2.1 Single-component parametric survival models

For the purpose of survival analysis, interest lies on the nonnegative random variable $T_i$ , which denotes the duration of a policy $i$ before termination (cancelation). The adoption of a probability density function $f(t)$ for $T_i$ implies a survival function given by:

(1) \begin{equation} S(t)= P(T_i \gt t) = \int _{t}^{\infty } f(u) du, \end{equation}

which describes persistency probability evolution, through time. $S(t)$ is monotonic and decreasing function, starting at $S(0) = 1$ and converging numerically to zero, since $S(\infty ) = \lim \limits _{t \to \infty } S(t) = 0$ .

The hazard function $\lambda (t)$ describes the instantaneous potential for cancelation and is given by:

(2) \begin{equation} \lambda (t) = \lim _{\Delta t \to 0} \frac{P(t\lt T_i\leq t+ \Delta t\mid T_i\gt t)}{\Delta t} \approx \frac{f(t)}{S(t)} = -\frac{S'(t)}{S(t)}, \end{equation}

where $\Delta t$ is a small-time increment. In particular, $\lambda (t)\Delta t$ is the approximate probability of a cancelation occurring in the interval $(t,t+\Delta t)$ , that is, the lapse rate, given that the policy has survived until time $t$ (for more details, see Ibrahim et al., Reference Ibrahim, Chen and Sinha2001; Kalbfleisch & Prentice, Reference Kalbfleisch and Prentice2002).

Survival analysis data typically contain censored observations, that is, data on sample units for which the event of interest was not observed during the study time. In the context of the present work, we define censored data as current (not canceled) policies, claim occurrences (death of the insured), and terminations of contracts.

To illustrate the larger surrender rates at the first months of a contract and a smaller rate later in time, we consider an example and assume an usual single-component parametric survival model, which proves to be inadequate for this kind of insurer portfolio behavior. An artificial database was simulated in order to emulate the cancelation behavior in insurance products with 1,000 policies, 40% censored data, and considering a dichotomous covariate $x$ (0/no attribute, 1/yes attribute). We let the observed failure-time data be denoted by:

(3) \begin{equation} d_i= (t_i, \delta _i, x_i), \quad i=1, \ldots, 1,000, \end{equation}

where the $t_i$ is the time recorded for the ith policy, $x_i$ is a covariate associated with the i-th policy, and $\delta _i$ is an indicator of the censoring status, given by:

\begin{equation*} \delta _i = \begin {cases} 1 & \text {if} \quad t_i \quad \text {is an observed lapse time}\\[5pt] 0 & \text {if} \quad t_i \quad \text {is a censored time}. \end {cases} \end{equation*}

Thus, $\{\delta _i=1\}$ is the event representing the occurrence of lapse or cancelation (with $f(t_i)$ probability density), whereas $\{\delta _i= 0\}$ denotes censored lapse time or, equivalently, nonoccurrence of lapse during the study period (with survival probability $S(t_i)$ ), that is, lapse/cancelation time is only known to be greater than $t_i$ . The observations on different policies are assumed to be independent. The data were simulated considering variable $Y_i= \log(T_i)$ as the logarithmic the duration of a policy $i$ before termination from a mixture of Gaussian distributions. For a more detailed discussion about data generation, see Appendix B.

Fig. 1 presents the empirical Kaplan–Meier survival curve (dashed line) adjusted to the simulated data, for both attributes, and it is clear that there is a heterogeneous behavior between the levels of the dichotomous covariate. The absence of the attribute described by the covariate is associated with an increased premature risk of cancelation. In addition, for both levels of the covariate, there is a difference in the survival behavior in the initial times when compared to the following ones.

Figure 1

Simulated dataset: empirical Kaplan–Meier survival curves (dashed line) and Exponential, Weibull and Log-Normal (solid line) fitted models.

In the context of survival analysis, parametric models play a key role in modeling the phenomenon of interest. If $T_i$ follows an Exponential model, then $\lambda (t) = \lambda$ is constant over time. If a Weibull model with parameters $\lambda$ and $\kappa$ is considered for $T_i$ , then $\lambda (t)= \lambda \kappa t^{\kappa -1}$ , resulting in constant ( $\kappa =1$ ), growing ( $\kappa \gt 1$ ), or decreasing ( $\kappa \lt 1$ ) hazard through time. The Log-Normal model with parameters $\mu$ (mean of the logarithmic failure time) and $\sigma$ (standard deviation of the logarithmic failure time) is an usual choice when the hazard function is not monotonous, reaching a maximum point and then decreasing. Structured regression models based on covariates are usually considered, relating the parameters in the sampling model with covariates responsible for the description of each contract profile. The simulated data were fit by the Exponential, the Weibull, and the Log-Normal survival regression models.

Panels in Fig. 1 (see solid line) make it clear that usual survival models are not flexible enough to accommodate the behavior of the empirical survival function, per stratum, over time. Even though the Log-Normal model produces a good performance when compared with the competing models in the initial instants, it fails to adapt in later times as in the other competing models. That is, the rates tend to decrease faster over time, and usual parametric survival models are not able to accommodate this behavior. Our working premise is that simple parametric models can serve as block builders of more flexible models, via mixtures.

