Modelling prognostic factors in advanced pancreatic cancer.

Abstract

Pancreatic cancer is the fifth most common cause of cancer death. Identification of defined patient groups based on a prognostic index may improve the prediction of survival and selection of therapy. Many prognostic factors have been identified often based on retrospective, underpowered studies with unclear analyses. Data from 653 patients were analysed. Continuous variables are often simplified assuming a linear relationship with log hazard or introducing a step function (dichotomising). Misspecification may lead to inappropriate conclusions but has not been previously investigated in pancreatic cancer studies. Models based on standard assumptions were compared with a novel approach using nonlinear fractional polynomial (FP) transformations. The model based on FP-transformed covariates was most appropriate and confirmed five previously reported prognostic factors: albumin, CA 19-9, alkaline phosphatase, LDH and metastases, and identified three additional factors not previously reported: WBC, AST and BUN. The effects of CA 19-9, alkaline phosphatase, AST and BUN may go unrecognised due to simplistic assumptions made in statistical modelling. We advocate a multivariable approach that uses information contained within continuous variables appropriately. The functional form of the relationship between continuous covariates and survival should always be assessed. Our model should aid individual patient risk stratification and the design and analysis of future trials in pancreatic cancer.

Full text

Page 1

Modelling prognostic factors in advanced pancreatic cancer
DD Stocken*,1, AB Hassan2, DG Altman3, LJ Billingham1, SR Bramhall4, PJ Johnson1and N Freemantle5
1Cancer Research UK Clinical Trials Unit, University of Birmingham, Birmingham, UK;2Weatherall Institute of Molecular Medicine, University of Oxford,
Oxford, UK;3Cancer Research UK Medical Statistics Group, Centre for Statistics in Medicine, University of Oxford, Oxford, UK;4Queen Elizabeth
Hospital, Birmingham, UK;5Department of Primary Care and General Practice, University of Birmingham, Birmingham, UK
Pancreatic cancer is the fifth most common cause of cancer death. Identification of defined patient groups based on a prognostic
index may improve the prediction of survival and selection of therapy. Many prognostic factors have been identified often based on
retrospective, underpowered studies with unclear analyses. Data from 653 patients were analysed. Continuous variables are often
simplified assuming a linear relationship with log hazard or introducing a step function (dichotomising). Misspecification may lead to
inappropriate conclusions but has not been previously investigated in pancreatic cancer studies. Models based on standard
assumptions were compared with a novel approach using nonlinear fractional polynomial (FP) transformations. The model based on
FP-transformed covariates was most appropriate and confirmed five previously reported prognostic factors: albumin, CA19-9, alkaline
phosphatase, LDH and metastases, and identified three additional factors not previously reported: WBC, AST and BUN. The effects
of CA19-9, alkaline phosphatase, AST and BUN may go unrecognised due to simplistic assumptions made in statistical modelling. We
advocate a multivariable approach that uses information contained within continuous variables appropriately. The functional form of
the relationship between continuous covariates and survival should always be assessed. Our model should aid individual patient risk
stratification and the design and analysis of future trials in pancreatic cancer.
British Journal of Cancer (2008) 99, 883–893. doi:10.1038/sj.bjc.6604568
Published online 26 August 2008
& 2008 Cancer Research UK
www.bjcancer.com
? ????????????????????????????????????????????
Pancreatic ductal adenocarcinoma is a common cause of cancer
death and is difficult to treat because clinical presentation is often
