
phase I
 few healthy volunteers, also terminal HIV/cancer for toxic drugs, or refractory patients
 pharmacodynamics, kinetics, toxicity, safety

phase II
 few patients with placebo if you want
 see if the drug works and the best dose

phase III
 many patients with control being best treatment/placebo
 is it better than the standard of care?

phase IV
 every drug on the market is in this stage
 delayed effects

sensitivity
 NOT affected by prevalence
 SNNNOut (Negative, rule out, increases Negative predictive value)
 TP/(TP+FN) (looking at everyone with the disease)

specificity
 NOT affected by prevalence
 SPPPIn (Positive, rule in, increases Positive predictive value)
 TN/(TN+FP) (looking at everyone without the disease)

positive predictive value
 affected by prevalence!
 probability that someone with a positive test has the disease
 TP/(TP+FP) (looking at everyone who tested positive)
 increases with specificity (SPPPin)

negative predictive value
 affected by prevalence!
 probability that someone with a negative test does not have the disease
 TN/(TN+FN)
 (increases with sensitivity (SNNNout)

positive likelihood ratio
 NOT affected by prevalence
 used to determine the odds of having the disease given a positive test
 SN/(1SP) (sensitivity on top because you have the disease, 1SP refers to false positives b/c the test was positive)
 use odds!
 LR+>1 is a test that will increase the odds of having the disease if positive
 LR+<1 a test decreases the odds of having the disease if positive

negative likelihood ratio
 NOT affected by prevalence
 used to determine the odds of HAVING the disease despite a negative test
 (1SN)/SP (sensitivity on top because you have the disease, 1SN because we are looking for false negatives)
 use odds!
 LR<1 is a test that will increase the odds of having the disease if positive
 LR+>1 a test decreases the odds of having the disease if positive

odds
 yes/no (opposed to yes/total or no/total)
 o=p/(1p)
 p=o/(1+o)

incidence
 new cases/number at risk
 can increase with a cure if the cure doesn't reduce their risk of reacquiring a disease (prevalence would be down because you are curing chronic cases)

prevalence
 number of cases/population
 can reduce prevalence by cure or increased mortality
 treatments that prolong your life with the disease increase the prevalence but do not increase the incidence

odds ratio
 casecontrol studies
 odds that the disease group was exposed divided by the odds that the healthy group was exposed
 approximates relative risk in rare disease assumption

relative risk
 cohort studies
 percent of exposed group who develops the disease divided by percent of unexposed group who develops the disease
 RR=EER/CER
 Relative Ratio

attributable risk
 difference in percent of exposed group who develops the disease and percent of unexposed group who develops the disease
 AR=EERCER
 used in number needed to hARm

absolute risk reduction
 difference in percent of untreated group who develops the disease and the percent of treated group who develops the disease
 ARR=CEREER (assuming the exposure is treatment)
 used in number needed to treat (treatments deal in ABSOLUTES!)
 used in relative risk reduction

relative risk reduction
 absolute risk reduction divided by percent of untreated group who develop the disease
 RRR=ARR/CER (assuming the exposure is treatment)
 2% of vaccinated patients get flu, 8% of unvax get flu
 ARR=8%2%=6%
 RRR=6%/8%=.75

number needed to treat
 how many patients are treated before one benefits who wouldn't have benefited without treatment
 NNT=1/ARR

number needed to harm
 how many patients are exposed before one gets the disease who wouldn't have gotten the disease if unexposed
 NNH=1/AR
 number needed to hARm

precision (aka and error)
 reliability
 consistency
 reproducibility
 repeatability
 reduces random error to reduce the SD and increase statistical power (1beta)

accuracy aka and how it is reduced
 validity
 trueness
 truthiness
 reduced by systematic errors

Berkson bias
 type of selection bias
 study population selected from hospital is less healthy than general population
 reduced by choice of appropriate control group
 reduced by randomization

healthy worker effect
 study population is healthier than general population
 reduced by choice of appropriate control group
 reduced by randomization

nonresponse bias
 those who participate are different from those who don't in a meaningful way
 reduced by randomization
 reduced by choice of appropriate control group

recall bias
 patients with disease more likely to recall exposure
 reduced by decreased time from exposure to study

measurement bias
 not using good tools or methods when assessing control AND study population
 reduced by objective, standardized, previously tested methods planned ahead of time

procedure bias
 study population or control population receive meaningfully different treatments because of participants' or researchers' bias
 reduced by blinding, placebo

observerexpectancy bias
 researchers' belief in treatment changes the outcome, like not classifying someone as having depression because you know they are in the treatment arm
 reduced by blinding, placebo

confounding bias
 a factor related to both the actual exposure and the disease is not responsible for causing all of the disease, like coal miners smoking more than the general population when looking at lung cancer
 reduced by multiple/repeated studies, crossover studies, matching, restriction, randomization

leadtime bias
 detecting a disease earlier so that they "live longer with the disease" even though they don't live longer
 reduced by measuring back end survival (survival according to disease severity)

positive outliers
 most affect mean
 least affect mode
 positive skew

negative outliers
 most affect mean
 least affect mode
 negative skew

standard deviation
 SD
 deviation from the mean within the study

standard error
 SEM
 deviation between the calculated mean and the actual real life mean
 SEM=SD/(rad[n])
 SEM decreases with a larger studied population
 used to calculate confidence interval

normal distribution SDs
 1 to +1 = 68%
 2 to +2 = 95%
 3 to +3 = 99.7%
 < +1 = 84%
 > +1 = 16%
 < +2 = 97.7%
 > +2 = 2.3%

H_{0} and error
 "null hypothesis"
 there is no association between the disease and exposure
 if H_{0} is true in reality but study indicates H_{1}, then type I error (alpha)
 alpha (usually .05) is chance of type 1 error
 aka false positive error

H_{1} and error
 "alternative hypothesis"
 there is an association between the disease and exposure
 if H_{1} is true in reality but study indicates H_{0}, then type II error (beta)
 beta is related to power power=1beta
 beta is reduced by increase in sample size, expected effect size, and precision of measurement

confidence interval
 range in which you are X confident that the real life number exists
 CI = +/ Z(SEM)
 CI of 95% > Z=1.96
 CI of 99% > Z=2.58
 for absolutes, if CI includes 0, do not reject the null hypothesis
 for ratios, if CI includes 1, do not reject the null hypothesis
 for two groups, if CIs overlap, do not reject the null hypothesis

ttest
 looks for differences between the means of 2 groups
 tea for 2
 ex: comparing mean blood pressure of men and women

analysis of variance
 ANOVA
 looks for differences between the means of 3 or more groups
 3 words for 3 or more groups
 ex: comparing mean blood pressure of 3 ethnic groups

chisquare
 looks for difference between 2 or more percentages of categorical outcomes
 chitegorical
 ex: comparing percentage of members of three ethnic groups who have HTN

Pearson correlation coefficient
 r
 between 1 and +1
 the closer to abs(1), the stronger the linear correlation is
 has nothing to do with the steepness of the slope
 1 for any negative slope
 +1 for any positive slope

coefficient of determination
 r^{2}
 usually the value reported
 the closer to 1, the stronger the linear correlation
 always positive even for negative slopes

