Microsoft Word search phase 3 Title Page Amendment

Mediation Modeling 
For some of the analyses described above we will need to carefully assess the potential 
effects of mediators on the relationships we are examining.  We will use mediation 
models as described by Baron & Kenny 
(1)
 in which a series of ANCOVA models are fit 
regressing: 1) outcome on exposure (Model 1); 2) mediator on exposure (Model 2); and 
3) outcome on both mediator and exposure (Model 3).  A change in the estimate of the 
exposure effect from Model 1 to Model 3 is evidence of mediation.  We will use 
bootstrap methods for significance testing of the mediation effect and for calculating 
confidence intervals
 (2)
. 
What is particularly exciting about using the SEARCH Cohort study is that potential 
mediator data will be measured at least at two time points.  For instance, we can examine 
potential differences in access to care (a potential mediator) at particular points in time in 
relation to outcomes of interest.  In addition we can form models that incorporate several 
potential mediators simultaneously (e.g., access to care and physical activity).  
Procedures, such as examining eigenvalues, discussed in Belsley, Kuh, and Welch 
(3)
 and 
Kleinbaum, Kupper and Muller 
(4)
 will be used to identify problems with multicollinearity 
and appropriate steps will be taken to reduce the problem. 
6.1.5.
 
Missing Data Considerations 
We utilize the sequential data collection of the SEARCH protocol to statistically adjust 
for potential selection bias related to response rates or other reasons for missing data.  
Variables determined to predict loss to follow-up will be included in our predictive 
models in order to satisfy the conditions for the data to be considered “Missing at 
Random” (MAR) 
(5)
.  Estimation techniques such as maximum likelihood will be used to 
estimate parameters.  If the MAR assumption is untenable, one must assume that 
“informative censoring” has occurred.  For example, biased estimates can result if 
participants with adverse experiences are more likely to withdraw (or, conversely, tend to 
be relatively less likely to withdraw).  A growing body of literature describes two 
alternative approaches to handle this potential problem: explicit modeling of the 
censoring mechanisms 
(6 - 12)
 and pattern-mixture models 
(13)
.  We have experience with 
these approaches for handling non-ignorable non-response and will analyze the data 
using several of these methods which incorporate varying assumptions about the missing 
observations 
(14)
.  This will provide useful information to allow us to understand the 
potential limitations we may have to interpret results in the presence of informative 
censoring. 

---

Section 6B - Statistical Considerations (Phase 3 - 12/2010) 
Section 6B - Page 4 
 Cohort 
Study
 
 
6.2.
 
POWER ANALYSIS 
Based on our calculations from SEARCH 1 and 2 we anticipate that at least 3,288 subjects 
will participate in the SEARCH Cohort Study in-person visit, and therefore the calculations 
below use that sample size as the starting value for estimating power and detectable 
differences. 
6.2.1.
 
Incidence Rate Estimation 
To estimate power for this component we first had to estimate the proportion of 
participants who would be free of the condition at their initial in-person visit during 
SEARCH 1 and 2.  Table 6-1 shows the expected sample sizes available for comparing 
incidence rates between subgroups under two scenarios: 1) proportion in subgroups are 
86% versus 14% (the proportions of T1D and T2D), and 2) proportion in subgroups are 
65% versus 35% (the proportions of NHW and all others).   
 
Table 6-1: Expected Sample Sizes Available for Comparing Incidence Rates between Subgroups 
Under Two scenarios 
 
Scenario 1 
Scenario 2 
Outcome (% of participants free of 
condition at initial SEARCH visit) 
Subgroup A 
(86%) 
Subgroup B 
(14%) 
Subgroup A 
(65%) 
Subgroup B 
(35%) 
 
Expected number of participants free of outcome at initial visit 
Hypertension (92%) 
2504 
423 
1893 
1019 
Obese (bmi-z > 95
th
%ile) 
(76%)  2069 349 1564 842 
High LDL (> 100) (57%) 
1511 
262 
1173 
631 
High ACR (>30 
(90%) 
2450 414 1852 997 
Hypoglycemia in last 6 months (91%) 
2477 
418 
1872 
1008 
DKA in last 6 months (85%) 
2314 
391 
1749 
941 
 
Using this table we can determine detectable differences for each outcome/group 
comparison for a variety of plausible scenarios for incidence rates.  Table 6-2 illustrates 
detectable differences assuming a two group continuity corrected chi-square test for a 
variety of scenarios with alpha=0.05 (2-sided). 
 
Table 6-2. Detectable Differences in Selected Outcomes Assuming a Two-Group Continuity 
Corrected Chi-square Test for a Variety of Scenarios with Alpha=0.05 (2-sided) 
 
Scenario 1:  Example T1 vs. T2 
Scenario 2:  NHW vs. Other 
Outcome  
Incidence Rate 
for T1 
Detectable Rate 
for T2 (Power) 
Incidence Rate 
for NHW 
Detectable Rate 
for Other 
Race/Ethnic 
Group (Power) 
Hypertension 
6%** 10% 
(80%) 6%** 10% 
(96%) 
 
12% 18% 
(88%) 12% 17% 
(95%) 
Obese 
5% 10% 
(90%) 9% 14% 
(95%) 
 
10% 16% 
(86%) 19% 25% 
(91%) 
High LDL  
23% 
32% (84%) 
14%  
20% (89%) 

---

Section 6B - Statistical Considerations (Phase 3 - 12/2010) 
Section 6B - Page 5 
 Cohort 
Study
 
 
 
33% 43% 
(85%) 24% 31% 
(87%) 
High ACR  
9% 
15% (93%) 
5% 
8% (86%) 
 
19% 26% 
(87%) 10% 14% 
(87%) 
Hypoglycemia in last 6 months  
11% 
17% (90%) 
23% 
28% (82%) 
 
21% 28% 
(86%) 33% 39% 
(88%) 
DKA in last 6 months  
15% 
21% (81%) 
9% 
13% (87%) 
 
25% 32% 
(80%) 19% 24% 
(84%) 
** First incidence rate reflects observed incidence rate in SEARCH 1 and 2 for follow-up visits already measured. 
 
Based on this table, we see, for instance, that there is 90% power to detect a difference 
between Type 1 and Type 2 participants on their rate of incident obesity if the rate of 
incident obesity is 5% in the Type 1 group and 10% or higher in the Type 2 group.  
Likewise, there if the rate of obesity were 9% in the NHW group then there is 95% power 
to detect a “Other” race/ethnic group rate of 14% or higher.  The above calculations 
should be conservative since when we adjust for participant level characteristics in our 
models we should reduce variability and increase precision as we estimate the difference 
in incidence rates between groups. 
6.2.2.