late, and the disease is resistant to conventional chemotherapy.
Long-term survival remains poor with a 5-year survival rate of
0.4–4% (Bramhall et al, 1995; Jemal et al, 2003). Multivariable
prognostic models are important for grouping patients into risk
sets for predicting survival and treating appropriately. There is
currently no prognostic tool in routine use to identify subgroups of
pancreatic cancer patients for selection and stratification of
treatment and prediction of survival.
Because of its poor prognosis, few prognostic factors may be
expected for patients with advanced pancreatic cancer; however,
many possible factors have been identified. The majority of
prognostic factor studies are questionable in terms of sample size
and statistical methods, most based on small retrospective
analyses. Literature searches (ISI Web of Science and Ovid
Technologies databases) identified 36 prognostic factor studies
reporting a total of 34 possible prognostic factors for advanced
pancreatic cancer patients (Table 1) grouped as surgical, clinical,
laboratory or demographic. Four studies (Johnson et al, 2001;
Berlin et al, 2002; Ducreux et al, 2002; Maisey et al, 2002) were
randomised controlled trials reporting five prognostic factors from
multivariate analyses, namely metastases, tumour site, perfor-
mance status, alkaline phosphatase and treatment. The remaining
32 studies (Friedman and van den Eeden, 1993; Yasue et al, 1994;
Falconer et al, 1995; Lundin et al, 1995; Ishii et al, 1996;
Rothenberg et al, 1996; Shibamoto et al, 1996; Cubiella et al,
1999; Storniolo et al, 1999; Halm et al, 2000; Terwee et al, 2000;
Trigui et al, 2000; Ueno et al, 2000; Ikeda et al, 2001; Ridwelski
et al, 2001; Tas et al, 2001; Tsuruta et al, 2001; Saad et al, 2002;
Bachmann et al, 2003; Engelken et al, 2003; Fujino et al, 2003;
Karayiannakis et al, 2003; Micke et al, 2003; Ohigashi et al, 2003;
Paillaud et al, 2003; Stemmler et al, 2003; Talar-Wojnarowska et al,
2003; Ziske et al, 2003; Gupta et al, 2004; Kuhlmann et al, 2004;
Watanabe et al, 2004; Ni et al, 2005) were based on consecutive
series of patients, often retrospective, often single-centre, of which
15 studies were based on fewer than 100 patients (Yasue et al, 1994;
Ishii et al, 1996; Rothenberg et al, 1996; Halm et al, 2000; Ikeda
et al, 2001; Tsuruta et al, 2001; Saad et al, 2002; Karayiannakis
et al, 2003; Micke et al, 2003; Ohigashi et al, 2003; Paillaud et al,
2003; Stemmler et al, 2003; Talar-Wojnarowska et al, 2003; Ziske
et al, 2003; Gupta et al, 2004). One was based on five observational
studies with varied inclusion criteria, inconsistent results and no
prospective verification (Terwee et al, 2000). The largest series
(2380 patients) identified factors based on univariate analyses and
data containing a large proportion (57%) of censored patients
(Storniolo et al, 1999).
An important issue in prognostic factor studies is the nature of
the relationship between the factor and survival (functional form).
Continuous variables are often simplified at analysis by assuming a
linear relationship with log-hazard or by introducing a step
function through categorisation (frequently dichotomisation). If
the linearity assumption is not correct, the final prognostic model
could be misspecified. Misspecification of the functional form may
lead to inappropriate conclusions but has not been previously
investigated in pancreatic cancer studies. Many researchers avoid
this problem by dichotomising, with a consequent loss of power.
There is also the risk of important bias when the choice of cutoff is
data-driven and the use of different cutoff points across multiple
studies hinders direct comparisons.
Received 3 September 2007; revised 8 July 2008; accepted 23 July 2008;
published online 26 August 2008;
*Correspondence: DD Stocken; E-mail: d.d.stocken@bham.ac.uk
British Journal of Cancer (2008) 99, 883–893
& 2008 Cancer Research UKAll rights reserved 0007– 0920/08$32.00
www.bjcancer.com
Clinical Studies

