Misclassification simulation extrapolation method for a Weibull accelerated failure time model

Abstract

The problem of misclassification in covariates is ubiquitous in survival data and often leads to biased estimates. The misclassification simulation extrapolation method is a popular method to correct this bias. However, its impact on Weibull accelerated failure time models has not been studied. In this paper, we study the bias caused by misclassification in one or more binary covariates in Weibull accelerated failure time models and explore the use of the misclassification simulation extrapolation in correcting for this bias, along with its asymptotic properties. Simulation studies are carried out to investigate the numerical properties of the resulting estimator for finite samples. The proposed method is then applied to colon cancer data obtained from the cancer registry at Memorial Sloan Kettering Cancer Center.

Keywords

Accelerated failure time model measurement error misclassification survival analysis weibull distribution

Introduction

Misclassification of binary covariates occurs frequently for survival analysis in epidemiological data. For example, dietary intake is misclassified based on food frequency questionnaires;^1,2 diagnosis of dental caries may be misclassified due to geographical differences in clinical practice;³ socioeconomic status may be misclassified based on recall data as opposed to the use of historical records;⁴ prenatal smoking status may be misclassified in gestational studies.⁵ It is well known that the analysis based on the misclassified covariates will lead to biased estimation, especially when the misclassification rate is high.

Several methods have been investigated for handling misclassification for survival data analysis, such as probabilistic bias analysis,⁶ weighted least squares method, corrected score function method, pseudo-partial likelihood method,^7–9 pooled estimation method^10,11 the regression calibration approach^11,12 and misclassification simulation extrapolation (MC-SIMEX) method. Additionally, purely numerical studies have also been conducted.¹³ Among these methods, the MC-SIMEX method gained popularity recently due to its flexibility, easy implementation, robustness and minimal assumption requirement.¹⁴ It is an extension of the simulation extrapolation (SIMEX) method¹⁵ for measurement error in continuous covariates and first proposed by Kuchenhoff et al.³ for regression models for complete data. Then it is studied for different models in survival data analysis. Bang et al.,¹¹ studied the performance of the MC-SIMEX method in a Cox model. The performance of MC-SIMEX in log-normal accelerated failure time (AFT) models was studied by Slate and Bandyopadhyay¹⁴ and in log-logistic AFT models by Sevilimedu et al.¹⁶ However, it has not been explored for Weibull AFT model although the Weibull distribution is a very popular distribution to model monotonic hazard functions in survival data analysis. In addition, none of the existing works^11,14,16 provided details about the asymptotic properties of the MC-SIMEX estimator in the corresponding survival models. Furthermore, they did not investigate the form of the extrapolation function used in the MC-SIMEX method.

Our work is motivated by colon cancer data from the Memorial Sloan Kettering Cancer Center (MSKCC) registry, where the survival times of patients may follow a Weibull distribution. We are interested in the effect of misclassification caused by faulty categorization of pathological nodal (pathn) status, which plays a pivotal role in determining survival in colon cancer patients. In this paper, we explore the performance of the MC-SIMEX³ method for misclassification correction in Weibull survival data, while also investigating its asymptotic properties and the form of the extrapolation function. These asymptotic properties of the MC-SIMEX estimator and the form of the extrapolation function are applicable to general AFT models, in which the data can follow any specified distributions. In the simulation studies, we also investigate the robustness of the proposed method to different scenarios such as misspecification of survival time distribution, dependent censoring, presence of multiple categories in covariates and the presence of multiple covariates that may be subject to misclassification.

2. Methodology

2.1. Weibull AFT model and misclassification matrix

Let T_i and C_i denote the failure and censoring times respectively. The observations are ${t_{i}, δ_{i}}, i = 1, \dots, n,$ in which $t_{i} = min (T_{i}, C_{i})$ is the observed survival time and $δ_{i} = I (T_{i} \leq C_{i})$ is the censoring indicator. We assume that Y_i, which is $l o g (T_{i})$ is linearly associated with covariates X_i and $Z_{i}$ . The Weibull AFT model can then be written as

\begin{aligned} Y_{i} = β_{0} + β_{x} X_{i} + β_{z} Z_{i} + ε_{i} / σ \end{aligned}

(1)

where

σ

is the scale parameter and ε represents random error which follows standard extreme value distribution. We denote the parameter vector in model (1) by

θ = (β_{0}, β_{x}, β_{z}, σ)^{'}

. The binary covariate

X_{i}

is the variable that is subject to non-differential misclassification and Z _i is the vector of correctly measured confounding variables. Here we only include one misclassified covariate

X_{i}

for clarity of the notations and convenience. More misclassified covariates can be added to model (1) and their inference procedure are same as the following inference procedure for

X_{i}

. The observed version of covariate

X_{i}

, which is denoted by

W_{i}

, is related to

X_{i}

through misclassification matrix

Π

, which is defined as:

Π = [\begin{array}{cc} π_{00} & 1 - π_{11} \\ 1 - π_{00} & π_{11} \end{array}]

where

π_{00}

and

π_{11}

are the specificity and sensitivity of classification respectively.

When the true covariate X_i is observed, the likelihood function of the $i th$ observation can be written as:¹⁷

L_{i} (θ; t_{i}, X_{i}, Z_{i}) = [f_{T} (t_{i}; X_{i}, Z_{i}, θ)]^{δ_{i}} [1 - F_{T} (t_{i}; X_{i}, Z_{i}, θ)]^{1 - δ_{i}},

where

f_{T} (t_{i}) = \frac{1 / σ}{\exp (β_{x} X_{i} + β_{z} Z_{i})} {[\frac{t_{i}}{\exp (β_{x} X_{i} + β_{z} Z_{i})}]}^{\frac{1}{σ} - 1} \exp [- {(\frac{t_{i}}{\exp (β_{x} X_{i} + β_{z} Z_{i})})}^{\frac{1}{σ}}],

and

F_{T} (t_{i}) = 1 - \exp [- {(\frac{t_{i}}{\exp (β_{x} X_{i} + β_{z} Z_{i})})}^{\frac{1}{σ}}],

are the density function and distribution function for T respectively. Then the log-likelihood function for the data is

l (θ; t_{i}, X_{i}, Z_{i}) = \sum_{i = 1}^{n} log (L_{i} (θ; t_{i}, X_{i}, Z_{i})) .

The maximum likelihood estimator of $θ$ is then derived by solving

U (θ) = \partial l (θ; t_{i}, X_{i}, Z_{i}) / \partial θ = 0.

The use of the observed covariate $W_{i}$ instead of $X_{i}$ in the Weibull AFT model can result in a biased estimate for $β_{x}$ , which is most likely attenuated.¹⁸ That is, the estimated (attenuated) effect size— $exp (β_{x})$ , is more likely to be closer to 1.

2.2. Misclassification simulation extrapolation

Here, we provide an overview of the MC-SIMEX procedure for the estimation of $β_{x}$ . The general idea of the MC-SIMEX algorithm is to increase the amount of noise in a misclassified covariate W through the use of an error inflation factor λ, estimate the coefficient at each level of added noise, and then extrapolate to a point where there is presumably no misclassification in the observed covariate W. The parametric form of this extrapolation function can be written as:

λ \to {\hat{β}}_{x} \approx D (1 + λ, Γ)

where D is the extrapolation function and

Γ

is the corresponding vector of parameters.³ We now briefly describe the individual components of MC-SIMEX namely, simulation and extrapolation.

