Assessing Surrogacy in CKD*

1
Assessing Surrogacy in CKD
Tom Greene, Ph.D. University of Utah
Joint work with Marshal Joffe, Liang Li,
Andrew Levey, and Lesley Stevens
2
Concept of Surrogate Endpoint

• The most relevant clinical outcome may be
  difficult to use in an RCT because
• measurement is costly or invasive
• long follow-up required
• large N required
• A surrogate endpoint is an alternative outcome
  that is substituted for the true clinical outcome
  in an attempt to reduce these difficulties

3
Difficult to Evaluate ESRD in Trials of Early CKD
Window for feasible RCTs with renal endpoints
GFR (ml/min/1.73m2)
ESRD
14
0
12
10
8
6
4
2
Years from ESRD
4
Past Problems with Surrogate Endpoints
Disease Intervention Surrogate Clinical outcome
Arrythmia Encainide, Flecainide Suppressed arrhythmias Increased mortality
CHF Milrinone Improved cardiac output Increased mortality
Osteoperosis Sodium Fluoride Increased bone mineral density More Fractures
5
Failures of Mean Change in GFR as Surrogate for
Renal Endpoints
Change in GFR from Baseline
Incidence of ESRD or Death
8
Amlodipine
40
Amlodipine
Ramipril
Ramipril
4
HR 0.51, p lt 0.001
30
0
Mean (SE) ? GFR (ml/min/1.73m2)
20
- 4
10
- 8
0
-12
0
12
24
36
48
60
0
6
12
24
36
48
Follow-up Month
Follow-up Month
6
Endpoints for Progression of CKD

• Target Clinical Endpoint
• Time-to-ESRD ( Death)
• Time-to-Event Surrogate Endpoints
• 50 reduction in GFR or ESRD ( Death) ?
• Doubling of serum creatinine (SCR) or ESRD
• ( Death) ?
• Slope-Based Surrogate Endpoints
• Slope of GFR vs. time ??
• Slope of estimated GFR (from SCR) ??
• Proteinuria Based Surrogate Endpoints ???

7
Statistical Approaches to Validation of Surrogate
Endpoints
Can be applied in singe study but subject to
confounding

• Individual level association
• Prentice criterion variations
• Trial-level association
• Uses randomized comparisons to determine if
  treatment effect on UP predicts the treatment
  effect on the clinical endpoint

Truly based on randomized inference but usually
requires analyses of multiple RCTs effect
modification
8
Statistical Approaches to Validating Proteinuria
(UP) as Surrogate Endpoint

• Evaluate individual level association of UP with
  the clinical endpoint
• Relationships of initial UP and ?UP with
  progression endpoints for individual patients
• Evaluated within treatment groups in RCTs

9
Association of Baseline Albuminuria with
progression in RENAAL
Renal endpoint is composite of doubling of SCR,
ESRD, or Death
10
Early Decline in UP Predicts Better Renal
Outcome
RCTs showing Individual-level association of
baseline UP ? UP with clinical endpoints

• CSG Trial treatment with captopril
• MDRD Trial blood pressure lowering
• REIN Trial treatment with ramipril
• AIPRD - 11 trials, treatment with ACEi
• IDNT Trial treatment with irbesartan
• RENAAL Trial treatment with losartan
• AASK Trial treatment with blood pressure
  lowering, ramipril, metoprolol, or amlodipine

11
Limits of Individual Level Association
(Fundamental Limitation of Causal Inference)
T(1) True endpoint if assigned to
treatment T(0) True endpoint if assigned
to control T(1) T(0) Causal effect of
treatment on true endpoint
UP(1) UP given treatment UP(0) UP given
control UP(1) UP(0) Causal effect of
treatment on UP
Treatment effect on UP predicts treatment effect
on T if UP(1) UP(0) accurately predicts T(1)
T(0)
UP(1) UP(0) or T(1) T(0)
But cannot observe either
12
Implicit Causal Model of Individual Association
Approach
Treatment
UP
T
13
Implicit Causal Model of Individual Association
Approach

• Key implicit assumptions of individual
  association approach
• UP on causal pathway between treatment and T
• No common causes of UP and T other than the
  treatment
• No direct effect of treatment on T

14
Prentice Criterion (1989)
Under regularity assumptions,
H0 No effect of the Treatment on the Surrogate
is equivalent to H0 No
effect of the Treatment on T
if and only if

• P The true clinical endpoint is unrelated to the
  treatment after controlling for surrogate

15
Variations on Prentice Criterion

• Related index (Freedman)
• Proportion of the treatment effect on the true
  clinical endpoint explained by surrogate (PTE)

