See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/233994854

Behavioral Support Intervention for Uncontrolled Hypertension A Complier

Average Causal Effect (CACE) Analysis

Article in Medical Care · December 2012

DOI: 10.1097/MLR.0b013e31827da928 · Source: PubMed

CITATIONS READS

6 508

4 authors, including:

Yuanyuan Liang Christopher S Hollenbeak

123 PUBLICATIONS 4,554 CITATIONS

Pennsylvania State University
447 PUBLICATIONS 11,344 CITATIONS
SEE PROFILE
SEE PROFILE

Barbara J Turner
Keck School of Medicine USC
280 PUBLICATIONS 13,565 CITATIONS

SEE PROFILE

All content following this page was uploaded by Christopher S Hollenbeak on 11 October 2017.

The user has requested enhancement of the downloaded file.

 APPLIED METHODS
Behavioral Support Intervention for Uncontrolled
Hypertension
A Complier Average Causal Effect (CACE) Analysis
Yuanyuan Liang, PhD,*wzy8 Benjamin R. Ehler, MS,* Christopher S. Hollenbeak, PhD,z
and Barbara J. Turner, MD, MSEDwy#**
Background: The complier average causal effect (CACE) analysis
addresses noncompliance with intervention and missing end-point
measures in randomized controlled trials.
Objectives: To conduct a CACE analysis for the Peer Coach and
Office Staff Support Trial examining the intervention’s effect
among “compliers,” defined as subjects who would have received
an effective dose of the intervention had it been offered, and to
compare with an intention-to-treat analysis.
Research Design and Subjects: A randomized controlled trial of
280 African American patients aged 40–75 with sustained uncon-
trolled hypertension from 2 general internal medicine practices.
Measures: Change in 4-year coronary heart disease (CHD) risk
(primary) and in systolic blood pressure (SBP) (secondary) from the
baseline to the end of the 6-month intervention.
Results: Of 136 intervention subjects, 68% were compliers who had
significantly more end points measured (86% vs. 34% for CHD risk;
99% vs. 57% for SBP) and lower baseline CHD risk (5% vs. 7.5%)
and SBP (139 vs. 144 mm Hg) compared with noncompliers. In the
intention-to-treat analysis, the effect of offering the intervention
was nonsignificant for 4-year CHD risk (P = 0.08) but significant for
SBP (P = 0.003). CACE analyses showed that receipt of an effective
dose of the intervention resulted in a 1% greater reduction in 4-year
CHD risk (P < 0.05) and at least 8.1 mm Hg greater reduction in
SBP compared with compliers in the control group (P < 0.05).
From the *Department of Epidemiology and Biostatistics, University of
Texas Health Science Center at San Antonio (UTHSCSA); wCenter for
Research to Advance Community Health; zDepartment of Urology,
University of Texas Health Science Center at San Antonio, San Antonio;
ySchool of Public Health, University of Texas Health Science Center at
Houston, Houston; 8Cancer Therapy & Research Center, University of
Texas Health Science Center at San Antonio, San Antonio, TX; zDe-
partments of Surgery and Public Health Sciences, Penn State College of
Medicine, Hershey, PA; #University Health System; and **Department
of Medicine, University of Texas Health Science Center at San Antonio,
San Antonio, TX.
This project received an unrestricted supplementary grant from Pfizer Inc. to
the University of Pennsylvania.
The authors declare no conflict of interest.
Reprints: Yuanyuan Liang, PhD, Department of Epidemiology and Bio-
statistics, University of Texas Health Science Center at San Antonio,
7703 Floyd Curl Drive, San Antonio, TX 78229. E-mail: liangy@
uthscsa.edu.
Copyright r 2012 by Lippincott Williams & Wilkins
ISSN: 0025-7079/12/000-000
Medical Care  Volume 00, Number 00, ’’ 2012
Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Conclusions: Among compliers, an effective dose of peer coach
and office-based support resulted in significant reductions in 4-year
CHD risk and SBP. More intensive interventions are likely to be
required for noncompliers.
Key Words: complier average causal effect, noncompliance,
missing data, coronary heart disease, uncontrolled hypertension
(Med Care 2012;00: 000–000)
T hrough randomization of treatment assignment,
randomized controlled trials (RCTs) provide more
credible evidence of causal treatment effects than do ob-
servational studies.1,2 However, treatment randomization
does not ensure that subjects will comply with treatment
assignment or the outcome assessment. Indeed, these prob-
lems frequently occur concurrently in RCTs. Ignoring the
combined effect of these problems can bias estimates of the
treatment effect and lead to invalid causal inferences.3–6