Page 2

The aim of this study was to evaluate potentially important
baseline prognostic factors for survival in advanced pancreatic
cancer using prospective data from two randomised controlled
trials and a total of 653 patients (Bramhall et al, 2001, 2002). The
study investigated clinical, histological, biochemical and demo-
graphic variables. A multivariable approach was used accounting
for the functional form of the relationship between continuous
factors and survival. Models were developed either on the basis of
standard assumptions of log linear or step functional relationships
with survival, or a novel approach based on nonlinear relation-
ships, using more complex fractional polynomial (FP) transforma-
tions: a flexible, parametric method for modelling nonlinear
Table 1Literature review
Type (number)
of studies
Number of patients
per studyPrognostic factors reported (frequency of reporting)
U or MV
analysis
Randomised controlled
trial (n¼4)3–6
207–322SurgicalMetastases
Tumour location
Performance status
Treatment
Alkaline phosphatase
(3)
(1)
(3)
(1)
(1)
4 MV
Clinical
Laboratory
Consecutive series
4500 patients (n¼3)7–9
782–2380SurgicalMetastases
Stage of disease
Operation
Performance status
Diabetes
Pain
Appetite/weight
Jaundice
Treatment
Albumin
Age
Specialist centre
(1)
(1)
(1)
(1)
(1)
(1)
(1)
(1)
(1)
(1)
(1)
(1)
2 MV
1 U
Clinical
Laboratory
Demographic
Consecutive series 100–500
patients (n¼14)10–23
102–450SurgicalMetastases
Stage of disease
Tumour location
Operation
Tumour size
Duodenal invasion
Peridissemination
Ascites
Performance status
Diabetes
Pain
Appetite/weight
Symptom onset
Treatment
CA242
CA19-9
Leukocytes
Gamma GT
Albumin
LDH
CRP
Iron
Age
(4)
(2)
(1)
(2)
(1)
(1)
(1)
(1)
(2)
(1)
(1)
(3)
(1)
(2)
(2)
(2)
(1)
(1)
(1)
(1)
(3)
(1)
(1)
13 MV
1 U
Clinical
Laboratory
Demographic
Consecutive series
o100 patients (n¼15)24–38
28–95Surgical Metastases
Stage of disease
Grade of disease
Nodal status
Operation
Tumour size
Fibrosis
Performance status Inflammation
Appetite/weight
Treatment
CA19-9
VEGF
CEA
Phase angle BIA
SCA
(1)
(1)
(1)
(1)
(1)
(2)
(1)
(4)
(1)
(1)
(1)
(7)
(1)
(1)
(1)
(1)
8 MV
7 U
Clinical
Laboratory
MV¼multivariate; U¼univariate.
Prognostic factors in pancreatic cancer
DD Stocken et al
884
British Journal of Cancer (2008) 99(6), 883–893
& 2008 Cancer Research UK
Clinical Studies

