Keywords
censoring, time to failure analysis, non-proportionality, splines, piecewise exponential models, piecewise exponential additive models, accelerated failure time, oncology
This article is included in the Oncology gateway.
censoring, time to failure analysis, non-proportionality, splines, piecewise exponential models, piecewise exponential additive models, accelerated failure time, oncology
Cancer causes a large disease burden worldwide, among which breast cancer is the most frequent cause of cancer deaths in women. Pakistan, being a lower-middle-income country, has a greater number of breast cancer patients compared to its neighboring countries. It is the country with the highest age-standardized death rate in 2019 globally. Risk of death increases after early breast cancer recurrence in the first three to five years of primary treatment. Time after recurrence to death is analyzed through time to failure techniques, incorporating recorded prognostic factors before recurrence such as age, tumor grade, molecular subtype and treatment.1 In previous research, age has not been proven to be a significant influence on breast cancer deaths.2,3 To further explore its role, age at diagnosis and age at recurrence are included in this study with other covariates.
Time to failure data have incomplete information about exact event occurrence time, which is known as censoring. Three common types of censoring are encountered in time to failure studies: right, left and interval censoring. The most common is right censoring, which is classified into three types: fixed type 1, type 11 and random type 1. In fixed type 1, right censoring occurs for all understudy subjects, who do not observe the event of interest during the predefined study time. Type 11 censoring is named for all subjects who do not observe a specific event after a pre-specified number of events have occurred. In random type 1 right censoring, censored subjects have different censoring times, as not all have same entry time into the study.4 Non-parametric, semi-parametric and parametric modelling techniques are amenable to analyze such types of time to failure disease studies.5
Kaplan-Meier (KM) is the simplest method, used to estimate survival function by a non-parametric maximum likelihood estimator (NPMLE), which has the limitation of studying only one factor at a time. Therefore, it is not suitable for multivariate studies.6 The Cox proportional hazard (PH) models, a semi-parametric approach, does not assume the shape of the baseline hazard function, so distributions of regression parameters’ outcomes remain unknown.7 Cox PH models incorporate multivariate predictors by holding the PH assumption, which assumes a fixed proportion of hazard for individuals. In case of right censoring where upper bounds of event occurrences are not specified, regression parameters are estimated through dividing the likelihood function of the PH model into two parts: one comprises of the baseline hazard and unknown parameters, while the other has only unknown parameters to be estimated, which is called partial-likelihood. Breslow8 and Efron9 introduced approximations in partial-likelihood to handle ordered ties in uncensored event times, while exact and discrete methods are also available, in which non-ordered tied survival times are applied through a partial likelihood approach.5
Validity of PH assumption can be checked through a standard global test suggested by Grambsch and Therneau.10 Furthermore, graphical ways of plotting residuals versus predictors are also discussed in their research.10 In case of non-proportionality, extended Cox PH models can apply, which account for the effects of time varying predictors on survival times.11 Spline-based methods are a good choice to estimate effects of unknown nonlinear predictors on continuous response through penalized partial likelihood, they also explore the functional form of non-proportional predictors.12 Piecewise models are a good choice for long length follow-up studies, where predictors’ effects are checked at different finite time intervals to obtain in-depth information about disease progression.13
Parametric models rely on a fully maximum likelihood approach, parametric estimates are more efficient and precise if conducted through correctly specified forms. In parametric modelling, time to failure is assumed to follow any distribution, such as exponential, Weibull, gamma, generalized gamma, log-normal, log-logistic, Gompertz and Generalized F.14 By building a linear relationship between the logarithm of failure time and predictors, data can be analyzed through the accelerated failure time (AFT) model. In AFT models one-unit changes in predictors explain a proportional change in survival time, as illustrated by Lee and Go,15 while in PH parametric form assumes a proportional change in hazard due to a one-unit change in predictors.
The aim of this paper is to review and apply the above stated modelling techniques to time to failure data, and to evaluate their performances through statistical measures, to investigate the best fitted one for right censored data, while fulfilling limitations and assumptions.
