Pro gradu –thesis Tuija Hevonkorpi
Basic of survival analysis Weibull model Frailty models Accelerated failure time model Case study
Analysis of data from a given time origin until occurence of a specific point in time Two main difficulties: Observed survival time are often incomplete Specifying the true survival time
Occurs when the favoured endpoint is not observed Complicates the exact distribution theory and the estimation of quantiles Special statistical models and methods for analysing data arises
Moment in time when the patient was recruited until endpoint occurs Only calculated for those who encounter the endpoint Survivor function summarises the distribution of the survival times Censoring time and survival time are statistically independent random variables
Describes patient´s probability to survive from the time origin t 0 over a specific time t. The probability that survival time is less than t is described with the distribution function of T, F(t).
The approximate probability for a patient encountering the endpoint in the next point in time t i+1, on condition that the endpoint has not been encountered at time t i Connection b/w the hazard and the survivor function can be easily made, where
Useful connections between the functions used in the analysis of survival data
Survivor function of the Weibull distribution is at the same time a proportional hazards model and an accelerated failure time (AFT) model Mathematically easy to handle Characterised by the scale,, and the shape,, parameter The hazard function:, for 0≤ t < ∞ The proportional hazards model for a patient i is
hazard decreases monotonically hazard increases monotonically reduces to constant exponential hazard
Often survival times are not independent More than one endpoint occuring for one patient – repeated event times within a patient The random effect is refered to as frailty Frailty is unobserved variation between patients - the most frail encounter the endpoint earlier than those not so frail
An alternative way to model failure time data Hazard function does not have to follow a specific distribution Regression parameters are robust towards the neglected covariates Best described by the survivor function The per cent of patients in the group A that live longer than t, is equal to the per cent of patients in the group B that live longer than t The survival time is speeded up or slowed down by the effect of the explanatory variable
Main objective is to evaluate the time to significant pain relief with the active medication group compared to placebo group Three models: A proportional hazards model with Weibull distributed event times and gamma frailty term A proportional hazards model with Weibull distributed event times and log- normal frailty term An AFT model with log-normal distributed event times and log-normal frailty term
Patients were randomised in 3:1 ratio in the two treatment groups 113 patients experienced two pain episodes, 6 patients only one. Pain episode Treatment group N% FirstPlacebo Active SecondPlacebo Active
Pain episodeTreatment group Time to pain relief N% FirstPlacebo5 min min min min14.55 Active5 min min min min min23.13
The model for the log-likelihood function with gamma frailty effect which in the NLMIXED- procedure can be written for patient i as
Kaplan-Meier estimate and the population survivor function for the two treatment groups separately
The population survivor function is calculated as The subject specific estimated survivor functions are obtained from where u i is the predicted frailty term and H 0 (t) the Weibull baseline hazard,
Individual and population survivor function estimate for the active treatment group
There is no explicit form for the the marginal likelihood Instead of integrating out the frailty, numerical integration is done using the NLMIXED-procedure in SAS software The log-likelihood function is of from
AFT model with log-normally distributed event times and log-normal frailty term The log-likelihood function is where, in where is the cumulative distribution function of the standard normal distribution.
The functions cross because different treatment is given to different patients when the usual re- parametrisation of the survivor function of the AFT model does not occur necessary
Model- 2 log-likelihoodAIC No frailty Gamma frailty Log-normal frailty AFT model with frailty All but log-normal frailty and AFT with frailty differ from each other statistically significantly In all analyses, the hazard ratio, or the accelerator factor in AFT model, was calculated, and the difference between the two models was not statistically significant
Questions?