Statistical Review

1
Statistical Review

• Chul H. Ahn, Ph.D.
• Biostatistician
• CDRH/FDA

2
Population and Sample

• Population - a collection of objects of interest
• Sample a subset of the population
• Want to know about the population using
  sample.
• Should be representative of the population.

3
BTT Study (original submission)

• Two-arm, non-randomized study (225 pts)
• Either the patient received the Novacor LVAS
  (Treatment) or was treated with medical
  management (Control group)
• 190 LVAS patients, and
• 35 Controls (28 retrospective, 7 concurrent)
  
• Primary endpoint - survival time, with transplant
  considered as censoring

4
Current Study

• Using the subset of data from BTT study
• - Seek to expand the indication to include
  the
• patients with relative contraindication
  (RC).
• Of the 225 patients, the sponsor found 87
  patients who are considered to have RC.
• Of the 87 patients, 75 are LVAS, and 12,
  Controls.
• The sponsor claims devices effectiveness based
  on these 87 patients.

5
75 RC
BTT Population
Population With Relative Contraindication
190 LVAS
87 RC
35 Con
12 RC
6
75 RC
BTT Population
Population With Relative Contraindication
190 LVAS
87 RC
35 Con
12 RC
7
Retrospective Subgroup Analysis

• Can the sponsor extend the result based on the
  87 patients to the entire population with
  relative contraindications?
• 4Problematic, because this subgroup may
• not be representative of the population
• with relative contraindications

8
75 LVAS vs. 12 CONTROLs

• Are they comparable?
• The year of implant
• Baseline covariates
• Propensity scores

9
Year of Implants
10
Differences in Baseline Covariates

• Factor Control LVAS
  p-value
• Milrinone 0.0 0.3
  0.000
• Pre. Antihyper. 83.3 33.3 0.001
• Age 41.8 49.9
  0.038
• Prev. TIA 16.7 1.3 0.047
• Dobutamine 12.7 9.1
  0.065
• Creatinine 2.5 1.9
  0.076
• Bleeding 16.7 2.7 0.084

11
Treatment Comparisons

• Two treatment groups are not comparable
• Imbalance of the year of implant
• Imbalance in multiple baseline covariates
• Any direct treatment comparisons on effectiveness
  endpoints are problematic
• So, all p-values from direct treatment
  comparisons are uninterpretable

12
Treatment Comparisons (cont.)

• What about treatment comparisons adjusting for
  imbalanced covariates?
• Traditional covariate analysis
• Propensity score analysis
• Example of adjustment for one covariate, e.g.,
  health condition

13
One Covariate Health Condition

• Suppose the event rate is higher in the control
  group where there are sicker patients.
• Then, the lower event rate in the treatment group
  may not be due to the treatment, but simply
  because there are healthier patients in the
  treatment group
• Should Compare patients with similar
• health condition!

14
What If There Are Many Covariates?

• One solution Replace the collection of
  covariates with one single number (propensity
  score)
• Age, Gender,
  Propensity
• Health condition,
  Score (PS)
• PS The conditional probability of receiving the
  LVAS, given a patients observed baseline
  covariate values, e.g., age, gender, prior
  cardiac surgery,

15
Propensity Scores

• When the propensity scores are balanced across
  the treatment and control groups, the
  distribution of all the covariates are balanced
  in expectation across the two groups
• So, we can use the propensity scores as a
  diagnostic tool to measure treatment group
  comparability
• The two treatment groups will be comparable if
  there is enough overlap between them.

16
Propensity Score Analysis

• Adjusted for all imbalanced and/or clinically
  important baseline covariates
• The propensity score model accurately predicted
  the treatment group membership
• However, the two treatment groups did not overlap
  enough to allow a sensible treatment comparison

17
Propensity Scores
57
(76)
6
3
5
6
7
(9)
3
18
Distributions of Propensity Scores
4
8
18
17
17
13
10
lt.749
lt.922
lt.983
lt.996
lt1.00
19
Propensity Score Analysis

• So, any treatment comparisons adjusting for
  imbalanced covariates are problematic

20
Sponsors Survival Curves for BTT
21
Sponsors Survival Curves for RC
22
Concerns with Sponsors Survival Curves

• Two treatment groups are not comparable
• Survival time is not independent of censoring
• p-values may not be interpretable
• 4If we assume that the two treatment groups were
  comparable, then we may also find a subgroup with
  significant difference in survival curves between
  two treatments

23
Survival Curves for Subgroup X
24
BTT Population
Population With RC
X
RC
RC
Y
25
Statistical Summary

• Subgroup with RC may not be representative of
  intended patient population
• Two groups (75 LVAS and 12 Controls) are not
  comparable so that any direct or covariate
  adjusted treatment comparison is problematic