Our data consists of 1,028 women diagnosed with breast cancer in Lahore, Pakistan. All women observed recurrence between February 2011 and February 2018 after initial treatment. They were treated at the same hospital (Institute of Nuclear Medicine & Oncology Lahore, Pakistan). The primary endpoint of this study is death due to breast cancer. Exclusion criteria were: incomplete or missing information, women diagnosed with another disease or another cancer before breast cancer, and bilateral carcinomas. Women who were still alive (survived) at the end of the study, or died due to another reason than breast cancer, are considered right-censored. Age at diagnosis, age at recurrence, estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (Her2), tumor grade, radiotherapy and chemotherapy are the predictors included in this study, all were chosen with the help of clinicians and oncologists.
In the data, age at diagnosis and age at recurrence (in years) are continuous variables, estrogen receptor, progesterone receptor and Her2 are represented by binary variables (0 = Negative, 1 = Positive), tumor grade is represented categorically (). In addition, chemotherapy and radiotherapy are indicated by dummy variables (0 = No, 1 = Yes), here indicated the patients who did not receive the treatment, while meant they received treatment. Survival time was considered from recurrence to death or drop out, and censoring status was coded with for censored, and for death due to breast cancer. Along with the aim of this study, which is the comparison of different parametric models, it is also a major interest to find out how the two treatments (radiotherapy and chemotherapy), in combination with the other predictors, affect the survival of breast cancer patients after recurrence.
Proportional hazards have been checked by scaled Schoenfeld residuals (Extended data). Furthermore, the statistical tests revealed that not all covariates are statistically significant (p < 0.05), but the global test is statistically significant. Therefore, we can assume that the proportional hazard assumption holds.
In this work, right censored time-to-event data are considered. Survival time is denoted by T and censoring time by , where are women diagnosed with invasive breast cancer who observed first recurrence after primary treatment. The event time is defined by , while time for censored and uncensored events is denoted by . The general relationship form, between and is given as
The hazard rate represents a probability of an instantaneous failure per unit of time given that an individual patient has survival after time .5 The main objective of time to failure studies is to estimate the hazard function accurately. For this purpose, different modelling approaches are applied.
In the multivariate approach, the semi-parametric Cox PH model is most popular for the analysis of time to failure data. The general form is written as
FPs provide flexible parameterization of continuous predictors, were first used for modelling families of curves by Royson and Altman,17 with polynomial of degree, it can be written as
is a positive integer and are regression parameters, while is divided into two parts: one consists of when , and another of for .17 FP models have a wide variety of shapes based on different transformations. The main issue with fractional polynomials is in choosing a suitable power for polynomials, as this has a direct positive relationship with flexibility. To increase flexibility, a greater power can be used, but with the major threat of non-locality. That means that the fitted function at a given point of depends on data points which are very far from that reference point.18
Spline regression is an improved technique, used to overcome the non-locality of fractional polynomials. In spline models, the dataset is divided into multiple parts and these parts are joined with knots. In time to failure regression analysis, spline modelling is extensively used to smooth non-linear effects of continuous predictors. Spline has a smooth function. The major problem occurs in choosing the number of knots, as no hard and fast rule is available to apply the suitable number of knots.19
Under RCS, the best way of choosing the number of knots is using the quotient of the difference between the largest and average uncensored log survival time and the largest and smallest uncensored log survival time.20 Royston suggested another method in this respect; according to him, a good way to choose the suitable number of knots is to randomly apply different number of knots every time, and select the best model with measure of information criteria.21 Flexible parametric models have scaling for proportional hazards or proportional odds, which are usually based on transformation of survival function by a link function
is the baseline survival function, is the vector of unknown parameters for predictors Piecewise exponential models (PEMs) are also a reasonable approach to estimate hazard ratios more accurately.21 Under the PEM modelling approach, follow-up time is divided into intervals, by assuming a constant baseline hazard in each interval, so that simplifies to,
For the cut points of follow-up times, minimum to maximum time is divided into finite intervals, Here, to are time intervals, and the hazard rate of exponential distribution is constant over time intervals. Censoring and all unique event times can be used as time interval cut points, but no hard and fast rule is available to choose cut points, which is a point to ponder; too small or too large cut points may cause under- or overfitting.
