Professional Documents
Culture Documents
University of Sydney
AE Adverse event
AFT Accelerated failure time
CR Complete response
CRF Case report form
DSMB Data safety monitoring board
ECOG Eastern cooperative oncology group
ESMO European society for medical oncology
HER2 Human epidermal growth factor receptor 2
HT Study medication taken by patients in treatment groups A and B
LD Longest diameter
LVEF Left ventricular ejection fraction
PD Progressive disease
PH Proportional hazards
PR Partial response
RECIST Response evaluation criteria in solid tumors
SAS Statistical analysis system commercial software
SD Stable disease
STATA Commercial statistical software
TTP Time to progression
X Study medication taken by patients in treatment group A
Preface
The two workplace projects are tied to one clinical trial conducted in Roche Products
Pty Ltd looking at the efficacy and safety of HT with or without X in patients with
human epidermal growth factor receptor 2 (HER2)-positive advanced or metastatic
breast cancer:
The primary objective of this study was to compare the overall response rate of
patients randomized to HT with overall response rate of patients randomized to HT
plus X.
As a SAS programmer in Roche Products Pty Ltd, assigned to work on this Phase II
clinical trial, I was responsible for the creation of report objects required for abstract
preparation, manuscript writing and final statistical report. Working closely with the
statistician responsible for the study, and having gained theoretical background
during my coursework over the past few years, I had been given the opportunity to
take on additional tasks as part of my workplace projects in our Biostatistics
Department.
2. Scope of work
The first project encompassed the evaluation of the study data collection and
making sure that the available data are suitable for the planned statistical analyses. It
also evaluated the design of the study and the use of the Response evaluation
criteria in solid tumors (RECIST) in comparing the overall response in the two
treatment arms.
Since there were set timelines for the presentation of the first efficacy and safety
results of the clinical trial, I had to ensure that data collection, data cleaning and data
transfer processes are all well planned and coordinated. Given the remote location of
the statistics and data management team working on this clinical trial, communication
was a challenging task, mainly carried out by email, as well as teleconferences on a
regular basis for issues resolution.
The second project involved the selection of the appropriate statistical models for
the expanded analyses of the clinical trial data, writing the SAS code to perform the
The use of logistic regression in a real-life workplace setting example, which was part
of the second project, gave me an insight on how challenging it can be to try to
reliably interpret the possible effects of the covariates on a primary efficacy variable,
particularly when there are missing data, and also in the presence of confounding.
The workplace projects undertaken as part of this Portfolio provided me with the
opportunity to get hand on experience of what it means being involved in a range of
professional duties related to the design and analysis of clinical research.
Working in collaborative manner with the Data Management team, the first project
emphasized the development of communication and negotiation skills through team
work and activities like attending Study Management Team teleconferences and face
to face meetings.
The two projects I was involved in during the last eight months at my workplace are
mostly related to the Design of Experiments and Randomised Clinical Trials (DES),
Data Management and Statistical Computing (DMC), Survival Analysis (SVA),
Categorical Data and GLMs (CDA) units of study that I’ve undertaken in my
coursework.
By working on the projects defined for this unit of study, I realized how interrelated
are the many concepts acquired in different units of the coursework undertaken, and
how these different units of study like puzzles address different aspects of the same
clinical trial.
Most of the practical examples and exercises during the graduate coursework were
done using the STATA software. I found it challenging but at the end rewarding to
learn and apply the theoretical knowledge that I’ve gained over years using the SAS
v8.2 software which is widely used at my workplace.
As part of my second workplace projects, for which I did not have time constraints
apart from the deadline to hand in the portfolio for this unit of study, I started on
several different things at a time. I got easily carried away with different approaches
and possibilities, and had not finished one thing before jumping onto something else.
At the end I realized that though it is interesting to do exploratory analysis, one has to
keep focused on the objective of the analysis in order to respect given timelines. This
is a lesson from which I can learn for future work projects.
