Interpreting Epidemiological Findings (2) Flashcards
What is confounding?
The effect of an extraneous variable that wholly or partially accounts for the apparent effects of the study exposure, or that masks an underlying true association
a confounding variable is a third variable which leads to bias in the estimate of association between outcome and exposure
What can confounding lead to?
Biased estimates
How can you identify confounding?
Knowledge of the subject matter: you could undertake an evidence review and see what other people have proposed or found
Test by the three criteria: associated with exposure in the source population; associated with the outcome in the absence of the exposure; not a consequence of the exposure
Stratification: we can look at the difference of apparent effect within different population strata
Compare the crude and adjust statistical estimate
What are the three conditions for confounding?
Associated with the exposure in the source population
Associated with the outcome in the absence of the exposure
Not a consequence of the exposure
How do you conduct stratification?
Compare stratum specific estimates with the estimate you get when you analyse the date from the study
What after conducting stratification might indicate confounding?
Pooled estimate is considerably different from what you expect from stratum specific estimates
What is effect modification?
Where the magnitude of the effect of an exposure variable on an outcome variable differs depending on a third variable
- not a problem. It is a natural phenomenon.
Effect modification – exists when the strength of an association varies over different levels of a third variable
Shouldn’t be controlled for + happens naturally
Take into account when reporting
When detected – conduct stratified analysis
How can you test for effect modification?
- Breslow-Day test
- Q test
- Interaction terms in regression models
What is synergism in relation to effect modification ?
The effect modifier potentiates the affect of the exposure
What is antagonism in relation to effect modification?
The effect modifier demeans the effect of the exposure
What are adjusted models used for?
Identification of potential confounding
And used to account for it
What is a crude model?
Univariate analysis of exposure vs outcome
What does the crude model do?
Simply looks at the impact of the exposure on the outcome – with no consideration of anything else
What is an adjusted model?
Multivariate analysis of a range of exposures vs. outcome
What is multivariate analysis?
Multiple potential exposures have been included
The inference is that the outputs of these analyses mean that holding all other adjusted variables equal, X is the association between exposure and outcome
What are other ways of naming crude vs adjusted models?
Hazard ratio (HR) vs Adjusted Hazard Ratio (Adj. HR) Odds ratio (OR) vs Adjusted odds ratio (AOR) Univariate model vs multivariate model
Going back to the ICSM Lifestyle Tracking Study last year, 20% of the participants reported using a smart watch.
Do you think this is representative of the first-year student body overall? Why?
Probably not. People with smartwatches may differ systematically from those who do not: it may be that they’re more interested in lifestyle and health behaviours. They may also be more comfortable and confident with technology. This is an indirect form of selection bias.
Do you think participants with a smartwatch (the exposure) will report more, less or the same number of steps (the outcome) on average than those participants without a smartwatch? Why?
I’d hypothesise that participants with a smartwatch will report more steps than average for two reasons:
Measurement error among those without smartwatches: The smartwatch is always worn and therefore will capture more steps than smartphone which may be left behind during physical activity. This is an example of differential misclassification.
Actual differences: People with smartwatches are more likely to be physically active: and this might be the driver of them buying a smartwatch, or the smartwatch may cause them to be more physically active.
Interestingly the actual results showed the precise opposite which is that those with smartwatches on average reported less time being physically active than those without smartwatches. It’s possible that people without the technology over-estimate their physical activity.
four ways to potentially identify confounders
- Knowledge of the subject matter: you could undertake an evidence review and see what other people have proposed or found.
- Test by the three criteria: associated with exposure in the source population; associated with the outcome in the absence of the exposure; not a consequence of the exposure.
- Stratification: we can look at the difference of apparent effect within different population strata.
- Compare the crude and adjusted statistical estimates.
study showed birth order associated with downs syndrome..what is the confounding variable
- maternal age is the confounding variable.
- Further analysis (shown in the 3D column plot) demonstrates that actually birth order is not associated with Down’s if you account for maternal age.
Study showed alcohol consumption associated with lung cancer…what is the confounding variable?
- alcohol is associated with tobacco consumption = tobacco consumption is associated with lung cancer - people who drink are more likely to smoke
confounder: cigarette use.
Several observational studies have demonstrated an association between playing games of cognition (such as Sudoku) and reduced risk of dementia in later life. what is the confounding variable?
It’s possible (indeed likely) that people who choose to undertake cognitively-stimulating exercises may be people who are at lower risk of dementia in many other ways: through educational level, occupation or other socioeconomic characteristic.
What do you think is the effect modification of the following:
In the Million Women Study,1 post-menopausal women were observed in two groups: those using continuous combined hormone replacement treatment (HRT) and those not using HRT. The overall estimate suggested a risk reduction for endometrial cancer of approximately 30% for women using HRT.
breaking down the results by BMI category, the association was only observed among women with a BMI ≥30kg/m2. There was no apparent difference for women in the other BMI groups.
BMI effect modifier
In a study of alcohol consumption and associated harm, what do you think is the effect modifier
A Scottish research group has identified socioeconomic status as a powerful effect modifier. Irrespective of alcohol consumption (volume or behaviour), socioeconomic status seems to be the most influential determinant of coming to alcohol-related harm. This is described by some as the alcohol harm paradox: even though more-affluent people drink more on average than those less-affluent, it is the socioeconomically more vulnerable who suffer the greatest harm.