One approach to deal with this cut point problem is an extension of PEM, in which a large number of cut points are used and the hazard is estimated semi-parametrically. This is called the Piecewise exponential additive mixed model (PAMM), in which a hazard is modeled through a smooth nonlinear function. In PAMM, predictors contribute to the hazard additively, imposing a quadratic penalty on the basis coefficients:
In fully parametric PH modelling, baseline hazard function is assumed to follow a specific distribution and coefficients are estimated via maximum likelihood. A number of different parametric PH models are derived by applying distributions, such as exponential, Weibull, and Gompertz.
An alternative to parametric PH is AFT models, the corresponding log-linear form of the AFT model with respect to time is given as
The likelihood estimates are maximized using the Newton Raphson procedure,15 which may be time consuming and tricky without computer programming. The freely available R software is used to implement all modelling techniques.
Comparison of fitting models is done via measures of fit, which describe accuracy of fitted models for a given data set, usually called goodness of fit measures. Model fitting accuracy has nothing to do with the predictive ability for external data prediction. The Akaike Information Criterion (AIC)25 and the Bayesian Information Criterion (BIC)26 are two of the most common measures which are used to compare models’ performances. For the PH model AIC and BIC are based on log-partial likelihood
According to the Ethical Guidelines for Epidemiologists (IEF-EGE) and the regulations of the ethics committee located at the Advanced Studies and Review Board, University of the Punjab, Lahore (Pakistan), no ethics approval is needed, because the analysis is based on routine data. The study was critically cleared by the Advanced Studies and Review Board of Punjab University. The letter of support written by the departmental head was submitted to the selected hospital. Prior to data collection, written consent was obtained from the head of oncology department and confidentiality was maintained by coding from data collection to analysis.
In the present study, women’s age was collected for: diagnosis and recurrence time. The median age at diagnosis of breast cancer was 47 years (range: 18–59); while the median age at recurrence was 49 years (range: 21–62). Median survival time after recurrence was 3 years, and just half (54.1%) of cancer cases were ER-negative. The majority of patients were PR-positive (64.6%) and had a positive human epidermal growth factor receptor 2 (52.9%). Overall, 207 women (20.1%) had tumor grade 1, whereas 821 (79.9%) had a higher level of malignancy. Chemotherapy (36.4%) and radiotherapy (87.4%) were given as primary treatments (Table 1).
There were 447 deaths among the 1,028 women included in the study. As shown in Table 1, 78.5% of deaths occurred due to breast cancer within three years after recurrence, while 20.8% between 3 to 6 years, and 0.7% of patients died due to breast cancer after 6 years of its recurrence. The molecular markers among women who died due to breast cancer were distributed as follows: 59.3% ER-positive, 66.7% PR-negative, and 65.5% human epidermal growth factor receptor 2-positive. Breast cancer death was positively associated with higher tumor grade (11 and 111; 97.3%) and no chemotherapy (67.8%).
Table 2 presents information measure results. Low values of AIC, AICc, BIC and BICc are considered good; if a model’s fitting values for AIC, AICc, BIC and BICc are smaller than others, that model is considered a good fitted one. To make results less lengthy and meaningful, we only discuss here AIC and AICc values of first three good fitted accelerated failure time distributional models. From the fully parametric models, Weibull is the best fitted one (AIC = 7269.5, AICc = 7271.9) among others, generalized gamma (AIC = 7269.8, AICc = 7272.1), gamma (AIC = 7270.2, AICc = 7272.5), and Generalized F (AIC = 7271.7, AICc = 7274.0) come next in terms of preferences, respectively. We also presented BIC and BICc in Table 2.