4. Ethical consideration
This clinical trial was conducted in accordance with the Declaration of Helsinki, and
International Good Clinical Practice principles as described in the International
Conference of Harmonization (ICH) guidelines for Good Clinical Practice, as well as
in accordance with all local ethical and regulatory requirements.
The projects undertaken as part of this portfolio were following the same principles of
confidentiality with respect to documentation related to the clinical trial and subject
records.
Some details related to the protocol, study conduct or results are omitted from this
portfolio for reasons of confidentiality.
Since the results of the clinical trial have not been presented yet to a wide audience,
and to protect proprietary information by the sponsor, the treatment information is
The above restriction should not affect the understanding of the concepts and
methods used during my workplace project work.
Apart from having regular consultations with Peter Button, the lead statistician for this
clinical trial who was also my supervisor for the workplace projects, I had a chance to
contact and consult colleagues from the Roche Welwyn office in UK, who had
experience with the use of the accelerated failure time (AFT) model for the analysis
of time-to-event data.
Project Report I
6. REFERENCES ..............................................................................................................................14
This workplace project comprised of the evaluation of data collection and data
management on an open-label randomized phase II study of HT and X in
combination, versus HT, in patients with advanced and/or metastatic breast cancers
that overexpress HER2. This project was geared towards making sure that the
available data are suitable for the primary study analysis defined in the Analysis Plan
[1]. In this clinical trial, overall, 222 patients were enrolled and received study drug
from January 2002 to September 2005.
The results of the primary efficacy analysis of the clinical trial investigating the
efficacy and safety of the triple combination of HT and X compared with HT as 1st-
line treatment for HER2-positive advanced/metastatic breast cancer, were to be
presented in Basel, Switzerland at the beginning of June 2006 for the first time to a
small audience consisting of Study Management Team, Clinical science leader,
International medical leaders and Documentation specialist.
Following this the results were to be used for presentation at the European Society
for Medical Oncology (ESMO) in September 2006.
The above milestones determined also quite strictly the timelines for this workplace
project.
Both biostatistics and data management activities for the study have been performed
by the Roche Basel teams until October 2005, at which time Roche Australia took
over the biostatistics activities – approximately six months prior to the planned
database lock.
The aim of this project was to evaluate the data collection and data management
process that was undergoing in this study at the stage when our biostatistics team
took over the responsibility of data analysis for the final study reporting event, and to
give guidance and input into further practice. This project also took on coordinating
and giving advice about particular data collection and data management issues that
arose during the busy data cleaning period preceding the clinical cut-off date for the
primary study analysis.
The clinical trial I worked on as part of this workplace project was an open-label,
randomized, multi-center, comparative phase II study of HT administered
intravenously once every 3 weeks, with or without oral X administered twice-daily for
2 weeks followed by a 1 week rest period. Treatment in both arms will continue until
progressive disease, unmanageable toxicity or patient request.
Screening and Day 1 until After cycle 8 After one year 28 days after
Baseline Day progressive until disease from the last dose of
disease progression or enrollment study drug
one year from
enrollment
From -28 to +1 Drug treatment. Tumor Disease Safety follow-up
Tumor assessment assessment
assessment every 12 weeks according to
every 6 weeks routine practice
until disease
progression.
Survival status
until 18 months
after the last
patient enrolled
Clinical cut-off for the primary study analysis was six months after the last patient
enrolled. However, survival data will be collected until 18 months after the last patient
enrolled. These results will be reported a year later in an addendum to the study
report.
This clinical trial provided me with an example for understanding the methods and
rationale for determining whether a study should be stopped early [1]. This included
getting familiar with the role and findings of the Data and Safety Monitoring Board,
Recruitment on this Phase II clinical trial began in January 2002 and was suspended
in February 2003, after enrolment of 65 patients, following concerns from the Data
Safety Monitoring Board (DSMB).
