Statistical Modeling:

1

• Statistical Modeling
• The (Fictitious) PAD Stent Study
• Ralph OBrien, PhD

2
Get .ppt .jmp files

• www.bio.ri.ccf.org/robrien/ BiostatLite06fall
• Get stuff in StatModels directory.

3
Read Brotman, Walker, Lauer, OBrien (2005)

• In Search ofFewer Independent Risk Factors

4
(No Transcript)
5
Activities

• Linear models for continuous responses.
• R2 as measure of fit.
• Correlation among predictors.
• Diagnostics to assess model adequacy.
• Linear models with many predictors, some
  continuous, some nominal (with multiple
  categories).
• Interpreting and testing the coefficients.
• Full vs. reduced models.

6
Activities

• Linear models for continuous responses.
• Model construction.
• Formulating a "correct" statistical model.
• All models are wrong, some are useful. -G. Box
• Stepwise methods rope to hang yourself
• Ineraction terms.

7
Peripheral Artery Disease (PAD)

• similar to coronary artery disease and carotid
  artery disease.
• fatty deposits build up in the inner linings of
  the artery walls.
• blockages restrict blood circulation, mainly in
  arteries leading to the kidneys, stomach, arms,
  legs and feet.
• often also fatty buildup in the arteries of the
  heart and brain.
• increased risk of death from heart attack and
  stroke.

8
Stent Treatment

• cylindrical, wire mesh tube placed in the
  narrowed artery with a catheter.
• expands and locks open.
• keeps diseased artery open.

9
Dataset

• Patient characteristics
• subject ID
• age in years
• gender
• severity of peripheral artery disease (PAD) of
  repair.
• Surgeon's assessment 1least severe to 4most
  severe.
• LDLoxi12 (contrived), a marker of LDL oxidation.
• Hypothesized to be risk factor for PAD, higher
  values indicating greater risk.

10
Dataset

• Intervention
• type of stent "CCF76X" vs. "ChemoStent IV
• selected by surgeon (non-random).
• CCF76X is the eXperimental stent. ChemoStent IV
  dominates the current market.

11
Dataset

• Outcome
• time in days to "failure" or to censoring.
• "failure" restinosis at site or related
  vascular problem or death by any cause)
• enrollment period of 2 years study period of 4
  years so follow-up is 2-4 years so lots of
  "benign" censoring

12
Dataset

• Number of patients 900

13
Unfocused Essential Question

• In general, what predicts "failure" rate?

14
Background Questions

• How is LDLoxi12 associated with age, gender, and
  PAD severity at time of surgery? (ordinary
  regression)
• Is the surgeon's choice of stent associated with
  these patient characteristics? If so, what is the
  nature of those associations? (logistic
  regression)

15
Background questions

• What are the survival curves for the two stent
  groups? (Kaplan-Meier plots)
• Are they "significantly" different? (log-rank,
  Wilcoxon)
• What does this comparison tell you (in this
  non-randomized design)?
• Does LDLoxi12 appear to be a risk factor in these
  patients? (Cox hazard modeling, quasi-likelihood
  ratio testing)
• If so, how much increase/decrease in risk is
  associated with a doubling of LDLoxi12?
  (interpreting Cox model coefficients)

16
Final (Primary) Question

• Our main goal is to answer
• Adjusting for patient characteristics, does
  the CCF76X stent have lower failure rates?

17
Dale Goodhands

• Adjusting for patientcharacteristics, does
• the CCF76X stenthave lower failure ratesthan
  the "ChemoStent IVstent?

from Edvard MunchThe Scream
18
The rest is all live and interactiveanalyses
with JMP.