Page 3

relationships (Royston and Altman, 1994; Altman and Lyman,
1998).
PATIENTS AND METHODS
Data
Two international phase III British Biotech studies (BB128,
Bramhall et al, 2001; BB193, Bramhall et al, 2002) randomised
414 and 239 patients with advanced pancreatic cancer, respec-
tively: BB128 randomised patients between marimistat and
gemcitabine; BB193 randomised patients between marimistat with
gemcitabine and gemcitabine alone. The studies had similar
eligibility criteria: histologically or cytologically unresectable
pancreatic cancer, within 8 weeks of diagnosis or disease
recurrence and Karnofsky performance status of X50% (BB128)
or X60% (BB193). Previous therapy for metastatic or locally
advanced disease was an exclusion criterion. The primary outcome
measure in both studies was survival time calculated from the date
of randomisation to the date of death from any cause.
Randomisation was stratified by cancer stage (stage I/II, III or
IV), Karnofsky performance status (50–70%, 80–100%), sex and
study centre. The first stage of data reduction was considering only
factors that were clinically relevant and available within an NHS
outpatient clinic. Eighteen baseline clinical, histological, bio-
chemicaland demographicvariables
randomised treatment group) were considered appropriate for
analysis as possible prognostic factors (Table 2).
(including trialand
Statistical analysis
We followed a strategy aimed at maximising model performance
and avoiding poorly fitted and overfitted regression models
in the development (Harrell et al, 1996) and reporting (McShane
et al, 2005) of multivariable prognostic models. Initial analysis
was based on standard methodology comparing Kaplan–Meier
survival estimates using the log-rank test and estimating univariate
hazard ratios for levels of each factor. The hazard of death
was assessed in the multivariable setting using Cox proportional
hazardsregression modelling
backward elimination. The proportional hazards assumption
was investigated for each covariate using log cumulative
hazard (Collett, 1994) and martingale residual plots, and
incorporatingatime-dependent
(survival)?LN (mean survival))) and did not indicate any
significant violation.
Ten of the 18 possible prognostic factors were collected as
continuous measurements. Continuous data were investigated by
assessing three different assumptions of the underlying relation-
ship between survival and predictor, they are: (a) a linear
relationship between the predictor and log hazard, (b) a step
functional relationship using dichotomised covariates (laboratory
measures based on central laboratory reference ranges) and (c) a
nonlinear relationships based on either a simple log or more
complex nonlinear FP transformation (Royston and Altman, 1994).
For the third model, the functional form of each variable was
assessed univariately comparing the Akaike’s Information Criter-
ion (AIC) (Collett, 1994) of a model based on the simple log
transformation with the AIC of a model based on the best fitting
FP transformation. First- and second-degree FP transformations
(Meier-Hirmer et al, 2003) were considered using a selection level
of 0.05 for input of variables based on power values of the
polynomial ranging (?2, ?1, ?0.5, 0 (log), 0.5, 1, 2, 3). The best FP
for each predictor was selected if it resulted in a significantly better
fit (significantly smaller AIC) than the log transformation. The
most appropriate (log or FP) transformation, if any, was applied to
each variable and all variables were considered multivariately
withvariablereductionby
covariate(X¼factor(LN
using Cox proportional hazards regression based on a backward
selection method using a nominal significance level of 0.05 for
elimination and including trial, sex, cancer stage (stratification
factors at randomisation) and randomised treatment in each
model.
The majority of variables had p5% missing values (Table 2).
Tumour stage, CA19-9 and WBC had 5–10% missing values, and
lymph node status was missing for 24% of patients. Metastases or
lymph node status was considered in the analysis as dummy
variables using ‘negative’ as a reference level. Primary analysis was
based on complete cases and a secondary analysis used multiple
imputation to investigate the possible influence of variables with
larger amounts of missing data (Rubin, 1987; Schafer, 1997) and
provided valid inferential alternative results.
Model fit was assessed comparing AIC statistics, deviance
residuals and Kaplan–Meier survival statistics for four predictive
groups. The four predictive groups were based on quartiles of
linear predictor scores, assessed comparing median survival
estimates and hazard ratios. The bootstrap resampling approach
described by Harrell et al (1996) was applied to assess the extent of
overfitting in the final model, using 200 bootstrap resamples. This
approach repeats the model selection methods used in the original
model development in a series of bootstrapped resamples, freezing
the derived model and applying to the original sample. Model
optimism (overfitting) is described by the difference in the rank
correlation coefficient relating predicted and observed survival
times between the model derived in the bootstrap resample and
that from the frozen model applied to the original sample averaged
over 200 resamples. This provides an honest estimate of internal
validity penalised for overfitting (Harrell et al, 1996).
Analyses were carried out using SAS and R using a two-sided
significance level of 0.05 throughout.
RESULTS
Patient characteristics
A total of 653 patients were randomised. The eighteen clinically
appropriate factors for analysis are presented in Table 2 and
appear balanced across the two studies. On average, patients in the
two trials were randomised 20 and 15 days after diagnosis and
started treatment the day following randomisation. The average
age of patients was 63 years (range 29–89), 368 (56%) were male,
439 (68%) had cancer stage IV disease, 436 (67%) presenting with
metastases and 251 (39%) had lymph node involvement.
Survival
The majority of patients (612, 94%) had died by the time of
analysis with a median follow-up time of 21 months for the 41
patients still alive (Table 2, Figure 1). The median survival estimate
for the group is 4.7 months (95% CI: 4.2, 5.1) with 12-month
survival estimate of 17% (Figure 1). Hazard functions estimated for
1-monthly time intervals to 18 months from trial entry were
similar for both trials and reasonably constant over time. No
significant survival benefit for marimastat was identified in the
BB128 trial (P¼0.19) when compared with gemcitabine (Bramhall
et al, 2001). Similarly, no significant survival benefit was seen for a
combination of gemcitabine and marimistat when compared with
gemcitabine alone in the BB193 trial (P¼0.95) (Bramhall et al,
2002).
Univariate analyses
Log-rank analyses (Table 3) indicated that potentially important
factors were age (split at median, P¼0.036), nodal status
(P¼0.035), cancer stage (I/II vs III/IV), metastases (both
Po0.001) and laboratory measures (split as normal/abnormal
Prognostic factors in pancreatic cancer
DD Stocken et al
885
British Journal of Cancer (2008) 99(6), 883–893
& 2008 Cancer Research UK
Clinical Studies