Simulation component: For a fixed set of values of $λ$ , namely, $Λ = {λ_{1} \dots . . λ_{m}}$ , B pseudo datasets are generated at each level of $λ$ by using the misclassification equation,

W_{b, i} (λ_{k}) = M C (Π^{λ_{k}}) W_{i},

where

i = 1 \dots . n, k = 1 \dots . m and b = 1 \dots . B

. In other words, for a fixed value b,

W_{b, i} (λ_{k})

are generated from

binom (n, p_{2})

where

p_{2} = 1 - π_{00} + p_{1} δ^{λ_{k}}

p_{1}

represents the mean of

W_{i}

δ

is the determinant of the misclassification matrix

Π

and

λ_{k}

is the error inflation factor. The misclassification equation

M C (Π^{λ_{k}}) W_{i}

represents the simulation of a given variable

W_{i}

with a misclassification matrix

Π^{λ_{k}}

, which represents the misclassification matrix raised to the power of

λ_{k}

where

λ_{k}

is the error inflation factor which is greater than 0. Let

{\hat{θ}}_{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})

denote the maximum likelihood estimator of

θ (λ_{k})

for the

b^{t h}

pseudo dataset and the corresponding variance estimate be denoted by

{\hat{V}}_{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})

. Then the naïve estimator

{\hat{θ}}_{n a} (λ_{k})

at a level of misclassification

λ_{k}

is obtained as

{\hat{θ}}_{n a} (λ_{k}) = B^{- 1} \sum_{b = 1}^{B} [{\hat{θ}}_{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})] .

Extrapolation component: Let ${\hat{β}}_{x}^{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})$ denote the component in ${\hat{θ}}_{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})$ for $β_{x} (t_{i}, W_{b, i} (λ_{k}), Z_{i})$ , ${\hat{V}}_{x}^{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})$ denote the corresponding variance component in ${\hat{V}}_{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})$ and ${\hat{β}}_{x}^{n a} (λ_{k})$ denote the component in ${\hat{θ}}_{n a} (λ_{k})$ for the naïve estimator of $β_{x} (λ_{k})$ . A parametric regression model $D (1 + λ, Γ)$ is fitted by regressing the dependent variable ${\hat{β}}_{x}^{n a} (λ_{k})$ on the independent variable $1 + λ$ . The least squares method is used to obtain the estimator $\hat{Γ}$ of the coefficient vector $Γ$ . ${\hat{β}}_{x}^{s i m e x}$ is then obtained by extrapolating the naïve estimates to a point where misclassification in the binary covariates does not exist, i.e., where $λ = - 1$ . That is,

{\hat{β}}_{x}^{s i m e x} = D (0, \hat{Γ}) .

To calculate the variance of

{\hat{β}}_{x}^{s i m e x}

, we first calculate the sample variance of each of the naïve estimates

{\hat{β}}_{x}^{n a} (λ_{k})

{\hat{V}}_{x}^{s a m} (λ_{k}) = B^{- 1} \sum_{b = 1}^{B} ({\hat{β}}_{x}^{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i}) - {\bar{β}}_{x}^{n a} (λ_{k}))^{2} .

where

{\bar{β}}_{x}^{n a}

is the sample mean of

{\hat{β}}_{x}^{n a} (t_{i}, W_{b, i} (λ_{k}), Z_{i})

, with

{\hat{V}}_{x}^{s a m} (0) = 0

. Next, the variance of each naïve estimator

{\hat{β}}_{x}^{n a} (λ_{k})

is obtained as

{\hat{V}}_{x}^{n a} (λ_{k}) = B^{- 1} \sum_{b = 1}^{B} {\hat{V}}_{x}^{n a} [(t_{i}, W_{b, i} (λ_{k}), Z_{i})]),

where

{\hat{V}}_{x}^{n a} [(t_{i}, W_{b, i} (λ_{k}), Z_{i})])

represents the variance of the naïve estimator in the

b th

replication and is obtained using the information matrix.

{\hat{V}}_{x}^{n a} (λ_{k})

is the average of the variances for the naïve estimators across B replications. We then calculate the variance

{\hat{V}}_{x}^{S T}

{\hat{β}}_{x}^{s i m e x}

by utilizing the method proposed by Stefanski and Cook,¹⁹ which extrapolates the difference between

{\hat{V}}_{x}^{n a} (λ_{k})

and

{\hat{V}}_{x}^{s a m} (λ_{k})

λ = - 1

,¹⁸

{\hat{V}}_{x}^{S T} = lim_{λ \to - 1} ({\hat{V}}_{x}^{n a} (λ) - {\hat{V}}_{x}^{s a m} (λ)) .

3. Properties of the Mc-SIMEX estimator

In this section, we will provide the properties of the MC-SIMEX estimator. They are applicable to general AFT models, in which the survival time can follow any specified distributions. The Weibull AFT model, in which the survival time T is specified to follow Weibull distribution, is one of the general AFT models.

3.1. Asymptotic properties of ${\hat{β}}_{x}^{s i m e x}$

Under regularity conditions for the maximum likelihood estimator, the MC-SIMEX estimator, ${\hat{β}}_{x}^{s i m e x}$ , can be shown to be consistent and asymptotically normally distributed. It is evident that ${\hat{θ}}_{n a} (λ)$ is a consistent estimator of $θ_{n a} (λ)$ based on the theory of maximum likelihood estimation and Slutsky's theorem. In addition, we know that $\hat{Γ}$ is a consistent estimator of $Γ$ based on the least squares estimation theory. Suppose that $D (1 + λ, Γ)$ is the true extrapolation function, then ${\hat{β}}_{x}^{s i m e x}$ is consistent for $β_{x}$ based on the continuous mapping theory. Asymptotic normality can be proven along the lines of proof provided by He et al.²⁰ The authors provide proof for asymptotic normality of the SIMEX estimator which corrects for measurement error in continuous variables in an AFT model. However, the same theory can also be applied to MC-SIMEX which handles categorical variables with misclassification, because both SIMEX and MC-SIMEX use maximum likelihood theory for estimation in the simulation step and least squares method in the extrapolation step. Note that in the simulation step, the proof only depends on the maximum likelihood estimator of the simulated data but does not depend on how the simulated data is generated. Therefore, it can be stated that

\sqrt{n} ({\hat{β}}_{x}^{s i m e x} - β_{x}) \sim N (0, D_{Γ}^{'} (- 1; Γ) Q (Γ) D_{Γ} (- 1, Γ)),

where

D_{Γ} (λ; Γ) = \partial D^{'} (λ; Γ) / \partial Γ

;

Q (Γ) = [A (Γ)]^{- 1} S (Γ) Ω_{x} (Λ; θ) S^{'} (Γ) [A (Γ)]^{- 1}

, in which

S (Γ) = (D_{Γ} (λ_{1}; Γ), \dots, D_{Γ} (λ_{m}; Γ))

A (Γ) = S (Γ) S^{'} (Γ)

Ω_{x} (Λ; θ) = Ω (Λ; θ) [1 : m, 1 : m]

is a submatrix of

Ω (Λ; θ)

formed by first m row and first m column and

Ω (Λ; θ) = [V (Λ; θ)]^{- 1} Σ (Λ; θ) [V (Λ; θ)]^{- 1} .

V (Λ; θ)

is the diagonal matrix of

E (\partial U_{i} (θ (λ); 1, λ) / \partial θ), λ = 1, \dots, m,

where

U_{i} (θ (λ); b, λ) = \partial log (L_{i} (θ; t_{i}, W_{b, i} (λ), Z_{i})) / \partial θ

and

Σ (Λ; θ)

is the covariance matrix of the vector

U_{i} (θ (λ); B, λ) = B^{- 1} \sum_{b = 1}^{B} U_{i} (θ (λ); b, λ), λ = 1, \dots, m

3.2. The form of extrapolation function in Mc-SIMEX method

To investigate the form of the extrapolation function in the MC-SIMEX method, we prove the relationship between the naïve parameter $β_{x}^{n a}$ and true parameter $β_{x}$ for general AFT models, which can be shown to be (proof provided in appendix):

\begin{aligned} β_{x}^{n a} = β_{x} \frac{δ (1 - π_{x}) π_{x}}{(1 - π_{00} + δ π_{x}) (π_{00} - δ π_{x})} \end{aligned}

(2)

where

δ

is the determinant of the misclassification matrix

Π

, given by

π_{00} + π_{11} - 1.