Multivariate fractional polynomial (MFP) is the best fitted model incorporating time dependent covariates (AIC = 5167.5, AICc = 5163.6), while the generalized additive model (GAM) (AIC = 5171.3, AICc = 5173.8) is also a good choice for analyzing non-linear continuous predictors in a multivariate setting, having the advantage of small numbers of parameters in non-integer, which is due to shrinkage during parameter estimation. Royston and Palmar’s flexible parametric models have been applied with different scales: we considered flexible parametric models on hazard and odd scales, by including time-dependent effects for the age at diagnosis and age at recurrence covariates. Hazard generalized gamma (AIC = 7211.3, AICc = 7214.8) outperformed odd generalized gamma (AIC = 7212.5, AICc = 7216.0). Although the subjective approach of knot selection may be criticized, sensitivity analyses studies showed insignificant differences in results while changing positions of knots.29–31
Figure 1 shows cumulative hazard graphs for all parametric and flexible parametric models with Nelson Aalen cumulative hazard as a reference. The Nelson Aalen estimator is represented by a step function, which starts at zero. It provides an estimate of the expected number of deaths observed for a given amount of time. Visually, all models provide fitting accuracy for the right censored breast cancer failure time understudy data, with some slight variations existing to capture the fluctuations. The wider lines show a greater confidence interval, which is indicative of a poor fit, while narrower lines show good model fitting.32
Time dependent covariates, age at diagnosis and age at recurrence are modeled using 4 degrees of freedom for splines. PEM and PAMM are applied to the understudy data to get baseline hazard estimates, where finite time intervals are considered as factors to maintain constant hazard for each interval. The age at diagnosis and recurrence are estimated using P-splines with the same 4 degrees of freedom. PEM and PAMM results are compared with Nelson Aalen estimator, graphical displays showed close agreement of good model fitting in Figure 2.
In this paper, different parametric models are compared in terms of theoretical aspects and application. Our findings suggested that progesterone receptor negative, human epidermal growth factor receptor 2 positive, higher tumor grade, and no chemotherapy increase the risk of death after recurrence. The most surprising result is regarding radiotherapy treatment, which depicted no reduction in breast cancer time to death. This might be due to a higher level of physical impairment of patients receiving radiotherapy treatment. However, patients treated by radiotherapy at an early stage have a larger survival time.33
We applied distributional parametric models which are known as standard parametric models, with a full maximum likelihood estimator to estimate unknown parameters. AFT models make practical sense to study the influence of covariates, which may accelerate breast cancer mortality. The AFT model is the best choice for the analysis of time to failure data when hazards are non-proportional, as it provides efficient estimates and an estimate of the median failure time.
The exponential distribution having one rate parameter is often used in experiments to account for the amount of time until an event occurs. The Weibull distribution is a special case of the exponential distribution with shape and scale parameters. It provides a better fit than exponential with one extra degree of freedom. The same is true for gamma distribution, which has two parameters and has close results to Weibull. Generalized gamma has mean, location and scale parameters. Log-normal is a probability distribution, with a normally distributed logarithm. It is widely used in lifetime data analysis. The two parameters of mean and standard deviation have a more stable behavior than log-logistic distribution. Generalized F distribution is a good alternative to generalized gamma with one extra parameter. From the interpretation point of view, the AFT model’s results are easy to interpret and help clinicians to make wise decisions related to the patients’ conditions.
