP
0.8879 0.9041 0.9394 0.9470AL
1.0555 1.1255 1.2576 1.3275
Simulation Study – EL vs. NA
Results for model 1:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
60%
30CP*
0.8136 0.8382 0.8760 0.8865AL**
2.6101 2.7787 3.1099 3.2616
50CP
0.8482 0.8492 0.9008 0.9028AL
1.9459 1.9870 2.3186 2.3588
75CP
0.8700 0.8669 0.9213 0.9162AL
1.5701 1.5890 1.8708 1.8744
100CP
0.8738 0.8807 0.9284 0.9275AL
1.3462 1.3824 1.6040 1.6210
Simulation Study – EL vs. NA
Results for model 2:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
15%
30CP*
0.8539 0.9082 0.9120 0.9504AL**
2.3067 2.4421 2.7485 2.8872
50CP
0.8753 0.9162 0.9293 0.9612AL
1.7432 1.8874 2.0770 2.2510
75CP
0.8850 0.9181 0.9374 0.9627AL
1.4096 1.5134 1.6795 1.8260
100CP
0.8880 0.9122 0.9409 0.9626AL
1.2158 1.2851 1.4486 1.5596
Simulation Study – EL vs. NA
Results for model 2:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
30%
30CP*
0.8520 0.9002 0.9065 0.9458AL**
2.4348 2.5429 2.9010 2.9912
50CP
0.8760 0.9063 0.9238 0.9514AL
1.8430 1.9749 2.1960 2.3563
75CP
0.8829 0.9125 0.9377 0.9606AL
1.4885 1.5895 1.7735 1.9085
100CP
0.8832 0.9091 0.9380 0.9587AL
1.2805 1.3528 1.5257 1.6343
Simulation Study – EL vs. NA
Results for model 2:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
45%
30CP*
0.8434 0.8828 0.8985 0.9328AL**
2.6690 2.8711 3.1801 3.4428
50CP
0.8698 0.8959 0.9207 0.9415AL
2.0308 2.1348 2.4197 2.5526
75CP
0.8804 0.9074 0.9326 0.9499AL
1.6297 1.7280 1.9418 2.0526
100CP
0.8875 0.9077 0.9367 0.9543AL
1.4077 1.4885 1.6773 1.7755
Simulation Study – EL vs. NA
Results for model 2:
1-α=0.90 1-α=0.95
CR n Wald EL Wald EL
60%
30CP*
0.8319 0.8634 0.8869 0.9093AL**
3.0160 3.1968 3.5935 3.7967
50CP
0.8705 0.8783 0.9179 0.9222AL
2.2770 2.3433 2.7130 2.7909
75CP
0.8818 0.8944 0.9300 0.9397AL
1.8232 1.8974 2.1723 2.2437
100CP
0.8808 0.8995 0.9330 0.9422AL
1.5662 1.6452 1.8661 1.9422
Simulation Study – EL vs. NA
Summary:
As the sample size increase, the coverage probabilities (CP) for both methods increase.
As the censoring rate increase, the coverage probabilities (CP) for both methods decrease.
When the sample size is small, the CP for EL is better than NA, for very heavy censoring rate, both are not good enough though.
Simulation Study – EL vs. NA
Summary:
Average length for the EL is a little longer than the NA in all cases.
A little over-coverage problem with the EL.
Under-coverage problem with the NA.
Simulation Study – Kendall vs. others
Consider the following AFT model:
We observe and
Model 3: Z ~ Normal distribution as N(1, 0.52). The censoring time C ~ Normal distribution as N(µ, 42),
where µ produce samples with censoring rate equal to 10%, 30%, 50%, 75%.
