COMPARING THE GOODNESS OF FIT AND ESTIMATES OF THE COX PROPORTIONAL HAZARDS AND THE EXPONENTIAL AND WEIBULL PARAMETRIC MODEL IN THE MULTIVARIATE SURVIVAL ANALYSIS OF THE PHILIPPINE LUNG CANCER DATA
BY: MARIANNETTE T. INOBAYA (APRIL 2008)
ABSTRACT:
IN BIOMEDICAL RESEARCH WHICH AIMS TO DETERMINE FACTORS ASSOCIATED WITH SURVIVAL. COX PROPORTIONAL HAZARDS MODEL HAS BEEN THE POPULAR METHOD USED IN DATA ANALYSIS. HOWEVER, IT IS KNOWN TO BE LESS EFFICIENT THAN THE PARAMETRIC METHODS WHEN SURVIVAL TIMES FOLLOWS A THEORETICAL DISTRIBUTION OF THE PROPORTIONAL HAZARDS ASSUMPTION IS NOT VALID. THIS STUDY AIMED TO COMPARE COX PROPORTIONAL HAZARDS WITH TWO PARAMETRIC METHODS, THE WEIBULL AND EXPONENTIAL MODELS, USING THE LUNG CANCER DATA FROM THE PHILIPPINE CANCER SOCIETY-MANILA CANCER REGISTRY. THE COMPARISON FOCUSED ON THE GOODNESS OF FIT USING THE ANALYSIS OF RESIDUALS AND AKAIKE INFORMATION CRITERION (AIC), PRECISION OF ESTIMATES AS ASSESSED BY THE STANDARDIZED VARIABILITY AND WIDTH CONFIDENCE INTERVALS, AND THE POWER TO IDENTIFY EFFECT MODIFIERS AND CONFOUNDERS.
A TOTAL OF 281 LUNG CANCER PATIENTS WERE INCLUDED IN THE MULTIVARIATE SURVIVAL ANALYSIS TO DETERMINE THE ASSOCIATION BETWEEN TREATMENT AND SURVIVAL. THE THREE MODELS EXHIBITED GOOD FIT WITH THE DATA BASED ON THE ANALYSIS OF RESIDUALS BUT THE WEIBULL AND EXPONENTIAL MODELS WERE BETTER FITTED BASED ON AIC. THE EXPONENTIAL MODEL PRODUCED ESTIMATES WITH THE LOWEST STANDARDIZED VARIABILITY AND THE NARROWEST GAP IN THE 95% CONFIDENCE INTERVAL. THIS MODEL ESTIMATED THE HAZARD RATIO OF TREATMENT AT 0.48 (95% CI: 0.12 - 1.99, P=0.313) FOR PATIENTS DIAGNOSED IN PRIVATE HOSPITALS AND 4.54 (95% CI: 2.30 - 8.94, P LESS THAN 0.001) AMONG THOSE IN GOVERNMENT HOSPITAL.
THE PARAMEDIC MODELS PERFORMED JUST AS GOOD AS THE COX MODEL IN TERMS OF GOODNESS OF FIT WHILE THE EXPONENTIAL MODEL PRODUCED THE BEST ESTIMATES IN THE MULTIVARIATE ANALYSIS OF SURVIVAL DATA. THESE RESULTS SHOWED THAT THE PARAMETRIC MODELS CAN PERFORM AS WELL AS THE COX PROPORTIONAL HAZARDS MODEL IN THE MULTIVARIATE ANALYSIS OF SURVIVAL DATA. METHODS OF COMPARISON ARE AVAILABLE TO HELP RESEARCHERS IN DECIDING WHICH MODEL IS MOST APPROPRIATE FOR THEIR DATA AND THE PURPOSE.
ABSTRACT:
IN BIOMEDICAL RESEARCH WHICH AIMS TO DETERMINE FACTORS ASSOCIATED WITH SURVIVAL. COX PROPORTIONAL HAZARDS MODEL HAS BEEN THE POPULAR METHOD USED IN DATA ANALYSIS. HOWEVER, IT IS KNOWN TO BE LESS EFFICIENT THAN THE PARAMETRIC METHODS WHEN SURVIVAL TIMES FOLLOWS A THEORETICAL DISTRIBUTION OF THE PROPORTIONAL HAZARDS ASSUMPTION IS NOT VALID. THIS STUDY AIMED TO COMPARE COX PROPORTIONAL HAZARDS WITH TWO PARAMETRIC METHODS, THE WEIBULL AND EXPONENTIAL MODELS, USING THE LUNG CANCER DATA FROM THE PHILIPPINE CANCER SOCIETY-MANILA CANCER REGISTRY. THE COMPARISON FOCUSED ON THE GOODNESS OF FIT USING THE ANALYSIS OF RESIDUALS AND AKAIKE INFORMATION CRITERION (AIC), PRECISION OF ESTIMATES AS ASSESSED BY THE STANDARDIZED VARIABILITY AND WIDTH CONFIDENCE INTERVALS, AND THE POWER TO IDENTIFY EFFECT MODIFIERS AND CONFOUNDERS.
A TOTAL OF 281 LUNG CANCER PATIENTS WERE INCLUDED IN THE MULTIVARIATE SURVIVAL ANALYSIS TO DETERMINE THE ASSOCIATION BETWEEN TREATMENT AND SURVIVAL. THE THREE MODELS EXHIBITED GOOD FIT WITH THE DATA BASED ON THE ANALYSIS OF RESIDUALS BUT THE WEIBULL AND EXPONENTIAL MODELS WERE BETTER FITTED BASED ON AIC. THE EXPONENTIAL MODEL PRODUCED ESTIMATES WITH THE LOWEST STANDARDIZED VARIABILITY AND THE NARROWEST GAP IN THE 95% CONFIDENCE INTERVAL. THIS MODEL ESTIMATED THE HAZARD RATIO OF TREATMENT AT 0.48 (95% CI: 0.12 - 1.99, P=0.313) FOR PATIENTS DIAGNOSED IN PRIVATE HOSPITALS AND 4.54 (95% CI: 2.30 - 8.94, P LESS THAN 0.001) AMONG THOSE IN GOVERNMENT HOSPITAL.
THE PARAMEDIC MODELS PERFORMED JUST AS GOOD AS THE COX MODEL IN TERMS OF GOODNESS OF FIT WHILE THE EXPONENTIAL MODEL PRODUCED THE BEST ESTIMATES IN THE MULTIVARIATE ANALYSIS OF SURVIVAL DATA. THESE RESULTS SHOWED THAT THE PARAMETRIC MODELS CAN PERFORM AS WELL AS THE COX PROPORTIONAL HAZARDS MODEL IN THE MULTIVARIATE ANALYSIS OF SURVIVAL DATA. METHODS OF COMPARISON ARE AVAILABLE TO HELP RESEARCHERS IN DECIDING WHICH MODEL IS MOST APPROPRIATE FOR THEIR DATA AND THE PURPOSE.