Analyses restricted to persons who complied with the
assigned intervention may compromise the balance afforded
by randomization.7,8 Therefore, most clinical trials use an
intention-to-treat (ITT) analysis that includes all study sub-
jects on the basis of their original random assignment re-
gardless of whether or not they received the intervention. To
address missing outcome measures, multiple imputation and
likelihood-based estimation methods can be integrated with
an ITT analysis.9,10 However, the ITT approach has been
criticized as being too conservative, because it examines
only the effect of offering an intervention but not the effect
of receiving the intervention.3,8 Further, the ITT analysis
may lead to biased estimates by failing to account for
confounding between noncompliance and missing outcome
measures.
The complier average causal effect (CACE) method, a
variation of the instrumental variable approach, offers a
complementary approach to an ITT analysis, because it can
address noncompliance with the intervention in conjunction
with methods to address missing end-point data.4–6,11–15 The
concept of CACE analysis was introduced in the late 1990s,
and research on developing statistical methodologies under
various model assumptions is still ongoing.8,11,16–19 The
CACE analysis evaluates the effect of an effective dose of an
intervention among “compliers” who are individuals who
www.lww-medicalcare.com | 1
Liang et al
would have received an effective dose of the intervention if it
had been offered. To perform a CACE analysis, 2 key as-
sumptions must be met: (1) study subjects are well balanced
by the randomization process so the proportion of observable
compliers in the intervention group can be expected to be the
same as the proportion of unobservable compliers in the
control group; and (2) noncompliers in the intervention
group showed a similar benefit from the intervention as the
noncompliers in the control group.5,14 In a CACE analysis,
randomization is effectively an instrumental variable that has
an effect on outcome, but only through its effect on treatment
receipt, and is independent of all confounders.17
The appeal of the CACE analysis is that, although it
estimates the causal intervention effect in a subpopulation of
study subjects (ie, compliers who would have received an
effective dose of the intervention if it had been offered), it
uses data from all the subjects in the trial and hence retains
the balance afforded by original randomization as in an ITT
analysis.14 In the context of missing outcome data, the
CACE analysis can be modified under different missing-data
assumptions to increase the validity and efficiency of causal
inferences.4–5,13,15 Despite the complementary information
afforded by the CACE analysis, relatively few recent trials
have taken advantage of this potentially valuable ap-
proach.4,5,8,20,21 It is particularly underutilized in heath care
research, possibly because it can only be performed on trials
that meet several assumptions and the results can be difficult
to explain. Nevertheless, CACE analyses can offer unique
additional insights into the benefits of a trial for the subset of
participants who complied with the intervention.
In this paper, we compare results of an ITT analysis
with those from a CACE analysis for the Peer Coach and
Office Staff Support Trial. In this RCT, we studied the effect
of a behavioral support intervention on reduction in 4-year
coronary heart disease (CHD) risk and systolic blood pres-
sure (SBP).22 In this trial, both noncompliance with the in-
tervention and nonresponse with the outcome measurement
occurred and were handled analytically. We use this example
to demonstrate how the overlooked but useful CACE method
can contribute valuable information that is complementary to
results using the ITT method.
METHODS
Design and Participants
Data from the Peer Coach and Office Staff Support
Trial were used in the analysis. The detailed study protocol
and results have been described elsewhere.22 Briefly, the
study included African American patients aged 40–75 years
with uncontrolled hypertension, defined as an average blood
pressure over a 2-year period above normal with at least
1 value Z10 mm Hg above normal.22 A total of 280 African
American patients were randomized into an intervention
group (n = 136) and control group (n = 144). Both inter-
vention and control subjects received “usual care” from their
primary care providers and American Heart Association
brochures offering information on healthy recipes and ways
