SURVIVAL DURING TREATMENT PERIOD OF PATIENTS WITH SEVERE

HEART FAILURE ADMITTED TO INTENSIVE CARE UNIT AT GONDAR

UNIVERSITY HOSPITAL, GONDAR, ETHIOPIA

BY

AZMERA HAILAY GIDEY

THESIS SUBMITTED TO

DEPARTMENT OF STATISTICS

COLLEGE OF NATURAL AND COMPUTATIONAL SCIENCES

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE

OF MASTER OF SCIENCE IN BIOSTATITICS

UNIVERSITY OF GONDAR, GONDAR, ETHIOPIA

MARCH 2015

1
SURVIVAL DURING TREATMENT PERIOD OF PATIENTS WITH SEVERE HEART
FAILURE ADMITTED TO INTENSIVE CARE UNIT AT GONDAR UNIVERSITY
HOSPITAL, GONDAR, ETHIOPIA

BY

AZMERA HAILAY GIDEY

ADVISOR

ESSEY KEBEDE (PhD)

UNIVERSITY OF GONDAR, GONDAR, ETHIOPIA

MARCH 2015

1
APPROVAL SHEET -2

We, the undersigned, members of the Board of Examiners of the final defense by Azmera Hailay
have read and evaluated her thesis entitled “Survival during treatment Period of Patients with
Severe Heart Failure Admitted to Intensive Care Unit at Gondar University Hospital,
Gondar, Ethiopia” and examined the candidate. This is therefore to certify that the thesis has
been accepted in partial fulfillment of the requirements for the degree of Master of Science in
Biostatistics.

____________________ ________________ ____________________

Name of Chair Person Signature Date

____________________ ________________ ____________________

Name of Main Advisor Signature Date

____________________ ________________ ____________________

Name of Co-advisor Signature Date

____________________ ________________ ____________________

Name of Internal Examiner Signature Date

____________________ ________________ ____________________

Name of External Examiner Signature Date

2
 ACKNOWLEDGEMENTS

First, and foremost, I thank almighty God and his mother Saint Mary for giving me the
opportunity to pursue my graduate study at the Department of Statistics, University of Gondar.

I would like to gratefully and sincerely thank my thesis advisor and instructor, Dr. Essey Kebede,
for his patience in repeatedly reading the draft manuscript of this study and for making
constructive comments and suggestions from which I have benefited a lot. It is a palpable
fact that, without his closer follow-up and continuous encouragement with valuable comments,
this thesis would not have been finalized in its present structure.

I am also indebted to take this opportunity to appreciate my family specially my father Mr.
Hailay Gidey for being with me closely throughout my study.

I am grateful to Gondar University Hospital for permission to use the survival time during
treatment period of patients with severe heart failure dataset for this study and intensive care unit
ward staff for cooperation during data collection.

My special thanks also go to Mekelle University for sponsoring me during my study years.

i
Table of Contents
ACKNOWLEDGEMENTS..........................................................................................................................i
LIST OF ABBREVIATIONS.....................................................................................................................iv
ABSTRACT................................................................................................................................................v
CHAPTER ONE..........................................................................................................................................1
INTRODUCTION.......................................................................................................................................1
1.1 Background of the Study...................................................................................................................1
1.2 Statement of the Problem...................................................................................................................4
1.3 Objective of the Study.......................................................................................................................4
1.4 Significance of the Study...................................................................................................................5
1.5 Limitations of the Study....................................................................................................................5
CHAPTER TWO.........................................................................................................................................6
REVIEW OF LITERATURE......................................................................................................................6
2.1 Etiologies and Causes of Death of Patients with Heart Failure..........................................................6
2.2 Correlates of Survival of Patients with Heart Failure.........................................................................7
2.3 Cox Proportional Hazards versus Parametric Models......................................................................12
3.2 Survival Analysis.............................................................................................................................18
3.2.1 Survivor Function and Hazard Function...................................................................................18
3.2.2 Non-Parametric Methods..........................................................................................................20
3.2.3 Regression Models for Survival Data.......................................................................................23
3.2.3.1.1.....................................................................................................................................................24
3.3 Comparison of Cox PH and Parametric Models..............................................................................38
CHAPTER FOUR.....................................................................................................................................40
RESULTS AND DISCUSSION................................................................................................................40
4.1 Descriptive Summaries and Non-Parametric Analysis....................................................................40
4.2 Cox Proportional Hazards Model....................................................................................................43
Figure 4: Index plots of dfbeta for variables in multivariable Cox regression...........................................48
4.3 Parametric Proportional Hazards Model..........................................................................................49
4.4 AFT Models.....................................................................................................................................51
4.5 Comparison of Models....................................................................................................................56
4.6 Interpretation and Discussion of the Results....................................................................................58
CHAPTER FIVE.......................................................................................................................................62
CONCLUSIONS AND RECOMMENDATIONS.....................................................................................62

ii
5.1 Conclusions.....................................................................................................................................62
5.2 Recommendations...........................................................................................................................63
5.3 References.......................................................................................................................................65
Appendixes............................................................................................................................................70
Appendix 1........................................................................................................................................70
Appendix 2........................................................................................................................................71
Appendix 3........................................................................................................................................71

iii
 LIST OF ABBREVIATIONS

ACEI Angiotensin Converting Enzyme Inhibitors

AIC Akaike’s Information Criterion
AFT Accelerated Failure Time
CAD Coronary Artery Disease
CHD Coronary heart disease
CHF Chronic or Congestive Heart Failure
CI Confidence Interval
CKD Chronic Kidney Disease
CS Cox-Snell
CVD Cardiovascular disease
DM Diabetes Mellitus
GG Generalized- gamma
HF Heart Failure
HR Hazard Ratio
HTN Hypertension
ICU Intensive Care Unit
ICDS International Cardiovascular Disease Statistics
IHD Ischemic heart disease
K-M Kaplan-Meier
NYHA New York Heart Association
OR Odds Ratio
PH Proportional Hazard
QQ Quantile-Quantile
RHD Rheumatic Heart Disease
RR Risk Ratio
SD Standard Deviation
SIGN Scottish Intercollegiate Guidelines Network
TR Time Ratio
VHD Valvular Heart Disease

iv
 ABSTRACT

Background: Heart failure (HF) is progressive and irreversible which occurs more slowly
because of damage to the heart muscle, building up through time due to disease of the heart or a
blood vessel leading from the heart as a result of various diseases. In nearly all regions of the
world HF is both common and increasing. Predictions for the next two decades include tripling
of heart disease and stroke mortality in Latin America, the Middle East, and even sub-Saharan
Africa. This study is an attempt to study the survival during treatment period of severe heart
failure patients admitted to intensive care unit ward at Gondar university hospital.

Methodology: Data were collected from 147 severe heart failure patients admitted to the ICU
ward of Gondar University Hospital, Gondar, Ethiopia during January 2012- June 2014. Non-
parametric, semi-parametric PH, parametric PH and AFT models were used for data analysis.
Different statistical techniques were used to compare performances of semi-parametric PH,
parametric PH and AFT models.

Results: Descriptive statistical results show that predominant causes of HF were coronary heart
disease and valvular heart disease which causes 38.1% and 29.9% of the total population
respectively. In this study (26.09%) of the deaths were attributed to respiratory failure, (15.22%)
due to cardiac arrest (15.22%) multi-organ failure, (10.87%) end stage renal failure. The
variables “History of HF”, “Duration of HF”, and “Department the patient seen at first” were
found to be significant predictors of the survival during treatment period of patients with severe
heart failure admitted to ICU ward by the multivariable Cox PH model and the multivariable
exponential and Weibull PH, and multivariate Log-Logistic AFT models. The Cox PH model
was a better fit than the other models.

Conclusion: Being seen at chronic illness follow-up first, diagnosed as patient with heart failure
for the first time and lower number of years stay with heart failure significantly decrease hazard
of in-ICU mortality. Special attention to patients discharged from emergency department,
diagnosed as patient with heart failure before current admission, with higher number of years
stay with heart failure, advanced age and with comorbidity condition coronary artery disease is
recommended as possible interventions to improve in-ICU survival of patients with severe HF.

v
Key words: Intensive Care Unit, Heart Failure, Survival analysis, Cox Proportional Hazards
model, Parametric Proportional Hazards model, Accelerated Failure Time model.

vi
 CHAPTER ONE

INTRODUCTION

1.1 Background of the Study

Heart failure is often used to mean chronic heart failure or congestive heart failure (CHF).
Clinically it is impossible to define HF by a single term as a result it is defined as a complex
clinical syndrome in which there is a functional or structural impairment in the heart. This can
result from any functional or structural cardiac disorder and it impairs the ventricle’s ability to
fill with or eject blood and to deliver oxygenated blood corresponding to the requirements of the
metabolizing tissues of the body and/or doing so at increased filling pressures (Agval 2014).

Heart failure (HF) is progressive and irreversible which occurs more slowly because of damage
to the heart muscle, building up through time due to disease of the heart or a blood vessel leading
from the heart as a result of various diseases, accordingly it is a serious clinical condition which
represents the end-stage of numerous other cardiac diseases (Ponikowski et al. 2014).

A healthy heart pump 60% to 65% of the blood in the ventricle in one beat a failing heart pumps
only 40% or less, in severe failure it may drop as low as 5% of the blood in the ventricle in one
beat. Consequently, patients with heart failure have slow circulation which causes excess fluid to
be retained in the body (SIGN 2007). The hemodynamic consequences of these disturbances may
explain symptoms typically shortness of breath at rest or during exertion and/or fatigue, signs of
fluid retention such as pulmonary congestion or ankle swelling and objective evidence of an
abnormality of the structure or function of the heart at rest (Agval 2014).

Patients with HF experience frequent and distressing symptoms that can adversely impact and
decrease their quality of life and survival. Based on symptoms their condition can be categorized
into one of four functional classes developed by the New York Heart Association (NYHA).
Patients in NYHA class one have symptoms of heart failure only at levels of exertion that would

1
limit normal individuals. As a patient’s class increases, so does their severity of symptoms such
that patients in NYHA class IV have symptoms like shortness of breath at rest and are unable to
perform any physical activity without having a great deal of discomfort (SIGN 2007).

Physical symptoms are key determinants of hospitalizations or emergency department visits in

patients with heart failure. Almost all patients experience dyspnea before their hospitalizations
for heart failure. And dyspnea may be the primary reason for visits to the emergency department
in above half patients with heart failure. It is estimated that once hospitalized for HF, the
hospitalization rate ranges from 20 to 59% (Hjelm 2013). The occurrence of frequent
hospitalizations also characterizes the illness trajectory for patients with HF (Howie-Esquivel
and Dracup 2012).

Treatment goals for patients with HF are to relieve symptoms and signs, prevent hospital
admission, and improve survival. Basic HF pharmacological treatment with strong scientific
evidence includes diuretics, angiotensin converting enzyme inhibitors/angiotensin receptor
blockers, beta blockers, and aldosterone antagonists (Hjelm 2013). Angiotensin receptor blockers
(ARB) is recommended when there are adverse reactions to ACEI and their effect in HF is
equivalent to ACEI. Despite the evident effect of loop diuretics on diuresis and in reducing HF
symptoms, there is no documentation on if diuretics affect morbidity or mortality. On the other
hand, beta blockers reduce mortality when added to an ACE inhibitor (Agval 2014).

Heart failure is a major clinical problem worldwide, reaching an epidemic level in the developed
world with no known cure at this time. Approximately 26 million people worldwide are living
with heart failure, and nearly 1 million new cases are diagnosed annually worldwide, making it
the most rapidly growing cardiovascular disorder. In economically developed countries, up to
one person in five is expected to develop heart failure at some point in their life (Ponikowski et
al. 2014) and it affects1-3% of the general population (Owusu and Boakye 2013; McMurray and
Stewart 2002). It is predominantly seen in the geriatric population in these countries, with almost
80% of cases occurring in patients over the age of 65. Thus, prevalence of heart failure has been
shown to follow an exponential pattern, which rises with age and affects 6-10% of people over
age 65 (Owusu and Boakye 2013).

Despite improvements in care over the past 20 years, the outlook for patients with heart failure
remains poor, and it has a higher mortality than many of the common malignancies (Agval 2014;

2
Ponikowski et al. 2014). One year mortality in developed countries is approximately 20% while
the 5-year mortality is approximately 50-65% in population-based studies (Agval 2014). Across
the globe, 17–45% of patients admitted to hospital with heart failure die within 1 year of
admission and the majority die within 5 years of admission. Besides survival rates are better for
those treated in outpatient clinics, who typically have less severe symptoms than those treated in
hospital (Ponikowski et al. 2014). About 2–17% of individuals admitted to hospital with heart
failure die while in hospital (Mayosi 2007).

Though review conducted in 2000 revealed that population-based studies have not been
reportedly conducted to estimate the proportion of the population living with heart failure in
Africa (Bennett et al. 2012), hospital based studies have shown that, cardiovascular diseases
accounts for 7-10% of all medical admissions with chronic heart failure contributing 3-7% of
these admissions (Mayosi 2007) and in internal medicine services HF as fifth to sixth cause of
hospital admission (Cabral et al. 2011; Owusu and Boakye 2013).

In a study conducted in Ghana the high prevalence of heart failure of 76% seen in the study
supports the fact that HF is a major contributor to cardiovascular disease burden in sub-Saharan
Africa. Similar findings have been reported from Cameroon where heart failure is found to be the
fifth to sixth cause of hospital admissions. In other parts of sub-Saharan Africa, heart failure has
been found to account to 5% to 10% of hospital admissions (Owusu and Boakye 2013).

Compared to studies from other parts of the world, heart failure in Africa tends to occur at a
much younger age with most cases recorded around the 5th and 6th decade and it is not a disease
of the elderly in sub-Saharan Africa (Ponikowski et al. 2014). This young age reflect the major
contribution of rheumatic valvular disease to heart failure, but could also be accounted for
infections as it remain a common cause of heart failure in many parts of the world including
Africa and can strike at any age. Hospital case fatality among those with heart failure in Africa
ranges from 9% to 12.5%. This consistent death rate ranks heart failure among the major causes
of death of cardiovascular origin in Africa (Cabral et al. 2011; Kengne et al. 2008).

In Ethiopia, Misganaw et al. in their systematic review, aimed to document the prevalence and
mortality associated with major non-communicable diseases in Ethiopia found cardiovascular
disease accounted for 3%-12.6% hospital admission and found to have increased between 1970s
and 2000s. This systematic review found congestive heart failure reported to have caused 2.5%

3
of deaths among all age-groups in a sampled hospital-based mortality study (Misganaw et al.
2014). Similarly a study conducted by analyzing surveillance data on causes of death in Addis
Ababa found that, the leading cause of death was cardiovascular disease causing 24% of all
death. Congestive heart failure is found to be the third cause of death following hypertension and
stroke among the cardiovascular disease deaths (Misganaw et al. 2012).

1.2 Statement of the Problem

In nearly all regions of the world HF is both common and increasing and it is predicted that the
number of patients with HF to increase in countries with ageing populations. In the USA, there
were 5.8 million patients living with heart failure in 2012, and this is expected to rise to 8.5
million by 2030 (Cook et al. 13).

Nowadays cardiovascular disease has become one of the major causes of premature death and
disability in low- and middle-income countries. Moreover, predictions for the next two decades
include tripling of heart disease and stroke mortality in Latin America, the Middle East, and even
sub-Saharan Africa, a rate of increase that exceeds than for any other region, except for Asian
and Pacific Island countries (ICDS 2004). However, heart failure as cardiovascular complication
remains unexplored largely in Africa (Bennett et al. 2012).

The long-term prognosis associated with HF is also poor (Cook et al. 13). Reis et al. found that
the three-year mortality from CHF related to severity of symptoms and ranges from 40% in
patients in NYHA functional class I to 82% in those in class IV (Reis et al. 1997). A Hospital-
based study conducted in Italy also found the highest mortality rates in the ICU and emergency
department with 19 and 78% deaths, respectively (Grigioni et al. 2002).

But little research work has been reported on HF patients admitted to Intensive Care Units
(ICUs), clinical department in which only critically ill patients with NYHA class IV were
treated, worldwide. Furthermore, all the population based follow-up studies in survival of HF
patients are based on the semi-parametric proportional hazards model. To fill these gap this
thesis is aimed to investigate the survival during treatment period of severe heart failure patients
admitted to intensive care unit ward of Gondar University Hospital using the Cox Proportional
Hazards model, Parametric Proportional Hazards model and Accelerated Failure Time models.

4
1.3 Objective of the Study

The general objective of the study is to assess the survival experience during treatment period of
patients with severe heart failure admitted to intensive care unit ward at Gondar university
hospital and to identify factors that influence their survival using semi-parametric Proportional
Hazards, parametric Proportional Hazards and Accelerated Failure Time models.