2.2 Bayesian mixture of parametric survival models

In this section, we present our proposal of a mixture of parametric models for censored survival data. Finite mixture models are described in detail in Frühwirth-Schnatter (Reference Frühwirth-Schnatter2006). As seen in the illustration with artificial data presented in section 2.1, the competing fitted models apparently do not reflect the empirical distribution of the data. A possible alternative is to use more flexible structures such as mixture models, which allow the incorporation of behavioral change in the probability distribution of the data. Our proposal is based on the classical finite mixture model, where observations are assumed to arise from the mixture distribution given by:

(4) \begin{equation} f(t_i) = \sum _{j=1}^{K} \eta _j f_j(t_i), \end{equation}

where $f(t_i)$ is the probability density function of $T_i$ and $f_j(t_i)$ denotes $K$ component densities occurring with unknown proportions $\eta _j$ , with $0 \leq \eta _j \leq 1$ and $\sum _{j=1}^{K} \eta _j=1$ . It follows that the survival function $S(t)$ has the mixed form:

(5) \begin{equation} S(t) = \sum _{j=1}^{K} \eta _j S_j(t) = \sum _{j=1}^{K} \eta _j \int _{t}^{\infty } f_j(u)du, \end{equation}

and the hazard function has the mixed form:

(6) \begin{equation} \lambda (t)= -\frac{S'(t)}{S(t)}=- \frac{\sum _{j=1}^{K} \eta _j S'_j(t)}{S(t)} =\frac{\sum _{j=1}^{K} \eta _j \lambda _j(t)S_j(t)}{S(t)}=\sum _{j=1}^{K} \eta _j h_j(t), \end{equation}

where $h_j(t)=\lambda _j(t)S_j(t)/S(t).$

For each policy $i$ , we define a latent group indicator $I_i \mid \boldsymbol{\eta } \sim Categorical(K, \boldsymbol{\eta })$ , with $\boldsymbol{\eta }=(\eta _1, \ldots, \eta _K)$ , following a categorical distribution given by $P(I_i \mid \boldsymbol{\eta })= \prod _{j=1}^{K} [\eta _j]^{I_{ij}}$ , where $I_{ij}=1$ if observation $i$ is allocated to group $j$ and is null, otherwise. These auxiliary non-observable variables aim to identify which mixture component each observation has been generated from and their introduction in the formulation makes it simple to express the likelihood function for observation $i$ :

(7) \begin{equation} f_{T_i}(t_i \mid I_i) = \prod _{j=1}^{K} [f_j(t_i)]^{I_{ij}}, \quad i=1, \ldots, n. \end{equation}

Notice that in the example shown in section 2.1, $K=2$ components are assumed in addition to a regressor with two levels, for the generation of the artificial lapse risk data. In this case, the lapse risk can be decomposed into two overlapping processes in time. The first process is associated with the period that immediately follows the contracting of the policy. In general, it covers the first three months when the lapse risk is relatively high. The second process refers to the subsequent period when persistence decreases smoothly over time. Thus, the mixture distribution is written as $f_{T_i}(t_i \mid I_i) = [f_1(t_i)]^{I_{i1}}[f_2(t_i)]^{(1-I_{i1})}$ . It is worth noting that single-component models are obtained as a particular case of the mixture structure when $\eta _1=1$ , $\eta _j=0, j\gt 1.$

For the estimation process in this work, the number $K$ of mixture components is assumed to be known and subjective to a sensitivity analysis. As usual in applied contexts, $K$ is subject to uncertainty. Frühwirth-Schnatter (Reference Frühwirth-Schnatter2006, Chap 4) provides a good discussion of informal methods seeking to identify the number of components in a mixture, including the evaluation of the predictive behavior of a statistic $T(Y)$ for future realizations of the response, conditional on its past values $y_1,\ldots, y_n$ and on a model with fixed $K$ components, for different specifications of $K$ . The choice of $K$ is then indicated by the specification that leads to the best value of the predictive performance evaluation statistic, among the proposed mixture structures. We performed sensitivity analysis to choose the number of components, observing the quality of the fit of the survival and hazard curves, compared to their respective empirical versions, which can be summarized by statistics indicating the performance of the fit. If $K$ is not fixed, its estimation would lead to models with variable-dimensional parametric space. Details on Bayesian methods in this context can be seen in Frühwirth-Schnatter (Reference Frühwirth-Schnatter2006, Chap 5). As shown in the following subsections, we offer a computationally efficient method for estimating a model with fixed $K$ , and we consider that the sensitivity analysis to different specifications of $K$ , which can be performed in reduced time, is sufficient for our purposes.

2.2.1 Mixture of Log-Normal components

In this paper, interest lies in modeling $Y_i = \log(T_i)$ , the logarithmic duration of a policy $i$ before termination (cancelation), such that $Y_i \sim \mathcal{N}(\mu _i,\sigma ^2)$ , implying that in the original scale $T_i \sim \mathcal{LN}(\mu _i,\sigma ^2)$ , $i=1, \ldots, n$ . Survival times associated with a different outcome than the lapse, such as survival times past the end of our study and deaths, are assumed to be censored for policies, $i= h+1, \ldots, n$ ., then

(8) \begin{equation} f(t_i\mid \mu, \sigma ^2) = (2\pi )^{-\frac{1}{2}}(t_i\sigma )^{-1}\exp \!\left \{-\frac{1}{2\sigma ^2}(\log (t_i)-\mu _i,)^2\right \}. \end{equation}

The resulting survival function is given by:

(9) \begin{equation} S(t_i \mid \mu _i, \sigma ^2)= 1 - \Phi \!\left (\frac{\log (t_i)- \mu _i}{\sigma }\right ). \end{equation}

We can thus write the survival likelihood function of $(\mu, \sigma ^2)$ implied by a Log-Normal model and based on data $D$ as:

(10) \begin{equation} L(\mu _i, \sigma ^2 \mid D) = \prod _{i=1}^{n} f(t_i \mid \mu _i, \sigma ^2)^{\delta _i} S(t_i \mid \mu _i, \sigma ^2)^{(1- \delta _i)}. \end{equation}

Adopting the finite mixture approach and assuming $Y_i= log(T_i) \mid \mu _i, \sigma ^2 \sim \mathcal{N}(\mu _i, \sigma ^2)$ and latent variables $I_{ij}$ , $i=1,\ldots, n$ , it follows that:

(11) \begin{equation} f_{Y_i}\!\left (y_{i} \mid \mathrm{I}_{i}, \mu _i, \sigma ^2 \right )= \prod _{j=1}^{K}{[\mathcal{N}_j}(y_i \mid \mu _{ij}, \sigma ^2_j)]^{\mathrm{I}_{ij}}, \end{equation}

where $\mathcal{N}_j$ is a normal distribution for the component group $j$ , $j=1,\ldots,K$ . Then,

\begin{eqnarray*}{y}_{i}\mid \{\mathrm{I}_{i1}=1 \}, \boldsymbol{\beta }_1, \sigma ^2_1 &\sim & \mathcal{N}_{1}(\mu _{i1}(\boldsymbol{\beta }_1),\sigma ^2_{1}) \\[5pt] {y}_{i}\mid \{\mathrm{I}_{i2} = 1 \}, \boldsymbol{\beta }_2, \sigma ^2_2 &\sim & \mathcal{N}_{2}(\mu _{i2}(\boldsymbol{\beta }_2),\sigma ^2_{2}) \\[5pt] \vdots && \vdots \\[5pt] {y}_{i}\mid \{\mathrm{I}_{iK}=1 \}, \boldsymbol{\beta }_K, \sigma ^2_K &\sim & \mathcal{N}_{K}(\mu _{iK}(\boldsymbol{\beta }_K),\sigma ^2_{K}), \end{eqnarray*}

with $\mu _{ij}= \textbf{x}_{ij}^T \boldsymbol{\beta }_j$ and $\boldsymbol{\beta }_j = (\beta _{0j}, \beta _{1j}, \ldots, \beta _{pj})$ characterizing the unknown mean and $ \sigma ^2_{j}$ , the variance, respectively, for $j=1,\ldots, K$ and $i=1,\ldots, n$ .

Using the latent indicators of categorical allocation, the likelihood simplifies to

(12) \begin{eqnarray} f(\textbf{y}\mid \boldsymbol{\eta }, \boldsymbol{\beta }, \sigma ^2) &=& \prod _{j=1}^{K}\prod _{i=1}^{n}\eta _j^{I_{ij}} [\mathcal{N}_j(y_i \mid \boldsymbol{\beta }_{j}, \sigma ^2_{j})]^{I_{ij}} \nonumber \\[5pt] &=& \prod _{j=1}^{K} \eta _j^{n_j} \!\left [ \prod _{i: I_{ij}=1} \mathcal{N}_j(y_i \mid \boldsymbol{\beta }_{j}, \sigma ^2_j)\right ], \end{eqnarray}

where $n_j=\sum _{i}I_{ij}$ is the number of observations allocated to group $j$ and $n= \sum _{j=1}^{K}n_j$ , for $j=1,\ldots, K$ , $i=1,\ldots, n$ . Thus, the mixture survival likelihood function is given by:

(13) \begin{equation} f(\textbf{y} \mid \boldsymbol{\eta }, \boldsymbol{\beta }, \sigma ^2) = \prod _{j=1}^{K} \eta _j^{n_j} \!\left [ \prod _{i: I_{ij}=1} \mathcal{N}_j(y_i \mid \boldsymbol{\beta }_{j}, \sigma ^2_j)^{\delta _i} S(y_i \mid \boldsymbol{\beta }_j, \sigma ^2_j)^{1-\delta _i}\right ], \end{equation}

where $\delta _i$ in the censorship indicator, as previously seen in section 2.1.

From a Bayesian point of view, we are interested in the posterior $p(\boldsymbol{\eta }, \boldsymbol{\beta }, \sigma ^2 \mid \textbf{y})$ . The posterior distribution are generally not available analytically, and numerical integration and simulation are considered, in particular, MCMC methods (Gamerman & Lopes, Reference Gamerman and Lopes2006) are used in this paper. Notice that to compute posterior distributions, we need to take into account the censored quantities, which in practice can be computationally prohibitive, depending on the percentage of censored data and the large dataset. Thus, inference is facilitated through data augmentation.

2.2.2 Inference based on data augmentation

The presence of censored data is a common feature when considering time data until the occurrence of an event and the likelihood function takes this fact into account, as seen in equation (10). Following the Bayesian approach, the estimation procedure can be based on MCMC algorithm using the data augmentation technique (see Tanner & Wong, Reference Tanner and Wong1987). Suppose we observe logarithmic survival times $\textbf{y}^{obs}= (y^{obs}_1, \ldots, y^{obs}_h)$ . Then the idea is to define the logarithmic survival times for the $n-h$ censored policies as missing data which we denote as $\textbf{z}= (z_{h+1}, \ldots, z_{n})$ .