The form of the extrapolation function can be visually evaluated by fitting a regression line for the

β_{x}^{n a}

(λ) over various values of λ with the aid of spectral decomposition. Spectral decomposition allows us to estimate the misclassification matrix at a particular level of error inflation

λ

as follows:

Π^{λ} = \frac{1}{1 - δ} [\begin{array}{cc} 1 - π_{11} + (1 - π_{00}) δ^{λ} & (1 - π_{11}) (1 - δ^{λ}) \\ (1 - π_{00}) (1 - δ^{λ}) & 1 - π_{00} + (1 - π_{11}) δ^{λ} \end{array}] .

The components of $Π^{λ}$ can then be inserted into (2) to obtain the relationship between $β_{x}^{n a}$ (λ) and $β_{x}$ at the specified value of λ. Figure 1 shows the function $β_{x}^{n a}$ (λ), which is calculated based on equation (2), for different values of λ. It shows that the quadratic extrapolation function is appropriate for MC-SIMEX method under general AFT models.

4. Simulation studies

4.1. Settings

In this section, we compare the new proposed estimator, referred to as the Weibull-Q MC-SIMEX estimator (Q denoting a quadratic form of extrapolation), with the Cox MC-SIMEX estimator proposed by Bang et al.,¹¹ Weibull-L MC-SIMEX estimator (L denoting linear form of extrapolation), Weibull-C MC-SIMEX estimator (C denoting cubic form of extrapolation), true estimator which uses the correct value of the covariate and the naïve estimator which does not correct for misclassification for finite samples. We consider a wide range of scenarios. The objective of the first simulation study is to assess the performance of the Weibull-Q MC-SIMEX estimator with different values of scale parameter ( $σ$ ) and varying sample sizes. In the subsequent simulation studies, we investigate the robustness of the proposed estimator to misspecification of the distribution of survival times, the presence of dependent censoring (the distribution of censoring time is dependent upon a covariate in the model), the existence of multiple categories in the misclassified covariate, and presence of two categorical covariates measured with error.

For the proposed Weibull-Q MC-SIMEX estimator, Cox MC-SIMEX estimator, Weibull-L MC-SIMEX estimator and Weibull-C MC-SIMEX estimator, a total of 50 pseudo-datasets are generated (B) at each level of error inflation with $λ = {0.5, 1, 1.5, 2}$ . The covariate Z is generated from a standard normal distribution. For each setting, a total of 1000 simulated datasets are analyzed. The rate of censoring is adjusted to 30% and two sample sizes, n = 500 or 1000 are considered. We present the results for n = 500 in the main article and for n = 1000 in the supplementary materials. We report bias, empirical variance, estimated variance, and coverage probability of the 95% confidence interval for each method. The R code for the Weibull MC-SIMEX method is available from the first author upon request

4.2. One misclassified covariate

In the first set of simulation studies, we investigate the performance of the Weibull-Q MC-SIMEX estimator in the following AFT model:

\begin{aligned} Y_{i} = β_{0} + β_{x} X_{i} + β_{z} Z_{i} + ε_{i} / σ \end{aligned}

(3)

where σ is the scale parameter and ε represents the error term which follows standard extreme value distribution and survival time follows Weibull distribution. The true values of the binary covariate X are generated from a random binomial distribution with a mean of 0.5. It is subject to non-differential misclassification with the misclassification matrix

[\begin{array}{cc} 0.9 & 0.3 \\ 0.1 & 0.7 \end{array}]

and

Z_{i}

is the correctly measured confounding variable. The values of

β_{x}

and

β_{z}

are both set to

log (2)

, while the value of

β_{0}

is set to 0.

Y_{i}

is the log of survival time. The scale parameters σ are set as 0.5, 1.5, or 2. The censoring time is generated from uniform (0, upper), where the upper bound is adjusted to yield 30% censoring. In this section, it is assumed that censoring is independent of covariates. The results for n = 500 are shown in Table 1.

Table 1.

Table showing the performance of the true $({\hat{β}}_{t r u e})$ , naïve $({\hat{β}}_{n a i v e})$ , Cox misclassification simulation extrapolation (MC-SIMEX) $({\hat{β}}_{C O X - M C})$ , Weibull-Q MC-SIMEX $({\hat{β}}_{W Q - M C})$ , Weibull-L MC-SIMEX $({\hat{β}}_{W L - M C})$ and Weibull-C MC-SIMEX $({\hat{β}}_{W C - M C})$ estimators with misclassified covariate X.

	${\hat{β}}_{t r u e}$	${\hat{β}}_{n a i v e}$	${\hat{β}}_{C O X - M C}$	${\hat{β}}_{W Q - M C}$	${\hat{β}}_{W L - M C}$	${\hat{β}}_{W C - M C}$
σ = 0.5
Estimated variance	0.048	0.052	0.053	0.064	0.064	0.064
Empirical variance	0.050	0.114	0.068	0.068	0.037	2.793
Bias	0.002	−0.248	−0.075	−0.071	−0.183	0.026
Coverage	0.949	0.808	0.878	0.880	0.967	0.211
σ = 1.5
Estimated variance	0.005	0.007	0.010	0.008	0.008	0.008
Empirical variance	0.005	0.083	0.026	0.018	0.041	0.380
Bias	−0.002	−0.275	−0.135	−0.095	−0.200	−0.042
Coverage	0.945	0.088	0.662	0.720	0.249	0.204
σ = 2
Estimated variance	0.003	0.004	0.011	0.005	0.005	0.005
Empirical variance	0.003	0.084	0.036	0.018	0.046	0.264
Bias	0.002	−0.283	−0.177	−0.109	−0.214	−0.094
Coverage	0.951	0.008	0.581	0.597	0.088	0.208

Sample size is set to 500. Scale parameter is denoted by σ.

From the perspective of bias and empirical variance, the true estimator generally has the smallest value as it uses the true covariate values, while the naïve estimator has the largest value as it uses the wrong covariate values. Both the Weibull-Q MC-SIMEX estimator and the Cox MC-SIMEX estimator have smaller bias and empirical variance than the naïve estimator, which shows that the MC-SIMEX method demonstrates a much-improved performance. The Weibull-Q MC-SIMEX estimator has a smaller bias and empirical variance than the Cox MC-SIMEX estimator, which indicates Weibull-Q MC-SIMEX method has a better correction for the misclassification. The estimated variance is usually smaller than the empirical variance for the Weibull-Q MC-SIMEX method and the Cox MC-SIMEX method, especially when the scale parameter is large. This may be due to the unknown extrapolation function for the variance. In terms of the coverage probability, the true estimator has coverage close to the nominal level as expected. The Weibull-Q MC-SIMEX method always has a better coverage probability than the Cox MC-SIMEX method.

Overall, the Weibull-Q MC-SIMEX method performs better than the Cox MC-SIMEX method in terms of bias, empirical variance, and coverage probabilities. Comparing the Weibull MC-SIMEX estimator with different extrapolation functions, the Weibull-C MC-SIMEX estimator always has the smallest bias and the Weibull-L MC-SIMEX estimator always has the largest bias. The Weibull-C MC-SIMEX estimator always has a much larger empirical variance than the other two estimators. The Weibull-Q MC-SIMEX estimator is better than the Weibull-L MC-SIMEX estimator except for the case with a small-scale parameter. As for the coverage probability, the Weibull-Q MC-SIMEX estimator performs robustly well than the other two estimators, except that it has a slightly smaller coverage than the Weibull-L MC-SIMEX estimator for the case with a small scale parameter. This indicates that the quadratic extrapolation function provides a good compromise from the perspective of bias-variance trade-off, which is supported by the findings in Section 3.2.

4.3. Misspecified distribution of survival times

In this setting, we explore the performance of the proposed estimator when the distribution of survival time is misspecified as a Weibull distribution while its original distribution is exponential, log-normal or log-logistic. The survival times for the mentioned distributions are generated using model (3) with scale parameter σ = 1. We simulate the log normally distributed survival data by generating the error term from the normal distribution (0,1) or normal distribution (0,2). The log-logistic survival data is simulated by generating the error terms from the logistic distribution (0,1) or logistic distribution (0,2). The censoring times are generated in a similar manner to that described in Section 4.2. The true values of the binary covariate X are generated from a random binomial distribution with a mean of 0.5. The results are given in Table 2.