to reduce CHD risk. These materials were specifically de-
signed for African Americans. Intervention subjects received
2 | www.lww-medicalcare.com
Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Medical Care  Volume 00, Number 00, ’’ 2012
phone calls from trained peer coaches on months 1, 3, and 5,
whereas on months 2 and 4, subjects had individual in-office
counseling during which medication adherence and barriers
to making healthy lifestyle changes were addressed. Study
outcomes included change from baseline to 6 months in: (1)
4-year CHD risk using a measure based on Framingham data
that assesses risk of a primary or secondary CHD event23;
and (2) absolute change in SBP. At 6 months, 69% (94/136)
of the intervention subjects and 82% (118/144) of the control
subjects completed the end-point 4-year CHD risk assess-
ment, and 85% (116/136) of the intervention subjects and
91% (131/144) of the control subjects had completed the
SBP assessment.
Statistical Analyses
The ITT analysis of the trial outcomes has been pre-
viously reported, in which a multiple imputation with a
chained equations (MICE) approach was used to handle
missing data.22,24 Before performing the CACE analysis, we
hypothesized that an “effective dose” of the intervention
would require exposure to both components of the inter-
vention defined as at least 1 office visit and at least 2 peer
coach calls. In the intervention group, “compliers” were
subjects who received the hypothesized effective dose of the
intervention, whereas “noncompliers” were those who did
not. In the control group, “compliers” were defined as the
patients who would have received an “effective dose” if it
had been offered, and “noncompliers” were defined as the
patients who would not have received an effective dose of
the intervention regardless of whether or not it had been
offered.5,25 This trial satisfies 2 key assumptions for a CACE
analysis: (1) the 2 study arms must be well balanced by the
randomization method22 so the proportions of compliers in
the 2 arms are assumed to be the same (assumption A1); and
(2) intervention subjects who did not receive an effective
dose showed the same (lack of) benefit as did the non-
compliers in the control group (assumption A2).5,14
Figure 1 illustrates the logic for a CACE analysis.
More specifically, for the 4-year CHD outcome, the CACE is
defined as
CACE¼u1c u0c ; ð1Þ
where u1c is the mean change from baseline at 6 months in
4-year CHD risk for compliers in the intervention group and
u0c is the mean change for compliers in the control group. On
the basis of the definition for “compliers” given above, the
mean change for compliers is directly observable only for the
intervention group but not the control group. Instead, we
observe the mean change for the entire control group, u0,
which is a mixture of both compliers (u0c) and noncompliers
(u0n). That is,
u0 ¼p0c u0c þð1p0c Þu0n ; ð2Þ
where p0c is the proportion of compliers in the control group.
Assumption A1 allows us to assume that the control
and intervention group have the same proportion of com-
pliers. For simplicity, we use pc to denote the proportion of
compliers in the study population, which is directly estim-
able from the intervention group. Assumption A2 allows us
to assume that the mean change in 4-year CHD risk for
r 2012 Lippincott Williams & Wilkins
Medical Care  Volume 00, Number 00, ’’ 2012
FIGURE 1. u1c indicates observable mean for compliers in the
intervention group; u0c, unobservable mean for compliers in
the control group; u1n, observable mean for noncompliers in
the intervention group; u0n, unobservable mean for non-
compliers in the control group; u0, observable overall control
group mean; pc, proportion of compliers that is observable
from the intervention group under assumption A1. Assump-
tion A1: the control and intervention groups have the same
proportion of compliers because of the well-balanced
randomization. Assumption A2: noncompliers in the inter-
vention group showed the same benefit as would the non-
compliers in the control group, that is, u1n = u0n.
noncompliers in the intervention group (u1n), directly
estimable from observed data, is the same as the un-
observable mean change for the noncompliers in the control
group (ie, u1n = u0n). On the basis of these 2 assumptions
and Eq. (2), we are able to estimate the mean change for
compliers in the control group by adjusting the control group
mean for noncompliance using the data from the intervention
group—that is,
u0 ð1pc Þu1n
u0c ¼ :
pc
Hence, the CACE in Eq. (1) can be rewritten as
 