Recruitment remained suspended for three months while follow-up safety data were
collected on the enrolled study population.
The study investigators were informed of the results of the Interim Safety Analysis
conducted in April 2003, (data cutoff for the 65 patients Safety analysis was April 7th
2003), and their awareness raised regarding the higher than expected incidence of
febrile neutropenia seen in the study.
After detailed assessment in June 2003, the DSMB were satisfied, and requested
amendments to the protocol and patient information sheet to address the findings.
The protocol changes were designed to ensure selection of patients better able to
tolerate the therapy, and monitoring the course of any neutropenia.
The DSMB requested a further safety data review of the study when another 45
patients had been enrolled, giving a total of 110 patients, half the expected overall
recruitment.
Following the second Interim Safety Report, recruitment into the study continued as
of the end of June 2004.
In familiarizing myself with the project that I was assigned to, I had to make sure that
I understand the implications of the stopping and re-opening of the study. Protocol
amendment included changes to the eligibility criteria such as widening the inclusion
criteria for the enrollment of patients with lower Eastern cooperative oncology group
(ECOG) performance status, and excluding patients with dyspnea at rest due to
malignant or other diseases and patients on chronic concomitant steroids. Due to
such amendments the baseline characteristics of patients before and after stopping
were slightly different. We had to monitor whether these differences in baseline
characteristics of patients could potentially impact on the outcome. No such
The primary measure of treatment efficacy was the overall tumor response rate.
Evaluation of target lesions and non-target lesions was in accordance with the
Response evaluation criteria in solid tumors (RECIST) [2].
To be able to work with the data collected for this clinical trial it was of importance to
understand the elements of the RECIST criteria for the measurement of tumor
response.
The following table shows how overall response is determined at each treatment
cycle, taking into account the tumor responses in target and non-target lesions and
the appearance of new lesions during that cycle.
CR CR No CR
CR Non-CR/Non-PD No PR
PR Non-PD No PR
SD Non-PD No SD
PD Any Yes or No PD
Any PD Yes or No PD
One of the problems of the definition that I see is that according to the study protocol,
partial and complete responses have to be confirmed at consecutive tumor
assessments at least 4 weeks apart in order to be considered as responses, whereas
confirmation of progressive disease is not required. A possible consequence of this is
also that without the confirmation of progressive disease, some patients could be
taken off study medication too early.
Another problem that I see is related to the definition of progressive disease for target
lesions itself. Though partial response is measured in relation to the baseline tumor
measurement, progressive disease is measured in relation to the smallest lesion at
any time during the study. According to this definition it could happen that the sum of
longest diameter at some tumor assessment during the study is smaller than the
baseline value, and the patient still to be classified based on this assessment as
having a progressive disease.
For example: baseline sum of longest diameter is 20 mm, value at cycle 4 is 10 mm,
and value at cycle 8 is 13 mm, which classifies as progression and as a
consequence the patient is being taken off the study drug without the confirmation of
the latest tumor measurement.
The framework for data collection for this clinical trial was defined in the protocol to
ensure quality data at the end of the study.
As this particular clinical trial had protocol amendments mentioned in section 2.1
above, and since our biostatistics team in Dee Why took over the responsibility for
this study from our colleagues in Basel, Switzerland, an evaluation of data collection
was necessary.
The study data for this clinical trial were collected on Case report forms (CRFs) and
entered into Oracle Clinical on a Unix platform.
The data collected in the Oracle Clinical database were mapped into SAS general
A regular overnight batch job of data transfer was set up between the Oracle Clinical
Administration side and the Biostatistic Computing Environment side stored on
different Unix platforms.
Additional information needed for the analysis was captured in value added datasets
for demography and efficacy variables.
Occasionally, upon request there were additional data transferred arranged. For the
receipt of the final snapshot data a special arrangement was made and the autocopy
lock procedure ensured that it will not be overwritten.
As mentioned already in section 2.1 above, this study had protocol amendments.