The specific objectives are

To model survival during treatment period of patients with severe HF admitted to ICU ward

To study the influence of predictors in-ICU ward survival of patients with severe HF

To examine etiologies of heart failure of patients with severe HF admitted to ICU ward

To investigate cause of death of patients with severe heart failure

To compare Cox PH, parametric PH and AFT models in explaining the data set

1.4 Significance of the Study

In order to make a reasonable recommendation in solving the problem of high in-ICU mortality
rate of patients with severe heart failure it is necessary to understand the in-ICU mortality rate
and factors influencing in-ICU survival in the hospital. From that concrete recommendations
may be given to the hospital society and health program managers for improving the health
policies and in-ICU care strategy of patients with severe heart failure. Thus, the findings from

this research are hoped to be useful in providing information about the risk factors or the
most influential covariates that have significant impact on survival during treatment period of
patients with severe heart failure during ICU ward admission at Gondar University Hospital and
to identify death risk extent of patients under these significant factors.

1.5 Limitations of the Study

The major limitation of the study goes with the problems related to the use of secondary data.
There are different factors that are expected to have impacts on the survival of heart failure
patients such as body mass index, alcoholism and smoking status. However, these variables are
not found in the patients’ charts. This may be considered as the limitation of this study.

5
 CHAPTER TWO

REVIEW OF LITERATURE

2.1 Etiologies and Causes of Death of Patients with Heart Failure

The most common causes of functional deterioration of the heart are damage or loss of heart
muscle, acute or chronic ischemia, valve dysfunction, increased vascular resistance with
hypertension, or the development of a tachyarrhythmia such as atrial fibrillation. These problems
are caused by only a limited number of ways in which the function of the heart can be affected
and cause heart failure (Ågren 2010).

The most common underlying cause of heart failure in high-income countries is coronary artery
disease. Available data suggest that the causes of heart failure in developing world differ from
those seen in the developed world. Infections remain a common cause of heart failure in many
parts of the developing world such as rheumatic fever due to preventable bacterial infections is a
prominent cause of heart failure in Africa, Asia, Australasia and Latin America (Ponikowski et
al. 2014).

Jingi et al. (2013) found hypertensive heart disease, cardiomyopathies and valvular heart diseases
to account respectively for 54.5%, 26.3% and 24.5% of causes of adult chronic heart failure in an
urban Cameroon. Similarly Cabral et al. (2011) investigated that reported etiology of congestive
heart failure as valvulopathies (35%), cardiomyopathies (32%), hypertension (15%), pericarditis
(7%), chronic obstructive pulmonary disease (8%) and congenital heart diseases (3%) in
Cameroon. Grimaldi et al. (2014) reported the major causes of HF as RHD (31%), CHD (29%),
hypertensive cardiomyopathy (16%), highly suspected IHD (10%), endomyocardial fibrosis
(6%) and right ventricular failure (5%) in Uganda. Onwuchekwa and Asekomeh (2009) found
that hypertension as the most frequent etiology (56.3%) in south Nigeria. Cardiomyopathies
(10.17%), chronic renal failure (7.80%), severe anemia (4.72%), and rheumatic heart diseases
(4.26%) were following in this study.

6
Regarding causes of death Volpe et al. (2010) in Naples, Italy found that deaths attributed to
progression of CHF (56.9%), sudden death (11.7%), acute myocardial infarction (6.9%),
cardiogenic shock (3.5%), acute pulmonary edema (2.3%), pneumonia (2.3%), end-stage renal
disease (1.2%), stroke (1.2%), pulmonary embolism (1.2%), other non-cardiac causes (10.5%),
and non-defined causes (2.3%). Henkel et al. (2008) in Olmsted County, Israel showed
cardiovascular death as dominant (57%) and coronary heart disease deaths represented 63% of
cardiovascular deaths. Of the 43% non-cardiovascular deaths, the most common causes of death
were pulmonary disease (28%) and cancer (25%), followed by central nervous system disease
(12%), gastrointestinal disease or genitourinary disease (12%), and diabetes mellitus or
endocrine disorders (9%).

Merlo et al. (2001) in Sweden also found that CVD as the most frequent cause of death, with
three diagnoses (chronic ischemic heart disease, acute myocardial infarction, and HF) accounting
for about 53% of the cases. Nine other conditions caused the deaths of 21% of the patients
(arteriosclerosis, valve disorder, complications of heart disease, old myocardial infarction,
arrhythmias, pulmonary embolism and infarction, stroke, pneumonia, and diabetes mellitus).
Cowie et al. (2000) in west London found that the single most common cause as coronary heart
disease (79 cases (36%)). The majority of deaths were related to CVD (83 (92%)) in this study.

2.2 Correlates of Survival of Patients with Heart Failure

Many studies have been reporting that male patients with heart failure to be at increased risk of
mortality than females using Cox proportional-hazards regression model. Barlera et al. (2012) in
Italy found that male sex to be highly significant predictor of death (HR=0.75 95% CI (0.66 -
0.85)). A Community based study by Henkel et al. (2008) in Olmsted County, Israel showed
male sex to be associated with the increased risk of all-cause mortality (HR: 1.28; 95% CI: (1.10,
1.49)) and cardiovascular mortality (HR: 1.26; 95% CI: (1.03, 1.54)). Pocock et al. (2005), from
patients in the Candesartan in Heart Failure, reported that 17% hazard reduction for females
(HR=0.83 95% CI 0.76 - 0.91). In a related study by Gustafsson et al. (2003) in Denmark found
that male gender to have negative influence on survival of the patients with CHF, after admission
to hospital (RR=1.26 95% CI (1.17–1.36)). Similarly Ho et al. (1993) from the Framingham

Heart Study data found survival after the development of heart failure to be better in women than
in men (HR=0.64; 95% CI, 0.54-0.77).
7
Similarly age was found to be the most powerful predictor of death in various studies. Barlera et
al. (2012) found a linear increase of 4% in hazard for every year (HR=1.04 95% CI (1.03 -
1.04)). A Population-based retrospective study conducted by Capell et al. (2012) in Catalonia
(north-east Spain) confirmed that an increased risk of death with older age (HR=1.06; 95% CI,
1.06-1.07), patients aged above 65 including were at increased risk of death than those aged less
than 65. In a related 15-year prospective follow-up study at the university hospital of Naples
(Italy), Volpe et al. (2010) found that age as the most significant predictor of death (unadjusted
HR=1.06; 95% CI: 1.03–1.09; adjusted HR=1.05 95% CI: 1.02–1.08). Henkel et al. (2008) from
Israel found advanced age to be associated with the increased risk of all-cause mortality
(HR=1.057; 95% CI: (1.049, 1.064)) and cardiovascular mortality (HR=1.057; 95% CI: (1.049,
1.064)). A study conducted by Levy et al. (2006) using previously collected data in 6 cohorts of
patients from different countries, found in the univariate analyses older age (measured in
decades) was associated with increased mortality (HR: 1.134; 95% CI: (1.034–1.245)). Pocock et
al. (2005) found that 46% increase in hazard for every 10 years of age>60 (HR=1.46; 95% CI:
1.38-1.54).

Abraham et al. (2008) employed logistic regression to identify significant predictors of in

hospital mortality in U.S. and they found older age (measured in decades) to be a significant
predictor of in-hospital mortality (OR=1.34; 95% CI, 1.26–1.41). Gustafsson et al. (2003) from
Denmark found increasing age to have negative influence on survival of the patients with CHF,
after admission to hospital (Age (per year) RR=1.04; 95% CI: 1.03–1.04). In a study in the
Bologna University Hospital of S. Orsola-Malpighi, Italy by Grigioni et al. (2002) age has found
to be the most significant predictor of in-hospital mortality (adjusted RR=1.04; 95% C.I.: 1.03-
1.06 per year of age). Similarly MacIntyre et al. (2000) used data gathered through Scotland
hospitals and found age to have a powerful effect with the 1-month case-fatality rate, increasing
from 10.4% in those aged, <55 years to 25.9% in those aged >84 years. A population based study
in west London conducted by Cowie et al. (2000) also found age to be associated with
cardiovascular mortality (unadjusted HR=1.48; 95% CI:1.10-1.79 and adjusted HR=1.26; 95%

CI: 1.01-1.57). Ho et al. (1993) from a population-based prospective epidemiological study also
found that advancing age to be a predictor of shortened survival in individuals with CHF, with a

8
27% increase in mortality per decade of advancing age in men and a 61% increase per decade in
women.

Ischemic etiology was also found to be highly significant predictor of death in some studies.
Barlera et al. (2012) and Levy et al. (2006) found ischemic etiology to be associated with
increased mortality (HR=1.21; 95% CI: 1.10 - 1.33) and (unadjusted HR=1.483; 95% CI: 1.194–
1.841, adjusted HR=1.354; 95% CI: 1.074–1.707) respectively. Capell et al. (2012) also found an
increased risk of death with ischemic heart disease (HR=1.18; 95%CI: 1.02-1.36). Similarly, Ho
et al. (1993) found that among men with CHF, survival was better in cases attributed to CHD or
HTN than in those attributed to VHD (hazards ratio, 1.68 relative to CHD cases) or other and
unknown causes (hazards ratio, 1.69 compared with CHD cases).

Most of the patients with heart failure are found to have other medical conditions, such as kidney
disease, lung disease, diabetes or others. These conditions are termed as comorbidities and are
particularly common in individuals with heart failure who are elderly or have been hospitalized.
Many heart failure medications may not be suitable for patients who also have other medical
conditions. This is because the drugs used to treat different diseases may interfere with one
another, or a drug suitable for treating one disease may worsen the other (Ponikowski et al.
2014). Consequently, a number of studies have been reported these conditions to have influence
on survival of patients with HF.

Capell et al. (2012) and Barlera et al. (2012) found an increased risk of death for patients with
diagnosis of diabetes mellitus (HR=1.53; 95% CI, 1.33-1.76) and (HR=1.25; 95% CI: 1.12 -
1.38) respectively. Similarly Henkel et al. (2008) found that diabetes mellitus to be associated
with the increased risk of all-cause mortality and cardiovascular mortality (HR: 1.44; 95% CI:
1.21-1.71). Pocock et al. (2005) found that diabetes as the third most important predictors of
mortality following age and ejection fraction. They reported that diabetes was associated with
around a doubling of risk of either death or the composite outcome when insulin-treated
(HR=2.03 95% CI 1.80-2.29), and a 50% increase in risk in generally less severe, non-insulin-

treated diabetes

(HR=1.58 95% CI 1.43-1.74). Gustafsson et al. (2003) found diabetes mellitus to have negative
influence on survival of the patients with CHF, after admission to hospital (adjusted RR=1.42,

9
95%CI: 1.30–1.56). Ho et al. (1993) found that among women with CHF, diabetes mellitus was
harbinger of shortened survival (HR: 1.70; 95%CI: 1.21-2.38), with a 70% increase in mortality.

Barlera et al. (2012) found presence of chronic obstructive pulmonary disease (COPD) to be
highly significant predictor of death with very powerful impact on mortality associated with 43%
increase in risk (HR: 1.48; 95%CI: 1.15 – 1.90). Abraham et al. (2008) found increased risk of
in-hospital mortality was associated with chronic obstructive pulmonary disease (unadjusted OR:
1.233; 95% CI: 1.115–1.363). In the same way Gustafsson et al. (2003) found history of COPD
to have negative influence on survival of the patients with CHF, after admission to hospital
(adjusted RR: 1.36; 95% CI: 1.25–1.47).

A study by Capell et al. (2012) found an increased risk of death for patients with diagnosed
chronic kidney disease (adjusted HR=1.73; 95% CI, 1.45-2.05). Henkel et al. (2008) also found
chronic kidney disease to be associated with the increased risk of all-cause mortality and
cardiovascular mortality (adjusted HR=2.38; 95% CI, 1.76-3.23). In their study severe chronic
kidney disease was associated with more than a 2-fold increase in the risk of overall and
cardiovascular death adjusting for other clinical characteristics.

Grigioni et al. (2002) found that presence of systemic hypertension to be associated with the risk
of hospital mortality (adjusted RR: 0.46; 95%CI, 0.29-0.73; P=0.001). On the contrary Capell et
al. (2012) found hypertension to had a protective effect on risk of death (adjusted HR=0.73; 95%
CI, 0.64-0.84). In the same way MacIntyre et al. (2000) using logistic regression analysis,
reported prior admission with hypertension reduced the short-term case-fatality rate (adjusted
OR=0.78 95% CI; 0.57–1.06).

Abraham et al. (2008) found cerebrovascular disease to be associated with increased risk of in-
hospital mortality (adjusted OR=1.32; 95% CI, 1.13–1.54). A study conducted by MacIntyre et
al. (2000) used both Logistic regression analysis and Cox proportional hazards regression model.
They reported cerebrovascular disease significantly increased the short-term case fatality rate

(adjusted OR=1.17; 95% CI, 1.04–1.32 in men and adjusted OR=0.97; 95%CI, 0.87–1.09 in
women), and long-term mortality (adjusted HR=1.31; 95% CI, 1.23–1.40 in men and adjusted
HR=1.33; 95% CI, 1.26–1.41 in women).

10
Barlera et al. (2012) perceived atrial fibrillation to be highly significant predictor of death and to
increase risk of death (HR=1.19; 95% CI, 1.06-1.34). In opposition, a study by MacIntyre et al.
(2000) reported prior admission with atrial fibrillation reduced the short-term case-fatality rate
by 32% in men and 24% in women (adjusted OR=0.68; 95% CI, 0.56–0.82 in men and adjusted
OR=0.76; 95% CI, 0.64–0.89 in women).

Gustafsson et al. (2003) found that decreasing renal function to have negative influence on
survival of the patients with CHF, after admission to hospital (adjusted RR=0.73 95%CI, 0.69–
0.77). A related study by Grigioni et al. (2002) reported that presence of renal insufficiency to be
predictors of in-hospital mortality in the multivariate analysis (adjusted RR=1.90 95% CI, 1.27-
2.86). MacIntyre et al. (2000) in their logistic result showed that renal disease prior admission
increased the short-term case fatality rate (OR=1.58; 95% CI, 1.17–2.12 in men and OR=1.16;
95% CI, 0.82–1.63 in women). And Cox Proportional Hazards results confirmed renal disease
prior admission increased long-term mortality (HR=2.12; 95% CI, 1.80–2.50 in men HR=1.58
95% CI, 1.32–1.88).

Some studies also reported that physical examination findings of patients with heart failure have
strong association with their survival. Barlera et al. (2012) found that a 2% increase in risk per 1
Kg/m2 decrease in body mass index (HR=0.98; 95% CI, 0.97- 0.99). Similarly Levy et al. (2006)
found lower body mass index to be associated with increased mortality (unadjusted HR=0.960;
95% CI, 0.943–0.979). Pocock et al. (2005) showed an inverse association between body mass
index and mortality, while appeared confined to patients whose body mass index is below the

median of 27.5 kg/m (HR=1.07; 95% CI, 1.05-1.09). Such low weight patients had a 9.3%
increase in mortality hazard for each 1 kg/m2 reduction. Thus, compared with body mass index
of>=27.5 kg/m2 a BMI of 20 kg/m2 carries a 67% increased risk.

Barlera et al. (2012) and Volpe et al. (2010) confirmed that increasing heart rate to be significant
predictor of death (HR=1.004; 95% CI, 1.001-1.007) and (unadjusted HR=1.03 95% CI, 1.01–
1.04) respectively. Similarly Abraham et al. (2008) found increased heart rate measured per 10
beats/min increase between 65 and 110 beats/min, to be most discriminative variable for in-
hospital mortality (OR=1.094; 95% CI, 1.062–1.127).

11
Systolic blood pressure has also been reported to have impact on survival of patients with heart
failure. Barlera et al. (2012) found systolic blood pressure to be independent, highly significant
predictor of death (HR=0.99, 95% CI, 0.987 - 0.994). They reported one unit increase in systolic
blood pressure caused a 1% decrease in hazard only for systolic blood pressure < 140 mmHg. A
related study by Abraham et al. (2008) found that low systolic blood pressure to be associated
with in-hospital mortality (OR=0.83 95% CI, 0.80–0.86; P<0.0001). They reported increased
SBP at admission, up to a threshold of approximately 160 mm Hg, was associated with a lower
risk of in-hospital mortality: each 10-mm Hg increase up to 160 mm Hg was associated with a
17% reduction in in-hospital mortality.

Other factors such as time and duration of diagnose as having heart failure and department the
patient seen first have been reported to be associated with risk of death in patients with heart
failure. Abraham et al. (2008) found patients were at lower risk if HF was diagnosed for the first
time during the index admission. Their logistic regression result showed significant association
between no known HF before this admission and survival (OR=0.524 95% CI, 0.434–0.632).
Gustafsson et al. (2003) found that duration of heart failure, measured per month, to have
independent negative influence on survival of the patients with CHF, after admission to hospital
(RR=1.002; 95% CI, 1.001–1.003). And Grigioni et al. (2002) found that discharge of patients
directly from the emergency department to be associated with an increased risk of hospital
mortality (P<0.001).

