2023/9/5 5 epidemiology

epidemiological paradigms

disease causation

causality & association (no association, no causality) (with association, not always causality) association: must discuss 1 bias 2 confounding 3 chance

genetics (can remember 1 e.g. for each type)

diagnosis & screening

click to edit

RR 1, Cl, Cl width - powerful?

The chance of people who have food poisoning from eating pork is XXX times higher than those who have food poisoning.

Pork = 4.8

The risk of food poisoning among people who have eaten pork is 4.8 times higher than who have not.

Bradford Hill criteria (almost all have: temporality)

biological gradient (threshold & ceiling effect)

bonus mark (limitation): Kenneth Rothman arguments

quality of study

bias, confounding (residual confounding, only RCT - all known and unknown confounding factors)

power (Cl width too big)

RCT vs cohort vs cross-sectional vs ecological study (no population, quicker but ecological fallacy, proxy of the exposure, only useful for getting a hypothesis and do the next study)

risk factor: increase the likelihood of outcome (but not a casality)

life-course (no critical period)

programming (David Barker), adult risk (only the 2 inner circles of social determinants of health)

click to edit

sensitivity & specificity (overlap, cutoff pt, no both 100%)

primary & secondary

screening (already have the disease, but no symptoms yet) Wilson and Jungner criteria: usually compare between diff screening, not just listing criteria (mass screening: inborn metabolism error)

e.g. post-exposure prophylaxis (not yet have the disease)

click to edit

given disease prevalence, and sensitivity & specificity

1 probaiilty to odds

2 likelihood ratio

most important: really have the disease? need to know: post-test probability = PPV

bigger area under curve, the better the test (A: perfect test, C: useless test, 50% have the disease) (every test between A and C)

implication, consider the disease (higher sensitivity: must treat, life-threatening) (high specificity: not life-threatening, lots of worry)

bias

1 volunteer effect

2 lead time bias

3 length time bias

never use survival time, use disease-specific mortality rates

solved by RCT, even out slow and fast growing cases

click to edit