Each such amendment potentially implies changes in the data collection process and
consequently deviations from the originally designed CRF.
• Dealing with duplicate observations entered due to CRF pages entered twice.
After the protocol amendment a new CRF was created which included additional
pages for higher cycle numbers than the initially planned 17 cycles. Recording of
data for some patients already in the study continued on old CRFs, but in some
cases a new CRF form was filled out as well. In these situations we ended up
having the same data entered twice in the database. These data entry problems
have been identified and reported to the data management team. The database
The scope of data management work during the conduct of a clinical trial is usually
defined at the beginning of the study conduct once the CRF design phase has been
finished.
For this clinical trial the requirements for data cleaning were defined in Data Review
Plan established by the Study Management Team [3]. To ensure appropriate actions
are taken in response to the review, the Data Review Plan lays out the requirements
defining:
• What data are to be reviewed
• Who reviews what
• The timing and frequency of the review
• The methods to be used for the review
The Data Review Plan is one of the components of a more comprehensive plan, the
Data Quality Plan which for this study includes the following information:
• Data listings used for review
• Protocol Violation, Deviation and Eligibility Rules
• Second pass data entry procedures
As the trial proceeds, data quality must be considered and maintained. For this study
a comprehensive validation check program was used to verify the data that were
The set of data validation checks defined by the Data Management at the beginning
of the study included a set of generic data consistency checks such as checks for
Demographic, Vital Signs and Adverse Events data, as well as study-specific
validation checks.
There were also validation checks defined outside the Oracle Clinical Database
Management system Some of these additional checks were generic data quality
checks programmed in SAS on the Unix operating system. Others were study-
specific statistical analysis oriented checks.
It is worth noting again here that all the above mentioned checks were defined much
earlier than our biostatistics team took over the responsibility for the statistical
analysis and reporting for the upcoming reporting event.
It was realized that the number of validation checks programmed over the course of
the study conduct built up to the point where discrepancy management required
excessive amount of time.
The evaluation of the number of relevant validation checks, and consequently the
amount of data cleaning needed was based on the safety parameters that were part
of our report objects, such as adverse event (AE) summaries by frequency counts,
contingency tables comparing the pattern of change for lab parameters, listings with
change from baseline for vital signs etc. It was equally based on the efficacy
In the evaluation of data cleaning needed it was also important to evaluate the
resource availability and make appropriate decisions about resource allocation in
order to meet the set timelines.
4.2 Teamwork
I decided to use an Excel workbook as a base for the management of data issues
found using the validation checks programmed in SAS. It was a practical tool that
was owned by both parties: data management team in Basel, Switzerland and
biostatistics team in Dee Why, and as such was mutually maintained.
Important communication was achieved via regular teleconferences that have been
set up to resolve ongoing issues. These teleconferences were attended by wider
audience, including also clinical scientists providing us with important input for both
safety and efficacy data issues.
Though data review for quality should be an ongoing activity as data collection
proceeds, it is most important before database lock in time for reporting the results of
the clinical trial.
To meet the set deadlines it was important to correct the identified discrepancies in
timely manner. For the Data Manager it meant sending a list of errors back to the
person collecting the data for them to correct.
It happened also that for some reason the data could not be corrected. In that case
the data manager escalated the problem to the study management team which
decided what to do with the unresolved discrepancy. An approach to resolution of
such problems was to claim the discrepancy as irresolvable, properly document it
and leave it as it is. This implied that we in Statistics department had to make
The realistic study management goal of meeting the data quality specification for this
clinical trial was met seven weeks after the clinical cut-off date, and the database
was locked and deemed ready for statistical analyses.
5. Report summary
Quality of data collection and data management activities are of high importance for
the accuracy of conclusions drawn from clinical trials.
