Final: Lectures 15-16 Flashcards
To be a confounder, a 3rd variable must be:
•Associated with the exposure AND the outcome of interest, yet independent of both
Confouding variable (a.k.a. Confounder)
•A 3rd variable that DISTORTS an observed relationship (assoctiation; RR/OR/HR) between the Exposure and the Outcome (disease)
Confounding stated other ways:
- A mixing of the effects; an association (between exposure and outcome) is DISTORTED due to them being mixed with another (3rd) factor which is associated with the outcome of interest at the same time.
- OR….
- Confusion of effects, where the effect of exposure is DISTORTED b/c the effect of an extraneous (3rd) factor is mistaken for the effect of exposure
- Can be over- or under-estimate an association and can change the apparent direction of an effect
Impact of Confounders (2 main ways)**
- Intesity/Magnitude/Strength: produces an estimate that is more or less extreme than the true association
- Direction: produces an estimate that moves the true association in a + or - direction (towards or away from the null associational (RR/OR/HR=1)
3 Steps of testing for Confounding
Step 1: calculate crude outcome measure of association (OR/RR/HR), commonly called “Unadjusted” association (ratio between just exposure and outcome)
Step 2: Re-calculate outcome measure of ass. while mathematically controlling for Confounder, removing effect of confounder from ass measurement between exposure and outcome of interest, commonly called “Adjusted” association (if OR doesn’t change, isn’t confounder)
Step 3: Compare the two measures of association, the point estimate of the ass. will be different by 10-15% if the IS confounding present, Mantel-Haenszel Test (of OR’s)
Purpose of controlling for confounders:**
•To get a more precise estimate of the true association between the exposure and the disease/outcome
Ways to control confounding
- Study design stage: randomization, restriction, matching
2. Analysis of data stage: stratification, multivariate or matched statistical analysis
Randomization
- Hopefully allocates an equal number of subjects with the known confounders into each intervention group
- Strength: with sufficient sample size (N), will likely be successful in serving its purpose (making groups “equal”)
- Weakness: Sample size may not be large enough to control for all known and unknown confounders, doesn’t guarantee successful, equal allocation, practical only for interventional studies
Restriction
- Study participation is restricted to only subjects who do not fall within pre-specified categories of confounder
- Strength: straight forward, convenient and inexpensive, doesn’t negatively impact internal validity
- Weakness: reduced sample size, if restriction criteria not sufficiently narrow it will allow the introduction of residual confounding, eliminates researchers ability to evaluate varying levels of the factor being excluded, negatively impact external validity (generalizability)
Matching
- Study subjects selected in matched-pairs related to the confounding variable to equally distribute confounder among each of the groups
- Strength: intuitive, some feel it gives greater analytic efficiency
- Weakness: Difficult to accomplish, very time consuming and expensive, doesn’t control for any confounders other than those matched on (over-matching possible, this will mask findings)
Stratification
- Statistical analysis of the data by evaluating the ass. between the exposure and disease within the various strata within confounding variables
- Strength: intuitive, straight-forward and enhances understanding of the data
- Weakness: impractical for simultaneous control of multiple confounders, especially those with multiple strata within each variable being controlled
Multivariate Analysis
- Statistical analysis of the data by mathematically factoring out the effects of the confounding variables
- Strength: can simultaneously control for multiple confounding variables, in Logistic Regression, beta coefficients can be directly converted to OR’s
- Weakness: Process can cause some individuals to not clearly understand the data, Ex. Linear & Logistics Regressions and Cox Proportional Hazards Modeling
Effect Modification (a.k.a. Interaction)
- 3rd variable, that when present, Modifies the Magnitude of effect of an ass. by varying it within different levels of a 3rd varaible (effect modifier)
- If an interaction is present, the researcher MUST report the measures of ass. for Each strata individually
- So, unlike confounding, EM should be described and reported at each level of the variable, rather than controlled for (changes layers/doses of that variable)
The most important thing to look for in Effect Modification is??
- LAYERS*
- EM is present b/c the OR changes substantially according to the different layers (strata) of the effect-modifying variable (birth weight)
Testing for Effect Modification
- Stratum-specific estimates are compared directly to see if they are different
- Point estimates (RR/OR/HR) for the ass. will be different by 10-15% between the lowest and highest strata if there IS effect modification present