Assume that the full data is given by:

(14) \begin{equation} \textbf{y}= (\textbf{y}^{obs}, \textbf{z}). \end{equation}

Let $\boldsymbol{\theta }$ be the parametric vector of interest. The data augmentation approach is motivated by the following representation of the posterior density:

(15) \begin{equation} p(\boldsymbol{\theta } \mid \textbf{y}^{obs}, \boldsymbol{\delta }) = \int _{\textbf{Z}} p(\boldsymbol{\theta } \mid \textbf{y}^{obs}, \textbf{z}, \boldsymbol{\delta })p(\textbf{z} \mid \textbf{y}^{obs}, \boldsymbol{\delta }) d\textbf{z}, \end{equation}

where the vector $\boldsymbol{\delta }$ is composed by censorship indicators $\delta _i \in \!\left \{0, 1\right \}$ ; $p(\boldsymbol{\theta } \mid \textbf{y}^{obs},\boldsymbol{\delta })$ denotes the posterior density of the parameter $\boldsymbol{\theta }$ given the observed data $\textbf{y}^{obs}$ ; $p(\textbf{z} \mid \textbf{y}^{obs}, \boldsymbol{\delta })$ denotes the predictive density of latent data $\textbf{z}$ given $\textbf{y}^{obs}$ ; and $p(\boldsymbol{\theta } \mid \textbf{y}^{obs}, \textbf{z}, \boldsymbol{\delta })$ denotes the conditional density of $\boldsymbol{\theta }$ given the augmented data $\textbf{y}$ .

In practice, we do not know a priori to which group a given observation belongs. Thus, in addition to the censored observations, whose outcome is unknown, the variable $I_i$ in equation (7) is also latent and is estimated in our inferential algorithm. Assuming that policies are independent, the likelihood function for the complete data can be written as:

(16) \begin{eqnarray} f(\textbf{y}^{obs}, \textbf{z}, \boldsymbol{\delta } \mid \boldsymbol{\theta }) &=&{ \prod _{j=1}^{K} \eta _j^{n_j}} \!\left [\prod _{i : \delta _i=1, I_{ij}=1}f{_{j}}(y_i^{obs} \mid \boldsymbol{\theta }_j) \prod _{i : \delta _i=0,I_{ij}=1}f{_j}(z_i \mid \boldsymbol{\theta }_j)\mathcal{I}(z_i \geq y_i^{obs}) \right ], \end{eqnarray}

where $\boldsymbol{\theta }_j= (\eta _j, \boldsymbol{\beta }_j, \sigma ^2_j)$ , $n_j= \sum _i I_{ij}$ , $\sum _{j=1}^{k} n_j=n$ , and $\mathcal{N}_j \sim f_j( \cdot \mid \boldsymbol{\theta }_j)$ . Following Bayes’ theorem, the posterior distribution of the model parameters and latent variables, given the complete data $\textbf{y} = (\textbf{y}_1, \ldots, \textbf{y}_{K})'$ , is proportional to

(17) \begin{eqnarray} p\!\left (\boldsymbol{\eta }, \boldsymbol{\beta }, \sigma ^2 \mid \textbf{y}, \boldsymbol{\delta } \right ) &\propto & f(\textbf{y}^{obs}, \textbf{z}, \boldsymbol{\delta } \mid \boldsymbol{\theta }) p(I \mid \boldsymbol{\eta }) \pi (\boldsymbol{\eta }, \boldsymbol{\beta },\sigma ^2) \\[5pt] &\propto &{ \prod _{j=1}^{K} \eta _j^{n_j}} \!\left [\prod _{i : \delta _i=1, I_{ij}=1}f{_{j}}(y_i^{obs} \mid \boldsymbol{\theta }_j) \prod _{i : \delta _i=0,I_{ij}=1}f{_j}(z_i \mid \boldsymbol{\theta }_j)\mathcal{I}(z_i \geq y_i^{obs}) \right ] \nonumber \\[5pt] &\times & \prod _{i=1}^{n} p(I_i \mid \boldsymbol{\eta }) \pi (\boldsymbol{\eta }, \boldsymbol{\beta },\sigma ^2). \nonumber \end{eqnarray}

The Bayesian mixture model is completed by the prior distribution specification. We assume independence in the prior distribution with $\boldsymbol{\eta } \sim Dirichlet(K, \alpha )$ , where $\sum _{j=1}^{K} \eta _j = 1$ and $\alpha =(\alpha _1, \ldots, \alpha _K)$ is a vector of hyperparameters, such that $\alpha _j \gt 0$ ; $\phi _j=\frac{1}{\sigma ^2_j} \sim Gamma(a_j,b_j)$ , the regression coefficients, $\boldsymbol{\beta }_j \sim \mathcal{N}_j(\boldsymbol{m}_j, \tau ^2_j\boldsymbol{I}_p)$ and $P(I_{ij}=1)=\eta _j$ , for $i=1, \ldots, n$ and $j=1, \ldots, K$ .

The resulting posterior distribution in equation (17) does not have closed form, and we appeal to MCMC methods to obtain samples from the posterior distribution. In particular, posterior samples are obtained through a Gibbs sampler algorithm, where the Markov chain is constructed by considering the complete conditional distribution of each hidden variable given the others and the observations. The scheme is presented in the following section.

2.2.3 Computational scheme for the mixture of Log-Normal components

Assuming $K$ groups, it is possible to consider the Gibbs sampler algorithm in order to overcome the numerical integration condition of the data augmentation techniques, resulting in a computationally efficient algorithm.