Treatment effect after controlling for surrogate
1 -
PTE
Treatment effect not controlling for surrogate
16
PTE for ACE/ARB Interventions
Study Intervention Index Outcome PTE
AIRPD (11 non-diabetic kidney disease studies) ACE ? UP ESRD alone Doubling SCR 18 26
AASK (hypertensive nephro.) Ramirpril vs. Amlodipine ? UP ESRD or Death 41
RENAAL (Type II diabetics) Losartan vs. Placebo ? ALB ESRD, SCR, Death ESRD alone 89 51
IDNT (Type II diabetics) Ibesartan vs. Placebo ? ALB ESRD alone 36
17
Implied Causal Model of Freedman/Prentice
Approach

• Key Assumption of Implicit Freedman/Prentice
  Model
• UP on the causal pathway between A and T
• No common causes of UP and T other than the
  treatment

18
Limited Extension of Prentice/Freedman Approach
Attempt to control confounding
Measured Covariates
Treatment
UP
T

• Key Assumptions of Expanded Framework
• UP on the causal pathway between A and T
• No unmeasured confounders of UP and T
• Can then estimate the causal direct effect and
  the indirect effect which is mediated by UP, and
  a properly defined PTE

19
Fundamental Limitation of Single-Trial Validation
Study

• No true assessment of generalizability
• Need to look at variations in treatment and in
  study population

20
Statistical Approaches to Validating a Surrogate
Endpoint

• Trial-Level Approach
• Relate treatment effects on true endpoint to
  treatment effects on surrogate
• Treatment effects on both UP and T estimated from
  randomized comparisons
• Usually requires joint analysis of multiple
  studies

Daniels and Hughes, Stat Med 1997
Molenberghs et al, 2000 2005
21
Trial-Level Approach Ideal Hypothetical Example
?
?
?
?
?
Treatment effect on ESRD (Log RR)
?
?
?
?
?
?
?
?
Treatment effect on ? UP
Points represent estimated treatment effects in
different RCTs
22
Trial-Level Approach Ideal Hypothetical Example
Points represent estimated treatment effects in
different RCTs
23
Trial-Level Approach Real Example
Burzykowski, JRSS A, 2004
24
Incorporation of Trial Level Can Obtain Direct
Indirect Effect In Broader Framework

• Trial level approach formulated under causal
  association framework just relates treatment
  effects on T to treatment effects on UP.
• No need to assume UP is directly on causal
  pathway.
• Under intermediate variable model, trial level
  approach estimates direct and indirect effects
  associated with UP even in presence of
  uncontrolled confounders

25
Limits of Trial Level Approach

• Logic for extrapolation to new studies works best
  if new study is similar to prior studies (e.g.,
  best for me-too studies)

26
Limits of Trial Level Approach

• Requires significant heterogeneity between
  studies in true treatment effects on UP
• Requires effect modification
• Opposite of typical situation in meta-analysis

27
Consistent Effects on both UP and Clinical
Endpoint Encouraging but Not Convincing from
Perspective of Trial-Based Approach
28
Looking for Heterogeneity

• Can look for effect modification between and
  within studies
• Between interventions
• ACE vs. Control
• CCB vs. Control
• Low vs. Usual BP
• Low vs. Usual Protein
• Immunosuppressive therapies
• Between patient subgroups
• Diabetics, Non-diabetics, Transplant recipients
• Low UP, High UP
• PKD, Glomerular disease, Hypertensive kidney
  disease, Interstitial kidney disease
• Younger, Older
• Black, white

29
Technical Difficulties for Trial-Level Approach
in CKD

• Huge variation in sample sizes between studies
• Observed treatment effects on clinical endpoints
  depend in part of arbitrary study characteristics
  e.g., distribution of GFR at entry, length of
  follow-up
• Variation in measurement of UP
• Unclear what is best index of ?UP (e.g., absolute
  change or change)

30
CKD-EPI Project

• Collaboration to analyze large databases of
  pooled individual-patient data, including most
  major RCTs of CKD-patients
• Assess validity of change in proteinuria as a
  surrogate marker using all three statistical
  approaches
• Lesley Stevens to later present
• very preliminary results for ACE/ARB studies

31
Potential Uses of Validated Surrogate Endpoints

• Early phase of development of new interventions
• Exploratory subgroup analyses
• Extension of established findings to related
  patient populations with less severe disease
• Extension of established findings to related
  interventions
• Establish benefit of new interventions

Less Risky
More Risky
32
Conclusions

• Use of surrogate endpoints is necessary component
  of clinical research
• Statistical formalisms for addressing validity of
  surrogate endpoints are still developing
• Two general statistical approaches
• Estimate direct indirect effects under models
  that attempt to control for all confounding
  between UP T
• Try to take advantage of effect modification to
  determine if treatment effects on UP predict
  treatment effects on T

33
Conclusions

• Biological evidence also very important
• Use of surrogates can be context specific
• Dont use BP for testing a lipid lowering drug
• All uses of surrogate endpoints entail
  extrapolation beyond the data