u0 ð1pc Þu1n
CACE¼u1c  ;
pc
where u0, u1c, u1n, and pc are all directly estimable from the
observed data and we utilize data from all study subjects.
To take missing outcome data into account, we
modified the CACE estimates using 3 approaches to handle
missing end-point data: (1) assuming that the response rate
for noncompliers is the same for both intervention and
control groups [Appendix: Assumption M2 and Eq. (A6)];
(2) assuming that the response rate for compliers is the same
for both intervention and control groups [Appendix: As-
sumption M3 and Eq. (A7)]; and (3) imputing missing data
by the MICE approach with 10 multiply imputed datasets,
computing a CACE estimate for each imputed dataset and
pooling the results together using Rubin’s rules.26 A sensi-
tivity analysis was conducted to examine CACE estimates
adjusting for baseline CHD risk and SBP.5,27 The CACE
analyses were performed using Stata software (version 11;
College Station, TX), and 95% confidence intervals (CIs) for
CACE estimates were computed using the bootstrapping
method in R (version 2.15.0).
r 2012 Lippincott Williams & Wilkins
Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Intervention Benefit Using CACE Analysis
RESULTS
In this trial, 136 of 280 patients were randomized to the
intervention group. An effective dose of the intervention (ie,
at least 1 office visit and at least 2 peer coach calls) was
received by 92 “compliers” (67.6%). Nearly half of the in-
tervention subjects (48.5%) received full dose of the inter-
vention (ie, 3 phone calls and 2 visits). Of the 44 patients in
the intervention group who were categorized as
“noncompliers,” 6 (13.6%) received none of the intervention
components (Table 1).
At baseline, the intervention and control arms did not
differ in any patient demographic, clinical, or health
behavior characteristic except for high-density lipoprotein
cholesterol (P = 0.047), supporting successful random-
ization.22 Specifically, baseline mean 4-year CHD risk was
6.1% (5.8% for intervention vs. 6.4% for control, P = 0.39),
and the mean SBP was 140.5 mm Hg for both study arms.22
However, compliers and noncompliers in the intervention
group differed, with the former group having significantly
lower baseline 4-year CHD risk and less severely elevated
SBP (Table 2). Also, a higher proportion of noncompliers
admitted to missing a medication recently compared with
compliers (36% vs. 23%, P = 0.15).
The proportions of patients with end-point data avail-
able to assess the 4-year CHD risk end point at 6 months
differed significantly for the 3 observable groups: 85.9% for
compliers in the intervention group, 34.1% for noncompliers
in the intervention group, and 81.9% for all control subjects
(P < 0.001, Table 3). Complete SBP end-point data also
differed: 98.9% for compliers in the intervention group,
56.8% for noncompliers in the intervention group, and 91.0%
for all control subjects (P < 0.001). Noncompliance with in-
tervention was significantly associated with missing end-
point data (P < 0.001).
In the previously reported ITT analysis, the reduction
in 4-year CHD risk was 0.7% greater for the intervention
than for the control group but not statistically significant
(P = 0.08). Given an overall 6.1% baseline 4-year CHD risk
in our study, this intervention effect was equivalent to an
11% (= 0.7%/6.1%) relative risk reduction. In contrast, the
ITT analysis showed that the reduction in SBP was 6.5 mm
Hg greater for the intervention than for the control group
(P = 0.003).
In the CACE analysis (Table 4), we examined the
effect of receiving an effective dose of the intervention using
3 methods to account for missing end-point data. All 3
methods showed that intervention consistently resulted in a
significant 1% greater reduction in 4-year CHD risk (ie,
relative risk reduction of 16% = 1%/6.1%) for the compliers
in the intervention group compared with the effect estimated
by the CACE analysis for the compliers in the control group
(P < 0.05). In addition, all 3 CACE methods showed that
receipt of an effective dose of the intervention resulted in
at least an 8.1 mm Hg greater reduction in SBP than the
compliers in the control group (P < 0.05). Adjustment for
baseline SBP and baseline CHD risk did not change the
CACE estimates substantively (CHD: CACE =  1.09%,
95% CI =  1.87%,  0.31%; SBP: CACE =  9, 95% CI =
 14.13,  3.86).
www.lww-medicalcare.com | 3
Liang et al Medical Care  Volume 00, Number 00, ’’ 2012