Table 2.
Table showing the performance of the true $({\hat{β}}_{t r u e})$ , naïve $({\hat{β}}_{n a i v e})$ , Cox misclassification simulation extrapolation (MC-SIMEX) $({\hat{β}}_{C O X - M C})$ , Weibull-Q MC-SIMEX $({\hat{β}}_{W Q - M C})$ , Weibull-L MC-SIMEX $({\hat{β}}_{W L - M C})$ and Weibull-C MC-SIMEX $({\hat{β}}_{W C - M C})$ estimators with misclassified covariate X, when the true underlying distribution is misspecified as a Weibull distribution. Sample size is set to 500.

${\hat{β}}_{t r u e}$ ${\hat{β}}_{n a i v e}$ ${\hat{β}}_{C O X - M C}$ ${\hat{β}}_{W Q - M C}$ ${\hat{β}}_{W L - M C}$ ${\hat{β}}_{W C - M C}$

Exponential

Estimated variance 0.012 0.013 0.015 0.016 0.016 0.016

Empirical variance 0.013 0.082 0.025 0.023 0.037 0.788

Bias 0.004 −0.261 −0.091 −0.073 −0.190 −0.069

Coverage 0.943 0.389 0.798 0.829 0.763 0.231

$L N (0, 1)$

Estimated variance 0.011 0.013 0.019 0.016 0.016 0.016

Empirical variance 0.015 0.100 0.023 0.032 0.051 0.755

Bias −0.050 −0.295 −0.035 −0.125 −0.223 −0.129

Coverage 0.916 0.270 0.880 0.744 0.490 0.223

$L N (0, 2)$

Estimated variance 0.036 0.039 0.019 0.047 0.047 0.047

Empirical variance 0.044 0.136 0.136 0.073 0.063 2.134

Bias −0.080 −0.311 −0.339 −0.152 −0.245 −0.131

Coverage 0.923 0.637 0.306 0.816 0.901 0.233

$L L (0, 1)$

Estimated variance 0.028 0.029 0.019 0.036 0.036 0.036

Empirical variance 0.031 0.119 0.091 0.051 0.055 1.636

Bias −0.068 −0.304 −0.267 −0.123 −0.229 −0.077

Coverage 0.935 0.566 0.467 0.829 0.882 0.214

$L L (0, 2)$

Estimated variance 0.096 0.101 0.019 0.120 0.120 0.120

Empirical variance 0.106 0.186 0.228 0.148 0.065 5.250

Bias −0.077 −0.297 −0.454 −0.101 −0.238 −0.040

Coverage 0.922 0.851 0.135 0.863 0.971 0.211

	${\hat{β}}_{t r u e}$	${\hat{β}}_{n a i v e}$	${\hat{β}}_{C O X - M C}$	${\hat{β}}_{W Q - M C}$	${\hat{β}}_{W L - M C}$	${\hat{β}}_{W C - M C}$
Exponential
Estimated variance	0.012	0.013	0.015	0.016	0.016	0.016
Empirical variance	0.013	0.082	0.025	0.023	0.037	0.788
Bias	0.004	−0.261	−0.091	−0.073	−0.190	−0.069
Coverage	0.943	0.389	0.798	0.829	0.763	0.231
$L N (0, 1)$
Estimated variance	0.011	0.013	0.019	0.016	0.016	0.016
Empirical variance	0.015	0.100	0.023	0.032	0.051	0.755
Bias	−0.050	−0.295	−0.035	−0.125	−0.223	−0.129
Coverage	0.916	0.270	0.880	0.744	0.490	0.223
$L N (0, 2)$
Estimated variance	0.036	0.039	0.019	0.047	0.047	0.047
Empirical variance	0.044	0.136	0.136	0.073	0.063	2.134
Bias	−0.080	−0.311	−0.339	−0.152	−0.245	−0.131
Coverage	0.923	0.637	0.306	0.816	0.901	0.233
$L L (0, 1)$
Estimated variance	0.028	0.029	0.019	0.036	0.036	0.036
Empirical variance	0.031	0.119	0.091	0.051	0.055	1.636
Bias	−0.068	−0.304	−0.267	−0.123	−0.229	−0.077
Coverage	0.935	0.566	0.467	0.829	0.882	0.214
$L L (0, 2)$
Estimated variance	0.096	0.101	0.019	0.120	0.120	0.120
Empirical variance	0.106	0.186	0.228	0.148	0.065	5.250
Bias	−0.077	−0.297	−0.454	−0.101	−0.238	−0.040
Coverage	0.922	0.851	0.135	0.863	0.971	0.211

$L N$ is used to denote log normal distribution and $L L$ is used to denote log logistic distribution.

For exponential survival data, the Weibull-Q MC-SIMEX method generally outperforms the Cox MC-SIMEX method, the Weibull-L MC-SIMEX method, and the Weibull-C MC-SIMEX method in terms of bias, empirical variance, and coverage probability. However, it has a slightly larger bias than the Weibull-C MC-SIMEX method. The estimated variance is usually smaller than the empirical variance for the MC-SIMEX-related methods.

For log-normal survival data, the performance of the Cox MC-SIMEX estimator is not stable. When the data variance is small, i.e., normal (0,1), it performs very well, that is, it has the smallest bias (even smaller than the true estimator), smaller empirical variance and better coverage probability than the Weibull-Q MC-SIMEX estimator. However, when the data variance is large, i.e., normal (0,2), it performs very poorly, that is, it has the largest bias (even larger than the naïve estimator), much larger empirical variance, and worse coverage probability than the Weibull-Q MC-SIMEX estimator. However, the performance of the Weibull-Q MC-SIMEX method is reasonably well and robust Similar to the results in Section 4.2, the estimated variance is usually smaller than the empirical variance for the Cox MC-SIMEX estimator and the Weibull-Q MC-SIMEX estimator due to the unknown extrapolation function of the variance. The Weibull-Q MC-SIMEX method always has smaller bias than the Weibull-L MC-SIMEX method and comparable bias to the Weibull-C MC-SIMEX method. It has a smaller empirical variance and better coverage probability than the Weibull-L MC-SIMEX method for large samples (n = 1000) and small sample (n = 500) with small variance and the Weibull-C MC-SIMEX method. This again supports the finding in Section 3.2.

For log-logistic survival data, the Cox MC-SIMEX estimator performs worse than for log normal survival data in terms of bias, empirical variance and coverage probability, while the Weibull-Q MC-SIMEX method still performs reasonably well. The Weibull-L MC-SIMEX method has a larger bias than the Weibull-Q MC-SIMEX method, although their performances are similar in terms of empirical variance and coverage probability. Although the Weibull-C MC-SIMEX method has a smaller bias than the Weibull-Q MC-SIMEX method, it has much worse empirical variance and coverage probability.

Overall, for the distributions considered in this section, the Weibull-Q MC-SIMEX method performs reasonably well and in a robust manner. Other methods, such as the Cox MC-SIMEX method, the Weibull-L MC-SIMEX method and the Weibull-C MC-SIMEX method have poor performance in some settings and perform generally worse than the Weibull-Q MC-SIMEX method.

4.4. Dependent censoring

In this simulation, we explored the performance of the proposed estimator when the censoring mechanism is dependent upon a covariate. The true values of the binary covariate X are generated from a random binomial distribution with a mean of 0.5. If the true binary covariate X took on the value of 1, then the censoring times were generated from a uniform (0,4) distribution. On the other hand, if the true covariate X took on the value of 0, then censoring times were generated from uniform (0, 9) distribution, which results in a censoring rate of approximately 30%. We still consider model (3) in Section 4.2. The results are shown in Table 3.

Table 3.
Table showing the performance of the true $({\hat{β}}_{t r u e})$ , naïve $({\hat{β}}_{n a i v e})$ , Cox misclassification simulation extrapolation (MC-SIMEX) $({\hat{β}}_{C O X - M C})$ , Weibull-Q MC-SIMEX $({\hat{β}}_{W Q - M C})$ , Weibull-L MC-SIMEX $({\hat{β}}_{W L - M C})$ and Weibull-C MC-SIMEX $({\hat{β}}_{W C - M C})$ estimators with misclassified covariate X, when the censoring mechanism is dependent upon X. Sample size is set to 500.

