Microsoft Word search phase 3 Title Page Amendment

6.
 
Statistical Considerations 
There will be four main analytic approaches: 1) estimation of incidence rates of 
complications (Aims 1 and 2); 2) estimation of prevalence of complications (Aims 1 and 2); 
3) longitudinal evaluation of predictors of outcomes and conditions, including consideration 
of potential mediators and moderators (Aims 1, 2 and 4); and 4) evaluation of mortality rates, 
including comparison with age-comparable non-diabetic youth (Aim 3).  Here we provide 
selected examples to illustrate analytic strategies and to provide key information regarding 
sample size and detectable differences.   
6.1.
 
STATISTICAL ANALYTIC METHODS 
6.1.1.
 
Incidence Rate Estimation 
Because all SEARCH Cohort Study participants will have had at least one previous 
SEARCH in-person visit, we will be able to define a group of participants who were free 
from the event of interest (i.e., normotensive) at “baseline”.  Multiple logistic regression 
methods will be employed to examine the incidence rates of binary measures (e.g., 
hypertension) of interest.  Predictors can include categorical or continuous variables.  A 
continuous variable that measures the time between visits for each participant (to account 
for the fact that individuals will have different lengths of follow-up) and the predictor- by- 
time interaction will be included.  Next, we will expand the logistic regression model to 
include other participant level characteristics (e.g., SEARCH clinical center, age, and 
gender [a “demographically adjusted model”]).  We will then expand the model to adjust 
for other covariates.  In addition, we will examine potential interactions; if significant 
interaction is present, analyses will be performed stratified by that characteristic. 
6.1.2.
 
Prevalence Estimation 
Some of the outcomes of interest will not have been measured during SEARCH 1 or 2, 
such as outcomes including retinopathy and neuropathy.  Therefore, prevalence of these 
outcomes will be estimated.  Models to evaluate cross-sectional associations of risk 
factors will use logistic regression and will proceed as described above to account for 
potential confounding or effect modification. 
6.1.3.
 
Longitudinal Models 
All participants in the SEARCH Cohort Study will have already had at least one in-
person visit during SEARCH 1 and 2, and ~75% of the 2002-2005 incidence cases have 
at least 2 in-person visits per the SEARCH 2 protocol.  Since SEARCH 2 also included 
longitudinal data (there are over 2000 SEARCH participants already with at least one 
follow-up visit), our team developed a plan for modeling longitudinal data.  Specifically, 
we will use longitudinal mixed effects analysis of covariance models that always include 
duration of diabetes as a time-varying covariate.  This approach correctly models the 
varying durations of disease prior to the initial SEARCH in-person visit, and the varying 

---

Section 6B - Statistical Considerations (Phase 3 - 12/2010) 
Section 6B - Page 2 
 Cohort 
Study
 
 
durations of time allowed via the SEARCH data collection windows between the initial 
and subsequent visits. 
The initial model will examine outcomes (measured previously between 1 (baseline) and 
4 times (baseline, 12, 24, 60 mo visits) and once during the SEARCH Cohort Study 
visit), the predictor of interest (e.g., DM type), the duration of diabetes at each 
measurement time and the predictor-by-diabetes duration interaction.  These models will 
then be expanded to include demographic information (e.g., sex) that would be 
considered as fixed/non-time varying effects.  In addition, based on our experience with 
performing these longitudinal analyses on the SEARCH 2 cohort, we also propose to 
consider treating the exposure (predictor) of interest as a time-varying covariate in these 
models as well.  This will allow the time-varying correlation of the predictor to the 
outcome of interest to be modeled correctly.  We will also consider adding other time-
varying covariates (e.g., BMI z-score) into these models as needed to examine the 
specific relationships being studied.  These mixed effects models also are flexible to 
allow for potentially non-linear relationships to be modeled over time, and permit random 
rates of progression, consistent with a perspective that different participants progress 
through time at different rates.  Use of random intercepts and/or slopes provides a source 
of autocorrelation between repeated measures.  More flexible structures for the 
correlation between repeated measures will be investigated using combination mixed 
models that allow the specification of separate parameters representing variation between 
experimental units, and serial correlation within units.  Our choice of methods for 
accounting for serial correlation depends on the plausibility of the model, and the number 
of outcomes relative to the number of participants.  For example, with many participants 
and few repeated measurements, an unstructured covariance matrix can often provide for 
the most efficient estimation of model parameters. 
For analysis of longitudinal discrete outcomes (e.g. transfer of care from a pediatric to 
adult provider), we will use the generalized estimating equation (GEE) approach to fit 
logistic or log-linear models that account for the dependency between repeated measures.  
GEE techniques allow estimation of model parameters and their standard errors from 
longitudinal data having continuous and categorical responses and potentially missing 
observations.  An advantage of this technique is that the assumptions required are weaker 
than those of maximum likelihood techniques: one need not specify the distribution of the 
dependent variable, just the relationships between the marginal mean and variance, and 
between the marginal mean and covariates. 

---

Section 6B - Statistical Considerations (Phase 3 - 12/2010) 
Section 6B - Page 3 
 Cohort 
Study
 
 
6.1.4.