Page 4

Table 2Patient characteristics by trial
Variable BB12839N¼414 (63%) BB19340N¼239 (37%)Total N¼653 (100%)
Demographics
Age at entry (years)a
Median
Range
63 6263
29–8932–8529–89
Ethnic race
White
Black
Oriental
Other
Missing
364 (88%)
27 (6%)
7 (2%)
15 (4%)
1
226 (95%)
8 (3%)
0
5 (2%)
0
590 (90%)
35 (6%)
7 (1%)
20 (3%)
1
Sex
Male
Female
228 (55%)
186 (45%)
140 (59%)
99 (41%)
368 (56%)
285 (44%)
Treatment
Gemcitabine
Marimistat
103 (25%)
311 (75%)
119 (50%)
120 (50%)
222 (34%)
431 (66%)
Tumour information
Cancer stage
I
II
III
IV
Missing
19 (4%)
45 (11%)
76 (19%)
268 (66%)
6
13 (5%)
27 (11%)
28 (12%)
171 (72%)
0
32 (5%)
72 (11%)
104 (16%)
439 (68%)
6
Distant metastases
M0
M1
Missing
129 (31%)
265 (64%)
20 (5%)
65 (27%)
171 (72%)
3 (1%)
194 (30%)
436 (67%)
23 (3%)
Regional lymph nodes
N0
N1
Missing
153 (37%)
164 (40%)
97 (23%)
90 (38%)
87 (36%)
62 (26%)
243 (37%)
251 (39%)
159 (24%)
Primary tumour T stage
T0
T1
T2
T3
T4
Missing
5 (1%)
114 (30%)
90 (24%)
167 (44%)
6 (1%)
32
3 (1.5%)
44 (20%)
54 (25%)
113 (53%)
1 (0.5%)
24
8 (1.5%)
158 (26%)
144 (24%)
280 (47%)
7 (1.5%)
56
Serum chemistry and haematology
Laboratory variablesMedian (range), missingMedian (range), missingMedian (range), missing
AST (SGOT)a
Total bilirubina
Alkaline phosphatasea
Albumina
LDHa
BUNa
CA19/9a
Haemoglobina
WBCa
24 (6–365), 17
13.7 (3.4–277.0), 16
136 (36–1660), 16
38 (22–47), 17
163 (77–1074), 21
9.2 (2.9–34.3), 17
686 (5–1000000), 17
12.5 (5.5–16.1), 28
7.6 (2.3–31.6), 28
26 (9–538), 12
13.7 (3.0–135.1), 8
157 (35–2064), 8
38 (24–47), 8
169 (29–1495), 11
9.3 (4.3–27.9), 16
800 (8–1000000), 30
12.4 (8.3–19.1), 13
8.3 (2.4–23.7), 13
25 (6–538), 29
13.7 (3.0–277.0), 24
140 (35–2064), 24
38 (22–47), 25
164 (29–1495), 32
9.3 (2.9–34.3), 33
710 (5–1000000), 47
12.4 (5.5–19.1), 41
7.9 (2.3–31.6), 41
Outcome
Event indicator
Alive
Dead
22 (5%)
392 (95%)
19 (8%)
220 (92%)
41 (6%)
612 (94%)
Follow-up of alive (months)a
Median
Range
20.119.420.7
0.9–24.6 1.9–23.30.9–24.6
aContinuous measurements.
Prognostic factors in pancreatic cancer
DD Stocken et al
886
British Journal of Cancer (2008) 99(6), 883–893
& 2008 Cancer Research UK
Clinical Studies