	${\hat{β}}_{t r u e}$	${\hat{β}}_{n a i v e}$	${\hat{β}}_{C O X - M C}$	${\hat{β}}_{W Q - M C}$	${\hat{β}}_{W L - M C}$	${\hat{β}}_{W C - M C}$
σ = 0.5
Estimated variance	0.049	0.051	0.052	0.059	0.059	0.059
Empirical variance	0.049	0.118	0.071	0.068	0.032	2.638
Bias	0.005	−0.253	−0.073	−0.067	−0.169	−0.094
Coverage	0.947	0.778	0.856	0.863	0.952	0.214
σ = 1.5
Estimated variance	0.006	0.007	0.010	0.008	0.008	0.008
Empirical variance	0.006	0.072	0.018	0.013	0.034	0.376
Bias	0.005	−0.257	−0.100	−0.067	−0.183	−0.049
Coverage	0.957	0.108	0.763	0.800	0.298	0.214
σ = 2
Estimated variance	0.003	0.004	0.011	0.006	0.006	0.006
Empirical variance	0.003	0.074	0.025	0.012	0.037	0.243
Bias	0.001	−0.264	−0.141	−0.080	−0.191	−0.059
Coverage	0.947	0.021	0.710	0.744	0.158	0.217

Scale parameter is denoted by σ.

The patterns of the results are similar to those in Section 4.2. It shows that dependent censoring mechanism does not affect the performance of these methods.

4.5. One misclassified covariate with three categories

In real data analysis, it is often the case that categorical covariates have more than 2 levels, in contrast to what has been described so far. In order to mimic this scenario and evaluate the proposed estimator in such a scenario, we assumed that the covariate X came from a multinomial distribution with three categories, which are 0,1, and 2, each with a probability of 0.33. The misclassification matrix was set to $[\begin{array}{ccc} 0.8 & 0.2 & 0.1 \\ 0.1 & 0.7 & 0.1 \\ 0.1 & 0.1 & 0.8 \end{array}]$ .

The resulting AFT model is therefore:

Y_{i} = β_{0} + β_{x 1} I (X = 1) + β_{x 2} I (X = 2) + β_{z} Z_{i} + ε_{i} / σ

where I is the indicator function with the corresponding coefficients

β_{x 1}

and

β_{x 2}

—both of which are set to

log (2)

. Censoring times are generated in a manner similar to what was described in Section 4.2. The results of this simulation analysis evaluating the numerical properties of the proposed estimator are shown in Table 4.

Table 4.
Table showing the performance of the true $({\hat{β}}_{t r u e})$ , naïve $({\hat{β}}_{n a i v e})$ , Cox misclassification simulation extrapolation (MC-SIMEX) $({\hat{β}}_{C O X - M C})$ , Weibull-Q MC-SIMEX $({\hat{β}}_{W Q - M C})$ , Weibull-L MC-SIMEX $({\hat{β}}_{W L - M C})$ and Weibull-C MC-SIMEX $({\hat{β}}_{W C - M C})$ estimators when the misclassified covariate X has two categories.

${\hat{β}}_{t r u e}$ ${\hat{β}}_{n a i v e}$ ${\hat{β}}_{C O X - M C}$ ${\hat{β}}_{W Q - M C}$ ${\hat{β}}_{W L - M C}$ ${\hat{β}}_{W C - M C}$

σ = 0.5

Estimated variance 1 0.068 0.071 0.069 0.082 0.082 0.082

Estimated variance 2 0.068 0.066 0.061 0.075 0.075 0.075

Empirical variance 1 0.070 0.138 0.098 0.098 0.043 3.827

Empirical variance 2 0.069 0.136 0.067 0.067 0.040 2.687

Bias 1 0.001 −0.263 −0.069 −0.067 −0.194 0.043

Bias 2 −0.005 −0.263 −0.070 −0.068 −0.186 −0.061

Coverage 1 0.946 0.827 0.840 0.863 0.958 0.212

Coverage 2 0.940 0.825 0.896 0.920 0.976 0.231

σ = 1.5

Estimated variance 1 0.007 0.009 0.014 0.011 0.011 0.011

Estimated variance 2 0.007 0.009 0.013 0.010 0.010 0.010

Empirical variance 1 0.007 0.096 0.028 0.022 0.050 0.532

Empirical variance 2 0.007 0.088 0.026 0.019 0.048 0.405

Bias 1 −0.006 −0.294 −0.130 −0.098 −0.222 −0.035

Bias 2 −0.001 −0.281 −0.131 −0.099 −0.217 −0.066

Coverage 1 0.944 0.140 0.731 0.753 0.268 0.197

Coverage 2 0.956 0.141 0.745 0.752 0.242 0.218

σ = 2

Estimated variance 1 0.004 0.006 0.016 0.007 0.007 0.007

Estimated variance 2 0.004 0.006 0.016 0.006 0.006 0.006

Empirical variance 1 0.004 0.098 0.035 0.021 0.058 0.325

Empirical variance 2 0.004 0.093 0.035 0.020 0.056 0.263

Bias 1 −0.001 −0.302 −0.170 −0.116 −0.240 −0.037

Bias 2 0.000 −0.295 −0.175 −0.121 −0.236 −0.053

Coverage 1 0.957 0.017 0.740 0.635 0.093 0.200

Coverage 2 0.950 0.028 0.739 0.590 0.069 0.237

Sample size is set to 500. Scale parameter is denoted by σ. Estimated variance 1 is used to denote the estimated variance of the first category of X and estimated variance 2 is used to denote the estimated variance of the second category of X. Notation is defined similarly for other parameters.

The comparisons of the Weibull-Q MC-SIMEX estimator with other estimators are similar to those in Section 4.2 in terms of bias and empirical variance. From the perspective of coverage probability, when sample size is small, i.e., n = 500, the Weibull-Q MC-SIMEX estimator is better than the Cox MC-SIMEX estimator for σ = 0.5; they perform similarly for σ = 1.5 and the Cox MC-SIMEX is better for σ = 2. When n = 1000 (Table s4 in Supplemental material), the Weibull-Q MC-SIMEX estimator has similar coverage probability to Cox MC-SIMEX estimator when σ = 0.5 and better coverage probability for σ = 1.5. Although the Cox MC-SIMEX estimator is still better for σ = 2, the difference of the coverage of these two methods is smaller than that when n = 500. This indicates that when the misspecified covariate has more categories, a larger sample size is required for the Weibull-Q MC-SIMEX estimator to have better coverage probability than the Cox MC-SIMEX estimator.

4.5. wo mismeasured covariates

In this section, we evaluate a setting wherein two binary categorical covariates are subject to misclassification. Therefore, the resulting model is:

Y_{i} = β_{0} + β_{x 1} X_{1} + β_{x 2} X_{2} + β_{z} Z_{i} + ε_{i} / σ

where both

X_{1}

and

X_{2}

follow a binomial distribution with a mean of 0.5. The values of both

β_{x 1}

and

β_{x 2}

are set to

\log (2)

. The misclassification matrices for both

X_{1}

and

X_{2}

are set to

[\begin{array}{cc} 0.9 & 0.3 \\ 0.1 & 0.7 \end{array}]

. The two misclassification processes are independent. The censoring rate is set to 30% by adjusting the upper bound of a uniformly distributed censoring time, which is independent of the covariates mentioned in the model immediately above. . The results of this analysis are shown in Table 5. Once again, the patterns of the comparisons among these estimators are similar to those in Section 4.2.

Table 5.
Table showing the performance of the true $({\hat{β}}_{t r u e})$ , naïve $({\hat{β}}_{n a i v e})$ , Cox misclassification simulation extrapolation (MC-SIMEX) $({\hat{β}}_{C O X - M C})$ , Weibull-Q MC-SIMEX $({\hat{β}}_{W Q - M C})$ , Weibull-L MC-SIMEX $({\hat{β}}_{W L - M C})$ and Weibull-C MC-SIMEX $({\hat{β}}_{W C - M C})$ estimators when there are two misclassified covariates X₁ and X₂.