We consider the following Bayesian Gaussian mixture survival model with data augmentation:

\begin{eqnarray*} \textbf{y}_j \mid I_i, \boldsymbol{\beta }_j, \phi _j &\sim & \mathcal{N}_j(x_i^T \boldsymbol{\beta }_j, \phi ^{-1}_j) \\[5pt] I_i \mid \boldsymbol{\eta } &\sim & Categorical(K, \boldsymbol{\eta }) \\[5pt] \boldsymbol{\eta } &\sim & Dirichlet(\alpha _1, \ldots, \alpha _K) \\[5pt] \boldsymbol{\beta }_j &\sim & \mathcal{N}_j(\boldsymbol{m}_j, \tau ^2_j\boldsymbol{I}_p) \end{eqnarray*}

\begin{eqnarray*} \phi _j &\sim & Gamma(a_j,b_j) \\[5pt] p(z_i) &\propto & \mathcal{I}\!\left \{ z_i \geq y_i\right \} \end{eqnarray*}

Algorithm 1 shows the scheme to estimate the parameters via data augmentation with censored observations. Details involved in obtaining the full conditional distributions can be seen in Appendix A.

Algorithm 1.

Gibbs sampler for a finite Gaussian mixture survival model with data augmentation.

2.2.4 Point estimation via EM

Optimization methods to obtain maximum likelihood estimates are less computationally expensive than Monte Carlo estimation because they depend uniquely on numerical convergence. In order to obtain point estimates efficiently, we consider the maximization of log-likelihoods for the mixture model.

Consider $K$ mixture components. In the classical context, the mixture model without censored data is described by equations (11) and (12), respectively. Furthermore, without considering censored data and latent indicators of the mixture components, the log-likelihood function to be maximized is given by:

(18) \begin{equation} l\!\left (\!\left \{\eta _j\right \}_{j=1}^{K}, \!\left \{\boldsymbol{\beta }_j\right \}_{j=1}^{K}, \!\left \{\sigma ^2_j\right \}_{j=1}^{K}\right ) = \sum _{i=1}^{n}\log \sum _{j=1}^{K} \eta _j f_j(Y_i \mid \boldsymbol{\beta }_j, \sigma ^2_j). \end{equation}

Notice that the expression depends on logarithms of sums, which cannot be simplified through logarithmic properties. The estimation, in this context, is exhaustive and without analytic or recursive forms for the maximum likelihood estimators (MLEs) of the model parameters.

In the mixture distribution context, it is very common to use the EM algorithm proposed by Dempster et al. (Reference Dempster, Laird and Rubin1977) which is an iterative mechanism to calculate the MLE in the presence of missing observations. Given the use of latent variables, and conditional on $I$ , equation (11) is valid. Thus, it provides a probability distribution over the latent variables together with a point estimate for parameters. If a prior distribution is assumed for the parameters, the joint posterior mode is obtained by the method.

Besides that, when we take into account the censored observed data, the data augmentation technique, as previously seen, can be applied by including a new latent variable vector $\textbf{z}$ . According to our model, censored observations are originated from a truncated normal distribution.

Assume that observation $y_i$ is censored. That is, there is an unobserved datum $z_i$ such that $z_i \mid \!\left \{I_{ij}=1 \right \} \sim f_j$ and $y_i^{obs} \mid z_i, \!\left \{I_{ij}=1\right \} \sim \mathcal{NT}_{(-\infty,z_i)}(x_{ij}' \boldsymbol{\beta }_j, \sigma ^2_j)$ . The strategy that we will adopt in the algorithm is to remove the truncation from the observed data $y_i^{obs}$ to obtain $z_i$ , at each iteration $(k)$ , so that $y_{1:n}^{(k)} = (y_1^{obs}, y_2^{obs}, \ldots, y_{h}^{obs}, z_{h+1}^{(k)}, \ldots, z_{n}^{(k)})$ , as previously seen in section 2.2.2.

Algorithm 2 is adapted for this context. For more details, see Jedidi et al. (Reference Jedidi, Ramaswamy and DeSarbo1993). Notice that $E(z_i \mid y_i^{obs}, \!\left \{I_{ij}=1\right \})$ and $Var(z_i \mid y_i^{obs}, \!\left \{I_{ij}=1\right \})$ denote the expected value and variance of a truncated Gaussian distribution, respectively. Although the computational cost (to obtain a point estimate) is smaller when compared to the proposal in sections 2.2.2 and 2.2.3, a disadvantage of this approach is that the EM algorithm is quite sensitive to the choice of initial parameters and does not take into account the uncertainty associated with parameter estimates. Besides, the EM algorithm will converge very slowly if a poor choice of initial value $\boldsymbol{\eta }^{(0)}$ , $\boldsymbol{\beta }_j^{(0)}$ , and $\sigma _j^{2(0)}$ is selected.

Algorithm 2.

Expectation–Maximization algorithm for a finite Gaussian mixture survival model with data augmentation.

3.1 Simulated dataset

In this section, we return to the illustrative dataset seen in section 2.1. Our aim is to compare the usual and mixture survival Log-Normal models under a Bayesian approach through our proposals described in sections 2.2.1 and 2.2.2.

We simulate three scenarios: (i) a dataset with 10% of censored data; (ii) a dataset with 40% of censored data; and (iii) a dataset with 60% of censored data, considering a mixture of $K=2$ components, with $\eta _1=\eta =0.6$ . We would like to assess whether our proposal is efficient in sampling from the posterior distribution as well as its computational efficiency, for the model of interest. In addition, we vary the sample size ( $n=1,000; 10,000; 50,000; 100,000$ ) in order to evaluate the computational cost through (a) our proposal with data augmentation for censored data. To allow fast computations, we perform part of the code implementation with the C++ programming language to build improved loops of complete conditional distributions through the RcppArmadillo package available in R software (see Eddelbuettel & Sanderson, Reference Eddelbuettel and Sanderson2014); and (b) without data augmentation via RStan package available in R (Stan Development Team, 2018; Carpenter et al., Reference Carpenter, Gelman, Hoffman, Lee, Goodrich, Betancourt, Brubaker, Guo, Li and Riddell2017), that is, considering the survival likelihood given by equation (13). Stan is a C++ library for Bayesian modeling and inference that primarily uses the No-U-Turn sampler (NUTS) (see Hoffman & Gelman, Reference Hoffman and Gelman2014) to obtain posterior simulations given a user-specified model and data.

