This document summarizes a presentation on statistical analysis of clinical cancer trial data. It discusses key topics like pivotal phase III cancer trial designs, endpoints, sample size considerations, statistical plans, interim analyses, and special issues. Superiority, non-inferiority, and equivalence designs are covered. Sample size calculations require understanding trial objectives, hypotheses, variability, and the minimum important difference to detect. Statistical plans address randomization, data collection and cleaning, and final analyses. Interim analyses balance early stopping for efficacy or futility while controlling type 1 errors. Lead statisticians and data safety monitoring boards oversee analyses.

1. Statistical Analysis of Clinical Data
A Cancer Trials Perspective
Presented at the Indian Statistical Institute, Kolkata on 06.02.2007
Dr. Bhaswat S. Chakraborty
VP, R&D, Cadila Pharmaceuticals Ltd.

2. Contents
 Research and Regulations of Cancer Trials
 Pivotal Cancer Trials (Phase III)
 Efficacy end points
 Merits and demerits
 Optimum Study Designs
 Superiority
 Non-Inferiority and other designs
 Sample Size Considerations
 Scientific questions
 Basics of sample size calculation
 Statistical Plan for a Cancer RCT
 Statistical Analysis of Cancer Data
 Lead Statisticians & DSMB
 Special Issues
 Conclusion

3. Cancer Research Today
 Research is conducted mainly on
 New Drugs
 New Combinations
 Radiotherapy
 Surgery
 In the West, research is usually done by large co-operative groups, in
addition to those mentioned for India
 In India
 Large Pharmaceuticals
 Co-operative Groups, e.g., ICON (Indian Co-operative Oncology Network)
 Regional Cancer Centres & Govt. sponsored studies
 Academia

4. What does FDA Look for?
 FDA approves a drug application based on
 Substantial evidence of efficacy & safety from
“adequate and well-controlled investigations”
 A valid comparison to a control
 Quantitative assessment of the drug’s effect
 (21 CFR 314.126.)
 The design of cancer trials intended to support
drug approval is very important

5. Study Design: Approaches
 Randomised Controlled Trials (RCT) most preferred
approach
 Demonstrating superiority of the new therapy
 Other approaches
 Single arm studies (e.g., Phase II)
 e.g., when many complete responses were observed or
when toxicity was minimal or modest
 Equivalence Trials
 No Treatment or Placebo Control Studies
 Isolating Drug Effect in Combinations
 Studies for Radio- and Chemotherapy Protectants

6. Randomized Clinical Trials
 Gold standard in Phase III
 Single centre CT
 Primary and secondary indications
 Safety profile in patients
 Pharmacological / toxicological characteristics
 Multi-centre CT
 Confirmation of the above
 Effect size
 Site, care and demographic differences
 Epidemiological determination
 Complexity
 Far superior to meta-analyzed determination of effect

7. Non-Inferiority Trials
 New drug not less effective by a predefined
amount, the noninferiority (NI) margin
 NI margin cannot be larger than the effect of the
control drug in the new study
 If the new drug is inferior by more than the NI
margin, it would have no effect at all
 NI margin is some fraction of (e.g., 50 percent) of
the control drug effect

8. Placebo Control Equality Trials
 No anticancer drug treatment in the control arm is
unethical
 Sometimes acceptable
 E.g., in early stage cancer when standard practice is to give
no treatment
 Add-on design (also for adjuvants)
 all patients receive standard treatment plus either no
additional treatment or the experimental drug
 Placebos preferred to no-treatment controls because they
permit blinding
 Unless very low toxicity, blinding may not be feasible
because of a relatively high rate of recognizable toxicities

9. Drug or Therapy Combinations
 Use the add-on design
 Standard + Placebo
 Standard + Drug X
 Effects seen in early phases of development
 Establish the contribution of a drug to a standard
regimen
 Particularly if the combination is more effective
than any of the individual components

10. What to Measure?
 Time to event end points
 Survival
 Disease free survival
 Progress (of disease) free survival
 Objective response rates
 Complete
 Partial
 Stable disease
 Progressive disease
 Symptom end points
 Palliation
 QoL

11. Cancer Trials – End Points
Endpoint Evidence Assessment Some Advantages Some Disadvantages
Survival Clinical benefit • RCT needed • Direct measure of • Requires larger and
• Blinding not benefit longer studies
essential • Easily • Potentially affected by
measured crossover therapy
• Precisely • Does not capture
measured symptom benefit
• Includes noncancer
deaths
Disease-Free Surrogate for • RCT needed • Considered to • Not a validated
Survival accelerated • Blinding be clinical benefit survival surrogate in most
(DFS) approval or preferred by some settings
regular • Needs fewer • Subject to assessment
approval* patients and bias
shorter studies than • Various definitions
survival exist