${\hat{β}}_{t r u e}$ ${\hat{β}}_{n a i v e}$ ${\hat{β}}_{C O X - M C}$ ${\hat{β}}_{W Q - M C}$ ${\hat{β}}_{W L - M C}$ ${\hat{β}}_{W C - M C}$

σ = 0.5

Estimated variance 1 0.046 0.050 0.053 0.065 0.065 0.065

Estimated variance 2 0.046 0.050 052 0.062 0.062 0.062

Empirical variance 1 0.044 0.116 0.076 0.077 0.038 2.663

Empirical variance 2 0.047 0.116 0.065 0.066 0.037 2.776

Bias 1 −0.001 −0.261 −0.080 −0.073 −0.184 −0.096

Bias 2 0.003 −0.257 −0.073 −0.066 −0.183 −0.001

Coverage 1 0.959 0.791 0.852 0.876 0.972 0.231

Coverage 2 0.947 0.793 074 0.874 0.970 0.221

σ = 1.5

Estimated variance 1 0.005 0.008 0.010 0.010 0.010 0.010

Estimated variance 2 0.005 0.008 0.010 0.010 0.010 0.010

Empirical variance 1 0.005 0.083 0.037 0.020 0.042 0.488

Empirical variance 2 0.005 0.083 0.036 0.020 0.041 0.475

Bias 1 −0.001 −0.273 −0.171 −0.098 −0.203 −0.080

Bias 2 0.003 −0.275 −0.166 −0.092 −0.202 −0.049

Coverage 1 0.951 0.138 0.531 0.746 0.298 0.212

Coverage 2 0.944 0.130 0.568 0.781 0.336 0.211

σ = 2

Estimated variance 1 0.003 0.005 0.011 0.008 0.008 0.008

Estimated variance 2 0.003 0.006 0.011 0.008 0.008 0.008

Empirical variance 1 0.003 0.087 0.054 0.019 0.045 0.339

Empirical variance 2 0.003 0.084 0.054 0.019 0.045 0.351

Bias 1 −0.001 −0.285 −0.221 −0.103 −0.211 −0.079

Bias 2 −0.002 −0.279 −0.222 −0.106 −0.212 −0.086

Coverage 1 0.946 0.032 0.412 0.686 0.185 0.219

Coverage 2 0.939 0.040 0.392 0.721 0.162 0.215

Sample size is set to 500. Scale parameter is denoted by σ. Estimated variance 1 is used to denote the estimated variance of X₁ and estimated variance 2 is used to denote the estimated variance of X₂. Notation is defined similarly for other parameters.

Overall, the Weibull-Q MC-SIMEX estimator generally performs better than the Cox MC-SIMEX estimator for Weibull or other survival data with one or more misclassified covariates. It is robust to all scenarios considered in the simulation studies. Moreover, the Weibull-Q MC-SIMEX estimator performs better than the Weibull-L MC-SIMEX estimator and the Weibull-C MC-SIMEX estimator in terms of the bias-variance trade-off, which supports the theoretical findings in Section 3.2.

5. Real data analysis

5.1. Real data analysis with one misclassified covariate

We illustrate an application of the Weibull-Q MC-SIMEX method to a colon cancer dataset containing survival information of 1327 patients who suffered from Stage I to Stage III colon cancer, for which they were surgically treated at MSKCC from January 1990 to December 2000. The dataset was obtained from the MSKCC registry. The dataset provides information on variables such as age and prognostic indicators such as T-stage, N-stage, lymphovascular invasion (LVI), and peri-neural invasion (PNI).²¹ In colon cancer patients it has been established that the pathn status of a patient is often misclassified and that the true pathn status (a binary covariate) is a function of the number of nodes examined. Therefore, we estimated the true pathn status of a patient using a statistical algorithm proposed by Gonen et al.²² We then used this true pathn status to estimate the misclassification matrix as $[\begin{array}{cc} 1 & 0.3 \\ 0 & 0.7 \end{array}]$ , where the row denotes pathn status assigned and column denotes the true pathn status. Therefore, it can be seen that the pathn status assigned (naïve covariate) is 100% specific and 70% sensitive.

Of the total of 1327 patients who underwent surgery, 12 were eliminated from the dataset due to ambiguous values on the total nodes examined. The median follow-up time in the cohort was 55 months with date of surgery being treated as the starting time point. 163 of the 1315 individuals died due to colon cancer and the rest were lost to follow-up or died of other causes. Disease-specific survival from colon cancer was the primary outcome of interest The distribution of survival times was checked using the fitdistrplus²³ package in R 4.2.²⁴ The Weibull distribution appeared to fit the data well based on preliminary visual examination of the QQ plots, probability plots, cumulative density plots, and probability density plots of survival times (Figure 2). Median survival time was not reached (Figure 3).

We regressed the survival time of individuals on the following covariates: age, T stage, LVI, PNI, and assigned pathn status using an AFT model. The variable pathn status is considered as the covariate which is subject to misclassification. The AFT model takes the form:

Y_{i} = β_{0} + {β_{a g e}}^{*} ag e_{i} + {β_{T}}^{*} T . stag e_{i} + {β_{l v i}}^{*} LV I_{i} + {β_{p n i}}^{*} P N I_{i} + {β_{n a}}^{*} path n_{i} + ε_{i} / σ

where

Y_{i}

is the log of the survival time, the

β s

represents the corresponding coefficients for the covariates in the model,

σ

is the scale parameter and ε is the error term. We use

β_{n a}

to denote the naïve coefficient for the naively measured pathn status. We evaluated interaction terms in the model and found that none of these were significant at a type I error rate of 0.05. Therefore, we eliminated interaction terms from the model.

Due to lack of proper validation data (as the true pathn status could only be estimated), we do not provide results for the true estimator. For the Weibull-Q MC-SIMEX, Weibull-L MC-SIMEX, Weibull-C MC-SIMEX and Cox MC-SIMEX procedures, we ran a total of 50 replications of pseudo-dataset generation for each level of $λ$ (i.e., $B = 50$ ). The grid of $λ$ values used in the MC-SIMEX procedure was $(0.5, 1, 1.5, 2)$ . The performance of the above-mentioned methods is evaluated and shown in Table 6.

Table 6.

Table showing the estimates of the coefficient associated with pathological nodal status obtained by fitting the AFT model to colon cancer data.

	Estimate of the coefficient associated with
Type of estimator	pathological nodal status	95% CI
Naïve estimator	−0.542	(−0.839, −0.244)
Cox MC-SIMEX	0.851	(0.433, 1.269)
Weibull-Q MC-SIMEX	−0.705	(−1.035, −0.375)
Weibull-L MC-SIMEX	−0.573	(−0.903, −0.243)
Weibull-C MC-SIMEX	−0.847	(−1.212,−0.482)

MC-SIMEX: misclassification simulation extrapolation; AFT: accelerated failure time.

All the estimators shown in Table 6 indicate the detrimental effect of a positive pathn status on survival. It can be deciphered from the Weibull-Q MC-SIMEX estimator that patients with a negative pathn status have a survival time which is 2 $(\frac{1}{e x p^{- 0.705}})$ times more than patients with positive pathn status. It can also be deciphered from the Cox MC-SIMEX estimator that the hazard ratio of those with positive pathn status is 2.34 ( $e x p^{0.851})$ .

Even though data fits Weibull distribution well as shown in Figure 2, it can be misspecified and the original distribution can be other distributions. Then the commonly used statistical criteria such as AIC, BIC, and loglikelihood are implemented to check all distributions considered in Section 4.3, The results are shown in Table 7.

Table 7.

Table showing the fit statistics of the accelerated failure time (AFT) regression model under different distributional assumptions using AIC, BIC, and log-likelihood.