The survival time can be analyzed according to $\log(T_i)= \beta _0 + \beta _1 x_i + \varepsilon _i$ , with $x$ the covariate that takes values ( $x=0$ or $x=1$ ) and error $\varepsilon _i \sim N(0, \phi ^{-1})$ . We assign vague independent priors to the parameters in $\boldsymbol{\theta }$ with $\phi _j \sim Gamma(0.01,0.01)$ , $\boldsymbol{\beta }_j \sim \mathcal{N}_j(\textbf{0}, 100\boldsymbol{I}_2)$ , and $\boldsymbol{\eta }\sim Dirichilet(\alpha _1=2, \alpha _2=2)$ , for $j=1,2$ . We run an MCMC chain for 20,000 iterations and consider the first 10,000 out as burn-in. The burn-in and lag for spacing of the chain were selected so that the effective sample size was around 1,000 samples.

Fig. 2 illustrates the fit of the survival curves by the competing models considering a sample with 1,000 policies and 40% rate of censorship (as seen in Fig. 1 (a)), the usual Bayesian Log-Normal (BLN) model (without mixture, like in Fig. 1 (d)) and the Bayesian mixture Log-Normal model (BMLN) (see panels (c) and (d) in Fig. 2). As can be seen, the proposed mixture model is able to accommodate different behaviors in the survival curves when compared to the usual Log-Normal model. In addition, the uncertainty associated with estimates is lower for our proposed mixture model. Panel (b), in Fig. 2, exhibits the point estimation via EM for the Log-Normal mixture modeling. The estimated survival curves via the EM algorithm follow the behavior of the empirical Kaplan–Meier curves. Point estimates of the parameters of interest are reasonable compared to those obtained via Gibbs sampler techniques. See more details about the simulated dataset in Appendix B.

Table 1 shows the posterior summaries for the survival Log-Normal model without mixture (Bayes LN) and considering Log-Normal mixtures via our proposal via data augmentation (BMLN) and via Stan (BMLN and SBMLN), respectively. As already mentioned, the non-mixture model is not able to capture the behavior of the survival curve. The structure of the non-mixture model does not allow the incorporation of mixture components in the coefficient estimates. On the other hand, the mixture Log-Normal model is capable of producing suitable estimates for the true parameters. Although the data augmentation proposal and the Stan method lead to similar point and interval estimates, the processing computational cost via Stan is much higher for all scenarios, as can be seen in Table 2. In this way, the use of Stan for large samples, highly censored observations, and considering more covariates in the survival model can be prohibitive. For the EM algorithm, 58 iterations were required until the parameters converged, which resulted in a computational time of 3.32 seconds. However, as already stated, the EM algorithm does not generate uncertainty measures associated with estimates.

Table 1.

Posterior summaries comparison: with mean and 95% credibility for the Bayes LN, the Bayes Mixture LN via Stan, and the Bayes Mixture LN; the point estimation via EM Mixture LN, for a simulated dataset with 40% censorship rate and considering $n=1,000$ policies.

Table 2.

Comparison of computational times (in seconds) involved in the Bayesian competing methods for simulated datasets with 10%, 40%, and 60% rates of censorship and considering $n$ (size) policies.

Computational system Ubuntu, Intel Core i5-10400F CPU 2.90 GHZ, 16 GB RAM.

3.2 A simulated dataset in insurance

In this section, we simulated a dataset aiming to emulate the behavior of a realistic portfolio in the private life insurance sector. We simulate a large dataset emulating 100,000 policies over 100 months, taking into account heterogeneous lapse rates and realistic censored data with an approximate $ 42.7\%$ censorship rate and two mixing components based on $\eta =0.6$ , via the mixture Log-Normal model previously seen in section 2.2.1. To illustrate, this dataset contains individual policyholder information as well as information about the subscription.

The factors considered in this study are gender (male and female), age group (18–29 years, 30–49 years, and 60+ years), policy type (standard and gold), where the level gold represents a segmentation of insureds that have a high insured capital and the premium payment mode (monthly, yearly, i.e., regular premium or single premium) of the policy. Time to churn is the response variable of interest.

Figure 3

Simulated dataset: a summary about the survival time with (a) survival time in log scale, (b) empirical survival curve, and (c) churn hazard curve.

Panel (a) in Fig. 3 presents the simulated survival times in the log scale, indicating the mixing of two distributions. In panels (b)–(c), the empirical survival curve and hazard rate behavior show that the lapse rate is higher and sharply falls for the first periods of time after subscription initiation, then it stabilizes for some time, exhibits a peak close to 20 months, and then gradually decreases.

Fig. 4 presents the marginal empirical Kaplan–Meier survival curve for the variables of this study. As we can be seen, categories for each variable present particular behavior and could be useful to understand the time to churn. There is a noticeable drop in active insured in the first periods after the policy subscription. Some reasons could be discussed, such as the policyholders who subscribe just to test a service or there may be some association with the premium payment mode. For the age group, there is no visible difference in patterns of survival probability between the 18–29 years and 30–59 years of age groups, except for a drop in during the first months of subscription and nearby the final portion of the survival curve.