2.3 Cox Proportional Hazards versus Parametric Models

A review of literature on survival analysis used in different journals reveals that the Cox PH
model is the most widely used way of analyzing survival data in clinical research. Researchers in
medical sciences often tend to prefer semi-parametric instead of parametric models because of
fewer assumptions. Nowadays, Cox proportional hazard versus parametric models attracted
considerable attention, because not only they do not need PH assumption but also thanks to
availability of standard statistical software maximum likelihood parameter estimation and testing
can be done readily (Bradburn et al. 2003).

The main drawback of parametric models is the need to specify the distribution that most
appropriately reflects that of the actual survival times. The appropriate use of these models offers
the advantage of being slightly more efficient; they yield more precise estimates (i.e. smaller
12
standard errors) and that in the parametric models we often use maximum likelihood procedures
to estimate the unknown parameters in which this technique and its interpretation are familiar for
researchers (Bradburn et al. 2003; Moghimi-Dehkordi et al. 2008). The parametric approach
offers more in the way of predictions, and the AFT formulation allows the derivation of a time
ratio, which is arguably more interpretable than a ratio of two hazards in Proportional Hazards
model.

Performance comparison between Cox PH and parametric survival models have also been made
on other data sets other than survival of patients with HF. The analysis of Survival of Patients
with Gastric Carcinoma of data from a historical cohort study of southern Iran with a diagnosis
of stomach cancer has been made using Cox PH and parametric Lognormal, Exponential,
Gompertz, Weibull, Log logistic and Gamma regression models (Pourhoseingholi et al. 2011).
The proportionality assumption is checked and found to be hold. The comparison of parametric
and semi parametric models were made based on AIC in which all parametric survival models
were performed better than the Cox model besides Log logistic is found to be an efficient model
by the same way. Another study on the survival of tuberculosis patients admitted in randomized
controlled clinical trial in India reported that AFT model gave smaller deviance than PH models
(Ponnuraja and Venkatesan 2010). Based on these results they concluded that AFT models had

seemed to be more appropriate modeling than the Cox PH model.

Similarly, a study conducted on prognostic factors of survival time after hematopoietic stem cell
transplant in acute lymphoblastic leukemia patients in Shariati Hospital assessed the performance
among AFT and Cox's models using explained variation and goodness of fit test methods. PH
assumption was not met for this model accordingly Cox with time-varying coefficients was
applied. Likelihood-ratio chi-square statistics was used to compare performance of Generalized-
Gamma family. Weibull distribution has the smallest AIC among the AFT models; therefore it is
the best fitted model on data. Explained variation and Cox-Snells residuals plot methods showed
that the Weibull AFT was the best model comparing with Cox PH and Cox with time-varying
coefficients. Cox-Snell residual shows Weibull AFT fitted to data better than other distributions
in multivariate analysis (Sayehmiri et al. 2008). The quantile-quantile (QQ) plot was used to
check the adequacy of AFT assumption and found to be straight line with slop one indicating
adequacy of the fitted Weibull distribution.

13
Moghimi-Dehkordi1 et al. 2008 compared parametric and semi-parametric methods in patients
with stomach cancer in southern Iran data. They used Akaike Information Criterion (AIC) to
evaluate among models. In this study the proportional hazard assumptions were hold. But the
Weibull and Exponential models were found to be the most favorable models for the data set.
Another study of patients with gastric cancer in Iran used AIC and standardized measure of
variability defined as to be analogous to the coefficient of variation for comparison among
parametric and semi parametric models (Pourhoseingholi et al. 2007). In this study the
proportional hazards assumption found to be hold and they reported both Cox and Exponential to
be similar in multivariate analysis. Their comparison methods revealed that there may not be a
single model that is substantially better than others, but in univariate analysis the data strongly
supported the log normal regression among parametric models and it can be lead to more precise
results as an alternative to Cox.

In contrast, in the analysis of survivals of breast cancer relapse time with different treatments
consistent results were obtained from both accelerated failure time model and Cox proportional
hazard model. But Cox PH is chosen over accelerated failure time model to calculate the

appropriate survival curves of relapse time for patients in different treatment groups. i.e., With
respect to predicting survival curve, Cox-PH model gives better fit than AFT models (Conge and
Tsoikos 2010). Similarly, in analysis of survival in acute severe illness, AFT models identified
the same predictors as the Cox model. In this study the proportionality hazards assumption is
checked and found to be hold. Cox PH was also chosen over accelerated failure time model
using explained variation and goodness of fit tests (Khanal 2009).

In this chapter we have reviewed the various factors that are associated with survival time of HF
patients, the statistical methods employed to address it and the comparison of the Cox PH and
parametric models on different datasets. The literature showed that different researchers used not
the same statistical measures to compare the performances of the Cox PH and Parametric
models. The performance of the survival models also vary depending on the data used and might
also be due to the statistical measures used to compare the performances. In addition, only a few
methods have tried to check the underlying model assumptions. In a review of survival analyses
in cancer journals only 5% of all studies used the Cox regression model checked the underlying
assumptions (Moghimi-Dehkordi et al. 2008).

14
15
 CHAPTER THREE

DATA AND METHODOLOGY

3.1 Data and Variables of the Study

The data used in this study is obtained from Gondar University Hospital, north western Ethiopia,
in Gondar city Amhara regional state of Ethiopia. This hospital is used as referral Centre for
different health centers in woredas and towns around the city. The data for this study is obtained
specifically from the intensive care unit (ICU) ward of the hospital. The intensive care unit ward
of the hospital has started in January 2012. From the day ICU ward started working until the data
collection date, June 2014, there were 147 heart failure patients admitted to the ward, all of
whom were in NYHA class IV. Thus, the data used in this study is collected from 147 patients.

All patients admitted to ICU ward within the indicated twenty nine months in the hospital had
information on their chart on demographic profiles, clinical variables, etiology, comorbidity and
other clinical variables at admission to the ward maintained by the ICU unit. Thus, we referred
the patients’ charts to gather information about the patient and we had had the warranty for doing
so from the concerned party and we had kept all the necessary information concerning the
privacy of the patient out of this thesis. All the information included in the thesis is not disclosed
to anyone except for us and data collectors.

Survival analysis always measures the time from a defined starting point to the occurrence of a
given event. In this study the response variable measures the length of treatment time in days
from ICU ward entry to event or censoring and the event of interest is death.

The most important expected correlates of the survival experience of patients with HF from
literature reviews and their theoretical justification included in this study are grouped as clinical
and demographic variables, and comorbid conditions. Categorical predictors included in this
study are given below in table 3.1. Continuous predictors are systolic blood pressure, weight,
heart rate, white blood cell count (WBCC) and duration of heart failure.

16
Table 3.1: Description and coding of categorical predictor variables

Variable Name Variable description Category

Agecat Age of the patient 0 = (0, 49]
1 = (49, 65]
2 = (65,∞)
Sex Sex of the patient 0 = Male
1 = Female
Residence Residence of the patient 0 = Rural
1 = Urban
MS Marital status of the patient 0 = Single
1 = married
2 = widowed or divorced
HFpatient History of heart failure 0 = New
1= HF patient before
Dptseenf Department seen first by the 0 = Emergency
patient 1 = medical OPD
2 = chronic illness follow-up
Causeofhf Etiology of heart failure 0 = CHD
1 = Hypertension
2 = Atrial Fibrillation
3 = VHD
4 = Anemia
5 = Unknown
DM Presence of Diabetes Mellitus 0 = No
1= Yes
HTN Presence of Hypertension 0 = No
1 = Yes
CKD Presence of Chronic Kidney 0 = No
Disease 1 = Yes
Stroke Presence of Stroke 0 = No
1 = Yes
HIV Presence of HIV 0 = No
1 = Yes
Pneumonia Presence of Pneumonia 0 = No
1 = Yes

17
 CAD Presence of Coronary Artery 0 = No
Disease 1 = Yes
TB Presence of Tuberculosis 0 = No
1 = Yes

3.2 Survival Analysis

Survival analysis is a collection of statistical procedures for data analysis for which the outcome
variable of interest is time until an event occurs. The event can be death, occurrence of a disease,
marriage, divorce, etc. The time to event or survival time can be measured in days, weeks, years,
etc. Hence, survival data often consists of a response variable that measures the duration of time
until a specified event occurs and a set of independent variables thought to be associated with the
event-time variable. The problem of analyzing survival data arises in several applied fields such
as medicine, biology, public health, epidemiology, engineering, economics, sociology and etc.

The key feature that distinguishes survival data from other types is that the event will not
necessarily have occurred in all individuals by the time the study ends, and for these patients,
their full survival times are unknown. These are called censored observations or censored times
and survival analysis methods correctly use both the censored and uncensored observations.
There are three types of censoring right censoring, left censoring and interval censoring.

Right censoring: is defined as a survival time recorded from its beginning to a defined time
before its end time. It is commonly recognized survival analysis and also considered in this
study.

3.2.1 Survivor Function and Hazard Function

The survival and hazard functions are key concepts in survival analysis for describing the
distribution of survival times. The survivor function S(t ); is the probability that the survival time
of a randomly selected subject is greater than some specified time t or the probability of an
individual being event-free beyond time t. In order to find the survival function, suppose T be
random variable associated with the survival times, t be the observed value of the random
variableT and f (t) be the underlying probability density function of the survival timet. The
cumulative distribution function F (t) represents the probability that an individual selected at

18
random will have a survival time less than or equal to the specified value t. Thus, the cumulative
distribution function and the survivor function are given by:

t
F ( t )=P ( T ≤ t ) =∫ f (u) du, t ≥ 0
0

S ( t ) =P ( T >t )=1−F (t ) , t ≥ 0 (3.1)

The relationship between S(t ) and f (t) is given as

d d −d
f ( t )= F ( t )= ( 1−S ( t ) )= S (t ), t ≥ 0
dt dt dt

Hazard function h(t)

The hazard function is generally denoted by h(t) and can be used to express the risk or hazard of
death at time t. It is obtained from the probability that an individual dies in an infinitesimally
small interval (t , ∆t ) given that the individual has survived up to time t i.e. P { t ≤ T < ∆t∨T ≥t }.

The hazard function is also known as the instantaneous death rate, force of mortality, conditional
mortality rate, and hazard rate as it measures the conditional probability of the occurrence of

P { t ≤T <∆ t∨T ≥ t }
death per unit time. Hence the hazard function is given ash ( t )= lim . The
∆ t →0 ∆t

hazard function varies from zero indicating no risk to infinity referring the certainty of failure at
that instant. In contrast to the survivor function, which focuses on surviving or not failing, the
hazard function focuses on failing, that is the event occurring. Thus, in a way, survival function
and hazard function are complementary to each other.

In practice, the hazard function is estimated as the proportion of subjects dying in an interval per
unit time given that they have survived to the beginning of the interval.

number of subjects dying ∈theinterval beginning at time t

h ( t )=
( number of subjects surviving at timet ) (interval width )

The hazard function can take different shapes. The hazard can remain constant throughout the
follow-up time. Such a function indicates Exponential survival model. Hazards function can also
be either increasing or decreasing monotonically. It is also possible that hazards increase initially

19
and then decrease which gives rise to a bell shaped hazard curve. Initially high hazards followed
by decreased hazards and finally increasing hazards can also be observed which give rise to a
bath tub shaped hazards curve. Another function called cumulative hazard function which
measures the total amount of risk that has been accumulated up to time t can also be used to
describe the survival experience.

There is a clearly defined relationship between S ( t ) and h ( t ) which is given by the formula

f (t) f (t) −d
h ( t )= = = lnS (t)
1−F( t) S (t) dt
(3.2)

[ ]
S ( t ) =exp −∫ h ( u ) du =exp (−H ( t ) ) , t ≥ 0
0
(3.3)

t
Where H ( t ) =∫ h (u ) du the cumulative hazard function, which can be obtained from
0

H ( t ) =−log ( S ( t ) ) (3.4)

The probability density function of T can be written as

f ( t )=h(t) S (t) (3.5)

3.2.2 Non-Parametric Methods

Non-parametric methods require no assumptions about the distribution of survival time and are
used to estimate survival function and to compare the survival experience of two or more groups.
These are Kaplan-Meier (KM) estimator to estimate the survival function and log-rank test to
compare survival experience of each group of categorical variables included in the study.

(i) Kaplan-Meier Estimator of the Survival Function

The Kaplan-Meier (KM) estimator proposed by Kaplan and Meier (1958) is the standard non
parametric estimator of the survival function S ( t ) . which is also called the Product-Limit
estimator incorporates information from all observations available, both censored and
uncensored, by considering any point in time as a series of steps defined by the observed survival
and censored times.

20
Supposet 1,t 2,… , t n be the survival times of n independent observations and t 1 ≤ t 2 ≤ …t m , m≤ n be
the m distinct ordered death times. Then the Kaplan-Meier estimator of the survivorship function
(or survival probability) at time t, S ( t ) =P(T > t) is defined as:

n j−d j d
^s ( t ) =∏
t j ≤t nj
=∏ 1− j
t ≤tj
nj [ ] (3.6)

Where, n j is the number of individuals who are at risk of dying at time t j , and d j is the number
of individuals who failed (died) at timet j . The variance of the KM survival estimator which
is also known as the Greenwood’s formula is given by

2 j d
var ( s^ ( t ) )=( s^ ( t ) ) ∑ n (n −d (3.7)
j )
j j

(ii) Comparison of Survivorship Functions

The simplest way of comparing the survival times obtained from two or more groups is to plot
the Kaplan-Meier curves for these groups on the same graph. However, this graph does not allow
us to say whether or not there is a real difference between the groups. Assessing whether or not
there is a real difference between groups can only be done by utilizing statistical tests. Thus, the
Mantel- Hanzel (1959), currently called the “log-rank” test is used for comparison of two or
more survival distributions in this thesis work.

Let t 1 ≤ t 2 ≤ …t m be the m distinct ordered death times across two groups. Suppose that d j failures
occur at t ( j ) and that n j subjects are at risk just prior to t ( j ) (j = 1, 2,..., m). Let d ij and nij be the
corresponding numbers in group i (i = 1, 2). Then the log-rank test compares the observed
number of deaths with the expected number of deaths for group i. Consider the null hypothesis:
S ( 1 )=S(2); i.e. there is no difference between survival curves in two groups. Given n j and d j the
random variable d 1 j has the hypergeometric distribution

( dd )( nn −d
j

1j 1j −d )
j j

1j

( nn )
j

1j

21
Under the null hypothesis, the probability of experiencing an event at t ( j )does not depend on the

dj
group, i.e. the probability of experiencing an event at t j is . So that the expected number of
nj
deaths in group one is

n1 j d j
E ( d 1 j ) =e 1 j=
nj

The test statistic is given by the difference between the total observed and expected number of
deaths in group one

m
U L =∑ ( d 1 j−e1 j ) (3.8)
j=1

Since d 1 j has the hypergeometric distribution, the variance of d 1 j is given by

n 1 j n2 j d j (n −d )
v1 j=Var ( d 1 j ) = j j
(3.9)
n j2 ( n j−1 )

So that the variance of U Lis given by

m
Var ( U L )=∑ v 1 j =V L
j=1

Under the null hypothesis, statistic (3.8) has an approximate normal distribution with zero mean

U L2 2
and varianceV L. This then follows x1
VL

The general form of the test statistic to test the equality of survival curves which can also be used
by several alternatives to the log-rank test, such as the Wilcoxon test, may be defined as follows:

∑ w j ( d ij− e^ 1 j)
Q= j=1 m
∑ w j2 v^ 1 j
j=1

Where: w j are weights whose values depend on the specific test

22
Log rank test is based on weights equal to one, i.e. w j=1. And it is appropriate when hazard
functions for two groups are proportional over time, i.e. h1 (t )=ψ h2 (t).

3.2.3 Regression Models for Survival Data

Both the non-parametric methods defined earlier are examples of univariate analysis; they
describe the survival with respect to the factor under investigation, but necessarily ignore the
impact of any others. In clinical investigations it is more common to have a situation where
covariates potentially affect patient prognosis. When investigating survival in relation to any one
factor, it is often desirable to adjust for the impact of others. Moreover, while the log-rank test
provides a P-value for the differences between the groups, it offers no estimate of the actual
effect size.

The use of a statistical model improves on these methods by allowing survival to be assessed
with respect to several factors simultaneously, and in addition, offers estimates of the strength of
effect for each constituent factor. Therefore, statistical models such as Cox proportional hazards
model, parametric proportional hazards model and accelerated failure time models are important
and frequently used tools which, when constructed appropriately, offer valuable insight into the
survival process.

