You are on page 1of 5

Effect of the Method of Cell Harvesting

on Multiple Events among Myeloid


Leukemia Patients

Homework 4: Survival Data Analysis (0383)


2022-2023

2nd year Master of Statistics and Data Science


Hasselt University

Group members:
Melvin Estolano (2159122)
Quynh Long Khuong (2159280)
Jessa Mae Lastimoso (2159120)
Jose Carlos Cortiñas Porras (2055631)

Submission Date: November 30, 2022

Lecturer:
Prof. Dr. Tomasz Burzykowski
Group 4: Survival Data Analysis Homework 5

1 Introduction
A clinical trial was performed to study the effect of the method of cell harvesting on bone marrow
transplantation in acute and chronic myeloid leukemia patients. We aim to analyze the data to evaluate
the difference between harvesting bone marrow from the donor’s pelvic bone or the peripheral blood in
each observed event.

2 Methodology
Competing risk refers to an event that hinders the observation of the event of interest or alters the proba-
bility of occurrence of the event of interest [1]. In this study, we are interested in evaluating the difference
between the two cell harvesting methods while controlling for the disease type and age of the patient.
The variables considered in the analysis were the disease type of the patient (acute or chronic myeloid
leukemia), age, and method of cell harvesting. The events of recurrence, cGVHD (chronic graft versus
host disease), and death were observed but the time-to-first event only was recorded.

The proportional sub-distribution hazard model was used to estimate the incidence of the outcomes over
time while taking into account the competing risk, Fine and Gray introduced this model in 1999. This
directly relates the variables to the cumulative incidence function (CIF) of the event [2]. In this paper,
the cell harvesting methods were compared for the events: recurrence, cGVHD, and death, while taking
into account the competing risks for each of these events. Several diagnostic checking tests were done
such as (1) the functional form of the continuous covariates which were determined through the plot of
Martingale residuals from the zero covariate model against the continuous covariates; (2) the propor-
tional hazards assumption which was measured through the Scaled Schoenfeld residuals versus time that
was formally tested using the Schoenfeld test; and (3) goodness-of-fit of the model assessed through the
comparison of the cumulative incidence function corresponding to the PH model for the sub-distribution
hazard functions for all event types versus the predicted values.

All the analyses were conducted in R version 4.2.1 [3] with packages survival and cmprsk at a significance
level of 5%.

3 Results and Discussion


3.1 Diagnostic Checking
Before interpreting the parameter estimates of the model, appropriate diagnostic checking tools were used.
Results are shown in Figure 1. Based on the visual inspection of the plot of martingale residuals against
the covariate age for all event types, suggesting that the functional form is acceptable. The proportional
hazards assumption was checked by plotting the scaled Schoenfeld residuals as seen in Figure 1 for each
covariate across all event types. The smoothed line shows a horizontal pattern for all the covariates
hence the proportional hazards assumption is supported. However, for the events recurrence and death,
5 observations were recorded, thus proportionality assumption is difficult to assess.

1
Group 4: Survival Data Analysis Homework 5

Figure 1: Scaled Schoenfeld residuals versus time

3.1.1 Results

One of the major caveats of the cause-specific model or sub-hazard model is that it assumes independent
censoring for subjects who are not actually censored but failed from competing events. Therefore, the
parameter estimates from the sub-hazard model do not have an informative interpretation since it relies
heavily on the independence censoring assumption. In addition, the cause-specific hazard function does
not have a direct interpretation in terms of survival probability [4].

The model considered in the analysis is the Fine and Gray model for sub-distribution hazard for each
event. Its model formulation is seen below :

hj (t) = h0j (t)eZj β
where, Z1 = β1 age + β2 treatment + β3 diagnosis

Table 1: Parameter estimates for the sub-distribution model for each event

Recurrence cGVHD Death


coef(se) exp(coef ) coef(se) exp(coef ) coef(se) exp(coef )
trt: PB -0.113 (0.9660) 0.893 0.319 (0.2079) 1.375 -1.573 (1.1069) 0.207
age -0.145 (0.0673)* 0.865 -0.012 (0.0092) 0.988 -0.092 (0.0506) 0.913
diag: CML -1.897 (1.2061) 0.150 0.540 (0.2284)* 1.715 -0.502 (0.8482) 0.606
Time-dependent coefficient
age 0.318 (0.1272)* 1.375 - - 1.221 (0.6077)* 3.389
∗p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

2
Group 4: Survival Data Analysis Homework 5

Based on the plot of Schoenfeld residuals on the sub-distribution hazard model, it may be deemed that
proportionality assumptions are not attained for recurrence and death event types. To account for this,
time-varying effects were considered in the model for recurrence and death. Table 1 gives the parameter
estimates for each event.