Data management tasks are usually labor intensive and time consuming. The set of
validation checks defined to run on a study data should be carefully selected in order
to target the data relevant for statistical reporting. For that reason the statistician’s
contribution to the specification of the Data Quality Plan and review of Data Quality
Checks is very important at the early stage of the clinical trial conduct.
Data cleaning should be balanced out carefully between data quality and resource
availability without compromising the precision of the scientific results.
Electronic data capture, that is an emerging technology for data collection and data
management will also introduce a shift in the definition of relevant validation checks
for assuring that data are suitable for statistical analysis.
There will be always issues originating from data collection or data management
activities. In resolving those issues, well established and constant communication
between the data management and biostatistics department is of utmost importance.
6. References
1. Statistical Analysis Plan for protocol MO16419, Roche Products Pty Ltd, 18th April 2006; 7
2. Therasse P, Arbuck SG, Eisenhauer EA et al, New guideines to Evaluate the Response to
Treatment in Solid Tumors, Journal of the national Cancer Institute, Vol 92, No 3, Feb 2000;
206-209
3. SMT Data Review Plan for protocol MO16419, Roche Products Pty Ltd, 28th October 2002;
4-10
Project Report II
5. REFERENCES ..............................................................................................................................19
This workplace project comprises of the exploratory analyses of data from an open-
label randomized phase II study of HT and X in combination (treatment A), versus H
plus T (treatment B), in patients with advanced and/or metastatic breast cancers that
overexpress HER2. Overall, 222 patients were enrolled from January 2002 to
September 2005 and received study drugs.
The exploratory analyses conducted as part of this workplace project have been
designed to complement (and extend) the analyses initially conducted to address the
primary and secondary objectives of the study. These analyses focus on logistic
regression of overall response and Cox regression on time to disease progression
with different potential prognostic factors explored in the models.
The primary analyses of the data on this clinical trial were conducted according to the
written requirements specification and in preparation for presenting the first efficacy
results at the European Society for Medical Oncology (ESMO) in September 2006.
The clinical cut-off date for the protocol-defined primary study analysis was 6 months
after the last patient enrolled. The Statistical Analysis Plan for this study [1] also
suggests conducting expanded analysis in an exploratory manner. These analyses
were to be done after the primary analyses if indicated by the data, and as such
deemed to be a good choice for applying my theoretical knowledge and skills learnt
during the coursework in the field of multivariate analysis.
After the analyses for the primary and secondary objectives of the clinical trial had
been completed, the aim of this project was to investigate the effects of various
possible baseline prognostic factors on the overall response rate and on the time to
disease progression (or tumor growth). [2].
The data were managed by Roche Data Management in Basel, Switzerland. The
data were collected by CRF, entered into a database and verified by a series of
This workplace project is based on the same data collected and cleaned for the
Project I of this Portfolio.
The primary and secondary efficacy endpoints of this clinical trial were the overall
response rate during the treatment period, time to disease progression and duration
of overall survival.
All efficacy analyses were done according to the intent-to-treat principle i.e as
randomized. All safety analyses related to adverse events and laboratory data were
done according to the treatment actually received. In this particular clinical trial these
two populations were in fact different (one patient randomized to A received
treatment B).
The results of the primary efficacy analyses 6 months after last patient enrolled
showed a high response rate in both treatment regimens, as shown in the following
table.
In investigating whether one treatment regimen is superior to the other the chi-square
test was used yielding a statistically non-significant result (p=0.72).
To test the null hypothesis that there is no difference between the treatment groups
in the probability of an event at any time point the logrank test was used. This test is
most likely to detect a difference between the treatment groups when the risk of an
event is consistently greater for one group than another [4]. The logrank test is purely
a hypothesis test. It cannot provide an estimate of the size of the difference between
the groups or a confidence interval.
From, the following Kaplan-Meier plot and the logrank test result for the equality of
the survivor functions we can see that treatment A showed longer time to progression
compared with treatment B reaching statistical significance (p=0.045).