3.2.3.1 Cox Proportional Hazards Regression Model

Cox (1972) proposed a semi-parametric hazards model for the survival data to see the effect of
explanatory variables on the hazard function. In Cox proportional hazards model, the hazard of
an individual consists of the product of two terms namely the baseline hazard function h0 ( t) i.e.
the hazard function when all covariates equal to zero and the exponential function of the
subject’s variables weighted by the regression coefficients. The hazard for the i th individual with
'
the set of covariates x i=( x 1 , x 2 , … , x p ) is defined as
'

hi ( t )=h0 ( t ) exp ( β 1 x 1 i + β 2 x 2 i+ …+ β p x pi )=h 0 (t)e β x i

(3.10)

Where, β ' =β 1 , β 2 , … , β p is a column vector of p regression coefficients

The above model is called as semi-parametric model for the reason that h0 ( t) is an arbitrary
function meaning no assumption is made about its actual form.

23
As can be seen in equation (3.10), we can write the hazard function as a multiplicative function
of the baseline hazard. This is the defining feature of all PH models which are sometimes
referred to as multiplicative hazard models for precisely this reason.

With time-fixed covariates x i, Cox regression model assumes that the hazards of any two
individuals are proportional over time i.e. the ratio of the hazards is the same at any time.
However, this does not preclude that the hazard may change over time. It only assumes that the
changes in the hazard of any patient over time will always be proportional to change in the
hazard of any other subject. This is due to the common baseline hazard function canceling out in
the ratio of the hazards. Hence the model is called as a proportional hazards model.

The hazard ratio (HR) or relative hazard for two individuals, one with covariate values x i and
the other with all covariate values zero is defined below.

hi ( t ) '

=e β x i
(3.11)
h0 ( t )

3.2.3.1.1 Estimation of Cox Regression Model

The β coefficients in the proportional hazards model can be estimated by maximum likelihood
estimation method. Which attempts to maximize the likelihood function for the observed data
simultaneously with respect to h0 ( t) and βs. A more popular approach is proposed by Cox (1972)
in which a partial likelihood function that does not depend on h0 ( t) is obtained for β in the
presence of nuisance parameterh0 ( t).

Let t (1) <t (2) <…< t (r) are ordered death times, so that t ( j ) is the j th ordered death time. Then R(t ( j) )

denotes the set of individuals who are at risk at time t ( j ). So that R(t ( j) ) is the group of individuals
who are alive and uncensored at a time just prior to t ( j )and it is called the risk set.

Probability of one death at time t ( j ) is the sum of the probabilities of death at time t ( j ) over all
individuals who are at risk of death at that time. If these individuals are indexed by Ɩ then the
partial likelihood function is given by

r
exp ( β ' x j )
LP ( β ) = ∏ (3.12)
j=1 ∑ exp ( β' x Ɩ )
Ɩ ∈R (t j )

24
The log partial likelihood function is

j =1 [
¿ P ( β ) =∑ β ' x j −ln
( ∑
Ɩ ∈R (t j )
exp ( β' xƖ )
)] (3.13)

The estimates of β parameters can be found by using Newton-Raphson procedure.

The partial likelihood derived above is valid when there are no ties in the data set. The simplest
approximation to the likelihood function proposed by Kalbfleisch and Prentice to be used in the
presence of tied observations is given by Breslow and Peto. This method is usually the default
procedure for handling ties in statistical software for survival analysis.

Let S jbe the vector of sums of each of the p covariates for those individuals who die at the j th

dj
death time, t j, j=1,2 ,… , r. If there are d j deaths at t j , the hth elements of S jis Shj =∑ x hjk , where
k=1

x hjk is the value of the hth explanatory variable, h=1,2 ,… , p for the k th of d j individuals,
k =1,2 , … ,d j, who die at the j th death time. Then the Breslow’s approximate likelihood is given
by

r
exp ( β' S j )
∏ dj (3.14)
exp ( β ' x Ɩ )
j=1
{∑
Ɩ ∈ R(t j ) }
3.2.3.1.2 Model Adequacy for Cox PH Model

Model-based inferences depend completely on the fitted statistical model. For these inferences to
be valid in any sense of the word, the fitted model must provide an adequate summary of the data
upon which it is based. The methods for assessment of a fitted proportional hazards model are
essentially the same as for other regression models (Hosmer and Lemeshow 1999).

For examining different aspects of the model different residuals have been proposed by different
authors for model diagnostics of Cox regression model. Goodness of fit of the final Cox PH
model can be tested using Cox-Snell residuals plot and measure of explained variation R2 type
can also be used.

Cox-Snell residuals

25
The Cox-Snell residual is given by Cox and Snell. The Cox-Snell residuals will not be
symmetrically distributed about zero and cannot be negative. The Cox-Snell residual for the i th
individual with observed survival time t i is defined as

r ci =exp ( β^ ' x i ) ^
H 0 ( ti )

^ 0 ( t i ) is an estimate of the baseline cumulative hazard function at time t i

Where H

If the final proportional hazards model is correct and the estimated regression coefficients are
close to the true values, the Cox-Snell residuals r ci can be regarded as a sample from a unit slope
exponential distribution, and therefore, the plot of H (r ci ) against r ci should be a 45°-line through
the origin (Hosmer and Lemeshow 1999).

R2 Type Statistic

A measure analogous to R2 in linear regression, as a measure of model performance of a Cox

model is also used. Hosmer and Lemeshow (1999) recommended a statistic based on the log-
likelihood of the model, which is defined as follows;

2
{ [
R p2=1− exp ( L −L p )
n 0 ]} (3.15)

where L0 the log-likelihood for null model i.e. the model with no covariates, L p the log-
likelihood for the fitted final Cox model with p covariates and n number of observations included
in the study. Higher the value of R2 naturally indicates the better fit of the model. However,
lesser the value of R2 may not mean that the model is poor fit.

Martingale Residuals: It is a slight modification of the Cox-Snell residuals and is defined as

^
M i=δ i−r i

where, δ i is the censoring indicator and r i is Cox-Snell residual for i th individual given as
^ i ( t )=−ln ^Si (t ), where ^
r i= H H i ( t ) and ^Si (t) are the estimated values of the cumulative hazard and
survivor functions of the i th individual at time t. It is interpreted as a difference between
observed and expected number of deaths until the moment t. As residuals of this type do not
have symmetric distribution, they can be transformed into deviance residuals that are supposed to

26
have a symmetric distribution with the mean equal to zero, assuming proper specification of the
model (Collett 2003).

Martingale residuals are useful while examining assumption of linear effect of covariates on
logarithm of hazard. They can be plotted against covariates to detect nonlinearity (Collett 2003).
Nonlinearity is not an issue for categorical variables, so we only examine plots of martingale
residuals against continuous covariate. LOESS smoothed curve can be superimposed on the
scatter plots to give interpretation. If the functional form observed in using the plots has some
pattern, which is nonlinear, the covariate can be transformed and the martingale residuals again
should be plotted against the transformed covariate. A horizontal straight line which is drawn as
a reference through zero would then confirm that the appropriate transformation has been used to
the covariate (Collett 2003).

Deviance Residuals: the deviance residuals help in identifying poorly fitted subjects, which is
defined in (Hosmer and Lemeshow 1999) as

D i=sign( ^ √
M i) −2 ( ^ ^)
M i +d i ) + log ( d i− M i

It is known that the deviance residuals are symmetrically distributed about zero when the fitted
model is adequate, and individuals with large positive or negative deviance residuals are poorly
predicted by the model. The function sign (.) is the sign function which takes the value 1 if ^
M iis

positive and -1 if ^
M iis negative.

The deviance residuals are a normalized transform of the martingale residuals. They also have a
mean of zero but are approximately symmetrically distributed about zero when the fitted model
is appropriate. Deviance residual can also be used like residuals from linear regression. The plot
of the deviance residuals against the covariates can be obtained. Any unusual patterns may
suggest features of the data that have not been adequately fitted for the model. Very large or very
small values suggest that the observation may be an outlier in need of special attention.

In a fitted Cox PH model, the hazard of experiencing the event for the i th individual at any time

depends on the value of exp ( β ' x i )which is called the risk score. A plot of the deviance residuals
versus the risk score is a helpful diagnostic to assess a given individual on the model. Potential

27
outliers will have deviance residuals whose absolute values are very large. This plot will give the
information about the characteristic of observations that are not well fitted by the model.

Schoenfeld Residuals

These are covariate-wise residuals and overcomes the problem that the above residuals depend
heavily on observed survival time and cumulative hazard function. They are computed for each
individual and covariate. Thus, the Schoenfeld residual for thei th individual and k thcovariate is
defined as:

∑ x ki exp ⁡( ^β ' x i )

[
^Ski =δ k x ki − k∈ R
∑
k∈ Rt
t (i)

(i)
exp ⁡( ^β ' x i ) ]
Where, x i is a vector of p fixed covariates for the i th individual, x ki is the value of k th covariate
for the i th individual and δ k is the censoring indicator of the k th covariate for the i th individual.
The Schoenfeld residuals sum to zero and they are uncorrelated with expected value zero
(Schoenfeld 1982).

Diagnostics for Influential Observations

Observations that have an undue effect on model-based inference are said to be influential. In the
assessment of model adequacy, it is important to determine whether there are any influential

observations. The most direct measure of influence is ^β j − ^β j(i) where ^β j is the j th parameter,

j=1 ,2 , ..., p in a fitted Cox PH model and ^β j (i ) is obtained by fitting the model after omitting
observation i. In this way, we have to fit the n+1 Cox models, one with the complete data and n
with each observation eliminated. This procedure involves a significant amount of computation
if the sample size is large. We would like to use an alternative approximate value that does not
involve an iterative refitting of the model. To check the influence of observations on a parameter

estimate an approximation to ^β j − ^β j(i) is the j th component of the vector r ' si Var ( β^ ) where r ' si is
the p ×1 vector of score residuals for the i th observation (Klein and Moeschberger 1997), which

are modifications of Schoenfeld residuals and are defined for all the observations, and Var ( ^β ) is
the variance-covariance matrix of the vector of parameter estimates in the fitted Cox PH model.

28
The j th element of this vector is called delta-beta statistic for the j th explanatory variable, i.e.

∆ i β^ j = ^β j− β^ j(−1)which tells us how much each coefficient will change by removal of a single
observation. Therefore, we can check whether there are influential observations for any
particular explanatory variable.

3.2.3.1.3 Checking Cox Proportional Hazards Assumption

The basic assumption of the Cox model is the assumption of proportional hazards. There are
several methods for verifying that a model satisfies proportionality assumption.

29
Graphical Methods

One can obtain Cox PH survival function by the relationship between hazard function and
survival function
p

S ( t , X ) =( S 0 ( t ) )
exp
(∑ )
i=1
β i xi

Where, X =(x 1 , x 2 , .. . , x p ) are the values of the vector of explanatory variables for a particular
individual. When taking the logarithm twice, we can easily get

p
log (−log ( S ( t , X ) ) ) =∑ β i x i+ log (−log ( ( S0 ( t )) ))
i=1

Then the difference in log-log curves corresponding to two different individuals with variables
x 1=( x 11 , x12 ,... , x 1 p ) and x 2=( x 21 , x 22 , ... , x 2 p ) is given by

p
log (−log ( S ( t , X 1 ) ) )−log (−log ( S ( t , X 2 ) ) )=∑ βi ( x 1 i−x 2i )
i=1

Which does not depend on t. This relationship is very helpful to identify situations where we
may have proportional hazards. By plotting estimated log (-log (survival)) versus survival time
for two groups we would see parallel curves if the hazards are proportional.

This method does not work well for continuous predictors or categorical predictors that have
many levels because the graph becomes "cluttered". Furthermore, the curves are sparse when
there are few time points and it may be difficult to tell how close to parallel is close enough.

Looking at the K-M curves and log(-log(survival)) is not enough to be certain of proportionality
since they are univariate analysis and do not show whether hazards will still be proportional
when a model includes many other predictors. But they support our argument for proportionality.
There are some other statistical methods for checking the proportionality.

Adding Time-Dependent Covariates in the Cox Model

We create time-dependent covariates by creating interactions of the predictors and a function of

survival time and including them in the model. For example, if the predictor of interest is x j , then
we create a time-dependent covariate x j (t ), x j ( t ) =x j∗gi ( t ); where gi ( t ) is some specified

30
function of time, usually gi ( t ) =ln ( t ) is used by most soft wares by default . Then the model
assessing PH assumption for x j adjusted for other covariates is:

h ( t , x ( t ) ) =h0 ( t ) exp ( β 1 x 1 + β 2 x 2 +…+ β j x j +…+ β p x p+ δ x j∗g i ( t ) )

Where, x ( t )=( x 1 , x 2 … , x j … , x p ); x j is the values of the vector of explanatory variables for a

particular individual. The null hypothesis to check proportionality is that δ= 0. The test statistic
can be carried out using either a Wald test or a likelihood ratio test. In the Wald test, the test

2
δ^
statistic is W = ( ) se ( δ^ )

The likelihood ratio test calculates the likelihood under null hypothesis, L0 and the likelihood
under the alternative hypothesis, La. The LR statistic is then

L0
LR=−2 log ( )
La
=−2 ( l 0−l a )

Where l 0 ,l a are log likelihood under two hypothesis respectively. Both statistics have a chi-
square with one degree of freedom under the null hypothesis. If the time-dependent covariate is
significant, i.e. the null hypothesis is rejected, and then the predictor is not proportional. In the
same way, we can also assess the PH assumption for several predictors simultaneously.

Tests Based on the Schoenfeld Residuals

This method is based on analysis of residuals proposed by Schoenfeld (1982). Schoenfeld

residuals are defined for each covariate separately. Each subject has a residual for each covariate
in the fitted final Cox regression model. The i th Schoenfeld residual for j th explanatory variable

is given by

rP ji =δ i ( x ji− a^ ji )

Where, x ji is the value of j th explanatory variable for the i th individual and a^ ji is given by

31
 ∑ x ji exp ( β ' x Ɩ )
Ɩ ∈R (t i )
a^ ji =
∑ exp ( β ' x Ɩ )
Ɩ ∈R (t i )

Where, R(t i) is the set of all individuals at risk at timet i.

Later, Grambsch and Therneau (1994) proposed the scaled version of Schoenfeld residuals
which is formulated as follows:

r ¿ Pi=r var ( ^β ) r Pi

Where, rstands for number of deaths among n individuals, var ( ^β ) the variance covariance matrix
'
of the parameter estimates in the fitted Cox regression model and r Pi=( r P1 i ,r P2 i ,… , r P pi ) is
the vector of the Schoenfeld residuals for the i th individual.

In this method, Pearson’s correlation coefficient (ρ) between the scaled and the survival time (or
function of it) is calculated for each covariate. The null hypothesis is that the correlation between
the Schoenfeld residuals and the function of survival time is zero. Rejection of null hypothesis
concludes that PH assumption is violated. This can be evaluated by plotting the rescaled
Schoenfeld residuals against time and should show no trend or should has zero slope in order to
the PH assumption be fulfilled. Another way is to see if the correlation coefficient is significant.
Significance shows that there is a deviation from proportional hazards assumption. This test can
also be used as a global test for the final model.

3.2.3.2 Parametric Proportional Hazards Model

Apart from semi-parametric proportional hazards regression model, other regression models can
be tried on the survival time data. Regression models assuming specific probability distribution
for the survival time are called as parametric survival models. Parametric model satisfying the
proportionality hazards assumption is called as a parametric proportional hazards model. In
Parametric proportional hazards model, there is specification of baseline hazard function.

There are three classical parametric proportional hazards model namely, Exponential, Weibull
and Gompertz parametric proportional hazards models. In both Weibull and Gompertz

32
parametric proportional hazards models; hazard function increases or decreases monotonically.
However Weibull PH model is the most widely used as it is mathematically more tractable.

Weibull Proportional Hazards model

Conceptually Weibull proportional hazards model including Exponential PH model as a special

case when the shape parameter γ is unity, is similar to Cox PH model. The hazard ratio is
interpreted similarly in both models. The effect of predictors in the Weibull PH model is also
multiplicative on hazard scale. The proportionality hazards assumption is required for Weibull
PH model as in Cox PH model. Major difference between the two models is that the baseline
hazard function is assumed to follow Weibull distribution while no assumption is required in
Cox PH model.