The error term has Normal distribution as N(0, 0.52). Sample Size: 50, 100 and 200 Repetition: 5000
Simulation Study – Kendall
Results for model 3:
Confidence Level = 90% Confidence Level = 95%
CR n B-J Logrank Gehan Kendall B-J Logrank Gehan Kendall
10%50 0.8924 0.8879 0.8832 0.9110 0.9406 0.9399 0.9356 0.9516
100 0.8888 0.8909 0.8904 0.9212 0.9404 0.9479 0.9446 0.9630
200 0.8810 0.9059 0.8938 0.9012 0.9458 0.9500 0.9446 0.9506
30%50 0.8866 0.8869 0.8804 0.9078 0.9374 0.9359 0.9290 0.9522
100 0.8936 0.8889 0.8870 0.9212 0.9472 0.9410 0.9382 0.9596
200 0.8922 0.9139 0.8958 0.9108 0.9468 0.9619 0.9440 0.9592
50%50 0.8838 0.8798 0.8650 0.8978 0.9324 0.9319 0.9226 0.9370
100 0.8926 0.8939 0.8820 0.9090 0.9414 0.9519 0.9370 0.9538
200 0.8952 0.8929 0.8968 0.9142 0.9482 0.9469 0.9424 0.9604
75%50 0.8420 0.8350 0.8030 0.8556 0.9042 0.8910 0.8628 0.8866
100 0.8818 0.8740 0.8536 0.8856 0.9344 0.9300 0.9118 0.9340
200 0.8928 0.8860 0.8788 0.9012 0.9438 0.9440 0.9358 0.9490
Simulation Study – Kendall
Results for model 3:
When the sample size is small (n=50) and the censoring rate is heavy, Kendall’s rank regression estimator is better an all the other estimators.
In other cases, Kendall’s rank regression estimator is also comparative.
Real Application
1. Bone marrow transplants are a standard treatment for acute leukemia.
2. Total of 137 patients were treated.
3. For simplicity, the model contains only one covariate at a time, which is where Ti is Time to Death.
4. The response variable Time to Death takes values from 1 day to 2640 days with mean equal to 839.16 days.
Real Application
We consider the following four variables:
1. Disease Group (3 groups)
2. Waiting Time to Transplant in Days (from 24 to 2616 days, mean=275 days)
3. Recipient and Donor Age (from 7 to 52 and from 2 to 56)
4. French-American-British (FAB): classification based on standard morphological criteria.
Real Application
FAB Group Age TimeToTrxβ_hat -0.8388 -0.4558 -0.4588 -0.2055
CI CI CI CI
Wald
1-α=0.90(-1.3485 -0.3290)
(-0.8415 -0.0700)
(-0.7770 -0.1406)
(-0.4379 0.0268)
Length 1.0195 0.7715 0.6364 0.4647
1-α=0.95(-1.4461 -0.2314)
(-0.9154 0.0039)
(-0.8379 -0.0797)
(-0.4823 0.0713)
Length 1.2147 0.9193 0.7582 0.5536
1-α=0.99(-1.6370 -0.0406)
(-1.0598 0.1483)
(-0.9571 0.0394)
(-0.5694 0.1583)
Length 1.5964 1.2081 0.9965 0.7277
EL
1-α=0.90(-1.3725 -0.2541)
(-0.8626 -0.0382)
(-0.9318 -0.1718)
(-0.4904 -0.0134)
Length 1.1184 0.8244 0.7600 0.477
1-α=0.95(-1.4933 -0.1240)
(-0.9442 0.0390)
(-1.0503 -0.1087)
(-0.5728 0.0211)
Length 1.3693 0.9832 0.9416 0.5939
1-α=0.99(-1.6189 0.1328)
(-1.1249 0.2233)
(-1.2549 0.0054)
(-0.7332 0.0774)
Length 1.7517 1.3482 1.2603 0.8106
Real Application
Results:1. Two methods show similar results.
2. Two exceptions may due to asymmetric CI of the EL.
3. Average lengths of the EL are a little longer than that of the NA. Same results with the simulation study.
Conclusion & Discussion
Average length of the CI by the EL are slightly longer than that by NA.
The coverage probabilities of the EL are closer to the nominal levels than NA, especially when the sample size is very small and censoring rate is heavy.
Kendall’s rank regression estimator is better than the Buckley-James, Logrank and Gehan estimators in terms of coverage probabilities.
Conclusion & Discussion
The combination of the Kendall estimating equation and the EL CI has strong advantages over the other considered approaches in the case of small sample size and heavy censoring rate.
The combination shows a problem of over-coverage.
A smoothing kernel is suggested to eliminate such a problem in the future work.
Thank you !