The above results show that patients receiving treatment A are 70% of the risk of
having a disease progression than those receiving treatment B. The difference
proves to be marginally significant.
0 10 20 30 40 50
analysis time
(months)
group A group B
Overall survival was measured as the time from start of treatment to the date of
death, irrespective of the cause of death. Patients for whom no death was captured
on the clinical database were censored at the most recent date they were known to
Table 3. Summary of the Cox regression for overall survival (time to death)
The confidence interval of the hazard ratio for the overall survival includes the value
1. Therefore the inferred result of patients in group A being at 82% of the risk of dying
at any time point compared to group B is not a statistically significant result.
0 10 20 30 40 50
analysis time
(months)
group A group B
Having only reached 25% of events for the Overall survival time to event analysis, it
is too early to draw conclusions yet. Follow-up is ongoing and survival data will be
collected until 18 months after the last patient enrolled.
The analysis population for the expanded analysis was the intent-to-treat population,
which included all patients who were randomized and received study medication at
least once. Groups are defined according to the study arm patients were randomized
to.
The covariates chosen for the secondary analysis of data are fixed covariates known
at baseline or entry to the study.
With the aim of providing a prognostic model for the exploratory analysis of the
overall tumor response, the logistic regression was used.
The logistic regression model is based on the logit link function and is fit for binary or
ordinal response data by the method of maximum likelihood.
The overall response is the indicator variable with a value of 1 for a complete or
partial overall response, and 0 otherwise. I was modelling here the probability of
having an overall response.The model included nine (including treatment) variables
as risk factors thought to be related to the overall tumor response. These variables
are listed in table 4 above. In developing the logistic regression model, initially a
logistic regression was performed for each covariate separately adjusting only for the
possible treatment difference. The results of these regressions are shown in the
following table.
Wald test
Covariates df statistics p>Chi-square
Age 1 4.61 0.032
Number of target tumor sites 1 2.24 0.13
Adjuvant Anthracycline Therapy 1 1.81 0.18
Liver metastases 1 2.34 0.13
Lung metastases 1 1.88 0.17
ECOG performance status 1 2.09 0.15
Progesterone status 1 4.29 0.038
Disease status 1 1.14 0.29
Subsequently, in identifying the prognostic factors for the overall tumor response, the
stepwise multivariate logistic regression was used [3]. A significance level of 0.2 was
set to allow a variable into the model, and a significance level of 0.2 was set for a
variable to stay in the model. I chose these variable entry end removal levels to be
relatively large so that the selection of covariates into the model is not bound too
much by the usual 0.05 level of statistical significance.
The backward stepwise logistic regression procedure automates the removal of
Result of the likelihood ratio test for testing the joint significance of the explanatory
variables has a chi-square statistics of 17.67 with 6 df, yielding a statistically
significant result (p=0.0071).
For the evaluation of how well the model represents the data, the Hosmer and
Lemeshow goodness-of-fit test produced a value of 5.71 with df=8, indicating no
evidence of a poor fit (p=0.68).
For the above analysis 27 observations were deleted form the analysis dataset due
to missing values for the progesterone status variable. The disposition of the
progesterone status levels including missing values relative to the overall response
variable is illustrated in the following table.
Progesterone status
negative positive missing
responder 85 56 18
non-responder 41 13 9
Interpretation of results
Although the main analysis of the clinical trial data for the Overall tumor response did
not provide evidence of the difference in Overall tumor response between the two
treatment groups, exploratory analysis was conducted in order to look at possible
predictors of the Overall tumor response.
Of all the variables included in the model, age, number of target tumor sites and
disease status at baseline were significant at alpha=0.1 level, as shown in Table 6.
Table 8. shows the odds ratio estimates for the baseline predictors of the Overall
tumor response. We can see that the 95% confidence intervals for the odds ratios
for number of target lesions and disease status at baseline do not include 1 indicating
that the odds ratios are statistically significant at the 0.05 level. There are two other
factors, progesterone status and especially age category for which the odds ratio
confidence interval merely touches that borderline.