TABLE 1. Distribution of the Receipt of the Components of the Intervention in the Intervention Group (n = 136)
# of Phone Calls # of Office Visits Count % Compliance Status*
0 0 6 4.4 No
1 0 7 5.1 No
1 1 1 0.7 No
1 2 8 5.9 No
2 0 7 5.1 No
2 1 2 1.5 Yes
2 2 9 6.6 Yes
3 0 15 11 No
3 1 15 11 Yes
3 2 66 48.5 Yes
*An adequate intervention dose is defined as at least 2 peer coach calls and at least 1 office visit.

CONCLUSIONS the intervention subjects who received the hypothesized

In this behavioral support trial for African Americans
with sustained poorly controlled hypertension, an ITT anal-
ysis found no significant effect of the intervention on the
primary outcome of 4-year CHD risk, whereas the inter-
vention produced a significant reduction in the secondary
outcome of SBP. One might conclude that this is a
“negative” trial, because the effect on the primary outcome
was not significant. However, the CACE analysis offers
valuable, additional information from the trial by estimating
the causal effect of the intervention among “compliers,”
while taking into account both noncompliance with the in-
tervention and missing outcome data. Among two thirds of
effective dose, we observed a significant 1% greater reduc-
tion in 4-year CHD risk than in their counterparts in the
control group as estimated by the CACE analysis. This in-
tervention effect was equivalent to a 16% relative risk re-
duction for compliers. The CACE analysis also found a
significant benefit of the intervention on SBP as in the
original ITT analysis but the effect size is even greater for
the compliers (8.1 vs. 6.5 mm Hg additional reduction, re-
spectively). Therefore, this CACE analysis offers support for
the value of this novel behavioral support intervention
among the majority of patients who were observed or esti-
mated to be compliers.

TABLE 2. Baseline Characteristics of African-American Patients With Uncontrolled Hypertension