12. Cancer Trials – End Points
Endpoint Evidence Assessment Some Advantages Some Disadvantages
Objective Surrogate for • Single-arm or • Can be assessed • Not a direct measure of
Response accelerated randomized in single-arm benefit
Rate (ORR) approval or studies can be studies • Usually reflects drug
regular used activity in a minority of
approval* • Blinding patients
preferred in • Data are moderately
comparative complex compared to
studies survival
Complete Surrogate for • Single-arm or • Durable CRs • Few drugs produce high
Response accelerated randomized represent obvious rates of CR
(CR) approval or studies can be benefit in some • Data are moderately
regular used settings (see text) complex compared to
approval* • Blinding • Can be assessed survival
preferred in in single-arm
comparative studies
studies

13. Design Concepts
Difference in Clinical Efficacy (Є)
Non-Inferiority
Superiority
+δ
0
Equivalence
-δ
Inferiority
Non-Superiority
Equality δ = Meaningful Difference

14. Phase III Cancer Trials
New Drug (or Regimen) is
Compared with a Standard
90
80 New
70 Standard
60
50
40
30
20
10
0
 Superiority Trials Survival DFS QoL

15. Phase III Cancer Trials
40
35 New
30 Standard
25
20
15
10
5
0
Survival DFS QoL
 Non-Inferiority or Equivalence Trials

16. Understanding Basics
 μ0 and μA
 Means under Null & Alternate Hypotheses
 σ02 and σA2
 Variances under Null & Alternate Hypotheses (may be the same)
 N0 and NA
 Sample Sizes in two groups (may be the same)
 H0: Null Hypothesis
 μ0 – μA = 0
 HA: Alternate Hypothesis
 μ0 – μA = δ
 Type I Error (α): False +ve
 Probability of rejecting a true H0
 Type II Error (β): False –ve
 Probability of rejecting a true HA
 Power (1-β): True +ve
 Probability of accepting a true HA

17. Basics of Sample Size Calculation
 Answer the scientific questions for the Trial size
 Understand the distribution and variability of the data
 Construct correct Null and Alternate hypotheses
 From the hypotheses derive formula for sample size
 Also make sure that this size trial has adequate power
to establish a true alternate

18. Five Key Questions
1. What is the main purpose of the trial?
2. What is the principal measure of patient outcome?
3. How will the data be analysed to detect a treatment
difference?
4. What type of results does one anticipate with standard
treatment?
5. How small a treatment difference is it important to detect
and with what degree of certainty?
 Answers to all of the five questions above enable us to
calculate the sample size and analyze the data with most
appropriate test of hypothesis.
Pocock SJ: Clinical Trials: A Practical Approach Chichester: Wiley; 1983

19. Start
Planning Reliable or historical
data available? No
Yes Use conventional
methods for analysis
Use bootstrap simulation for
sample size
Normally distributed
continuous data? Summary
Yes
measure: mean & mean
μT – μC difference
∆normal = Use parametric methods of
σ analysis, two sample ‘t’ or
ANOVA
Effect Size
No
2 [Z1-α/2 + Z1-β/2]2
Use non-parametric methods
Nnormal =
of analysis, Mann-Whitney U
∆2normal or Proportional Odds Model
Sample Size

20. Understanding Sample Size Determination
H0: μ0 – μA = 0 HA: μ0 – μA = δ
Critical Value
S.Error =σ(√2/N) S.Error =σ(√2/N)
Power = 1-β
β
α/2 α/2
0 δ
X0–XA
0+Z1-α/2σ√(2/N) δ–Z1-βσ√(2/N)

21. From the Previous Graph, We have
0+Z1-α/2σ√(2/N) = δ–Z1-βσ√(2/N)
Upon simplification,
2 [Z1-α/2 + Z1-β/2]2
Nnormal =
∆2normal

22. Sample Size: 2-Sample, Parallel
Superiority/Non-Inferiority Trial
(z+zβ)2 (p1 (1– p1) + p2(1 – p2))
N in each arm =
(Є – δ)2

23. Power: 2-Sample, Parallel Superiority/Non-
Inferiority Trial
Є –δ
Φ – z
√p1 (1– p1)/n1 + p2(1 – p2)/n2)
Where p1 and p2 are true mean response rates
from Test & Control & Φ is the cumulative
standard normal distribution function

