Download presentation
Presentation is loading. Please wait.
1
Testing Dose-Response with Multivariate Ordinal Data Bernhard Klingenberg Asst. Prof. of Statistics Williams College, MA Paper available at www.williams.edu/~bklingen In Collaboration with Aldo Solari, Luigi Salmaso and Fortunato Pesarin, University of Padova
2
Outline Introduction Safety and Toxicity Data Notation and hypothesis of interest Stochastic Ordering Theorem (SMH IJD) Testing SMH Simple Test statistics Permutation Approach Step-down methods for indiv. endpoint significance Increase power Example Parallel, 5 dose group study with rats (8 rats per dose group) 25 Adverse Events from exposure to Perchlorethylene
3
Introduction: Safety and Toxicity Safety and Toxicity Data: To capture large number of possible manifestations of a dose (exposure) effect on safety or toxicity: Multiple Endpoints One such collection of endpoints to evaluate neurophysiological effects: Functional Observational Battery (FOB) Others: Drug Safety, Disease progression
4
Introduction: FOB Goal: Evaluation of neurophysiological effects to a toxin (Perchlorethylene) Data 1 : 2 groups (No exposure vs. 1.5g/kg exposure) 8 rats in each group Each evaluated at 25 endpoints (various effects), grouped into 6 domains Response ordinal, on a scale from 1 (no effect) to 4 (most severe reaction) 1 Moser (1986) Journal of the American College of Toxicology
5
Introduction: FOB
6
Introduction: Notation k-dimensional response vectors: ControlTreatment Random Sample ControlTreatment Hypothesis of interest: “No dose effect” d st
7
Introduction: No Toxicity “ ”: For all response sequences Control Treatment “ “: Stochastically larger 2 Control Treatment Note: Rejection of H 0 does not lead to H 1 d st 2 Marschall & Olkin (1979) Inequalities: Theory of Majorization and Its Applications
8
SMH Usually only interested if k margins are equal or not. I.e., for each adverse event, Def.: Simultaneous Marginal Homogeneity (SMH) 3 : Vector of marginal probabilities are equal under the two exposures, for all adverse events simultaneously 3 Agresti and Klingenberg (2005) JRSS C, Klingenberg and Agresti (2006), Biometrics
9
SMH SMH with just two adverse events Control Treatment 1234 1212 … 3 4 Lacrimation Arousal Lacrimation Arousal 1234 1212 … 3 4
10
SMH Theorem: Prior assumption plausible when dealing with adverse events data (increase in exposure shift towards higher outcome categories) IJD SMH Cumulative marginal inhomogeneity:
11
Testing SMH Consequence of Theorem: If prior assumption plausible, can use permutation approach to test hypothesis of SMH Test for SMH: Modeling approach via cumulative logits (proportional odds form) 4 Estimation (ML, conditional ML, GEE,…) computationally impossible, Asymptotics invalid 3 4 Han, Catalano, Senchaudhuri, Metha (2004) Exact Analysis of Dose Response for Multiple Correlated Binary Outcomes, Biometrics.
12
Testing SMH Let Simple test statistic: Standardized differences in marginal sample proportions given by (from multinomial assumption):
13
Testing SMH To take advantage of ordinal nature: Consider scoring function Let be score matrix Look at difference in mean scores: Estimate covariance matrix 4 under SMH assuming working independence
14
Testing SMH Test statistic for sparse data, ignoring correlation among adverse events: with This gives global test of safety/toxicity Permutation approach: 16!/(8!8!) = 12870 possible permutations, many leading to identical values of Advantage of permutation approach: Incorporates dependence by resampling entire vectors; exact significance levels
15
Testing SMH Example: Arousal Endpoint Computation with equally spaced scores: Note:
16
Testing SMH Permutation Distribution: observed Perm. Distr. Asympt. Distr.
17
Testing SMH Identifying which individual adverse events are significant leads to multiple hypotheses testing: Use test statistic (standardized mean score difference) for individual tests Multiplicity adjustments via step-down approach of Westfall &Young (1993), using distribution of maximum test statistic
18
Testing SMH Permutation Distribution: Observed maximum Perm. Distr.
19
Testing SMH
20
How sensitive are results to assigned scores? Consider the scores that maximize (obtainable via isotonic regression; data-driven) Appropriate for safety/toxicity data; maximizes the contrast btw. the mean score differences Equally spaced scores: With
21
Testing SMH
22
Testing Domains Domain effects? Some endpoints may measure similar effects Multiplicity adjustment at the endpoint level may be too conservative, leading to some false negatives Adjusted P-value for domain less than or equal to smallest adjusted P-value within domain “Proof”: Let endpoint h be in the first domain Dom 1 :
23
Testing Domains Important Consequence (Robustness Property): Consonant domain test statistic: Reject only (at domain level) if at least one endpoint within domain significant If no significant endpoint, domain also not significant For domain significance, it is irrelevant how many, potentially non-significant endpoints are grouped into a domain! * * Provided the same test statistic is used for all intersection hypotheses
24
Testing Domains Dissonant domain test statistic: Accumulate effects over endpoints within domain Even though no individual endpoint is significant, several marginally significant ones can result in significant domain P-value
25
Summary Testing dose-response for multivariate ordinal data Correlated ordinal responses (typical for toxicity or safety data) are often sparse and imbalanced use permutation approach Instead of modeling dose-response, we focused on testing SMH vs. stochastic ordering SMH IJD, but SMH IJD under Test statistic: z h = (difference in mean scores) / s.e., for each endpoint, assuming working independence s.e. derived from multinomial model and estimated under SMH Multiplicity adjusted P-values for each endpoint from Westfall and Young’s step-down procedure st
26
Conclusions for FOB Neuromuscular domain showed significant effect Domain P-value 0.003 with dissonant test Domain P-value 0.025 with consonant test Adverse events in neuromuscular domain that show increased toxicity at the 1.5 g/kg exposure level when compared to control Sensorimotor domain also shows increased level of toxicity, although no individual adverse event is significant Gait (p= 0.025) Hindelimb (p=0.044) Forelimb (p=0.096)
27
Did not show How methods extend to several dose levels Effect of discreteness (used mid P-values throughout) Combining P-values instead of test statistics, with functions other than the maximum Thank you and Go Gators
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.