Control Intervention Complier Intervention Noncomplier
Characteristics* N = 144 N = 92 N = 44
Age (y), mean (SD) 62.6 (8.3) 60.9 (8.9) 61.8 (10.1)
Sex, N (%)
Female 88 (61.11) 64 (69.57) 31 (70.45)
Male 56 (38.89) 28 (30.43) 13 (29.55)
General health, N (%)
Excellent or very good 47 (32.64) 30 (32.61) 16 (36.36)
Good 64 (44.44) 43 (46.74) 21 (47.73)
Fair or poor 32 (22.22) 19 (20.65) 7 (15.91)
Missing 1 (0.69) 0 (0) 0 (0)
Clinical conditions, N (%)
Diabetes mellitus 75 (52.08) 51 (55.43) 25 (56.82)
Coronary artery disease or equivalent 30 (20.83) 11 (11.96) 9 (20.45)
Depressive symptoms 62 (43.06) 38 (41.3) 17 (38.64)
More stressed than others 65 (45.14) 34 (36.96) 21 (47.73)
Current smoker 25 (17.36) 20 (21.74) 7 (15.91)
Baseline blood pressure (mm Hg), mean (SD)
Systolicw 140.5 (8.9) 138.8 (7.1) 144 (12.2)
Diastolic 81 (6.7) 81.3 (7.1) 81.5 (9.2)
Baseline lipid levels (mg/dL), mean (SD)
LDL cholesterol 113.4 (29.3) 115.2 (28.8) 118.5 (31.1)
HDL cholesterol 54 (15.2) 58 (16.2) 56.3 (13.6)
Triglycerides 149.1 (93.8) 138.3 (80.3) 137.7 (61.9)
Total cholesterol 196.1 (33.8) 200.1 (33.5) 201.4 (34.3)
4-year CHD risk %, mean (SD)z 6.4 (6.5) 5 (5.6) 7.5 (10)
*Entries are mean (SD) and count (%) for continuous and categorical outcomes, respectively.
w
P = 0.049 on the basis of the Kruskal-Wallis test for comparing 3 groups; P = 0.01 on the basis of the Mann-Whitney test for comparing compliers and noncompliers in the
intervention group.
z
P = 0.047 on the basis of the Kruskal-Wallis test for comparing 3 groups; P = 0.02 on the basis of the Mann-Whitney test for comparing compliers and noncompliers in the
intervention group.
CHD indicates coronary heart disease; HDL, high-density lipoprotein; LDL, low-density lipoprotein.

4 | www.lww-medicalcare.com r 2012 Lippincott Williams & Wilkins

Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Medical Care  Volume 00, Number 00, ’’ 2012 Intervention Benefit Using CACE Analysis

TABLE 3. Intervention Receipt Status Versus Respondent at 6 Months

Respondent at 6 Months
Outcome Group Yes No Total
4-year CHD risk Compliers in the intervention 79 13 92
(pR1c = 79/92 = 85.9%) [pc = 92/(92+44) = 67.6%]
Noncompliers in the intervention 15 29 44
(pR1n = 15/44 = 34.1%)
Control 118 26 144
(pR0 = 118/144 = 81.9%)
Total 212 68 280
Systolic blood pressure Compliers in the intervention 91 1 92
(pR1c = 91/92 = 98.9%) [pc = 92/(92+44) = 67.6%]
Noncompliers in the intervention 25 19 44
(pR1n = 25/44 = 56.8%)
Control 131 13 144
(pR0 = 131/144 = 91.0%)
Total 247 33 280
pc indicates proportion of compliers; pR1c , response rate for the compliers in the intervention group; pR1n , response rate for the noncompliers in the intervention group; pR0 , response
rate in the control group; CHD, coronary heart disease.

The CACE analysis complements the standard ITT
analysis of an RCT, because the latter evaluates the effect of
offering a treatment but not the effect of receiving a treat-
ment. This may lead to missing clinically important effects
in RCTs. For example, the authors of a recent patient-level
meta-analysis of RCTs on vitamin D dose requirements for
fracture prevention criticized the ITT approach for under-
estimating the benefit of vitamin D supplementation.28 The
previous meta-analysis based on the ITT approach indicated
a nonsignificant reduction in the risks of hip or vertebral
fracture. Yet a meta-analysis examining actual receipt of
high-dose vitamin D showed clinically and statistically sig-
nificant reductions in both, which were independent of the
assigned treatment dose of vitamin D.28 Thus, trials need to
take advantage of multiple approaches to gain information
from this valuable research resource. The CACE analysis
provides a useful alternative tool, in conjunction with anal-
yses understanding the unique characteristics of the
“compliers” who would have received an effective dose had
it been offered.
Several potential limitations with CACE analysis
should be acknowledged. First, compared with the well-es-
tablished ITT approach, the CACE approach involves po-
tentially challenging statistical assumptions especially when
the subjects in the control group could access the inter-
vention, although research is ongoing to relax
them.3,5,6,8,13,15,17–19,29–31 Second, determining a minimally
effective dose of the intervention may need to be based on
subjective judgment when there are no data to guide this
selection. Before conducting the CACE analysis for our trial,
we hypothesized that subjects needed at least 1 office visit to
be exposed to the tailored educational program and at least 2
phone calls to develop a relationship with a peer coach on the
basis of clinical judgment. Other studies appear to have se-
lected a minimally effective dose post hoc. For example, in a
CACE analysis of the Johns Hopkins University Preventive

