stats:study_designs

see also Medical statistical analysis

- investigator assigns at random the independent variable (eg. drug or vaccine) & examines the effect on the outcome or dependent variable (eg. % pts surviving relevant illness - case fatality; or incidence rate of the relevant disease);
- better order of evidence of causality than both cohorts or case-control studies as more able to manipulate exposure to factor of interest & tend to have less confounding influences (3rd variable related both to exposure & to outcome of interest)

- similar to the above (incl. wrt sample size, cost, time taken to obtain statistically unambiguous answer), but allocation to the exposed & non-exposed group is not a random event;
- thus useful for studying problems where randomisation not possible such as in inherited susceptibility to disease;

- prominent since 1950's;
- usually offer potential advantage of small sample size, cost & duration over cohort studies;
- compare 2 samples - one with disease & one as control without the disease, then determine & compare the exposure status of each each sample;
- measure using odds ratio:
- (cases exposed H ctrls not exposed)/(ctrls exp. H cases not exp)
- if > 1 then exposure increased risk of disease;
- if 95% confidence interval overlaps value of 1 then chance but cannot definitely reject the hypothesis;
- if < 1 then exposure decreased risk of disease;
- then need to use statistical tests to see if any observed increase or decrease in odds ratio from unity is statistically significant - ie. is sample size sufficient to rule out chance for that result;

- controls must be representative of population from which cases arose:
- volunteers may not be representative;
- relatives/friends may have similar lifestyles, jobs, or residential histories as the cases;
- staff of hosp/lab. are also usually unsuitable;
- ⇒ ideally:
- controls would have entered same hosp. or practice had they developed the disease;
- eg. use of electoral rolls/ GP age-sex registers; random telephone dialling;

- recall bias - likelihood that cases will think more deeply & carefully about their previous experiences than will controls to “explain” their misfortune of their illness thus may need to add a third control group consisting of patients with similar but unrelated disease;
- ⇒ ideally:
- should have detailed exposure records both groups taken at time of exposure;

- intervewer should be blind to case/control else may contribute to systematic difference b/n the groups (eg. increased questioning of cases);
- may need to control for confounding variables either by stratification; logistic regression; or by matched case-ctrl study;

stats/study_designs.txt · Last modified: 2008/11/03 17:52 (external edit)