The baseline hazard function for Weibull PH model with scale parameter λ and shape parameter
γ is:

h0 ( t )=λγ (t)γ −1

Under the Weibull PH model, the hazard function of a particular i th individual with covariates

( x 1 , x 2 ,… , x p ) is given by

hi ( t )=h0 ( t ) exp ( β 1 x 1 i + β 2 x 2 i+ …+ β p x pi )
'

¿ λγ ( t)γ −1 e β x i
(3.16)

Where β 1 , β 2 ,… , β p are unknown regression coefficients

The cumulative hazard function of the Weibull PH model for the i th individual is given by:

H i ( t )=exp ⁡( β ' x i ) λ t γ

From equation (3.4), the corresponding survival function is given by

Si (t)=exp {−exp ( β ' x i ) λ t γ } (3.17)

3.2.3.2.1 Fitting Weibull Proportional Hazards Model

33
The Weibull PH model is estimated by maximum likelihood estimation method. The likelihood
function of n observations maximized with respect to unknown parameters β 1, β 2,…, β p, and λ
and γ . The likelihood function of any PH model is given by:

n δi
f (t i )
L=∏
i=1 { }
S (t i )
Si (t i ) (3.18)

utilizing equation (3.5), the likelihood function becomes

n
δi
L=∏ ( h i ( t i ) ) S i ( t i ) (3.19)
i=1

where the indicator variable δ i takes value 1 for uncensored and value zero for censored survival
time t i.

The logarithm of the likelihood function is as follows:

n
¿
L =∑ { δ i log hi ( t i ) + log Si ( t i) } (3.20)
i=1

For Weibull PH model, by substituting the values of hi ( t i ) and Si ( t i ) from equations (3.16) and
(3.17), the log likelihood function becomes

n n

∑[ δ i { β ' x i + log ( λγ )+ γ logt i }−λexp ( β ' x i ) t γ −∑ δ i log t i

] (3.21)
i=1 i=1

n
The term ∑ δi log ti does not involve any of the unknown parameters. Hence, after omitting
i=1

this term, the likelihood function for estimating the parameters of Weibull PH model is as
follows:

∑ [ δ i { β ' x i + log ( λγ )+ γ logt i }−λexp ( β ' x i ) tγ ] (3.22)

i=1

The maximum likelihood estimates of the unknown parameters μ , λ , γ , β 1 , β 2 . . . β P can be found

by maximizing this function using the Newton-Raphson procedure. For assessing the overall fit

34
of the fitted Weibull PH model, Cox-Snell residuals plot and R2 type statistic are applied in a
similar manner as applied in Cox PH model explained in earlier sections.

3.2.3.3 Accelerated Failure Time Model

In the analysis of survival data, survival models can also be used in addition to hazards model.
One advantage of such models is that the proportionality assumption of the hazards is not
required. The parametric survival models work analogous to the multiple linear regression of
logarithm of survival time on explanatory variables. Such survival models are termed as
parametric accelerated failure time models or simply AFT models. Because these models work
on survival, the complementary concept of hazard, the sign of the regression coefficients in an
AFT model will be opposite to those in PH models (Klein and Moeschberger 1997).

Most commonly used parametric AFT models are Exponential, Weibull, Log-Logistic, Log-
normal and Generalized Gamma. Exponential and Weibull parametric models can work both in
PH metric and in AFT metric. These models are equally appropriate viewed in either metric. And
one can transform regression coefficients computed in PH metric into the regression coefficient
in AFT metric or vice versa for Exponential and Weibull parametric survival models. Other
parametric survival models such as Log-Logistic, Log-normal and Generalized Gamma work
only in AFT metric as these models do not fit into the proportional hazards frame work.

AFT models work to measure the effect of covariate to “accelerate” or to “decelerate” survival
time meaning the effect of covariate is multiplicative on time scale. Under AFT models the
survival function of the i th individual with covariates ( x 1 , x 2 ,… , x p ) at time t is the same as the

t
survival function of an individual with a baseline survival function at a time ' , mathematically
eβ x i

it can be expressed with its corresponding hazard function as:

t
Si ( t ) =S 0
[ ]
eβ x
'
i
(3.23)

− β' x i t
hi ( t )=e h0
( )
eβ x
'
i
(3.24)

35
Where, β ' = ( β 1 , β 2 ,…, β p) is a vector of regression coefficients, S0 ( . ) and h0 ( .) are the baseline
survival and hazard functions respectively. The effect size for the AFT model is measured using
the time ratio (TR), which is a ratio of the survival time of an individual with an exposure to the
survival time of an individual without the exposure for a given survival probability.

Suppose T i is a random variable representing the survival time for the i th individual. Then the
general log-linear form of the AFT model shows the mathematical relation between the log of
time and the set of covariates expressed as follows:

log T i=μ+ β ' x i+ δ ε i (3.25)

Where,β ' =(β 1 , β2 , … , β p ), μ is intercept, δ is scale parameter and ε i is a random variable used to
model the deviation of values of log T i from the linear part of the model. ε i is assumed to have a
particular probability distribution supposed to be followed by the survival time under study.

Suppose a random variableT , representing survival time, follows Log-Logistic distribution with
shape parameter k and scale parameter θ with probability distribution

eθ k t k−1
f ( t )= 2 (3.26)
( 1+e θ t k )

The baseline hazard function is given by

eθ k t k−1
h0 ( t ) = for k > 0 (3.27)
1+ eθ t k

If k ≤ 1 the hazard function decreases monotonically and if k > 1, the hazard function has single
mode (Collett 2003). The corresponding baseline survival function is given as

1
S (t)= (3.28)
1+ eθ t k

When a random variable T is said to have a Log-normal distribution with parameters μ and σ the
probability density function is given as follows:
2

f ( t )=
1
t
{(
−¿exp −
( logt −μ )
2σ
2 )} ¿
for 0 ≤ t< ∞ , σ >0 (3.29)
σ √ (2 π )

36
From which the survivor and hazard functions can be derived. The survivor function is given by

s ( t ) =1−Ф ( log (σt )−μ ) (3.30)

Where Ф( .) is the standard normal distribution function. The hazard function can be found from
the relationh ( t )=f (t)/ S (t ). This function is zero whent=0, increases to a maximum and then
decrease to zero as t tends to infinity (Collett 2003).

The generalized gamma model with scale parameter λ and two shape parameters k and p have a
probability density function defined by

p λkp pk −1
f ( t )= t (−( λt) p) fort >0 , p> 0 , k >0 , λ>0 (3.31)
Γ (k)

The two shape parameters allow for quite a flexible hazard rate. A nice characteristic of the
generalized gamma model is that it nests several of the other parametric models as special cases:
Weibull, exponential, log-normal, and the standard gamma. Thus, this model is good for
adjudicating between (some) competing parametric models. The shape parameters work in the
following way:

If k = 1, then the Weibull distribution is implied.

If k = p = 1, the exponential is implied.

If k = 0, the log-normal is implied.

Discrimination among distributions of the GG family can be assessed using the Chi-square
which can also be used to test the hypotheses that k = 1 or that k = 0, or likelihood-ratio test
defined as below.

LR=−2(lnLR−lnLV )

WherelnLR is log likelihood for the restricted (less general) model and lnLV is log likelihood for
the unrestricted model (GG in this case)

3.2.3.3.1 Fitting an AFT model

37
AFT models are fitted using the maximum likelihood estimation method. The likelihood function
is derived from the log-linear function of the AFT model defined in equation (3.25). The
likelihood function of n observed survival times, t 1 , t 2 , … , t n for the log-linear form of the AFT
model is given by

n
δi ( 1−δ i )
L ( β , μ , σ )=∏ [ f i (t i ) ] [ Si (t i) ] (3.32)
i=1

Where f i (t i ) and Si (t i) are the density and survival functions for the i th individual at time t i and δ i
is the event indicator for the observation and has value zero for censored and one for uncensored
individuals. If f ε (z i) and Sε ( z i ) are probability density function and survival function
i i

respectively of the random variable ε i in equation (3.25) in such a way that

Si ( t i ) =S ε (z i )
i
(3.33)

1
and f i ( t i )= f (z ) (3.34)
σ ti ε i i

logt i−( μ+ β 1 x1 i + β 2 x 2 i+ …+ β P x Pi )
Where, z i= ( σ )
(3.35)

The resulting likelihood function using survival function and density function of assumed
probability distribution represented by random variable ε i is as follows:

n
δi ( 1−δ i )
L ( β , μ , σ )=∏ ( σ t i )−δ [ f ε ( zi )] [ S ε (z i) ]
i

i i
(3.36)
i=1

The log-likelihood function is:

n n

∑ {−δ i log ( σ )+ δi log f ε ( z i ) +( 1−δi ) log Sε (z i )}−∑ δi log ti

i i
(3.37)
i=1 i=1

n
The term ( )
−∑ δ i log t i is omitted as it does not involve any unknown parameters. Hence the
i =1

full log-likelihood function is given by

38
 n
¿
L ( β , μ , σ ) =∑ {−δ i log ( σ ) +δ i log f ε ( zi ) + ( 1−δ i ) log S ε ( z i) }
i i
(3.38)
i=1

The maximum likelihood estimates of the parameters are estimated by using iterative Newton-
Raphson procedure.

3.2.3.3.2 Checking the Adequacy of AFT Model

Like log-cumulative hazard plot used to examine the validity of the proportional hazards model,
the percentile-percentile plot or quantile-quantile (Q-Q) plot is an exploratory assessment of the
adequacy of the AFT model for each covariate with two levels. The percentiles of the
distribution of survival times in each of the two groups of a covariate can be estimated from the
Kaplan-Meier method. If the points of Q-Q plot fall reasonably on a straight line, then the AFT
model would be appropriate (Collett 2003). We can also use percentiles instead of quartiles.

3.2.3.3.3 Test of Goodness of Fit of AFT Model

Once the model has been finalized, it is necessary to test the overall fit of it. For assessing the
goodness of fit of an AFT model different methods can be applied. Cox-Snell residuals plot and
R2 type statistics are applied in this study.

Cox-Snell residuals plot

The overall fit of the AFT model is evaluated by using the diagnostic plot of Cox-Snell residuals
as described in Cox regression model. However, the calculation of Cox-Snell residuals in AFT
model is different from that of Cox regression model because of the difference in formulation
between these two families. The Cox-Snell residuals for AFT model are calculated by using
standardized residuals which are defined as

r S i=
{ log t −( ^μ + ^β
i 1 x 1 i+ β^ 2 x2 i +…+ β^ p x pi ) }
(3.39)
σ^

Where ^μ , ^β1 , ^β 2 , … , ^β p and σ^ are the maximum likelihood estimates ofμ, β 1 , β 2 ,… , β p and σ
respectively.

Cox-Snell residuals for Log-Logistic AFT model will be

39
 r Ci =log {1+exp ( r Si ) } (3.40)

And the Cox-Snell residuals for a Log-normal AFT model are computed as follows:

r Ci =−log {1−Ф ( r S i) } (3.41)

Where Ф( .) represents the cumulative distribution function of standard normal distribution

In the Cox-Snell residuals plot, if the plotted points lie on a line that has an intercept zero and
slope unity, then it indicates that the fit is good.

R2 Type Statistic

The formula for calculation of R2 type statistic based on log-likelihood values is the same in AFT
model as used in the case of Cox PH model. However it should be noted that it is a full log-
likelihood in AFT models as compared to the partial log-likelihood in Cox proportional hazards
model.

3.3 Comparison of Cox PH and Parametric Models

Different models can be compared on the basis of the variables selected and their coefficients in
each model, goodness of fit tests such as R2 type statistic and Cox-Snell residuals plot.

Comparison of Model Estimates

If the models being compared have a similar set of covariates that have entered in the respective
final models, it can be interpreted as all models are equally good or bad as far as the
identification of important covariates associated with the outcome. However, it is difficult to
interpret either way if the selected variables in the models being compared are different, as there
is no way of knowing the truth (Khanal 2009).

The precision of the regression coefficients is another criterion that can be used to compare
different models. The smaller the standard error, the more precise an estimate is expected to be.
A model with more precise coefficients can be considered as a more precise model.

But as discussed earlier, Cox regression model, Exponential and Weibull PH models work on PH
metric whereas Log-Logistic and Log-normal AFT models work only on AFT metric. Because of
this, regression coefficients of a Cox model, Exponential or Weibull PH model cannot be directly

40
compared with those of Log-Logistic and Log-normal AFT models. However regression
coefficients of Cox PH, Exponential PH and Weibull PH can be compared directly, as all works
on PH metric and similarly all AFT models are directly comparable for the precision of the
estimates, as all AFTs work on AFT metric (Khanal 2009).

Comparison of R2Type Statistic

R2 Type statistic is a relative measure for the improvement in log-likelihood value from the
model without any covariate to the final model. Hence two or more relative measures can be
compared irrespective of the type of log-likelihood estimated. The model with more R2 type is
indicative of a better model than a model with a low R2.

Comparison Based on Cox-Snell Residuals Plots

The construction of the Cox-Snell residuals plot is explained in the respective sections above.
Broadly, all models require the plot to be a straight line, passing through the origin to qualify for
a good fit. So the plots under each model can be visually assessed as to which one of them is
close to the requirements of a good fit.

Comparison based on Akaike’s Information Criterion (AIC)

Akaike’s Information Criterion (AIC) proposed by (Akaike, 1974) may also be used when
comparing the viability of different parametric models. The AIC of a model may be defined as

AIC=−2≪+ 2(c +k +1)

Where LL is the log-likelihood, c is the number of covariates and k the number of model-
specific ancillary parameters. A lower value of the AIC suggests a better model. Note, however,
that the likelihood computed in a Cox model is a partial likelihood, and so it is not possible to
compare Cox PH models to fully parametric ones in this manner (Bradburn et al. 2003).

41
 CHAPTER FOUR

RESULTS AND DISCUSSION

This section reports results from different statistical methods used to analyze correlates of the
survival time of patients with severe heart failure at ICU ward treatment period. The study was
based on 147 patients, which is obtained from Gondar university hospital ICU ward. Among the
patients considered, 31.3% (46) of them were dead while the rest 68.7% (101) were censored.
The median time of a patient stay in the ICU ward to treatment until death or censoring is 6 days.

4.1 Descriptive Summaries and Non-Parametric Analysis

The predominant causes of HF were coronary heart disease and valvular heart disease which
caused 38.1% and 29.9% of the total population respectively. Other etiologies found in this study
were hypertension 8.8%, anemia 6.1%, atrial fibrillation 4.0% and unknown causes 12.9%. For
the reason that HF is not a disease by itself patients with HF have other causes of death. In this
study deaths were attributed to respiratory failure (26.09%), cardiac arrest (15.22%), multi-organ
failure (15.22%), end stage renal failure (10.87%), cardiogenic shock (6.52%), septic shock
(4.35%), pneumonia (4.35%) and non-defined causes (17.39%).

The mean age of patients with severe heart failure were 48.04years for men (SD = 22.27) and
50.83 years for women (SD = 19.40). The mean baseline body weight were 60.18kg (SD = 9.69)
and 59.158kg (SD = 9.04) for male and female patients, respectively. The mean systolic blood
pressure were 116.45 for men (SD = 25.12) and 112.72 for women (SD = 25.47). The mean heart
rate were 95.74bpm (SD = 20.74) and 98.24bpm (SD =24.11) for male and female patients
respectively. The mean white blood cell count were 9.05 for men (SD = 2.12) and 11.79 for
women (SD = 15.46502). The mean duration of heart failure were 3.90years for men (SD = 4.52)
and 5.01years for women (SD = 5.69). Descriptive measures such as, median treatment time,
number of censored and death for each categorical predictor variables included in the study are
presented below in table 4.1.

42
Table 4.1: Results of descriptive measures of categorical predictor variables

Covariates Category Median Status

treatment Censored (%) Death (%)
time (n)
Age of patients (0, 49] 6 (72) 52(72.2) 20(27.8)
(49, 65] 5.5(36) 28(77.8) 8(22.2)
(65, ∞) 6 (39) 21(53.8) 18(46.2)
Sex of patients Female 6 (92) 67(72.8) 25(27.2)
Male 6 (55) 34(61.8) 21(38.2)
Residence of patients Urban 6(90) 57(63.3) 33(77.2)
Rural 6(57) 44(36.7) 13(22.8)
Marital status of patients Single 6(28) 22(78.6) 6(21.4)
Married 6(96) 63(65.6) 33(34.4)
Divorced/Widowed 7(23) 16(69.6) 7(30.4)

History of heart failure New 6(38) 23(60.5) 15(39.5)

HF patient before 6(109) 78(71.6) 31(28.4)

Department seen first Medical OPD 7(86) 64(74.4) 22(25.6)

Emergency 5(41) 17(41.5) 24(58.5)
Chronic illness follow up 6(20) 20(100) 0

Etiology of heart failure CHD 7(56) 40(71.4) 16(28.6)

VHD 6(44) 30(68.2) 14(31.8)
Hypertension 5(13) 10(76.9) 3(23.1)
Anemia 6(9) 4(44.4) 5(55.6)
Atrial Fibrillation 8.5(6) 3(50.0) 3(50.0)
Unknown 6(19) 14(73.7) 5(26.3)
Diabetes Mellitus Yes 6(14) 7(50.0) 7(50.0)
No 6(133) 94(70.7) 39(29.3)
Hypertension Yes 6(31) 22(70.9) 9(29.1)
No 6(116) 79(68.1) 37(31.9)
Chronic Kidney Disease Yes 8(17) 11(64.7) 6(35.3)
No 6(130) 90(69.2) 40(30.8)

43
 Stroke Yes 6(4) 2(50.0) 2(50.0)
No 6(143) 99(69.2) 44(30.8)
HIV Yes 8(9) 7(77.8) 2(22.2)
No 6(138) 94(68.1) 44(31.9)
Pneumonia Yes 6(49) 31(63.3) 18(36.7)
No 6(98) 70(71.4) 28(28.6)
Coronary Artery Disease Yes 4(11) 7(63.6) 4(36.4)
No 6(136) 94(69.1) 42(30.9)
Tuberculosis Yes 5(5) 3(60.0) 2(40.0)
No 6(142) 98(69.0) 44(31.0)

To get insight into the shape of the survival function for each group, we look at the Kaplan-
Meier curves for all the categorical predictors. In addition the test for equality of survivor
functions between groups is performed using log-rank test. The Kaplan-Meier survivor curves
for categorical variables are plotted in annex1. The Kaplan-Meier survivor curves for the three
department groups showed that patients first seen at chronic illness follow-up had higher survival
probability compared to patients first seen at other departments that is medical OPD or
emergency, whereas patients first seen at emergency department had lower survival probability
than patients first seen at other departments. Log-rank test showed significant difference in
survival experience of patients relating to department the patient seen at first (p=0.000). The
Kaplan-Meier survivor curves for the three age groups showed that patients aged above sixty-
five years had lower survival probability compared to patients aged below forty-nine and
between fifty and sixty-five, but log-rank test did not confirm the difference.