Fig. 5 shows the performance of the fitted curves for the competing models, the Log-Normal model, and the Log-Normal mixture model for some scenarios. Panel (a) represents the characteristics of the policyholder in scenario 1 (male, standard, 30–59 years of age, and monthly payment), panel (b) exhibits the fitted curves for scenario 2 (male, gold, 60+ years of age, and year), and panel (c) exhibits the estimated survival curves for scenario 3 (female, standard, 60+ years of age, and month). As we can see, policyholders in scenario 1 have lower persistence when compared to scenarios 2 and 3, respectively. This behavior is understood due to the fact of the standard group and 30–59 years of age group present survival empirical curves with more abrupt decay than 60+ years of age and gold categories.

Figure 4

Simulated dataset: marginal empirical survival curves via Kaplan–Meier for the covariates: gender, type policy, age, and payment.

Figure 5

Simulated dataset: summary survival curves with the empirical (dashed line), and the Bayes LN, the EM mixture LN, and the Bayes mixture LN with 95% credible interval (solid gray line): scenario 1 (first row), scenario 2 (second row), and scenario 3 (third row).

Our proposed Bayesian mixture survival model is able to capture the behavior of the empirical curves (see Fig. 5). Note that the BLN produces a poor fit due to the fact this model is not allowed to access distinct behavior of the curve at different times of the study. Besides that, the BMLN and EMMLN converge for the true values generated from the simulated dataset. In order to understand the lapse of the policyholders, especially in the initial periods of the contract, we consider obtaining some probabilities to profile the policyholder to customer retention via the BMLN. We hope that the longer customers are with the insurance company, the less likely they are to cancel a policy. Table 3 shows a summary of the probabilities conditional on the hypothesis that the policyholder has survived the first three months, that is, conditional on the event $L = \left \{T \gt 3 \right \}$ for the scenarios in Fig. 5. As one can see, the risk probability of a lapse occurring in the first year is higher for scenario 1 (0.394) when compared to scenarios 2 and 3 (0.247 and 0.123), respectively. On the other hand, the conditional probability that the policyholder maintains the contract term after 36 months, having survived to the first months ( $P(T \geq 36 \mid T\gt 3)$ ), is high for all considered scenarios. Other probabilities can be evaluated in order to understand the profile of the company’s policyholders. To illustrate, Table 4 exhibits the results obtained conditioning on other events, such as $\!\left \{ T\gt 12\right \}$ , $\!\left \{T\gt 24\right \}$ , and $\!\left \{T\gt 36\right \}$ . As expected, the probability of churn decreases, having the insured persisted in the insurance company for long periods, that is, the lapses reduce substantially with increasing policy age. Furthermore, when identifying insured profiles, we are able to understand which profiles need more attention; thus, the insurance company could develop retention strategies focused on these policyholders.

3.3 Case study: Telco customer churn

The Telco customer churn data is available on IBM Business Analytics Community Connect (IBM, 2019) and contains information about a Telco company that provides home phone and internet services to 7,043 customers in California in the United States. The dataset contains 18 variables about the profile of the customers and a variable indicating who has left or stayed in the service resulting in approximately 73.4% of censoring. The response variable tenure represents the number of months the customer has stayed with the company with an average of 32 months. Lifetimes equal to zero were removed resulting in a dataset of 7,032 customers. Demographic characteristics are included for each customer, as well as customer account information and service information. Following previous studies, we aim to predict customer churn rate behavior and identify the most important factors that contribute to high- or low-lapse risks.

Panels in Fig. 6 illustrate the pattern of the lifetime of the customers. Although a large admission of new customers is observed for the first months of contract, panel (c) indicates that there is a high lapse rate at the beginning, that is, there is a large portion of customers that will leave the company after just a few months of service. In the other months, the lapse risk rate remains around 0.005. Note that after 60 months of contract, there is a steep increase in the lapse rate. We expect that our proposed model is able to capture these sudden changes of pattern in the hazard function over time.

Table 3.

Simulated dataset: probability for the time to churn conditional on the event $L=\{T \gt 3 \}$ for the scenarios 1, 2, and 3.

Table 4.

Simulated dataset: probability for the time to churn conditional on the events $\{ T\gt 12\}$ , $\{T\gt 24\}$ and $\{T\gt 36\}$ for the scenarios 1, 2, and 3.

Figure 6

Case study: summary about the survival time with (a) lifetime distribution, (b) empirical survival curve, and (c) churn hazard rate.

Figure 7

Case study: model comparison compared with the empirical lapse curve for the Exponential model, the Weibull model, the Log-Normal model, and the Bayesian mixture Log-Normal model considering $K=2,3,4,5,6$ mixture components, respectively. The gray line represents the fitted model and the black line is the empirical curve.

Figure 8

Case study: model comparison compared with the empirical lapse curve (first row) and the hazard curve (second row) for the Exponential model, the Weibull model, the Log-Normal model, and the Bayesian mixture Log-Normal model considering $K=2,3,4,5,6$ mixture components, respectively. The gray line represents the fitted model and the black line is the empirical curve.