	Statistic
Weibull AIC (smaller is better)	2215.389
Weibull BIC (smaller is better)	2262.017
Weibull logLikelihood	−1098.695
Exponential AIC (smaller is better)	2221.309
Exponential BIC (smaller is better)	2262.756
Exponential logLikelihood	−1102.655
Log Normal AIC (smaller is better)	2205.09
Log Normal BIC (smaller is better)	2251.717
Log Normal logLikelihood	−1093.545
Log logistic AIC (smaller is better)	2206.59
Log logistic BIC (smaller is better)	2253.218
Log logistic logLikelihood	−1094.295

As there are only minor differences in AIC, BIC, and log-likelihood for the distributions considered in Section 4.3, any distribution for the data is justifiable. Under such conditions, we have demonstrated via simulations in Section 4.3 that a Weibull-Q MC-SIMEX estimator provides the most reliable parameter estimates in comparison to Cox MC-SIMEX estimator, except for log-normally distributed data with lower variance, for which the Cox MC-SIMEX estimator has less bias. We estimated the variance of the log of survival times in the current dataset to be 0.80. Given that this variance is close to that of standard log-normal distribution, it will be fair to say that if this data follows log-normal distribution, Cox MC-SIMEX estimator may have smaller bias than the Weibull-Q MC-SIMEX estimator as shown in the simulation study in Section 4.3. However, since the Weibull-Q MC-SIMEX estimator performs generally well for all distributions, we recommend using the results from Weibull-Q MC-SIMEX method to make conclusions for this data. Because the log-normal distribution may best fit the data based on the statistics in Table 7, the log-normal MC-SIMEX method may provide more accurate results. However, simulations are needed to compare the performance of log-normal MC-SIMEX method with other methods. Future research could investigate the performance of the MC-SIMEX method with different distributions, and therefore yield more guidelines for real dataanalysis.

5.2. Real data analysis with two misclassified covariates

In this section, we would like to illustrate the application of Weibull-Q MC-SIMEX method for the colon cancer data considered in Section 5.1 with two misclassified covariates. In order to examine the performance of the proposed estimator in this situation, we simulated a misclassified version of the variable LVI using the misclassification matrix $[\begin{array}{cc} 0.9 & 0.3 \\ 0.1 & 0.7 \end{array}]$ and used it as a covariate in the following model.

\begin{aligned} Y_{i} = β_{0} + {β_{a g e}}^{*} ag e_{i} + {β_{T}}^{*} T .stag e_{i} + {β_{n a 1}}^{*} path n_{i} + {β_{n a 2}}^{*} mLV I_{i} + {β_{p n i}}^{*} PN I_{i} + ε_{i} / σ \end{aligned}

(4)

where

β_{n a 1}

is the coefficient associated with misclassified pathn, mLVI is the misclassified version of the variable LVI with associated coefficient

β_{n a 2}

. All other notations are the same as in Section 5.1. Table 8 below shows the coefficient estimates obtained from model (4).

Table 8.

Table showing the estimates of the coefficient associated with pathological nodal status and mLVI.

Type of estimator	Covariate	Coefficient estimate	95% CI
Naïve estimator	Pathological nodal status	−0.601	(−0.891, −0.31)
	mLVI	−0.133	(−0.441,0.176)
Cox MC-SIMEX	Pathological nodal status	0.907	(0.498, 1.316)
	mLVI	0.132	(−0.294,0.557)
Weibull-Q MC-SIMEX	Pathological nodal status	−0.752	(−1.108, −0.396)
	mLVI	−0.104	(−0.459,0.25)
Weibull-L MC-SIMEX	Pathological nodal status	−0.631	(−0.987, −0.275)
	mLVI	−0.136	(−0.49,0.219)
Weibull-C MC-SIMEX	Pathological nodal status	−0.752	(−1.107,−0.396)
	mLVI	−0.123	(−0.477,0.232)

MC-SIMEX: misclassification simulation extrapolation; mLVI: misclassified version of lymphovascular invasion

The result shows the coefficient estimate for pathn status are very similar to what was seen in Section 5.1. With regard to the coefficient estimates for mLVI, we now already know that the Weibull-Q MC-SIMEX estimator is the most reliable method based on the results of the simulation studies. It concludes that LVI is not a significant covariate associated with survival time of colon cancer patients. Although other methods have different corresponding estimates when compared to that of the Weibull-Q MC-SIMEX estimator, they still reach the same conclusion as that of the Weibull-Q MC-SIMEX estimator.

6. Discussion

In this article, we investigate the performance of the Weibull-Q MC-SIMEX estimator in an AFT model in which the survival times followed a Weibull distribution and are subject to right censoring. We demonstrate that under varying scenarios of censoring and misclassification, the Weibull-Q MC-SIMEX estimator is robust and performs better than the naïve estimator and the existing Cox MC-SIMEX estimator.

Additionally, we also provide proof for the relationship between the parameter estimates obtained at each level of error inflation in the presence of a correctly measured confounder and demonstrate graphically and using simulation studies that the quadratic extrapolation function will offer the best fit in a Weibull AFT model. This finding highlights the importance of the choice of the extrapolation function when using the MC-SIMEX method. Colon cancer is one of the leading causes of death in the US. Therefore, it is important that while estimating hazards and time to survival among patients with risk factors, potential misclassification must be corrected in the covariates. Through this study, we show that the Weibull-Q MC-SIMEX estimator could be used to obtain accurate predictions of time to survival, by correcting for misclassification.

Potential future directions include investigating the performance of nonparametric extrapolation techniques such as splines in the MC-SIMEX method, application of the procedure to situations where the correctly measured confounder and the misclassified covariate interact with each other, application of the procedure to a competing risks model and extension of the method to other fields such as mediation analysis.

Supplemental Material

sj-docx-1-smm-10.1177_09622802231168248 - Supplemental material for Misclassification simulation extrapolation method for a Weibull accelerated failure time model

Supplemental material, sj-docx-1-smm-10.1177_09622802231168248 for Misclassification simulation extrapolation method for a Weibull accelerated failure time model by Varadan Sevilimedu, Lili Yu and Hani Samawi in Statistical Methods in Medical Research

Footnotes

Acknowledgments

The authors are grateful to Dr Mithat Gonen, Attending Biostatistician and Chief of Biostatistics Service, MSKCC, for providing us with the colon cancer data.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Varadan Sevilimedu

Lili Yu

Hani Samawi

Supplemental material

Supplemental material for this article is available online.

Appendix

Derivation of equation 2:

Derivation of $β_{0}$ :

We define:

Y = \log (T)

π_{00} = P (W = 0 | X = 0)

π_{11} = P (W = 1 | X = 1)

and

π_{x} = P (X = 1)

Then,

\begin{aligned} E (Y | W, Z) & = β_{0}^{*} + β_{n a} W + β_{z} Z = β_{0} + β_{x} * P (X = 1 | W) + β_{z} Z \\ => β_{0}^{*} + β_{n a} W = β_{0} + β_{x} * P (X = 1 | W) \end{aligned}

When W = 0,

\begin{aligned} β_{0}^{*} & = β_{0} + β_{x} * P (X = 1 | W = 0) \\ => β_{0}^{*} = β_{0} + β_{x} * \frac{P (X = 1, W = 0)}{P (W = 0)} \\ => β_{0}^{*} = β_{0} + β_{x} \frac{p (W = 0 | X = 1) p (X = 1)}{p (W = 0 | X = 1) p (X = 1) + p (W = 0 | X = 0) p (X = 0)} \\ => β_{0}^{*} = β_{0} + β_{x} \frac{(1 - π_{11}) π_{x}}{(1 - π_{11}) π_{x} + (π_{00}) (1 - π_{x})} \\ = β_{0} + β_{x} \frac{(1 - π_{11}) π_{x}}{π_{00} - δ π_{x}} \end{aligned}

Derivation of $β_{n a}$

Along similar lines, as previously stated in the derivation section for $β_{0}$ ,

E (Y | W = 1, Z) = β_{0}^{*} + β_{n a} + β_{z} Z = β_{0} + β_{x} * P (X = 1 | W = 1) + β_{z} Z

=> β_{0}^{*} + β_{n a} = β_{0} + β_{x} * P (X = 1 | W = 1)

= β_{0} + β_{x} * \frac{P (W = 1, X = 1)}{W = 1}

= β_{0} + β_{x} * \frac{π_{11} π_{x}}{π_{11} π_{x} + (1 - π_{00}) (1 - π_{x})}

Substituting for the value of

β_{0}^{*}

, we have

\begin{aligned} β_{n a} & = β_{x} \frac{π_{11} π_{x}}{π_{11} π_{x} + (1 - π_{00}) (1 - π_{x})} - β_{x} \frac{(1 - π_{11}) π_{x}}{π_{00} - δ π_{x}} \\ = β_{x} π_{x} [\frac{π_{11} π_{00} - π_{x} π_{11} δ - (1 - π_{11}) (1 - π_{00} + δ π_{x})}{(π_{00} - π_{x} δ) (1 - π_{00} + δ π_{x})}] \\ = \frac{β_{x} π_{x} (1 - π_{x}) δ}{(π_{00} - π_{x} δ) (1 - π_{00} + δ π_{x})} \end{aligned}

References

Dalen

Buonaccorsi

Laake

, et al. Regression analysis with categorized regression calibrated exposure: some interesting findings. Emerg Themes Epidemiol 2006; 3: 6.

