Epidemiology: 10) Cohort Analysis: Multivariate analysis
Terms
undefined, object
copy deck
 Regression models
 control for confounding using math models
 Choice of Regression
 (105)
 4 chars of Proportional Hazards (Cox's)

1) Binary outcome
2) Variable followup
3) Start and end time knows for individuals
4) Assumptions are satisfied  Effect estimate of Cox's
 exp(b) estimates the incidence rate ratio.
 3 chars of Logistic Regression

1) Binary outcome
2) Fixed/defined followup
3) Binomial assumptions are satisfied  Effect estimation of logistic regression
 exp(b) = odds ratio (risk ratio if disease is rare)
 4 chars of Poisson Regression

1) Binary outcome
2) variable followup
3a) start and end times known for individuals or
3b) data stratified into mututally exclusive groups according to E and confounder, number of disease cases and followup time are knows for these strata
4) Rare disease (and other Poisson assumptions)  Effect estimate of Poisson Regression
 exp(b) = incidence rate ratio
 3 chars of Linear Regression

1) Continuous outcome (BP, LDL)
2) Followup is fixed/defined
3) Linear regression assumptions satisfied  Effect estimation of Linear regression
 b, the estimated regression coefficient is a difference in mean values
 7 Issues to consider for model construction

1) Model Type
2) Independence observations
3) Disease variable scale (continuous)
4) Covariate Definitions
5) Building model (selection)
6) Additivity
7) Unknowns  Pitfalls for Model Type

logistic regression for variable followup
Assuming OR is RR when disease NOT rare  Pitfalls for Independence observations

wrong unit analysis (BP rather than people_
Ignores tight matching
time series/growth curve data  Pitfalls for Disease variable scale (continuous)
 Need to log transform or other transform
 3 issues of covariate definitions

1) Categorical vs. continuous
2) Categories
3) Scale (if continuous)  Pitfalls for Categorical vs. Continuous
 Continuous covars when relationship nonlinear (should inspect data)
 Pitfalls for Categories
 Torturing the data (pick quintiles)
 Pitfalls for scale (if continuous)
 covariates w/ extreme variability (consider logtransform)
 Pitfalls for building models

automatic selection algorithm
colinearity
(consider including demographics and known risk factors)  Pitfalls for Additivity
 effect modification
 Pitfalls for unknowns
 failure to consider