Based on the Kaplan-Meier survival curves female patients, rural residents, patients without
coronary artery disease, non-DM and newly diagnosed as patient with HF at current admission
had slightly higher survival probability than their counter parts.

Kaplan-Meier survival estimate and log-rank test for equality of survival functions revealed the
same results which has no difference in survival experience between the groups of the variable
for marital status, cause of heart failure, hypertension, chronic kidney disease, stroke, pneumonia
and tuberculosis.

44
4.2 Cox Proportional Hazards Model

The univariate Cox proportional hazards regression models are fitted for every covariate as
shown in table 4.2 to check covariates affecting survival of patients with severe heart failure,
admitted to ICU ward, before proceeding to higher models. Problem of single covariate approach
is that it ignores the possibility that a collection of variables, each of which is weakly
associated with the outcome, can become an important predictor of the outcome when taken
together. For this reason, we considered including the predictor in the multivariate model if the
test for the univariate analysis has a p-value less than or equal to 0.25.

Both the log-rank and the Cox model identified similar significant predictors at 25% level.
Consequently, the candidate variables for building a multivariate Cox model are department the
patient seen at first, place of residence, sex of the patient, history of HF, age category of the
patient, presence of DM as comorbidity, duration of heart failure and presence of coronary artery
disease as comorbidity. The hazards ratio, 95% confidence intervals and standard errors for each
variable is given below in table 4.2.

Table 4.2: Results of univariate Cox analysis

Covariates HR 75% CI P_value SE

Sex 0.62 (0.35 1.11) 0.109 0.18
Residence 1.75 (0.92 3.34) 0.086 0.58
Coronary artery disease 2.03 (0.72 5.72) 0.182 1.07
Diabetes mellitus 1.70 (0.76 3.80) 0.198 0.70
History of HF 0.68 (0.37 1.26) 0.219 0.21
White blood cell count 0.98 (0.92 1.03) 0.411 0.03
Department seen first
Medical OPD 1
Emergency 3.00 (1.67 5.36) 0.000 0.89
Duration of HF 1.04 (0.99 1.09) 0.144 0.03
Age category 0.097
Less than 49 years 1
50-65 years 0.93 (0.41 2.11) 0.863 0.39
Above 65 years 1.76 (0.93 3.33) 0.081 0.57
Marital status 1.18 (0.39 3.51) 0.792 0.26
Systolic BP 1.003 (0.99 1.01) 0.608 0.01
Weight 0.98 (0.95 1.02) 0.361 0.02
Heart rate 0.996 (0.98 1.009) 0.554 0.01

45
 Cause of HF 1.29 (0.37 4.44) 0.506 0.09
Hypertension 0.96 (0.46 1.98) 0.906 0.36
Chronic kidney disease 1.09 (0.46 2.58) 0.845 0.48
Stroke 2.04 (0 .49 8.51) 0.327 1.49
Tuberculosis 2.00 (0.48 8.31) 0.342 1.45
HIV 0.45 (0.11 1.88) 0.256 0.33
Pneumonia 1.22 (0.68 2.22) 0.502 0.37

The significant variables at 25% level were considered in a forward stepwise manner with an
entry probability 0.05 and removal probability 0.251. Among the candidate variables considered
for building multivariate Cox, stepwise procedure picked up three variables, history of HF,
department the patient seen at first and duration of heart failure. The HR, 95% confidence
intervals and standard error for each variable in the fitted model are shown in table 4.3.

Table 4.3: Multivariate Cox regression model for HF data

Covariates HR 95% CI P value SE

History of HF 0.32 (0.15 0.68) 0.003 0.12

Department seen first

Emergency 3.32 (1.85 5.98) 0.000 1.00

Duration of HF 1.08 (1.03 1.14) 0.003 0.03
Adequacy of the fitted model that is the assumption of proportional hazards and the goodness of
fit were assessed. We used the – (log (-log (survival))) plot versus survival time, which is called
as a log-cumulative hazard plot, to check the PH assumption for all the categorical variables
included in the fitted model (annex 2). The graphs for each of the categorical variable display
lines that appeared to be parallel implying that the proportional-hazards assumption among
groups of the categorical variable such as department the patient seen at first and history of heart
failure has not been violated.

Table 4.4: Test of proportional-hazards assumption based on Schoenfeld residuals

Variable Rho (ρ) chi2 DF P-value

46
 History of HF 0.11820 0.68 1 0.4092
Department seen first (emergency) -0.14262 0.93 1 0.3357
Duration 0.03327 0.06 1 0.8139
Global test 2.24 3 0.5240

The proportionality assumption was also tested based on the Schoenfeld residuals. Plots of scaled
Schoenfeld residuals versus analysis time with overlaid linear fit to check if the slope is zero are
given in figure (2). More or less, zero slopes are revealed by the plots. As graphical methods are
subjective the correlation between Schoenfeld residuals and survival time test is used. The
correlation (ρ) between Schoenfeld residuals and survival time for each covariate and its
summary is shown in table 4.4. The results indicate that all variables satisfied the proportional
hazards assumption as the correlation between Schoenfeld residuals and survival time is not
significant. The assumption of proportionality hazard was also checked for each covariate in the
Cox regression model by adding an interaction term with time. The results after adding the
interaction term with time are presented in table 4.5. The coefficient for interaction effect of each
covariate with time is found not significant indicating the hazard ratio is independent of time.

Table 4.5: Results for Cox after adding interaction term with time

Variable Coef P value

History of HF -1.643161 0.070

History of HF * (time) 0.1283902 0.416

Duration of HF 0.0932174 0.125

Duration of HF* (time) -0.0033428 0.769

Department seen first 0.4757832 0.283

Department seen first * (time) -0.0863136 0.261

Plot of the Cox-Snell residuals was applied to test the overall fit of the model. In this method
Cox-Snell residuals were plotted against the cumulative hazard of Cox-Snell residuals as shown
in figure (1). The figure reveals that the overall fit of the Cox model is good. However, there is

47
little evidence of a systematic deviation from the straight line at the right, this can be expected
even if we have a well-fitting Cox model because of the reduced effective sample size caused by
prior failures and censoring (Khanal 2009). The R2 type statistic was also calculated as follows:

2
R p2=1− exp{ [ 147 ]}
(−206.4725−(−190.75384 ) ) =0.192538029

Which conveys 19.25% of variation in the partial log-likelihood is explained by Cox PH model.

Figure 1: Cumulative hazard plot of the Cox-Snell residual for multivariate Cox PH model
3
2
1
0

0 1 2 3
Cox-Snell residual

H Cox-Snell residual

In the plot of deviance residuals against the risk score (figure 3) the deviance residuals seem to
be approximately symmetrically distributed about zero suggesting the fitted model is adequate
and there exists no clear poorly fitted observation. We also used delta-beta statistic to measure
the influential observations on the model as a whole (figure 4). Accordingly, the coefficient did
not change too much when the observations corresponding to the largest delta-beta statistics are
removed. That is, comparing the magnitudes of the largest dfbeta values to the regression
coefficients suggests that none of the observations is terribly influential individually. Therefore,
we do not remove them from the dataset and conclude that there are no influential observations.

In order to assess the linearity assumption of the covariates, we plotted martingale residuals
compared by excluding the covariate to be checked for linearity against the values of the
covariate. The smoothed curve using Lowess is used to check the linearity assumption. In our
model "duration of HF" is the only continuous variable among the three significant variables.

48
Accordingly, Martingale residuals for covariate duration of HF are plotted against the duration of
HF values along with a Lowess smooth curve in annex 3. There was no definite pattern in the
scatter plots and the smoothed curve and it was almost a horizontal line. This is an indicator of
approximately linearity in the covariate duration of HF. Thus, there are no signs of nonlinearity
in duration of HF indicating that the assumption of linearity is fulfilled by this variable.

Figure 2: Plots of scaled schoenfeld residuals versus analysis time with overlaid linear fit

.8
4

.6
2

.4
0

.2
-2

0
-4

-.2
-6

0 5 10 15 20 0 5 10 15 20
_t _t

scaled Schoenfeld - hfpatient Fitted values scaled Schoenfeld - duration Fitted values
4
3
2
1
0
-1

0 5 10 15 20
_t

scaled Schoenfeld - _Idptseenf_1 Fitted values

Variables from left to right: history of HF, duration of HF and emergency department
respectively.

Figure 3: Deviance residuals versus the risk score for Cox PH model

49
 2
deviance residual
0 -1
-2 1

0 2 4 6 8 10
Riskscore

Figure 4: Index plots of dfbeta for variables in multivariable Cox regression

.15

.01
.1

.005
DFBETA - hfpatient

DFBETA - duration
.05

0
0 -.05

-.005

0 50 100 150 0 50 100 150

.05
DFBETA - _Idptseenf_1
-.05 -.1 0

0 50 100 150

50
Variables from left to right: history of HF, duration of HF and emergency department
respectively.

4.3 Parametric Proportional Hazards Model

Univariate parametric PH model was applied in a similar manner as applied in a Cox PH model
for both Exponential and Weibull PH models. The results of univariate parametric PH models
are presented in table 4.6. In both models variables significant at 25% level in the univariate
analysis were taken as candidate variables for their multivariate analysis. Stepwise forward
selection procedure was also implemented for these models as used in multivariate Cox model.

Table 4.6: Results of univariate Weibull PH model and Exponential PH

Covariates Weibull Exponential

HR 75% CI P value SE HR 75% CI P value SE

Sex 0.57 (0.32 1.02) 0.059 0.17 0.63 (0.35 1.13) 0.121 0.19
Residence 1.76 (0.93 3.34) 0.084 0.58 1.72 (0.91 3.28) 0.096 0.56
CAD 2.34 (0.83 6.62) 0.109 1.24 1.71 (0.61 4.78) 0.303 0.90
Diabetes 1.64 (0.73 3.67) 0.228 0.67 1.69 (0.76 3.78) 0.201 0.69
mellitus
History of HF 0.75 (0.40 1.39) 0.243 0.24 0.71 (0.38 1.31) 0.270 0.22
WBCC 0.98 (0.93 1.03) 0.391 0.02 0.98 (0.92 1.03) 0.416 0.03
Department
Medical OPD 1
Emergency 3.14 (1.76 5.61) 0.000 0.93 2.84 (1.59 5.06) 0.000 0.84
Duration of HF 1.04 (0.99 1.09) 0.114 0.03 1.04 (0.99 1.09) 0.139 0.03
Age category
Less than 49Y 1
50-65 years 1.01 (0.44 2.29) 0.985 0.42 0.92 (0.41 2.09) 0.843 0.38
Above 65Y 1.75 (0.92 3.30) 0.086 0.57 1.75 (0.92 3.30) 0.086 0.56
Marital status 0.99 (0.61 1.63) 1.00 0.25 1.08 (0.67 1.76) 0.751 0.27
Systolic BP 1.00 (0.99 1.02) 0.467 0.01 1.00 (0.99 1.01) 0.676 0.01
Weight 0.98 (0.95 1.02) 0.398 0.02 0.99 (0.95 1.02) 0.443 0.02
Heart rate 0.99 (0.98 1.01) 0.408 0.01 0.99 (0.98 1.01) 0.586 0.01
Cause of HF 1.06 (0.91 1.25) 0.445 0.09 1.05 (0.90 1.24) 0.531 0.08
Hypertension 0.85 (0.41 1.77) 0.676 0.32 0.90 (0.43 1.87) 0.779 0.33
CKD 1.15 (0.49 2.71) 0.753 0.50 1.14 (0.48 2.68) 0.770 0.50
Stroke 2.33 (0.56 9.66) 0.253 1.69 1.87 (0.45 7.71) 0.387 1.35
Tuberculosis 2.39 (0.57 9.97) 0.252 1.74 1.79 (0.43 7.38) 0.421 1.29

51
 HIV 0.44 (0.11 1.81) 0.256 0.32 0.51 (0.12 2.11) 0.355 0.37
Pneumonia 1.25 (0.69 2.25) 0.465 0.38 1.24 (0.69 2.25) 0.469 0.38

Both final multivariate exponential and Weibull PH models picked up the same three variables
namely history of HF, duration of HF and department the patient seen at first as selected by
multivariate Cox PH model. The hazard ratio and corresponding 95% CI with standard error for
both models are given in table (4.7). More or less both models had same HR with almost
identical standard errors in estimating the variables history of HF and duration of HF, while in
estimating the variable department the patient seen at first exponential achieved considerably
smaller standard error.

Table 4.7: Results of multivariate Weibull PH model and Exponential PH model

Covariates Weibull Exponential

HR 95% CI P value SE HR 95% CI P value SE
History of HF 0.33 (0.15 0.72) 0.006 0.13 0.37 (0.17 0.79) 0.011 0.14
Duration of HF 1.09 (1.03 1.16) 0.002 0.03 1.08 (1.02 1.13) 0.006 0.03
Department seen
first Emergency 4.71 (2.60 8.51) 0.000 1.42 3.82 (2.14 6.82) 0.000 1.13
The overall goodness of fit of both models was also assessed using the Cox-Snell residuals plot
(Figure 5).

52
Figure 5: Cox-Snell residuals plot of Weibull PH model and exponential PH model

3
3

2
2

1
1
0

0 1 2 3 0 0 .5 1 1.5 2
Cox-Snell residual Cox-Snell residual

WH Cox-Snell residual EH Cox-Snell residual

where WH is for Weibull PH model and EH for Exponential PH model

Cox-Snell residuals plot for Weibull PH model shows deviation from the straight line passing
through origin, almost at residuals size 1 while the exponential shows deviation at size greater
than 1 as shown in figure 5.

R2 type statistic was also obtained for both models in the same way as obtained in Cox PH
model. R2 type statistic for Exponential PH model is calculated as follows:

2
{ [
R p2=1− exp
147
(−105.1385− (−90.26696 ) ) =0.183177836 ]}
Which conveys 18.32% variation in the full log-likelihood is explained by the fitted exponential
model. By the same fashion R2 type is found to be 0.1626 for Weibull which indicated that
16.26% variation in the full log-likelihood is explained by the fitted Weibull model. Results of
R2 type statistic and Cox-Snell residuals plot may indicate that the overall fit of the Weibull PH
model is less satisfactory than exponential.

53
4.4 AFT Models

The association between each exposure variable with survival time taking one at a time was
assessed by applying univariate Log-Logistic, Log-normal and Generalized-Gamma AFT
models. Time ratio (TR) with 95% C.I. and standard error for each variable computed from
univariate AFT models are presented in table 4.8.