* Odds ratios >1 mean that patients in lower category have more chance of responding compared to
those in higher category
After this exploratory analysis, we see a trend backed by the odds ratio point
estimates that women under 50 years of age, having metastatic breast cancer with
only 1 target tumor site at baseline assessment and positive progesterone status
have better prognosis for overall tumor response.
One could ask a question (as I did in questioning whether the above given
interpretation is sensible in a clinical context): What is the reason that women with
positive progesterone receptor may respond better to the cancer treatment than
those with the negative hormone receptor? Usually negative test results mean
something good, but in this case the hormone receptor assay at baseline test shows
that if the receptors are present on the surface for the breast cancer cells, then the
cancer is likely to respond better to this particular type of therapy [5].
Cox regression model, the most commonly used multivariate approach for
analyzing survival time data in clinical research was used for the expanded analysis
of the secondary endpoint in this clinical trial.
The Cox regression model describes the relationship between the risk of an event, in
this case time to progression (TTP), and a set of covariates. It provides an estimate
of the hazard ratio and its confidence interval.
The assumption of the proportional hazards model is that the hazard of the event of
interest in one treatment group is a constant multiple of the hazard in the comparator
group. In this case the hazard ratio indicates the relative likelihood of progression
between the two treatment groups at any given point in time.
The log-log plot is presented in Figure 3. The curves do not cross and the separation
between the two curves (one for each treatment group) is fairly constant with not
much deviation. This indicates that the PH assumption holds for these data.
The aim of the analysis is to investigate a collection of factors of known relevance for
their ability to predict time to progression. The covariates considered for the analysis
are the same as for the primary efficacy endpoint described in section 3.1 of this
Report (see table 4.)
The strategy used is the attempt to initially model all covariates, and then
subsequently fit the reduced model, provided the predictive ability of the model is not
compromised.
Table 9. contains the individual logrank test results for the variables considered for
the model. Each factor is assessed through separate univariate Cox regression. P
According to the logrank test for equality of survivor functions, the variables that are
associated with shorter time to progression are the following baseline assessments:
number of target tumor sites, prior anthracycline therapy for breast cancer, liver
metastases, lung metastases, ECOG performance status and disease status.
Table 9. Logrank test results for the potential predictor factors
Covariates df p>Chi-square
Age 1 0.39
Number of target tumor sites 1 0.017
Adjuvant Anthracycline Therapy 1 0.0055
Liver metastases 1 0.031
Lung metastases 1 0.0015
ECOG performance status 1 0.0002
Progesterone status 1 0.55
Disease status 1 0.0006
After fitting the initial full model with all the covariates included the following results
were obtained:
* Hazard ratios >1 mean that patients in lower category have reduced risk of progression compared to
those in higher category
According to the partial likelihood ratio test for the overall significance (G=36.66,
df=9, p<0.0001), we reject the null hypothesis, and conclude that overall the model is
significant. This means that one or more covariates in this model are significant
Looking at the individual Wald tests given in the output, the categorical variables age,
prior adjuvant anthracycline therapy, baseline ECOG performance status and
disease status are the ones contributing to the model at a 0.1 significance level.
I fitted than the reduced model, and the test for the significance of the variables
removed from the model (G=3.53, df=5, p=0.62) yields the conclusion that number of
target sites, liver or lung metastases and progesterone status are not significant
predictors of time to progression.
Interpretation of results
The factors: prior anthracycline therapy for breast cancer, ECOG performance status
and disease status at baseline are all significant determinants of time to progression.
The following table presents the hazard ratios and their confidence interval of the
factors contributing to the reduced model.
* Hazard ratios >1 mean that patients in lower category have reduced risk of progression compared to
those in higher category
The hazard ratio derived from the Cox model must be interpreted with caution. It
does not translate directly into information about the duration of time until disease
progression. It may be used for purposes of statistical hypothesis testing and
indication of the amount of benefit (e.g. a decrease in odds of time to progression),
but other measures must also be applied to understand the full importance of the
study.
