2. Arbitrary p-value cut off3. No information about range of effects consistent with observed data (only with null)4. No information about the power1. 100(1â€Î±)% of the time, it will include the true value of the parameter of interest2. gives the range of values of the parameter of interest that would not be rejected at the3. sense of the magnitude and direction of the parameter of interest that are consistent wNarrow = powerful studyWide = weak studythere is no difference between the treated and untreated groups with respect to the outcoa. unmeasured confounders orb.confounders that were measured but not adjusted for statistically

89238836a347e7e6e4bfddd6ab949c044ea64996Page 6H0, Ha, measure of association, definition of variables, assumptions.assuming no selection bias, residual confounding, or information bias.X was 72% higher among smokers compared to non-smokers (assuming no confoundingratios are inherently assymetricratios are inherently assymetricadd in more controls!information = 1/variancethe MH is an unbiased estimatorthe test statistic for stratified observational studiesall of the effect measures get equal weight across strataweight by the amount of Tp in each strata, even though the average is different, it is still because the number of cases is different across strata, person-time may not be the mosweight by amount of information, will be most efficient estimate with narrow intervals, andall of the weighting techniques are unbiased estimates, but the inverse variance and mansum(ai*Noi/Ti)/sum(bi*N1i/Ti)For stratified open cohort studiesthe estimated variance of ln (ÃŽRRmh)no - because the sample is not representative (we're not controlling for confounding)true - but WHY?effect measure modificationSUM[ln(Stratum-specific estimate of the effect measure) - ln(MH summary estimate assuwe can collapse the sparse data into one 2*2 or add .5 to each zero-sumdo not use a weighted summary; report the observed association at each level of the moprobability of being truly exposed and classified as exposed; critical when the majority haprobability of being truly unexposed and classified as unexposed; critical when minority herror in a continuous determinant (e.g. blood pressure, etc.)misclassifyication depends on the errors in measuring or classifying other variables - i.e.misclassification is not dependent on the probability of errors in classifying other variableInternal validity: The extent to which a particular variable, rather than extraneous Â Â -Threats to internal validity: maturation, effects of testing, subjectselection biases, third variablesThe degree to which findings can be generalized to other people, influences, accounts fÂ Â - Threats to external validity individual characteristics of participants, setting, and tthese data are consistent with rate ratios ranging from 1.39 to 2.12Â with 95% confidencthe TIMELINE!

#### You've reached the end of your free preview.

Want to read all 8 pages?

- Summer '14
- FrancisCook