TABLE 4. CACE Analysis Results Versus ITT Analysis Results

CACE Analysis
Outcome Method for Handling Missing Data CACE 95% CI* ITT Analysis Effect Size (95% CI)
4-year CHD risk M2w pR0n ¼pR1n [Eq. (A6)]  0.98% 1.75%,  0.28%z 0.73%
( 1.54%, 0.09%)
M3y pR0c ¼pR1c [Eq. (A7)]  0.94% 1.92%,  0.01%z
MICE8 Eq. (A3)  1.08% 2.12%,  0.03%z
MICE With covariatesz  1.09% 1.87%,  0.31%z
Systolic blood pressure (mm Hg) M2w pR0n ¼pR1n [Eq. (A6)]  8.10 13.04,  3.25z  6.47
(10.69, 2.25)z
y z
M3 pR0c ¼pR1c
[Eq. (A7)]  8.28 13.70,  3.00
MICE8 Eq. (A3)  9.56 12.13,  6.99z
MICE With covariatesz  9.00 14.13,  3.86z
*95% CIs were estimated using the bootstrapping method with 5000 bootstrap samples.
w
For noncompliers, the response rate is not affected by group assignment status; see Assumption M2 and Eq. (A6) in the Appendix.
z
95% CI does not contain 0; therefore the effect size is significant at the significance level of 0.05.
y
For compliers, the response rate is not affected by group assignment status; see Assumption M3 and Eq. (A7) in the Appendix.
8
MICE: missing data were imputed using the multiple imputation by chained equations (MICE) approach17 with 10 multiply imputed datasets. For each imputed dataset, CACE
was computed using Eq. (A3) in the Appendix, and the standard error for CACE was estimated using the bootstrapping method with 100,000 bootstrap samples. The results from 10
imputed datasets were combined using Rubin’s rules.19
z
Adjusted for baseline SBP and baseline 4-year CHD risk.
CACE indicates complier average causal effect; CHD, coronary heart disease; CI, confidence interval; ITT, intention to treat.

r 2012 Lippincott Williams & Wilkins www.lww-medicalcare.com | 5

Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Liang et al Medical Care  Volume 00, Number 00, ’’ 2012

Intervention Research Center, which involved a multi- Therefore, Eq. (A1) can be rewritten as
component intervention designed to improve academic ach-  
u0 ð1p0c Þu0n
ievement and to reduce early behavioral problems in first- CACE¼u1c  : ðA2Þ
grade children, the effective dose of the intervention was p0c
defined as receiving 45 of 66 activities.5,25 However, that Assumption 1 (A1): we assume that both study arms
appears to have been selected post hoc to include at least half are well balanced by randomization so the control group has
of the subjects. We propose that the minimally effective dose the same proportion of compliers as the intervention group.
should be prespecified to improve generalizability and var- Therefore, p0c = p1c is observable from the intervention
iations could be examined post hoc. group. For simplicity, we use pc to denote the proportion of
Third, there are a variety of options to deal with the compliers in the population, that is, pc = p0c = p1c.
situation in which noncompliance with the intervention is Assumption 2 (A2): we assume that there is no effect
associated with noncompliance with obtaining the end point; of group assignment on the outcome of interest for non-
however, it is unclear which is optimal, and most require compliers. That is, the population mean for noncompliers in
unverifiable assumptions.5,32,33 Therefore, we compared the intervention group is the same as the population mean for
CACE estimates from 3 different approaches that we used to noncompliers in the control group, u0n = u1n.
address missing end-point data to provide a plausible range Under the assumptions 1 and 2, the CACE in Eq. (A2)
of the intervention effect. Because our results using all 3 can be expressed as
approaches were similar (ie, a significant 1% additional re-  
duction in 4-y CHD risk, ie, 16% relative risk reduction, and u0 ð1pc Þu1n
CACE¼u1c  ; ðA3Þ
a significant additional reduction of at least 8.1 mm Hg in pc
SBP), our analyses provide more robust evidence that our
behavioral support intervention produced positive changes in where u0, u1c, u1n, and pc are all directly estimable from the
CHD risk and SBP among compliers and can be used to observed data.
guide future trials among similar subjects who constituted When missing data are present, let pR1 denote the re-
the majority of this study population. sponse rate in the intervention group, and pR0 the response
In summary, CACE and ITT analyses offer comple- rate in the control group. Further, pR0 can be thought of as a
mentary approaches to examining results from trials that mixture of compliers and noncompliers in the control group,
have the common problems of noncompliance with the in- which cannot be separately observed. That is,
tervention and missing end-point data. The CACE analysis
pR0 ¼pc pR0c þð1pc ÞpR0n ;
adds value by examining the effect of the intervention among
compliers while still using all the data from the RCT, in where pR0c is the response rate for the compliers in the control
contrast to subgroup and per-protocol analyses. By com- group, and pR0n is the response rate for the noncompliers in
paring the characteristics of compliers and noncompliers, it the control group. Therefore, the observed average outcome
is possible to refine eligibility criteria for future trials. Our of the control group is
CACE results will guide a future trial of this intervention, pR0c pR
although a more intensive intervention must be developed uobs
0 ¼ R pc u0c þ 0n ð1pc Þu0n :
p0 pR0
and tested for persons similar to our noncompliers who had
greater baseline elevations in CHD risk and more severely Therefore, u0c can be rewritten as
elevated SBP. uobs R R
0 p0 u0n p0n ð1pc Þ uobs pR u0n pR ð1pc Þ
u0c ¼ R ¼ 0 R0 R 0n :
p0c pc p0 p0n ð1pc Þ
APPENDIX Under assumption A2, u0n = u1n, therefore the CACE in Eq.
Following the notations from Jo et al,5 the CACE is (A3) can be rewritten as
defined at the average population level as
uobs R R
0 p0 u1n p0n ð1pc Þ
CACE¼u1c  ; ðA4Þ
CACE¼u1c u0c ; ðA1Þ R R
p0 p0n ð1pc Þ
where u1c is the population mean potential outcome (eg,
change from baseline at 6 months in 4-year CHD risk) for where u1c, u1n, uobs R
0 , p0 , and pc are directly estimable from
compliers in the intervention group and u0c is for compliers the observed data. However, further assumptions are neces-
in the control group. sary to identify pR0n , the response rate for the noncompliers in
Note that u1c is observable from the intervention group the control group, in order to fully identify CACE. We
data but u0c is nonobservable from the control group data. consider the following 3 missing-data assumptions.
Instead, we observe the overall control group mean, u0, Missing-data assumption 1 (M1): missing at random,
which is a mixture of the compliers (u0c) and noncompliers that is, missingness is not attributable to the unobserved
(u0n) in the control group. That is, compliance status in the control group. Therefore,
pR0n ¼pR0c ¼pR0 , meaning that compliers and noncompliers
u0 ¼p0c u0c þð1p0c Þu0n ; have the same response rate in the control group, but they
may have different response rates in the intervention group.
where p0c is the proportion of compliers in the control group. Under missing at random, the CACE in Eq. (A4) can be