Median time to progression is also a useful parameter for interpreting effects [8]. The
following table lists the median time to progression in months for the prognostic
factors.
Though the age category variable is not a statistically significant predictor, from the
above table we can see that women in the age group under 50 progress later, with
median time to progression 15.2 months for younger and 12.9 months for older
women.
Women who did not have prior anthracycline therapy for breast cancer have longer
median time to progression: the median duration for women in this category is 18.6
months which is significantly different from 11.6 months for women having prior
anhracycline therapy.
The effect of ECOG performance status is the strongest predictor in the model
according to the Cox regression: the median time to progression is 17.3 months for
women with performance status 0 and 10.6 for women with performance status 1 or
more.
Disease status has the largest hazard ratio according to the Cox regression results
as reported in Table 11, but with a fairly large confidence interval. The median time to
progression for the locally advanced cancer patients at baseline assessment is 38.4
months compared to 12.9 months for the women with a metastatic breast cancer at
the start of the clinical trial.
As can be seen from comparing the hazard ratio estimates in Table 11 with the
median ratio estimates in Table 12, the hazard ratio estimates for our data are
consistent with median ratios, and hence they are a reliable measure of the
association of the prognostic factors with the time to disease progression.
To assess whether the AFT model is suitable for the data we can plot the percentiles
of the KM estimated survivor function from one treatment group against the other. If
the model is appropriate, this would approximate to a straight line through the origin.
According to the regression results, the slope of the fitted straight line is 1.57
providing an unadjusted estimate of the multiplicative factor of 0.97, indicating that
the time to progression probability for a patient in treatment group A at time t is equal
to the survival probability in group B at time 0.97 t.
However since the Quantile-Quantile (QQ) plot in Figure 3. does not approximate
well enough to a straight line from the origin, the AFT model might not be suitable for
these data, and hence the results obtained might not provide a reliable estimate of
treatment differences. For this reason, and also because of time constraints, I did not
consider exploring further this approach during my project work.
4. Report summary
Although the clinical trial was unsuccessful in providing statistically significant result
for the overall tumor response in favor of the experimental treatment regiment, it
showed a high (above 70%) overall tumor response rate in both treatment arms.
The proportional hazard model applied for the analysis of time to progression showed
the existence of association between shorter time to progression and age under 50,
no prior adjuvant anthracycline therapy, fully active ECOG performance status and
locally advanced breast cancer at baseline.
This randomized clinical trial will continue to collect follow up data on the same study
population, and it will be interesting to see the results of overall survival in order to
draw other possible conclusions.
1. Statistical Analysis Plan for protocol MO16419, Roche Products Pty Ltd, 18th April 2006;
16-17
2. Clinical Study Protocol, Protocol number MO16419D, Roche Products Pty Ltd, 26th April
2005; 75
3. SAS/STAT User’s Guide Version 8, Cary, NC: SAS Institute Inc., 1999, Volume 2; 1975-
1988
4. Bland MJ, Altman DG, The logrank test, BMJ 2004; 328(7447) 1073
5. Are hormone receptors present?, breastcancer.org,
http://www.breastcancer.org/dia_pict_hormone.html
6. Bradburn MJ, Clark TG, Love SB, Altman DG, Survival Analysis Part III: Multivariate data
analysis - choosing a model and assessing its adequacy and fit, British Journal of Cancer
(2003) 89: 605-611
7. Patel K, Kay R, Rowell L, Comparing proportional hazards and accelerated failure time
models: an application in influenza, Pharmaceutical Statistics, 2006; 5: 213-224
8. Squance SL, Reid JE, Grace M, Samore M, Hazard Ratio in Clincial Trials, Antimicrobial
Agents and Chemotherapy, Aug 2004, 2787-2792