CCEB Modeling Quality of Life Data with Missing Values Andrea B. Troxel, Sc.D. Assistant Professor of Biostatistics Center for Clinical Epidemiology and.

CCEB

Modeling Quality of Life Datawith Missing Values

Andrea B. Troxel, Sc.D.Assistant Professor of

BiostatisticsCenter for Clinical Epidemiology

and BiostatisticsUniversity of Pennsylvania

School of Medicine

CCEB

Outline

• Why measure QOL in oncology?

• Types of missing data

• Possible modeling approaches

• Example: SWOG study of QOL in colorectal cancer

CCEB

QOL in Oncology

• Potentially debilitating effects of treatment

• Tradeoff between quantity and quality of life

• An increasingly chronic disease

• Important focus on survivorship

• Longitudinal measurements

CCEB

Missing Data - Examples

• Subject moves out of town• Researcher forgets to administer

questionnaire• Subject returns incomplete

questionnaire• Subject’s family refuses questionnaire• Subject is too sick to fill out

questionnaire• Subject dies

CCEB

Missing Data - Definitions

• Missing completely at random

• Missing at random

• Nonignorable

CCEB

Modeling Approaches

• Complete case approaches• Models for MAR data• Models for NI data• Sensitivity analyses• Extensions of failure-time

models• Imputation methods

CCEB

Models for MAR data

• Generalized linear models

• Generalized estimating equations

• Weighted methods

CCEB

Models for NI data

• Fully parametric models–Directly model the missingness

mechanism

–Estimate a nonignorability parameter

–Computationally difficult

–Untestable assumptions

CCEB

Sensitivity Analyses

• Vary aspects of model and determine effects on inference

• Local sensitivity analysis– ISNI (Troxel, Ma, and Heitjan, 2005)

–Assess sensitivity in the neighborhood of the MAR assumption

–Easy to compute and interpret

CCEB

Failure-time Models

• Take advantage of bivariate survival methods• Integrate clinical and QOL

data• Avoid primacy of one outcome

over the other• Partially handle missing data

due to death

CCEB

Multiple Imputation

• Use an appropriate method to create a series of “complete” data sets

• Use any appropriate method of analysis on each data set

• Combine the analyses to achieve one reportable result

CCEB

SWOG 9045

• Companion study to SWOG 8905– 599 subjects with advanced colorectal

cancer

–Seven arms (!) assessing effectiveness of 5-FU

CCEB

SWOG 8905

• Variations in–Route of administration

» Bolus injection (arms 1-3)

» Protracted 28-day continuous infusion (arms 4-5)

» Four weekly 24-hour infusions (arms 6-7)

–Biochemical modulation» None (arms 1, 4, 6)

» Low dose leucovorin (arms 2, 5)

» High dose leucovorin (arm 3)

» PALA (arm 7)

CCEB

SWOG 9045

• Five primary outcomes–Mouth pain–Diarrhea–Hand/foot sensitivity–Emotional functioning (SF-36)–Physical functioning (SF-36)

• Secondary outcome–Symptom distress scale (high scores = more distress)

CCEB

SWOG 9045

• 4 assessments–Randomization– 6 weeks– 11 weeks– 21 weeks

• 287 patients registered• 272 (95%) submitted baseline

questionnaire

CCEB

QOL Submission Rates

Week

0 6 11 21n 272 230 207 182

% of total 95 80 72 63

% of 272 100 83 76 65

% of alive 100 85 79 78

CCEB

Missing Data Patterns

and Reasons

11

16

21

26

0 6 11 21

Assessment Time (weeks)

SD

S

lost - death

lost - illness

lost - other

completefollow-up

CCEB

Submission Rates

• Restrict analysis to subjects who survived for 21 weeks

• N=227

Week

0 6 11 21

N 227 197 187 172% 100 87 82 76

CCEB

Missing Data Patterns

Time Pattern ( 1=submitted, 0=missing) Total

0 1 1 1 1 1 1 1 1 2276 1 1 1 1 0 0 0 0 197

11 1 1 0 0 1 1 0 0 187

21 1 0 1 0 1 0 1 0 172

n 150 26 8 13 9 2 5 14 227

CCEB

Models - SDS

• Normal GLM–Complete cases

–All available data, unweighted

–All available data, weighted

• NI model–Normal component for SDS data

– Logistic model for missingness probs.

0 1logit 0 1, 2,3it t t itP R Y t

CCEB

Results - SDS

12

14

16

18

20

22

24

0 6 11 21

Time (weeks)

SD

S s

core

NI

Wtd MAR

MAR

CC

CCEB

Sensitivity Analysis

• Assess sensitivity to nonignorability in the neighborhood of the MAR model

• Sensitivity of parameters depends on how the model is parameterized

CCEB

Sensitivity - SDS

Estimate SE ISNI*

T0(single) 17.0 .51 14.29

T6(single) 17.4 .53 1.24

T11(single) 17.3 .56 0.87

T21(single) 18.1 .59 0.73

T0(comb) 18.5 .57 4.26

T6(comb) 19.0 .60 1.10

T11(comb) 18.8 .62 1.21

T21(comb) 19.6 .64 1.02

CCEB

Frailty Model - SDS

• SDS>24 SDS “event”• Jointly assess survival and

SDS events• Estimate correlation• Estimate covariate effects• No special programming

required

CCEB

Frailty Model – SDS

• No significant effect of combination therapy

• Frailty variance estimated to be 0.54

• 95%CI (0.28, 0.92)

• Significant random subject effect (p < .0001)

CCEB

Models – Hand/Foot Sensitivity

0 1 2 3 4logit 6 11 21it iE Y I t I t I t X

• Yit is a binary indicator of bothersome or worse symptoms

• Xi is an indicator of continuous infusion vs bolus injection (arms 4,5 vs arms 1-3)

• N=154 (arms 1-5, alive for 21 weeks)

CCEB

Results – Hand/Foot Sensitivity

0

5

10

15

20

25

30

0 6 11 21

Time (weeks)

Est

imat

ed %

CC

Unwtd GEE

Wtd GEE

CCEB

Models – Hand/Foot Sensitivity

• Treatment effect OR estimates–CC: 3.1 (1.4 – 7.0)

–MAR: 2.5 (1.2 – 5.3)

–Wtd MAR: 2.5 (1.2 – 4.8)

CCEB

Conclusions

• Missing data is a pervasive problem• Standard approaches can lead

to misleading inferences• Sensitivity analysis is a key

component• Certain comparisons are more

susceptible than others