EBM Flashcards
Type I Error
FP; we reject Ho when we shouldn’t (if Ho is not false)
There’s no relationship but you found one.
Smoking does not cause cancer but you found that it did.
Type II Error
FN; we fail to reject Ho when we should
There’s a relationship there but you didn’t find one.
Smoking does cause cancer but you don’t find more cancer in smokers.
Alpha
Willingness to reject the null hypothesis when we shouldn’t (type I error); willingness to be wrong.
There really is no association but our results show an association just by chance.
Beta
willingness to make type II error (failing to reject false Ho)
Ho
Null hypothesis (no association)
HA
Alternative hypothesis
Palpha
probability of a false positive study
if Pa=0.05, there is a 5% chance we will report a difference that occured by chance alone
Pbeta
probability of a false negative
if Pbeta= 0.10, then 10% chance we will fail to detect a difference that is real
Power
ability to detect or verify a difference that is real
Power = 1 - Pbeta (or 1-beta)
Ability to avoid a type II error
If you reject Ho, then by definition you cannot lack power (even if n is small)
What to consider when Ho is not rejected
- likely that Ho is true
- suspect type II error (P >.05 but <.10)
- consider that the study was underpowered
- small n, esp if P is close to 0.05
What to consider when Ho is rejected
- consider possability of type I error, esp if P is close to 0.05
- if P<.01, be confident it is not a type I error
T-test
- compares the difference between 2 means
- divided by the variability in the two samples
- parametric: relies on parameters of the distribution (assumes equal variances)
t=
meanA - meanB / (varA + varB).5
df
df= nA-1) + (nB-1)
critical value
- using alpha and df to determine
- if t is calculated as les than critical value, we fail to reject Ho
-
ie: alpha=0.05, df=8, critical value is 2.30
- if t<2.30, we fail to reject Ho
Mann Whitney U test
- non-parametric
- distribution free
- ignores the mean and the median
Relative Risk
a/(a+b)
_______________
c/ (c+d)
Differential (regarding error)
misclassifies subjects differently depending on exposure and/or outcome - biases result in one direction or the other
Nondifferential
tends to bias results toward the null
can be systematic or random error
Confounding Variable
Must be associated with independent variable and the dependent varible; cannot be in the causal pathway btwn exposure and disease
Which designs would you use to answer the fundamental questions of:
- prevalence?
- risk of harm?
- treatment or prevention?
- prognosis?
- screening?
- prevalence? cross sectional
- risk of harm? cohort, case-control
- treatment or prevention? RCT, Cohort, Case-control
- prognosis? Cohort
- screening? RCT, Cohort, Case-control
Disease Prevalence
- point prevalence: instantaneous time period
- period prevalence: longer time period
Prevalence =
(# of ppl with diease at a given pt in time)/
(total # of ppl in pop)
- often called prevalence rate but time not in denominator
Disease Incidence
- how quickly ppl are being diagnosed with the disease
- considers only new cases
cumulatie incidence =
(# of new cases) / (# of ppl at risk of developing the disease over a defined time)
Incidence Rate
rate at which new diesease has occurd in the pop at risk per some time unit
Relative Risk (RR)
(Incidenceexposed) / (Incidenceunexposed)
risk ratio
Cohort Studies
Criteria:
- They do not have diease in question at time of assmbly
- Should be observed over a meaningful pd of time in the natural history of the diesease in question.
- Members of cohort should be observed over the full period of the followup.
- Upon entry into the study, ppl in the cohort are classified according to characteristics (possible risk factors) that might be related to outcome.
- Classified as exposed or not-exposed
- aka: longitudinal, prospective/historical/retrospective, incidence studies
Case-Cohort Studies
- all exposed people are included in the study and followed for some outcome of interest.
- however, only a small random sample of unexposed ppl is studied (group of unexposed is “enriched” with those who subsequently suffer the outcomes of interest)–results then adjusted to reflect the sampling fractions used to obtain the sample
- need large, computerized, medical database