54
 Table 4.8: Results of univariate AFT models

Covariates Log-Logistic Log-normal Gamma

TR 75% CI P SE TR 75% CI P SE TR 75% CI P SE
Sex 1.31 (0.92 1.86) 0.13 0.23 1.39 (0.98 1.98) 0.07 0.25 1.44 (0.99 2.12) 0.06 0.28

Residence 0.70 (0.48 1.01) 0.06 0.13 0.68 (0.47 1.00) 0.05 0.13 0.68 (0.47 1.00) 0.05 0.13

CAD 0.66 (0.36 1.22) 0.19 0.21 0.70 (0.36 1.36) 0.24 0.24 0.71 (0.35 1.41) 0.24 0.25

DM 0.75 (0.44 1.27) 0.29 0.20 0.70 (0.40 1.22) 0.21 0.20 0.66 (0.36 1.22) 0.19 0.21

History of HF 1.28 (0.87 1.88) 0.21 0.25 1.38 (0.93 2.02) 0.10 0.27 1.60 (1.01 2.51) 0.04 0.37

WBCC 1.01 (0.98 1.04) 0.36 0.01 1.01 (0.99 1.04) 0.33 0.01 1.01 (0.99 1.04) 0.32 0.01

Department
Emergency
0.51 (0.37 0.71) 0.00 0.09 0.50 (0.35 0.70) 0.00 0.08 0.49 (0.35 0.69) 0.00 0.08
Duration of HF 0.98 (0.95 1.01) 0.16 0.01 0.98 (0.96 1.02) 0.34 0.02 0.98 (0.95 1.02) 0.37 0.02

Age category
<=49 years
50-65 years
Above 65 1.00 (0.63 1.60) 0.98 0.24 1.02 (0.63 1.63) 0.93 0.24 1.03 (0.64 1.66) 0.90 0.25
0.70 (0.47 1.04) (0.46 1.05) (0.46 1.06) 0.09
years 0.08 .14 0.70 0.08 0.14 0.70 0.15

Marital status 0.99 (0.74 1.33) 0.96 0.15 0.98 (0.72 1.34) 0.91 0.16 0.98 (0.71 1.35) 0.89 0.16

Systolic BP 0.99 (0.99 1.00) 0.53 .004 0.99 (0.99 1.00) 0.48 .004 0.99 (0.99 1.00) 0.46 .004

Weight 1.01 (0.99 1.03) 0.46 0.01 1.01 (0.99 1.03) 0.52 0.01 1.01 (0.99 1.03) 0.55 0.01

Heart rate 1.00 (0.99 1.01) 0.58 .004 1.00 (0.99 1.01) 0.60 .004 1.00 (0.99 1.01) 0.61 .004

Cause of HF 0.97 (0.88 1.07) 0.56 0.05 0.98 (0.89 1.08) 0.66 0.05 0.98 (0.89 1.09) 0.72 0.05

Hypertension 1.02 (0.66 1.58) 0.92 0.22 0.97 (0.63 1.50) 0.90 0.22 0.93 (0.57 1.52) 0.76 0.23

CKD 0.96 (0.57 1.62) 0.87 0.26 0.94 (0.55 1.62) 0.83 0.26 0.95 (0.54 1.64) 0.85 0.27

Stroke 0.72 (0.31 1.69) 0.45 0.31 0.77 (0.28 2.13) 0.62 0.40 0.80 (0.27 2.37) 0.69 0.44

Tuberculosis 0.58 (0.25 1.36) 0.21 0.25 0.61 (0.25 1.49) 0.28 0.28 0.61 (0.25 1.53) 0.30 0.29

HIV 1.66 (0.77 3.58) 0.20 0.65 1.79 (0.81 3.98) 0.15 0.73 1.86 (0.82 4.21) 0.13 0.77

55
Pneumonia 0.92 (0.64 1.31) 0.65 0.17 0.93 (0.64 1.35) 0.71 0.18 0.94 (0.64 1.38) 0.76 0.18

In Log-Logistic univariate analysis sex, residence, CAD, history of HF, department the patient
seen at first, duration of HF, age category, TB and HIV were significant at 25% significance
level. Both Log-normal and generalized-gamma univariate analysis had variables sex, residence,
CAD, DM, history of HF, department the patient seen at first, duration of HF, age category and
HIV significant at 25% significance level. Those variables, which are significantly associated
with survival time at 25% level in each univariate AFT model, are considered in the
corresponding multivariate AFT model.

A stepwise procedure with forward entry and backward elimination picked up three variables
namely history of HF, department the patient seen at first and duration of HF in multivariate
Log-Logistic AFT model. These are same as those selected by Cox proportional hazards model,
exponential PH and Weibull proportional hazards model. Whereas, Log-normal and Generalized-
Gamma AFT models picked up same variables but different from set of variables selected by
earlier models, namely history of HF, department the patient seen at first, CAD and age category
in their multivariate analysis by the same procedure. The time ratios (TR) along with 95%
confidence intervals and their standard error for each variable in the multivariate AFT models are
presented in the following tables (table 4.9 and table 4.10).

Table 4.9: Results of multivariate Log-Logistic AFT model

Covariates TR 95% CI P value SE

History of HF 1.73 (1.13 2.65) 0.012 0.38
Department emergency 0.45 (0.32 0.63) 0.000 0.08
Duration 0.96 (0.93 0.99) 0.014 0.01

Table 4.10: Results of multivariate Log-normal and Generalized-Gamma AFT models

Log-normal Gamma
Covariates TR 95% CI P value SE TR 95% CI P value SE

History of 1.49 (1.05 2.11) 0.025 0.26 1.50 (1.06 2.12) 0.021 0.26
HF
Department 0.44 (0.32 0.62) 0.000 0.07 0.44 (0.32 0.62) 0.000 0.07
Emergency

56
 CAD 0.54 (0.31 0.96) 0.034 0.16 0.52 (0.29 0.92) 0.024 0.15
Age category
Above 65Y 0.64 (0.46 0.90) 0.011 0.11 0.63 (0.45 0.89) 0.008 0.11

Significance of distributions of generalized-gamma family was done using chi-square test. The
test for the shape parameter K=0 was insignificant (p=0.352) which suggests Log-normal
distribution as inappropriate distribution to fit the data set. Whereas both exponential (p=0.0000)
and Weibull (p=0.0089) distributions were highly significant with the set of variables selected in
the fitted generalized gamma model. But as can be seen in the earlier results exponential and
Weibull models select different set of variables in their fitted model than variables selected by
generalized-gamma. For this reason we assessed the significance of both models with the set of
variables in their fitted model for a second time. Then exponential distribution was significant
again (p=0.0000) while Weibull loses its significance (p=0.1244).

Only exponential AFT model was significant among the nested AFT models and it is taken as a
candidate model without competitor from the generalized-gamma family. As generalized-gamma
model and exponential AFT model selected different set of variables in both univariate and the
fitted models, comparison using log-likelihood is impossible.

Comparison of generalized-gamma and Log-Logistic AFT models using AIC was also
impossible because of the different set of variables selected in their univariate and fitted models.
We checked the goodness of fit of the candidate models in AFT metric, Log-Logistic and
generalized-gamma using the cumulative hazard plot of the Cox-Snell residuals as presented in
figure 6.

57
Figure 6: Cox-Snell residuals plot of Log-Logistic Vs Generalized- Gamma

2.5
2

2
1.5

1.5
1

1
.5

.5
0

0 .5 1 1.5 2 0 0 .5 1 1.5 2
Cox-Snell residual Cox-Snell residual

llH Cox-Snell residual GH Cox-Snell residual

llH for Log-Logistic and GH for Generalised gamma

The Cox-Snell residuals plot of generalized-gamma deviates from straight line at residuals size
of almost 0.6 while the plot for Log-Logistic deviates at size almost 1. R2 type statistic was
calculated for both Log-Logistic and generalized-gamma models. The R2 type for Log- Logistic
was calculated as follows:

2
R p2=1− exp{ [ 147
(−104.3941−(−90.0074 ) ) =0.17777188 ]}
In the same way R2 type is found to be 0.167414353 for generalized gamma. Results of R2 type

convey 16.74% variation in full log-likelihood was explained by the fitted generalized- gamma
model and 17.78% of variability in full log-likelihood is explained by the Log-Logistic AFT
model. The R2 type statistic and Cox-Snell residuals plot may be indicators of Log-Logistic
model as better performed model than generalized-gamma model.

The Q-Q plot of the variable with two levels in the fitted AFT models, history of HF, is used to
check the adequacy of AFT model. The Q-Q plot in Figure (7) approximates to a straight line

58
indicating that the AFT model may provide an appropriate fit for survival time data. Thus, HF
data might be modeled suitably by the AFT models.

Figure 7: Q-Q plot of history of HF

20
15
ne w
10
5
0

0 5 10 15
HF patient before

4.5 Comparison of Models

Although, exponential, Log-Logistic and generalized-gamma AFT models work in same metric
the difference in significant variables in the final and fitted models restricted us from comparison
based on model estimates. As it is difficult to interpret either way since there is no way of
knowing the truth. Comparison of exponential, Log-Logistic AFT and generalized-gamma AFT
models is made based on R2 type statistic and Cox-Snell residuals plots. According to R2
18.32%, 17.78% and 16.74% of variation in full log-likelihood is explained by exponential, Log-
Logistic AFT and generalized-gamma AFT models respectively.

The Cox-Snell residuals plot show that the fit of exponential model appear to be good as most of
the points of the plot fall on straight line passing through origin. Furthermore, Cox-Snell
residuals plot for generalized-gamma and Log-Logistic models shows deviation from straight

59
line at size of residuals almost 0.6 and 1respectively, while the exponential shows deviation at
size greater than 1 (figure 8).

Results of R2 type statistic and Cox-Snell residuals plot may indicate that the overall fit of the
exponential distribution is more satisfactory than the parametric models such as, Log-Logistic
and generalized-gamma models. Furthermore, exponential model was highly significant based on
the chi-square test of generalized-gamma family. Thus, it is taken to be as the only candidate
parametric model for comparison with Cox proportional hazards model.

Figure 8: Comparison of CS residuals: Cox PH vs. Exponential vs. Log-Logistic vs. GG

3

3
2

2
1

1
0

0 1 2 3 0 .5 1 1.5 2
Cox-Snell residual Cox-Snell residual

H Cox-Snell residual EH Cox-Snell residual

2.5
2

2
1.5

1.5
1

1
.5

.5
0

0 .5 1 1.5 2 0 .5 1 1.5 2
Cox-Snell residual Cox-Snell residual

llH Cox-Snell residual GH Cox-Snell residual

H is for Cox, EH for Exponential, llH for Log-Logistic and GH for Generalized-gamma

60
As Cox and exponential PH models can be in the same metric, comparison can be made based on
model estimates. Both Cox and exponential PH models selected the same eight variables in
univariate analysis (P<=0.25) that were considered as candidates for multivariate analysis. Same
three variables (history of HF, duration of HF and department the patient seen at first) were also
selected in the final multivariate analysis of each model. Therefore, as far as variable selection is
concerned, both models are found on the same footing.

The hazard ratio and corresponding 95% CI with standard error for each variable of both models
together are shown in table (4.11). Both models are the same in estimating duration of HF with
identical HR and SE. Hazard ratios of the two variables, history of HF and department the patient
seen at first, for exponential PH model are more than those of Cox model with corresponding
larger standard errors comparative with Cox. Therefore, Cox results are more precise than
exponential results. Based on the explained variation, 19.25% of variation in the partial log-
likelihood is explained by the Cox PH model which is greater than that of the variation explained
in the full log-likelihood by exponential model (18.32%). The Cox-Snell residuals plot of Cox
also deviates from straight line at size greater than 1.5 as shown in figure 8 which is greater than
that of exponential and it seems straight than exponential model. Thus, Cox proportional hazards
model is better performed than exponential Proportional hazards model.

Table 4.11: Comparison of model estimates: Cox PH model vs. exponential PH model

Covariates Cox Exponential

HR 95% CI P value SE HR 95% CI P value SE
History of HF 0.32 (0.15 0.68) 0.003 0.12 0.37 (0.17 0.79) 0.011 0.14
Duration of HF 1.08 (1.03 1.14) 0.003 0.03 1.08 (1.02 1.13) 0.006 0.03
Department seen first
Emergency 3.32 (1.85 5.98) 0.000 1.00 3.82 (2.14 6.82) 0.000 1.13

4.6 Interpretation and Discussion of the Results

Chi-square test shows only exponential is the significant fitted model among generalized-gamma
family. Explained variation and Cox-Snell residuals plot also show exponential AFT is the best
fitted model among AFT models. The quantile-quantile (QQ) plot used to check the adequacy of
AFT assumption was found to be straight line with slop one indicating adequacy of the fitted
exponential AFT distribution. There are two types of exponential model: exponential AFT and

61
exponential PH; but only for the latter is the assumption of PH indispensable; thereby when it
does hold we can use either exponential AFT, exponential PH or Cox PH (Kleinbaum and Klein
2005). The proportionality of hazards assumption was satisfied accordingly exponential model is
taken in its PH metric in order to ease comparison with Cox PH. Cox PH model estimates were
more precise than estimates of the exponential PH model. Explained variation and Cox-Snell
residuals plot also shows Cox PH fitted to data better than exponential distribution this is in line
with reports choose Cox over parametric models (Khanal 2009; Conge and Tsoikos 2010).

The significance of exponential model cannot be realized as irrelevant finding. As exponential

model assumes constant hazard over the study time it is acceptable that hazard of death during
treatment period of patients with severe heart failure admitted to ICU ward to be constant. This
can be obviously because of the very short time of ICU ward treatment period. Besides, the
hazard ratios and corresponding standard errors were similar but non-identical with that of Cox
estimates; as well explained variation and Cox-Snell residuals plot were not far from that of Cox
PH model. These all can be taken as an indication of the exponential model can also suitably fit
heart failure data as an alternative to Cox PH model.

Many studies have indicated demographic, clinical and patient characteristics as correlate
variables with respect to survival of HF patients. This study evaluated the association between
survival during ICU ward treatment period of patients with severe HF and several correlates such
as age, sex, history of HF, duration of HF, department the patient seen at first, white blood cell
count, heart rate, marital status, residence, etiologies of HF and comorbidities such as, diabetes
mellitus, coronary artery disease, chronic kidney disease, HIV, tuberculosis, hypertension, stroke
and pneumonia.

History of HF and emergency department were significant factors which affected the survival
probability of patients in multivariate analysis for all considered models. Based on Cox
multivariate analysis newly diagnosed patients as patient with heart failure at current admission
had an estimated 68% lower risk of mortality than patients diagnosed as patients with heart
failure before current admission (HR=0.32). This finding was similar to a study which pointed a
lower risk of in-Hospital mortality if diagnosed as HF patient for the first time (Abraham et al.
2009). Similarly based on Cox multivariate analysis, it is estimated that patients first seen at
emergency department experiences 3.32 times higher risk of mortality than patients first seen at

62
medical OPD (HR=3.32). This finding was similar to study which pointed discharge of patients
directly from the emergency department as associated with an increased risk of hospital mortality
(Grimaldi et al. 2014).

Duration of heart failure has been a superior prognostic discriminator in multivariate analysis of
almost all considered models. Based on Cox multivariate analysis as duration of the patient stay
with heart failure increases by one year, the hazard of death increases by 8.5% (HR= 1.085). Our
result, confirmed by other study (Gustafsson et al. 2003) which showed that duration of HF to
have negative influence on patient’s survival.

In multivariate generalized-gamma AFT analysis age was found to be a strong predictor variable
which revealed patients aged above 65 years were at increased risk of in-ICU mortality than
patients aged less than 49 years (TR=0.63), and our finding was similar with previous reports
(Abraham et al. 2009; Grigioni et al. 2002; Gustafsson et al. 2003) indicating advanced age
associated with higher in-hospital mortality and (Capell et al. 2013; Pocock et al. 2006) agreeing
patients aged above 65 were at lower survival experience. Multivariate generalized-gamma AFT
model was also discovered patients with presence of coronary artery disease as comorbidity were
at increased risk of in-ICU mortality than patients without coronary artery disease (TR=0.52).

Furthermore, the non-parametric Kaplan-Meier survival estimator curve showed female patients
had slightly higher survival probability than male patients this is in line with (Barlera et al. 2012;
Capell et al. 2012; Henkel et al. 2008; Pocock et al. 2005; Ho et al. 1993) revealing female sex
associated with higher survival experience in population based follow-up studies and
(Gustafsson et al. 2003) indicating male patients at higher risk of in-hospital mortality. Patients
with diabetes mellitus as comorbidity had lower survival experience than diabetes mellitus
absent patients based on the Kaplan-Meier survival estimator curve. our result is in line with
other studies (Barlera et al. 2012; Capell et al. 2012; Henkel et al. 2008; Pocock et al. 2005; Ho
et al. 1993) revealing diabetes mellitus absent patients to have higher survival experience in
population based follow-up studies and (Gustafsson et al. 2003) indicating patients with diabetes
mellitus as comorbidity were at higher risk of in-hospital mortality.

In this study white blood cell count, heart rate, marital status, residence, etiologies of HF and
comorbidities such as, chronic kidney disease, HIV, tuberculosis, hypertension, stroke and
pneumonia were insignificant variables in both univariate and multivariate analysis of all

63
considered models. This may be due to previous researches were population based follow-up
studies while this is ICU based study.