To check if usual parametric models are able to accommodate the different regimes of the survival curve over time, we consider fitting three alternative models: Exponential, Weibull, and Log-Normal and compare them with the BMLN model proposed in this paper. We present the Bayesian solution as follows instead of EM point estimates as it is desirable to access uncertainty in the curves. Figs. 7 and 8 present the general survival and hazard curves for model comparison, respectively. Indeed, the Exponential alternative seems naive, since it is a well-known fact that it assumes constant lapse rates over time. This hypothesis does not reflect reality and in this case, we will not take into account this model to model the profiles risk. Regarding our proposed mixture model, a sensitivity analysis for the number of components $K$ was performed. Thus, the time to churn is modeled based on the mixture model with $K$ components for different values of $K$ . Then, the best model can be selected based on model comparison measures computed for each value of $K$ . As we can see, the BMLN with $K=2$ and $K=3$ components return fits that are similar to the Weibull model, failing to capture the curves from 30 to 70 months of contract. For $K\gt 4$ , the mixture models provide better fits, with the general curve recovering the risk to churn in all periods of time including the first and last months of the contract. This indicates that our proposed model allows for more flexibility, besides being computationally efficient even for a large proportion of censorship in the data.

Figure 9

Case study: marginal empirical survival curve (first row) and hazard curves (second row) for covariates Payment Method, Contract, and Internet Service.

Although there is a large set of covariates available to explain the lapse risk, a group of them is not informative for the survival time under study. For example, the lapse rates for the two categories of gender in the data are nearly the same. Similar behavior is observed for some customer service information such as PhoneService, MutipleLines, and OnlineSecurity. Also, there is multicollinearity for groups of variables, in particular OnlineSecurity, TechSupport, OnlineBackup, and DeviceProtection are all dependent on the OnlineService variable. For instance, the covariates MonthlyCharges and TotalCharges present high collinearity. A detailed descriptive analysis for this data is presented at https://github.com/vivianalobo/LapseRiskAAS . We discuss ways to aggregate categories in each covariate considered in the analysis to achieve more interpretability of effects.

In the final models, we considered the covariates PaymentMethod, InternetService, and Contract resulting in eight risk profiles after rearranging the categories established in the descriptive analysis. Fig. 9 presents empirical survival and hazard curves for these risk profiles. For the customer account information, see that customers with Month-to-Month contracts have higher lapse rates compared to clients with One-year and Two-year contracts. Notice that the hazard curve of those with One-year payment is very similar to the one associated with more than one-year payment regime. Thus, we considered aggregating these categories in a unique group called “One-year +”. For the payment method, see that the behavior of survival curves and churn rates for Credit Card (automatic) and Bank Transfer (automatic) are quite similar. On the other hand, when we consider the instantaneous lapse curve, notice that customers that prefer an Electronic check as a payment method are the majority to leave the company, so we merged the other classes in a new category called “others.”

Figs. 10 and 11 present some scenarios for the profile risks. The first scenario (first row) refers to customers that have Fiber optics internet service and others (DSL, No) (for Month-to-Month and Electronic check). See that individuals that own Fiber optics internet service tend to have a higher propensity to churn when compared to others. Notice that the choice of the BMLN with $K=4$ components is enough to capture the trend of the lapse curve and the lapse instantaneous rate. The posterior mean mixture weights estimated for the BMLN with four components (and respective 95% credible interval) are: $\eta _1=0.603 (0.719, 0.888), \eta _2= 0.229 (0.174, 0.298), \eta _3= 0.102 (0.095, 0.107), \eta _4= 0.066 (0.048, 0.085)$ . A posterior summary of this model is detailed in Table 5 which gives some insights into the parameters of the model.

Figure 10

Case study: model comparison survival curves for the scenarios “varying Internet Service” (first row), “varying Payment Method” (second row), and “varying both Payment Method and Contract” (third row) for the Weibull model, the Log-Normal model, and the Bayesian mixture Log-Normal model with $K=4,5,6$ components. The solid line represents the fitted curve and the dashed line is the empirical curve. For the BMLN, the 95% credible interval is provided in gray.

Figure 11

Case study: model comparison hazard curves for the scenarios “varying Internet Service” (first row), “varying Payment Method” (second row), and “varying both Payment Method and Contract” (third row) for the Weibull model, the Log-Normal model, and the Bayesian mixture Log-Normal model with $K=4,5,6$ components. The solid line represents the fitted curve and the dashed line is the empirical curve. For the BMLN, the 95% credible interval is provided in gray.

Table 5.

Posterior summary for regressors of the BMLN model with $K=4$ mixture components: with mean, standard error, and 95% credible interval, for the case study.

As customers who have internet services via Fiber optics have the propensity to churn faster, we would like to understand whether this trend is maintained by varying the payment method. For this purpose, scenario 2 (second row) presents customers’ profiles that prefer to deal through Electronic versus customers with Bank Transfer, Credit Card, or Mailed Check (for Month-to-Month and Fiber optics). Again, the BMLN shows to be the best model, compared to the Weibull and Log-Normal. Notice that there is a significant difference between customers that prefer Electronic check and others: the churn risk for the profile Electronic check, Month-to-Month, and Fiber optics presents higher lapses over several months (see Fig. 10, second row, after Month 40).

In scenario 3, four profile risks are considered taking into account both types of contracts and payment regimes. As we can see, customers whose contract is for One year or more (One year +) have a lower chance of lapsing when compared to Month-to-Month contracts. This can be seen if one varies the payment method (for the ones who have Fiber optics). See that risk lapse rates are higher for Electronic check payment method.

Overall the proposed mixture models with $K\geq 4$ present the best fits when compared with well-known survival models such as the Weibull model. The flexibility of our proposal captures changes of regimes in the lapse rates over time, exhibited by the data.