Spiegelman

, et al. Functional and structural methods with mixed measurement error and misclassification in covariates. J Am Stat Assoc 2015; 110: 681–696.

Küchenhoff

Mwalili

Lesaffre

. A general method for dealing with misclassification in regression: the misclassification SIMEX. Biometrics 2006; 62: 85–96.

Kauhanen

Lakka

Lynch

, et al. Social disadvantages in childhood and risk of all-cause death and cardiovascular disease in later life: a comparison of historical and retrospective childhood information. Int J Epidemiol Aug 2006; 35: 962–968.

Buonaccorsi

Laake

Veierød

. On the effect of misclassification on bias of perfectly measured covariates in regression. Biometrics Sep 2005; 61: 831–836.

Ahrens

Lash

Louik

, et al. Correcting for exposure misclassification using survival analysis with a time-varying exposure. Ann Epidemiol Nov 2012; 22: 799–806.

Zucker

Spiegelman

. Inference for the proportional hazards model with misclassified discrete-valued covariates. Biometrics Jun 2004; 60: 324–334.

Zucker

. A pseudo–partial likelihood method for semiparametric survival regression with covariate errors. J Am Stat Assoc 2005; 100: 1264–1277.

Zucker

Spiegelman

. Corrected score estimation in the proportional hazards model with misclassified discrete covariates. Stat Med 2008; 27: 1911–1933.

10.

Spiegelman

Carroll

Kipnis

. Efficient regression calibration for logistic regression in main study/internal validation study designs with an imperfect reference instrument. Stat Med Jan 15 2001; 20: 139–160.

11.

Bang

Chiu

Y-L

Kaufman

, et al. Bias correction methods for misclassified covariates in the cox model: comparison of five correction methods by simulation and data analysis. J Stat Theory Pract 2013; 7: 381–400.

12.

Cole

Chu

Greenland

. Multiple-imputation for measurement-error correction. Int J Epidemiol 2006; 35: 1074–1081.

13.

Zhou

Pepe

. Auxiliary covariate data in failure time regression. Biometrika 1995; 82: 139–149.

14.

Slate

Bandyopadhyay

. An investigation of the MC-SIMEX method with application to measurement error in periodontal outcomes. Stat Med Dec 10 2009; 28: 3523–3538.

15.

Cook

Stefanski

. Simulation-extrapolation estimation in parametric measurement error models. J Am Stat Assoc 1994; 89: 1314–1328.

16.

Sevilimedu

Samawi

, et al. Application of the misclassification simulation extrapolation procedure to log-logistic accelerated failure time models in survival analysis. J Stat Theory Pract 2018; 13: 24.

17.

Liu

Lim

. Using the weibull accelerated failure time regression model to predict time to health events. bioRxiv 2018: 362186. doi:10.1101/362186

18.

Carroll

Ruppert

Stefanski

, et al. Measurement Error in Nonlinear Models: A Modern Perspective, Second Edition. New York: CRC Press, 2006.

19.

Stefanski

Cook

. Simulation-extrapolation: the measurement error jackknife. J Am Stat Assoc 1995; 90: 1247–1256.

20.

Xiong

. Accelerated failure time models with covariates subject to measurement error. Stat Med 2007; 26: 4817–4832.

21.

Weiser

Landmann

Kattan

, et al. Individualized prediction of colon cancer recurrence using a nomogram. J Clin Oncol Jan 20 2008; 26: 380–385.

22.

Gönen

Schrag

Weiser

Nodal staging score: a tool to assess adequate staging of node-negative colon cancer. J Clin Oncol 2009; 27: 6166–6171.

23.

Delignette-Muller

Dutang

. Fitdistrplus: an R package for fitting distributions. J Stat Softw 2015; 64: 1–34.

24.

R Foundation for Statistical Computing; 2022.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.04 MB

	${\hat{β}}_{t r u e}$	${\hat{β}}_{n a i v e}$	${\hat{β}}_{C O X - M C}$	${\hat{β}}_{W Q - M C}$	${\hat{β}}_{W L - M C}$	${\hat{β}}_{W C - M C}$
σ = 0.5
Estimated variance 1	0.068	0.071	0.069	0.082	0.082	0.082
Estimated variance 2	0.068	0.066	0.061	0.075	0.075	0.075
Empirical variance 1	0.070	0.138	0.098	0.098	0.043	3.827
Empirical variance 2	0.069	0.136	0.067	0.067	0.040	2.687
Bias 1	0.001	−0.263	−0.069	−0.067	−0.194	0.043
Bias 2	−0.005	−0.263	−0.070	−0.068	−0.186	−0.061
Coverage 1	0.946	0.827	0.840	0.863	0.958	0.212
Coverage 2	0.940	0.825	0.896	0.920	0.976	0.231
σ = 1.5
Estimated variance 1	0.007	0.009	0.014	0.011	0.011	0.011
Estimated variance 2	0.007	0.009	0.013	0.010	0.010	0.010
Empirical variance 1	0.007	0.096	0.028	0.022	0.050	0.532
Empirical variance 2	0.007	0.088	0.026	0.019	0.048	0.405
Bias 1	−0.006	−0.294	−0.130	−0.098	−0.222	−0.035
Bias 2	−0.001	−0.281	−0.131	−0.099	−0.217	−0.066
Coverage 1	0.944	0.140	0.731	0.753	0.268	0.197
Coverage 2	0.956	0.141	0.745	0.752	0.242	0.218
σ = 2
Estimated variance 1	0.004	0.006	0.016	0.007	0.007	0.007
Estimated variance 2	0.004	0.006	0.016	0.006	0.006	0.006
Empirical variance 1	0.004	0.098	0.035	0.021	0.058	0.325
Empirical variance 2	0.004	0.093	0.035	0.020	0.056	0.263
Bias 1	−0.001	−0.302	−0.170	−0.116	−0.240	−0.037
Bias 2	0.000	−0.295	−0.175	−0.121	−0.236	−0.053
Coverage 1	0.957	0.017	0.740	0.635	0.093	0.200
Coverage 2	0.950	0.028	0.739	0.590	0.069	0.237

Misclassification simulation extrapolation method for a Weibull accelerated failure time model

Abstract

Keywords

Introduction

2. Methodology

2.1. Weibull AFT model and misclassification matrix

3. Properties of the Mc-SIMEX estimator

3.1. Asymptotic properties of β ^ x s i m e x

3.2. The form of extrapolation function in Mc-SIMEX method

4.1. Settings

4.2. One misclassified covariate

5.1. Real data analysis with one misclassified covariate

Supplemental Material

sj-docx-1-smm-10.1177_09622802231168248 - Supplemental material for Misclassification simulation extrapolation method for a Weibull accelerated failure time model

Footnotes

Acknowledgments

Declaration of conflicting interests

Funding

ORCID iDs

Supplemental material

Appendix

References

Supplementary Material

3.1. Asymptotic properties of ${\hat{β}}_{x}^{s i m e x}$