Page 5

according to laboratory reference ranges as per current clinical
practise) as AST, alkaline phosphatase, albumin, LDH, WBC (all
Po0.001), bilirubin (P¼0.002), CA19-9 (P¼0.005) and haemo-
globin (P¼0.009). Trial, treatment (gemcitabine vs marimistat),
race (white vs not-white), sex, tumour stage (T0, 1, 2 vs T3, 4) and
BUN were not significantly related to survival.
Multivariate analyses
Three Cox proportional hazards regression models were developed
(Table 4) using 556 patients (520 deaths) with complete data
(excluding patients with missing data) based on the assumption of
(a) a linear relationship between continuous covariates and log
hazard, (b) a step functional dichotomisation of continuous
covariates and (c) a nonlinear transformation of continuous
covariates. All three models included trial, sex, cancer stage
(stratification factors at randomisation) and randomised treatment
group.
The ‘linear’ model (Table 4, Model 1) identified five highly
significant prognostic factors, namely albumin, alkaline phospha-
tase, LDH, WBC and metastases. The ‘categorical’ model (Table 4,
Model 2) identified six highly significant prognostic factors,
namely LDH, albumin, metastases, WBC, CA19-9 and bilirubin.
Univariate analysis of the 10 continuous variables identified that
nonlinear transformations were appropriate for 3 variables in their
relationship with survival: bilirubin and LDH both as log
transformations and CA19-9 as a second-degree FP transformation
(CA19-90.5þ(CA19-90.5?log(CA19-9))). The seven remaining
continuous covariates were analysed assuming a linear relation-
ship with log hazard, as in Model 1. The ‘transformed’ model
(Table 4, Model 3) identified eight prognostic factors. Five factors
were highly significant with Po0.01, namely albumin, CA19-9,
LDH, alkaline phosphatase and WBC with AST, BUN and
metastases being more borderline in the model (P¼0.023, 0.026
and 0.047, respectively).
Nonlinear transformations were appropriate for two variables,
LDH and CA19-9, and the estimated log hazard ratio functions are
shown graphically in Figures 2 and 3. The second-degree FP
function for CA19-9 estimates increasing risk up to an approx-
imate CA19-9 value of 14000 and then decreases with increasing
CA19-9. The log function for LDH estimates increasing risk for
increasing values of LDH.
Model comparison
In all three models, albumin, LDH and WBC were highly
statistically significant and influential prognostic factors. Meta-
stases were also an important variable but its parameter estimate
and overall significance were reduced in the ‘transformed’ model
when continuous covariates were included in a more appropriate
format. In both the ‘linear’ and ‘transformed’ models, alkaline
phosphatase was also a highly significant and influential prog-
nostic factor. CA19-9 was also a highly significant and influential
prognostic factor in both the ‘categorical’ and ‘transformed’
models. This variation is largely explained by the nonlinear
relation of CA19-9 to survival (Figure 2), which could explain why
it was considered important when dichotomised but not when
included as linear. When considered as a transformed second-
degree FP, its significance was much greater. Bilirubin was selected
as a highly significant factor in the ‘categorical’ model but was not
included in either the ‘linear’ or ‘transformed’ models. AST and
BUN were only selected as prognostic in the ‘transformed’ model.
Model performance
The AIC statistic was smallest for the ‘transformed’ model (Table 4,
Model 3), indicating a better fit to the data. Deviance residuals for
this model were plotted against the linear predictor and were
randomly scattered centred around a residual value of zero
ranging between ?3.86 and 3.33, which suggests the data have not
been mis-modelled.
Patients were split into four groups based on quartiles of the
distribution of linear predictor scores from the ‘transformed’
model. Kaplan–Meier survival estimates (Figure 4) show four
distinct predictive groups with descending median survival
estimates of 9.1 (95% CI: 7.4, 10.9), 7.0 (95% CI: 5.9, 8.3), 4.0
(95% CI: 3.4, 4.9) and 2.0 (95% CI: 1.6, 2.4) months. The hazard
ratios for groups 2, 3 and 4 using predictive group 1 as the baseline
were 1.35 (95% CI: 1.09, 1.66), 2.08 (95% CI: 1.64, 2.64) and 4.21
(95% CI: 3.11, 5.68), respectively.
When assessing model validity, the R2measure of model fit was
estimated as 0.30. The bootstrap resampled estimate of R2of 0.26
described model optimism (overfitting) under 5% and gives an
improved estimate of model accuracy.
Multiple imputation allowed all 653 patients to be included in
the modelling process and confirmed all the variables included in
the ‘transformed’ model with increased significance for metastases
(P¼0.001), and the model also included nodal status (P¼0.016)
that had been excluded from all models prior to imputation,
suggesting a strong link with other variables already in the model.
DISCUSSION
Large, prospective, phase III randomised controlled trials aim to
provide robust statistical evidence for new treatment combina-
tions. Stratification is important to control for known important
variability in the data. Generally, patients with pancreatic cancer
are not clinically separated into prognostic groups, with the
exception of surgical status, before consideration for treatment.
This study investigated potentially important baseline prognostic
factors for survival as possible stratification variables for
randomisation and analysis. Data from 653 patients included in
two international randomised controlled trials in advanced
pancreatic cancer (Bramhall et al, 2001, 2002) were analysed
investigating multiple clinical, histological, biochemical and
demographic variables in the form of both binary and continuous
measurements. Valid statistical analyses are necessary to make
best use of the data and optimise clinical results. As such, a
multivariable approach was used to account for the functional
form of the relationship between continuous prognostic variable
factors and survival. Misspecification of functional form may lead
to inappropriate conclusions but has not been previously
investigated in pancreatic cancer studies. Continuous variables
are often simplified by assuming a linear relationship between
predictor and log hazard, that is the log risk increases or decreases
100
BB128
BB193
Survival (%) 75
50
25
0
0
No. at risk
BB128 414
BB128 239
6
Months
148
103
70
42
33
18
12 18
Figure 1 Survival function by trial.
Prognostic factors in pancreatic cancer
DD Stocken et al
887
British Journal of Cancer (2008) 99(6), 883–893
& 2008 Cancer Research UK
Clinical Studies