Flexible parametric models have advantage of using restricted cubic splines, which incorporates time dependent effects of predictors on the log hazard and reduces the bias of non-proportionality. The Royston and Parmar generalized gamma flexible parametric model under hazard scale outperformed the odd scale model. Of course, one should not ignore the threat of overfitting, by including greater number of internal knots. The functional polynomial model has the advantage of only considering significant factors, so it gives better results than other spline-based models. GAM under spline basis function has the potential to provide a better fit of data than generalized linear flexible parametric models.34,35
The main strength of this study is that we described and applied different time to failure models, to right censored breast cancer data. In a piecewise exponential model the baseline hazard is modeled by step function with different intervals, estimation is done by including dummy variables for each interval. The major disadvantage of this technique is that data becomes too long, and parameter estimation becomes unstable. The piecewise additive mixed model overcomes this drawback. By adding a large number of basis functions and using P-splines between neighboring basis coefficients, parameters are estimated through restricted maximum likelihood (REML).13
There are several limitations to our study. First, the use of a single case study may be viewed as a limitation. However, a simulation study can also be designed to validate results. Second, the model comparisons are based on within sample information measures (AIC, AICc, BIC, BICc), while predictive performances of models can also be checked via different measures. Third, sensitivity analyses of choosing different numbers of knots in spline-based models can be performed to make firm conclusions.
Flexible parametric modelling of the hazard function is more efficient than standard parametric models, incorporating the complex patterns of the observed failure data. Generalized additive models provide more accurate estimates under spline-basis function, with time dependent covariates. For long follow-up studies and multiple time dependent covariates, which may have effects on hazard, penalized models are more suitable.
Data is available from the corresponding author, Dr. Florian Fischer (florian.fischer1@charite.de), upon reasonable request. Data can be used for research purposes, but cannot be published because it is taken from a hospital.
We thank the staff of the Institute of Nuclear Medicine & Oncology Lahore (INMOL), who supported in data collection. We also wish to thank Dr. Rab Nawaz Maken from INMOL cancer hospital, Lahore, Pakistan, for providing full support to conduct this research.
Views | Downloads | |
---|---|---|
F1000Research | - | - |
PubMed Central
Data from PMC are received and updated monthly.
|
- | - |
Is the work clearly and accurately presented and does it cite the current literature?
No
Is the study design appropriate and is the work technically sound?
No
Are sufficient details of methods and analysis provided to allow replication by others?
No
If applicable, is the statistical analysis and its interpretation appropriate?
No
Are all the source data underlying the results available to ensure full reproducibility?
No
Are the conclusions drawn adequately supported by the results?
No
Competing Interests: No competing interests were disclosed.
Reviewer Expertise: Multivariate analysis; Survival analysis; Study design; high-dimensional data.
Is the work clearly and accurately presented and does it cite the current literature?
No
Is the study design appropriate and is the work technically sound?
Partly
Are sufficient details of methods and analysis provided to allow replication by others?
No
If applicable, is the statistical analysis and its interpretation appropriate?
No
Are all the source data underlying the results available to ensure full reproducibility?
No
Are the conclusions drawn adequately supported by the results?
No
References
1. Kearns B, Stevenson MD, Triantafyllopoulos K, Manca A: Generalized Linear Models for Flexible Parametric Modeling of the Hazard Function.Med Decis Making. 39 (7): 867-878 PubMed Abstract | Publisher Full TextCompeting Interests: No competing interests were disclosed.
Reviewer Expertise: Methodological research; health economics, survival analysis, time-series analysis.
Alongside their report, reviewers assign a status to the article:
Invited Reviewers | ||
---|---|---|
1 | 2 | |
Version 1 13 Oct 21 |
read | read |
Provide sufficient details of any financial or non-financial competing interests to enable users to assess whether your comments might lead a reasonable person to question your impartiality. Consider the following examples, but note that this is not an exhaustive list:
Sign up for content alerts and receive a weekly or monthly email with all newly published articles
Already registered? Sign in
The email address should be the one you originally registered with F1000.
You registered with F1000 via Google, so we cannot reset your password.
To sign in, please click here.
If you still need help with your Google account password, please click here.
You registered with F1000 via Facebook, so we cannot reset your password.
To sign in, please click here.
If you still need help with your Facebook account password, please click here.
If your email address is registered with us, we will email you instructions to reset your password.
If you think you should have received this email but it has not arrived, please check your spam filters and/or contact for further assistance.
Comments on this article Comments (0)