24. Statistical Plan
 Primary outcome considerations
 Study Design
 Sample size calculation
 Randomization
 Statistical consideration in Inclusion/Exclusion criteria
(Homogeneity within centre and strata)
 Accrual of patients
 Cleaning of data
 Interim Analysis
 Go/No go criteria
 α Considerations
 Final analysis
 Final conclusions

25. Statistical Designs
Crossover Trials
A A
Baseline
B B
Parallel Arm Trials
A
Baseline
B

26. Accrual of Patients
 Study of the statistical trends in accrual patterns
 Seasonal
 Planned approaches
 Reasons for drop outs and loss to follow up
 Motivational factors
 Monitoring of recruitment progress and strategies
 Frequency
 Parameters
 Duration
 Understanding natural history and non-cancer, non-intervention deaths
 Changes in accrual after Interim Analysis

27. Randomisation
 Generation of randomisation scheme according to
 Centre
 Block
 Strata
 Patient
 Investigational Product to be given
 Measures of ensuring non-bias
 Allocations
 What should go on the labels
 Primary, secondary, tertiary packaging

28. Blinding
 Often difficult in oncology trials
 Test and control are of different characteristics
 Different routes of administration
 Different schedules
 New low toxicity oral treatments are relatively easy to
blind
 In other cases the end-point evaluating investigator
must be different from the one administering the drugs

29. Data Capture
 Manual or Electronic
 CRF is the main source of raw data capture
 Data must be quality assured
 Integrity, accountability, traceability
 Data must be validated
 All production and/or quality system software, purchased or
developed in-house
 Should document
 Intended use, and information against which testing results and other
evidence can be compared
 To show that the software is validated for its intended use

30. Data Cleaning & Locking
 Data are cleaned based on a good plan for interim or final analysis
 E.g.,
 Hundred percent data are made quality checked and assured
 Eligibility criteria for data selection
 Correction and editing
 Double data entry or other methods of data integrity
 Data will be locked after cleaning the data and resolving all the
queries
 SOP for data locking
 No change after locking
 Only locked data are used as input into data analysis program

31. Interim Analysis of Data
0.05
0.045
0.04
0.035
0.03
0.025
Nominal p
0.02
0.015
0.01
0.005
0
1 2 3 4
Looks

32. Interim Analysis of Data
2 Looks 3 Looks
How many times
0.05
0.05
can you look into
Nominal Pvalue
Nominal Pvalue
the data?
0.03
0.03
o pocock
o ob+fle
0.01
0.01
o fle+har+ob
0.0
0.0
1 2 1 2 3
Look Look
4 Looks 5 Looks
0.05
0.05
Type 1 error at kth
Nominal Pvalue
Nominal Pvalue
test is NOT the
0.03
0.03
same as the
0.01
0.01
nominal p value
0.0
0.0
for the kth test 1 2 3 4 1 2 3 4 5
Look Look

33. Considerations for IA
 Stopping rules for significant efficacy
 Stopping rules for futility
 Measures taken to minimize bias
 A procedure/method for preparation of data for analysis
 Data has to be centrally pooled, cleaned and locked
 Data analysis - blinded or unblinded?
 To whom the interim results will be submitted?
 DSMB
 Expert Steering Group
 What is the scope of recommendations from IA results?
 Safety? Efficacy? Both? Futility? Sample size readjustment
for borderline results?

34. Final Analysis and Conclusion
 Clinically meaningful margins must be well defined in Control trials
prospectively
 Superiority and non-inferiority margins must not be confused
 Two or one-sidedness of α should also be prospectively defined
 Power must be adequate
 Variance must be analysed using the right model
 Strategy for dealing with multiple end points must be prespecified
 Too many end points ot tests will increase the false positive (α) error
 Sometimes (e.g., in equality trials) statistically significant results may not
be medically significant
 Data censoring or skewed data
 E.g., time to event data

35. Equality Designs
(e.g., 2-Sample, Parallel)
H0 : Є = 0 HA : Є ≠ 0
Reject H0 when
p1— p2
ˆ ˆ
> z/2
√p1(1—p1)/n1 + p2(1— p2)/n2
ˆ ˆ ˆ ˆ
Where p1— p2 are true mean response rates
ˆ ˆ
from Test & Control

36. Superiority/Non-Inferiority Designs
(e.g., 2-Sample, Parallel)
H0 : Є ≤ δ HA : Є > δ … Superiority
H0 : Є ≥ δ HA : Є < δ … Non-Inferiority
Reject H0 when
p1— p2 – δ
ˆ ˆ
> z
√p1(1—p1)/n1 + p2(1— p2)/n2
ˆ ˆ ˆ ˆ