6 | www.lww-medicalcare.com r 2012 Lippincott Williams & Wilkins

Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Medical Care  Volume 00, Number 00, ’’ 2012 Intervention Benefit Using CACE Analysis

rewritten as
 
uobs
0 u1n ð1pc Þ
CACEM1 ¼u1c  ; ðA5Þ
pc
where all involved parameters are directly estimable from
the observed data.
Missing-data assumption 2 (M2): for noncompliers, the
response rate is not affected by group assignment status, that
is, pR0n ¼pR1n . Under M2, the CACE in Eq. (A4) can be re-
written as
 obs R  prevention trials: Treatment of no-shows using Rubin’s causal model.
u0 p0 u1n pR1n ð1pc Þ
CACEM2 ¼u1c  ; ðA6Þ
pR0 pR1n ð1pc Þ
where all involved parameters are directly estimable from
the observed data.
Missing-data assumption 3 (M3): for compliers, the
response rate is not affected by the group assignment status,
that is, pR0c ¼pR1c . Under M3, the CACE in Eq. (A4) can be
rewritten as
 obs R 
u0 p0 u1n ðpR0 pR1c pc Þ
CACEM3 ¼u1c  ; ðA7Þ
pR1c pc
where all involved parameters are directly estimable from
the observed data.
REFERENCES
1. Eccles M, Mason J. How to develop cost-conscious guidelines. Health
Technol Assess. 2001;5:1–69.
2. Reeves B, Deeks J, Higgins J, et al. Chapter 13: Including non-
randomized studies. In: Higgins J, Green S, eds. Cochrane Handbook
for Systematic Reviews of Interventions. John Wiley & Sons Ltd,
England: The Cochrane Collaboration; 2008;2008.
3. Hirano K, Imbens GW, Rubin DB, et al. Assessing the effect of an
influenza vaccine in an encouragement design. Biostatistics. 2000;1:69–88.
4. Horiuchi Y, Imai K, Taniguchi N. Designing and analyzing randomized
experiments: application to a Japanese election survey experiment. Am J
Polit Sci. 2007;51:669–687.
5. Jo B, Ginexi EM, Ialongo NS. Handling missing data in randomized
experiments with noncompliance. Prev Sci. 2010;11:384–396.
6. Esterling KM, Neblo MA, Lazer DMJ. Estimating treatment effects in
the presence of noncompliance and nonresponse: the generalized
endogenous treatment model. Polit Anal. 2011;19:205–226.
7. Frangakis C, Rubin D. Addressing complications of intention-to-treat
analysis in the combined presence of all-or-none treatment-noncompli-
ance and subsequent missing outcomes. Biometrika. 1999;86:365–379.
8. Connell AM. Employing complier average causal effect analytic
methods to examine effects of randomized encouragement trials. Am J
Drug Alcohol Abuse. 2009;35:253–259.
9. Schafer JL. Analysis of Incomplete Multivariate Data. London: CRC;
1997.
10. Little RJA, Rubin DB. Statistical Analysis With Missing Data. New
York: Wiley; 2002.
11. Angrist JD, Imbens GW, Rubin DB. Identification of causal effects
using instrumental variables. J Am Stat Assoc. 1996;91:444–455.
12. Frangakis CE, Rubin DB. Principal stratification in causal inference.
Biometrics. 2002;58:21–29.
13. Peng Y, Little RJ, Raghunathan TE. An extended general location
model for causal inferences from data subject to noncompliance and
missing values. Biometrics. 2004;60:598–607.
14. Hewitt CE, Torgerson DJ, Miles JNV. Is there another way to take
account of noncompliance in randomized controlled trials? Can Med
Assoc J. 2006;175:347–348.
15. Chen H, Geng Z, Zhou XH. Identifiability and estimation of causal
effects in randomized trials with noncompliance and completely
nonignorable missing data. Biometrics. 2009;65:675–682.
16. Imbens GW, Rubin DB. Estimating outcome distributions for compliers
in instrumental variables models. Rev Econ Stud. 1997;64:555–574.
17. Little RJ, Yau LHY. Statistical techniques for analyzing data from
Psychol Methods. 1998;3:147–169.
18. Jo B. Statistical power in randomized intervention studies with
noncompliance. Psychol Methods. 2002;7:178–193.
19. Jin H, Rubin DB. Principal stratification for causal inference with
extended partial compliance. J Am Stat Assoc. 2008;103:101–111.
20. Schochet PZ, Chiang HS. Estimation and identification of the complier
average causal effect parameter in education RCTs. J Educ Behav Stat.
2011;36:307–345.
21. Stanger C, Ryan SR, Fu H, et al. Parent training plus contingency
management for substance abusing families: A Complier Average
Causal Effects (CACE) analysis. Drug Alcohol Depend. 2011;118:
119–126.
22. Turner BJ, Hollenbeak CS, Liang Y, et al. Randomized trial of peer
coach and office staff support to reduce coronary heart disease risk in
african-americans with uncontrolled hypertension. J Gen Intern Med.
2012;27:1258–1264.
23. D’Agostino RB, Russell MW, Huse DM, et al. Primary and subsequent
coronary risk appraisal: new results from the Framingham study. Am
Heart J. 2000;139:272–281.
24. Raghunathan T, Lepkowski JM, Van Hoewyk J, et al. Multivariate
technique for multiply imputing missing values using a sequence of
regression models. Surv Methodol. 2001;27:85–95.
25. Ialongo NS, Werthamer L, Kellam SG, et al. Proximal impact of two
first-grade preventive interventions on the early risk behaviors for later
substance abuse, depression, and antisocial behavior. Am J Community
Psychol. 1999;27:599–641.
26. Rubin D. Multiple Imputation for Nonresponse in Surveys. New York:
John Wiley and Sons; 1987.
27. Dunn G, Maracy M, Dowrick C, et al. Estimating psychological
treatment effects from a randomised controlled trial with both non-
compliance and loss to follow-up. Br J Psychiatry. 2003;183:323–331.
28. Bischoff-Ferrari HA, Willett WC, Orav EJ, et al. A pooled analysis of
vitamin D dose requirements for fracture prevention. N Engl J Med.
2012;367:40–49.
29. Yau LHY, Little RJ. Inference for the complier-average causal effect
from longitudinal data subject to noncompliance and missing data, with
application to a job training assessment for the unemployed. J Am Stat
Assoc. 2001;96:1232–1244.
30. Barnard J, Frangakis CE, Hill JL, et al. Principal stratification approach
to broken randomized experiments. J Am Stat Assoc. 2003;98:299–323.
31. Jo B, Vinokur AD. Sensitivity analysis and bounding of causal effects
with alternative identifying assumptions. J Educ Behav Stat.
2011;36:415–440.
32. Jo B. Bias mechanisms in intention-to-treat analysis with data subject to
treatment noncompliance and missing outcomes. J Educ Behav Stat.
2007;33:158–185.
33. Little RJ, D’Agostino R, Cohen ML, et al. The prevention and treatment
of missing data in clinical trials. N Engl J Med. 2012;367:1355–1360.

r 2012 Lippincott Williams & Wilkins www.lww-medicalcare.com | 7

Copyright © Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
View publication stats