It is believed that when a suitable distribution can be found the parametric model is more
informative than the Cox model. Besides, the appropriate use of these models offers the
advantage of being slightly more efficient; they yield more precise estimates (Bradburn et al.
2003; Moghimi-Dehkordi et al. 2008). On the other hand, Cox regression does not require any
distributional assumption for baseline hazard function. However, it is able to provide good
estimates of regression coefficients, hazard ratios of interest and survival curves for a wide
variety of data situations. It can be considered as a robust model, in the sense that the results
from using the Cox model will closely approximate the results for the correct parametric model
(Kleinbaum and Klein 2005). As mentioned earlier the evaluation criteria considered in our study
such as explained variation, Cox-Snell residuals plot and precision of model estimates indicated
Cox regression model to be more powerful in comparison to both parametric PH model and AFT
model.

64
 CHAPTER FIVE

CONCLUSIONS AND RECOMMENDATIONS

5.1 Conclusions

The study has empirically investigated and identified the factors that are associated with the
survival during treatment period of patients with severe heart failure admitted to ICU ward of
Gondar university hospital using both the semi parametric and parametric survival methods.

In this study, Exponential model was significant and better fitted the data than other AFT
models. The adequacy of AFT assumption was checked using QQ plot and found to be straight
line with slop one indicating adequacy of AFT distribution. It was found that semi-parametric
Cox proportional hazards model was better fitted the data than parametric exponential
proportional hazards model and AFT models. The proportionality of hazards assumption was
found satisfied. Thus, it appears to be sensible to apply Cox PH model as a first choice and if the
Proportional Hazards assumption is not met and/or the fit is not good, then only one should
contemplate other models.

From the multivariable Cox regression model and exponential regression model it was found that
the variables patient first seen at emergency department, history of HF and duration of
HF were the significant correlates of survival during treatment period of patients with severe HF
admitted to ICU ward at Gondar university hospital. Other models and the non-parametric
Kaplan-Meier survival estimate curves found that age, sex, coronary artery disease and diabetes
mellitus to have effect on survival of the patients in the ward. However, residence, marital status,
white blood cell count, systolic blood pressure, heart rate, HIV, all the comorbidities were
found to be insignificant factors of in ICU-ward mortality of patients with severe heart failure.

The predominant causes of HF were coronary heart disease and valvular heart disease which
causes 38.1% and 29.9% of the total population respectively. Deaths were attributed to
respiratory failure (26.09%), cardiac arrest (15.22%), multi-organ failure (15.22%), end stage
renal failure (10.87%), cardiogenic shock (6.52%), septic shock (4.35%), pneumonia (4.35%)
and non-defined causes (17.39%).

65
5.2 Recommendations

Results of this study have implication of the following recommendations:

 Patients first seen at emergency department have higher risk of mortality than patients first
seen at other departments. Thus, attention should be given to patients directly discharged
from emergency department to the ward.
 Patients with known HF before have higher risk of mortality than newly diagnosed once.
Thus, attention should be given to patients diagnosed before current admission as patient
with heart failure in the ward.
 Patients stay higher years with heart failure have higher risk of mortality. Thus, attention
should be given to patients with higher number of years stay with heart failure in the ward.
 Patients aged above 65 have higher risk of in-ICU mortality. Thus attention should be given
to older patients with severe heart failure in the ward.
 Patients with comorbidity coronary artery disease were at higher risk of mortality than
patients without coronary artery disease. Thus, attention should be given to patients with
coronary artery disease as comorbidity in the ward.
 Male patients and patients with diabetes mellitus were at higher risk of mortality than
female patients and patients without diabetes mellitus respectively. Thus, attention should
be given to male patients and patients with diabetes mellitus as comorbidity in the ward.

66
5.3 References

1. Abraham WT, Fonarow GC, Albert NM, Stough WG, Gheorghiade M, Greenberg BH,
O’Connor CM, Sun JL, Yancy CW, Young JB (2008) Predictors of In-Hospital Mortality in
Patients Hospitalized for Heart Failure. J Am Coll Cardiol 52:347–56
2. Ågren S (2010) Supportive care for patients with heart failure and their partners Sweden,
PhD dissertation, Linköping University
3. Agvall B (2014) Heart failure in primary care with special emphasis on costs and benefits of
a disease management program, Sweden PhD dissertation, Linköping University
4. Barlera S, Tavazzi L, Franzosi MG, Marchioli R, Raimondi E, Masson S, Urso R, Lucci D,
Nicolosi GL, Maggioni AP, Tognoni G (2012) Predictors of Mortality in 6975 Patients with
Chronic Heart Failure in the GISSI-HF Trial: Proposal for a Nomogram. Circ Heart Fail
110-135; available from http://circheartfailure.ahajournals.org/
5. Bennett DA, Eliasz TK, Forbes A, Kiszely A, Khosla R, Petrinic T, Praveen D, Shrivastava
R, Xin D, Patel A, MacMahon S, Rahimi K (2012) Study protocol: systematic review of the
burden of heart failure in low- and middle-income countries. Bennett et al. Systematic
Reviews 1:59
6. Bradburn MJ, Clark TG, Love SB, Altman DG (2003) Survival Analysis Part II: Multivariate
data analysis – an introduction to concepts and methods. British Journal of Cancer 89: 605–
611.
7. Bradburn MJ, Clark TG, Love SB, Altman DG (2003) Survival Analysis Part III:
Multivariate data analysis – choosing a model and assessing its adequacy and fit. British
Journal of Cancer 89: 605 – 611
8. Cabral TT, Samuel AJ, Alessandro G, Sylvia C, Alessandro F, Gianfranco B (2011)
Occurrence, etiology and challenges in the management of congestive heart failure in sub-
Saharan Africa: experience of the Cardiac Centre in Shisong, Cameroon. Pan African
Medical Journal 8:11
9. Capell EF, Colet JC, Miralles JD, Saladich IJ, Wensing M, Rotellar JM (2013) Survival in
Mediterranean Ambulatory Patients With Chronic Heart Failure. A Population-based
Study. Rev Esp Cardiol 66(7):539–544

67
10. Collett D (2003) Modelling survival data in medical research (Second edition). Chapman
and Hall/CRC, London
11. Conge C, Tsoikos CP (2010) Statistical modeling of breast cancer relapse time with different
treatments. Journal of applied sciences 10:37-44
12. Cook C, Cole G, Asaria P, Jabbour R, Darrel P. Francis (2013) The annual global economic
burden of heart failure. Int J Cardiol: available from:
http://dx.doi.org/10.1016/j.ijcard.2013.12.028
13. Cowie MR, Wood DA, Coats AJ, Thompson SG, Suresh V, Poole-Wilson PA, Sutton GC
(2000) Survival of patients with a new diagnosis of heart failure: a population based study.
Heart 83: 505–510
14. Falk K, Patel H, Swedberg K, Ekman I (2009) Fatigue in Patients with Chronic Heart
Failure. Eur J Cardiovasc Nurs 8: 91–96
15. Grigioni F, Carinci V, Favero L, Reggiani L, Magnania G, Potena L, Barbieri A, Magelli C,
Branzi A, Magnani B (2002) Hospitalization for congestive heart failure: is it still a
cardiology business? European Journal of Heart Failure 4: 99-104
16. Grimaldi A, Ammirati E, Karam N, Vermi A, De Concilio A, Trucco G, Aloi F, Arioli F,
Figini F, Ferrarello S, Maria Sacco F, Grottola R, D’Arbela P, Alfieri O, Marijon E, Freers J,
Mirabel M (2014) Cardiac surgery for patients with heart failure due to structural heart
disease in Uganda: access to surgery and outcomes. Cardiovasc J Afr 25
17. Gustafsson F, Torp-Pedersen C, Brendorp B, Seibæk M, Burchardt H, Køber L (2003) Long-
term survival in patients hospitalized with congestive heart failure: relation to preserved and
reduced left ventricular systolic function. European Heart Journal 24: 863–870
18. Henkel DM, Redfield MM, Weston SA, Gerber Y, Roger VL (2008) Death in Heart Failure:
A Community Perspective. Circulation 1:91-97
19. Hjelm C (2013) Cognitive function in elderly patients with chronic heart failure, Sweden.
PhD dissertation, Linköping University
20. Ho KK, Anderson KM, Kannel WB, Grossman W, Levy D (1993) Survival after the onset of
congestive heart failure in Framingham Heart Study subjects. Circulation 88: 107-115
21. Hosmer DW, Lemeshow S (1999) Applied Survival Analysis. John Wiley and Sons, Inc.,
New York

68
22. Howie-Esquivel J, Dracup K (2012) Communication with hospitalized heart failure patients.
Eur J Cardiovasc Nurs 11: 216-22
23. International Cardiovascular Disease Statistics (2004) Statistical Fact Sheet. American Heart
Association Web Site, www.americanheart.org.
24. Jingi AM, Noubiap JJ, Kamdem P, Yonta EW, Temfack E, Kouam CK, Kingue S (2013) The
spectrum of cardiac disease in the West Region of Cameroon: a hospital-based cross-
sectional study. International Archives of Medicine 6:44
25. Kalbfleisch JD and Prentice RL (1980) The statistical analysis of failure time data. John
Wiley and Sons, Inc., New York
26. Kengne AP, Dzudie A, Sobngwi E (2008) Heart failure in sub-Saharan Africa: A literature
review with emphasis on individuals with diabetes. Vascular Health and Risk Management
4(1): 123–130
27. Khanal SP (2009) Comparison of Cox proportional hazard, parametric proportional hazard
and accelerated failure time models in the survival of acute liver failure. New Delhi, PhD
dissertation, AIIMS
28. Klein J, Moeschberger M (1997) Survival analysis: Techniques for censored and truncated
data; New York, Springer-Verlag
29. Kleinbaum D, Klein M (2005) Survival analysis: A self- learning text; New York, Springer-
Verlag.
30. Levy WC, Mozaffarian D, Linker DT, Sutradhar SC, Anker SD, Cropp AB, Anand I,
Maggioni A, Burton P, Sullivan MD, Pitt B, Poole-Wilson PA, Mann DL, Packer M (2006)
The Seattle Heart Failure Model; Prediction of Survival in Heart Failure. Circulation;
113:1424-33
31. MacIntyre K, Capewell S, Stewart S, Chalmers JW., Boyd J, Finlayson A, Redpath A, Pell
JP, McMurray JJ (2000) Evidence of Improving Prognosis in Heart Failure: Trends in Case
Fatality in 66 547 Patients Hospitalized Between 1986 and 1995. Circulation 102:1126-31
32. Mayosi BM (2007) Contemporary trends in the epidemiology and management of
cardiomyopathy and pericarditis in sub-Saharan Africa. Heart 93:1176–83
33. McMurray JJ and Stewart S (2002) The burden of heart failure. Eur Heart J Supplements 4
(Suppl D): D50-D58

69
34. Merlo J, Östergren PO, Broms K, Bjorck-Linné A, Liedholm H (2001) Survival after initial
hospitalization for heart failure: a multilevel analysis of patients in Swedish acute care
hospitals. J Epidemiol Community Health 55:323–329
35. Misganaw A, Haile-Mariam D, Ali A and Araya T (2014) Epidemiology of Major Non-
communicable Diseases in Ethiopia: A Systematic Review. J Health Popul Nutr 32(1):1-13
36. Misganaw A, Haile-Mariam D, Ali A and Araya T (2012) The Double Mortality Burden
among Adults in Addis Ababa, Ethiopia, 2006-2009. Prev Chronic Dis 9:110-142
37. Moghimi-Dehkordi B, Safaee A, Pourhoseingholi MA, Fatemi R, Tabeie Z, Zali MR (2008)
Statistical comparison of survival models for analysis of cancer data. Asian Pacific J Cancer
Prev 9: 417-20
38. Onwuchekwa AC, Asekomeh GE (2009) Pattern of heart failure in a Nigerian teaching
hospital. Vascular Health and Risk Management 5: 745–750
39. Owusu IK, Boakye YA (2013) Prevalence and etiology of Heart Failure in Patients Seen at a
Teaching Hospital in Ghana. J Cardiovasc Dis Diagn 1: 131
40. Pocock SJ, Wang D, Pfeffer MA, Yusuf S, McMurray JJ, Swedberg KB, stergren JO,
Michelson EL, Pieper KS, Granger CB (2006) Predictors of mortality and morbidity in
patients with chronic heart failure. European Heart Journal 27: 65–75
41. Ponikowski P, Anker SD, AlHabib KF, Cowie MR, Force TL, Hu S (2014) Heart failure:
Preventing disease and death worldwide. European Society of Cardiology; Available from:
http://spo.escardio.org/eslides/view.aspx?eevtid=59&f
42. Ponnuraja C, Venkatesan P (2010) Survival models for exploring tuberculosis clinical trial
data an empirical comparison. Indian Journal of Science and Technology 3: 755-58.
43. Pourhoseingholi MA, Pourhoseingholi A, Vahedi M, Moghimi Dehkordi B, Safaee A,
Ashtari S, Zali MR (2011) Alternative for the Cox regression model: using parametric
models to analyze the Survival of Cancer Patients. Iran J Cancer Prev 4: 1- 9
44. Pourhoseingholi MA, Hajizadeh E, Dehkordi BM, Safaee A, Abadi A, Zali MR (2007)
Comparing Cox Regression and Parametric Models for Survival of Patients with Gastric
Carcinoma. Asian Pacific J Cancer Prev 8: 412-416
45. Reis S, Holubkov R, Edmundowicz D, Mcnamare D, Zell KA, Detre KM, and Feldman AM
(1997) Treatment of Patients Admitted to the Hospital With Congestive Heart Failure:

70
 Specialty-Related Disparities in Practice Patterns and Outcomes. J Am Coll Cardiol 30:733–
8
46. Sayehmiri K, Eshraghian MR, Mohammad K, Alimoghaddam K, Foroushani AR, Zeraati H,
Golestan B, Ghavamzadeh A (2008) Prognostic factors of survival time after hematopoietic
stem cell transplant in acute lymphoblastic leukemia patients: Cox proportional hazard versus
accelerated failure time models. Journal of Experimental & Clinical Cancer Research 27:1-9
47. Schoenfeld D (1982) Chi-squared goodness-of-fit tests for the proportional hazards
regression model. Biometrika 7:145-153
48. Scottish Intercollegiate Guidelines Network (2007) Management of chronic heart failure: a
national clinical guideline. Edinburgh, K: Available from: http://sign.ac.uk/pdf/sign95.pdf
49. Turner B, Williams S, Taichman D, Goldberg LR (2010) In the clinic Heart Failure. Annals
of Internal Medicine available from; www.annals.org/intheclinic/
50. Volpe M, Francia P, Tocci G, Rubattu S, Cangianiello S, Rao MA, Trimarco B, Condorelli
M (2010) Prediction of Long-Term Survival in Chronic Heart Failure by Multiple Biomarker
Assessment: A 15-Year Prospective Follow-Up Study. Clin. Cardiol 33(11): 700–707

71
 Appendixes

Appendix 1: The Kaplan-Meier survival function estimates for department the patient seen first,
age category, residence, sex, presence of CAD, presence DM and history of HF respectively.

Kaplan-Meier survival estimates Kaplan-Meier survival estimates

1.00
1.00

0.75
0.75

0.50
0.50

0.25
0.25

0.00
0.00

0 5 10 15 20 0 5 10 15 20
analysis time analysis time

dptseenf = Medical OPD dptseenf = Emergency OPD agecat = lessthan 49 years agecat = 50-65 years
dptseenf = Chronic illness follow up agecat = above 65

Kaplan-Meier survival estimates Kaplan-Meier survival estimates

1.00

1.00
0.75

0.75
0.50

0.50
0.25

0.25
0.00

0.00

0 5 10 15 20 0 5 10 15 20
analysis time analysis time

residence = Rural residence = Urban sex = Male sex = Female

Kaplan-Meier survival estimates Kaplan-Meier survival estimates

1.00
1.00

0.75
0.75

0.50
0.50

0.25
0.25

0.00
0.00

0 5 10 15 20 0 5 10 15 20
analysis time analysis time

cad = No cad = Yes dm = No dm = Yes

72
 Kaplan-Meier survival estimates

1.00
0.75
0.50
0.25
0.00

0 5 10 15 20
analysis time

hfpatient = No hfpatient = Yes

Appendix 2: Plot of log (-log (survival)) versus log survival time for categorical predictors in
the fitted model, history of HF and department the patient seen first respectively.
5

4
-ln[-ln(Survival Probability)]
4
-ln[-ln(Survival Probability)]

3
3

1 2
2

0
1

-1
0

0 1 2 3
0 1 2 3 ln(analysis time)
ln(analysis time)
dptseenf = Medical OPD dptseenf = Emergency OPD
hfpatient = No hfpatient = Yes dptseenf = Chronic illness follow up

Appendix 3: Plot of Martingale residuals for covariate duration of heart failure

1 .5
martingale residual
-.5 0 -1
-1.5

0 10 20 30
duration

73