37. Survival Data – The Kaplan-Meier Estimator
1.0
Survival Percentage
0.75
0.50
0.25
~40% Patient
will survive
beyond 0.8 year
0.0
Time (Year)
0.0 0.2 0.4 0.6 0.8 1.0

38. Intent-to-Treat Principle
 All randomized patients
 Exclusions on prespecified baseline criteria permissible
 also known as Modified Intent-to-Treat
 Confusion regarding intent-to-treat population: define and agree
upon in advance based upon desired indication
 Advantages:
 Comparison protected by randomization
 Guards against bias when dropping out is realted to outcome
 Can be interpreted as comparison of two strategies
 Failure to take drug is informative
 Refects the way treatments will perform in population
 Concerns:
 “Difference detecting ability”

39. Per Protocol Analyses
 Focuses on the outcome data
 Addresses what happens to patients who remain on
therapy
 Typically excludes patients with missing or
problematic data
 Statistical concerns:
 Selection bias
 Bias difficult to assess

40. Intent to Treat & Per Protocol Analyses
 Both types of analyses are important for approval
 Results should be logically consistent
 Design protocol and monitor trial to minimize
exclusions
 Substantial missing data and poor drug compliance
weaken trial’s ability to demonstrate efficacy

41. Missing Data
 Protocol should specify preferred method for
dealing with missing primary endpoint
 ITT
 e.g., treat missing as failures
 e.g., assign outcome based on blinded case-by-case
review
 Per Protocol
 e.g., exclusion of patients with missing endpoint

42. Data Safety and Monitoring Board (DSMB)
 All trials may not need a DSMB
 DSMB Membership
 Medical Oncologist, Biostatistician and Ethicist
 Statistical expertise is a key constituent of a DSMB
 Three Critical Issues
 Risk to participants
 Practicality of periodic review of a trial
 Scientific validity of the trial

43. Practicality of a DSMB
 If the trial is likely to be completed quickly, the DMC
may not have an opportunity to have a meaningful
impact
 In short-term trials with important safety concerns,
however, a DMC may still be valuable

44. Outstanding Issues
 Problems in ‘discounting’ (raising the hurdle to declare non-
inferiority in order to account for inherent weaknesses) when
indirectly comparing an experimental treatment to a placebo.
Confronting weaknesses due to lack of assay sensitivity and
constancy.
 Selection of the primary analysis population (ITT population vs. per
protocol population). Inappropriate selection of analysis population
will lead to biased results.
 The impact of trial quality on the results of active controlled trials –
what the regulator will want to see substantiated on trial quality.
 Ethical issues if placebo controlled trials are used vs. statistical
issues if they are not.

45. Outstanding Issues
 Selecting and using surrogate endpoints – the determination of
statistical significance.
 Problems of designing clinical trials involving combination
therapies. Statistical analysis and interpretation of complex data.
 Issues with the choice of control group in dose-finding studies –
which doses to test and whether to include a placebo dose.
 Statistical inference with non-randomized controls
- Propensity scores
- Paired availability design for historical controls
 Statistical inference with randomized controls
- Adjusting for non-compliance
- Surrogate endpoints/missing outcome

46. Conclusions
 Clinical testing of new Oncology products is very sophisticated
and complex
 A Statistician’s role in Cancer trials is invaluable
 Statistical considerations must be thoroughly given attention and
built in while planning the study design and calculating the
sample size
 Cancer clinical data is very complex (censored, skewed, often
fraught with missing data point), therefore, proper
hypothesization and statistical treatment of data are required
 Prospective RCTs are usually the preferred approach for
evaluation of new therapies

47. Conclusions
 Survival as primary end point is preferred by regulatory agencies
 Randomisation and blinding offer a robust way to remove bias in end-point
estimations
 Data must be accurately captured without any bias and analysed by
prospectively described methods
 Interim analysis should carefully plan ‘ spending’ function
 Final analysis should be done carefully, independently and meaningfully
(medical as well as scientific)
 Choose clinically relevant delta
 Design, conduct, and monitor trials to minimize missing data and poor compliance
to drug
 Analysis
 Both intent-to-treat and per protocol analyses should be conducted
 Sensitivity analyses
 There are many oustanding statistical issues in Cancer trials that need no be
discussed and solved

48. Acknowledgements
 Dr. Nikunj Patel
 Dr. Sumit Goyal
 Dr. Manish Harsh
 Dr. Nilesh Patel
 Ms. Darshini Shah
Thank You Very Much