For the recurrence event, age and age with time dependence are significant. For death event type,
treatment, diagnosis and age were not statistically significant at 5% level of significance but age with
time-dependence is significant at α = 0.05.

Moreover, for cGVHD event type, it can be seen that diagnosis is statistically significant at 5% level of
significance while taking other clinical factors into account. Hence, there is a significant increase in the
hazard of having cGVHD for patients with acute myeloid leukemia as compared to patients with chronic
myeloid leukemia by 71.5%, in the presence of competing events.

For the three events, the effect of treatment is not significant, therefore the method of method cell har-
vesting does not have an effect on the hazard of an event occurring for recurrence, cGVHD, and death.

3.1.2 Goodness-of-fit

Figure 2 was plotted to assess the goodness-of-fit of the PH model for the sub-distribution hazard model.
By visual inspection of the plot, it suggests that the plot of the cumulative incidence function for cGVHD
event type captures similarly the plot for the predicted cumulative incidence function. Thus, the model
fits the data well.

Figure 2: CIF of PH models vs. CIF predicted sub-distribution hazard model

3
Group 4: Survival Data Analysis Homework 5

4 Conclusion
Competing risk is a phenomenon that occurs in survival data such that a competing event/s may influ-
ence the risk of the event of interest occurring. In this study, the Fine-Gray model was used to overcome
this problem. It was found that there is a significant difference in the hazard of having cGVHD between
acute and chronic myeloid leukemia patients by 71.5%, this is in the presence of competing events such
as recurrence and death. Moreover, the treatment showed no significant difference at the 5% level on the
hazard of an event occurring for all event types.

However, caution is posted to the results of parameter estimates for recurrence and death event type
since there were only a few patients observed with these events.

5 Appendix
R Codes (selected)

# ------ Goodness - of - fit of the model - - - -#


# Comparing the nonparametric estimates and FG model
event _ pred <- predict ( mod _ event , cov1 = rbind (0 ,1) )
plot ( event _ pred , lty =1:2 , color =2)
# ------ PH Assumption - - - - - - - - - - - - - - - - - - -#
plot ( mod _ event $ uft , mod _ event $ res [ , col ] , main = " Covariate " , xlab = " Failure time " ,
ylab = " Schoenfeld residuals " )
# ------ Fine and Gray Model - - - - - - - - - - - -#
# Recurrence
mod _ recur _ tv <- crr ( ftime = df $ first _t , fstatus = df $ recur , cov1 = design _ mat , failcode =1 ,
cov2 = cbind ( df $ age ) , tf = function ( t ) cbind ( t ) )
# cGVHD
summary ( mod _ cGVHD <- crr ( ftime = df $ first _t , fstatus = df $ cGVHD , cov1 = design _ mat ,
failcode =1) )
# Death
mod _ death _ tv <- crr ( ftime = df $ first _t , fstatus = df $ death , cov1 = design _ mat , failcode =1 ,
cov2 = cbind ( df $ age ) , tf = function ( t ) cbind ( t ) )

References
[1] Tomasz Burzykowski. Lecture notes in survival data analysis, September 2022.

[2] Peter C Austin, Ewout W Steyerberg, and Hein Putter. Fine-gray subdistribution hazard models to
simultaneously estimate the absolute risk of different event types: cumulative total failure probability
may exceed 1. Statistics in Medicine, 40(19):4200–4212, 2021.

[3] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical
Computing, Vienna, Austria, 2022.

[4] L1 Scrucca, A Santucci, and F Aversa. Regression modeling of competing risk using r: an in depth
guide for clinicians. Bone marrow transplantation, 45(9):1388–1395, 2010.

You might also like