EBM Day 2 Flashcards
Statistical Hypot
is there difference between groups,
Ho
no association between x and y
Fail to reject means
Reject null
fail to reject, never prove hypo because may be due to the 5% chance
consider possibility of type 1 error (unless p
statistically sig means
result where reject Ho at whatever alpha level we set
type 1 vs type 2 errors
t1-fail to reject null when true
t2- reject null when false
T1 and T2 in releation to power, alpha, and beta
and what is alpha, beta, power
alpha-willingness to be wrong (reject null when we shouldn’t)
beta-=t2 error=willingess to fail to reject a false Ho (typically .1 or .2)
Power-1-beta=power to correctly reject false null (80%)-ability to detect or verify difference is real
Sensitivity and Specificity
sensitivity=true positives/(true positives and false negs)
specificity=true negatives/(true negatives and false positives)
power determination
alpha (more stringent, less power), beta (too lax, less power), prevalence of condition, magnitude of effect, sample size (more subjects make more power)
Effect size
how big of a difference we look for
smaller effect size=larger the sample size needed
Bonferroni adjustments
adjusting for multiple comparisons-increases type 2 error and decreases power
T test
-compares difference between two means divided by variability in sample
assumes equal variances
Mann Whitney U
Non parametric-operates on ranks
ignores mean and median
Minimize T1 and T2 at same time?
there is always a tradeoff
Tolerate type 1 if false positive okay
Tolerate type 2 if if procedure may be serious danger to patient
Increase power
lower beta, raising alpha, raising sample size, testing a large difference
2x2 table
Draw=THERE ARE QUESTIONS AT END OF HIS SLIDES
Risk and how offset
Probabilty of an outcome
Offset by intervention, treatment, prevention
Primary, secondary, tertiary prevention
before disease
Catching early
treatment
Pathogenic Triangle
host, environment, agents
Risk factors
Anything which increase likelihood of disease
ex. other disease, environmental, genetic
Environmental risk factors (5) and examples
chemical (oxidation substances), physical (radioactivity), biological (pathogens), psychosocial (stress), mechanical (heavy lifting)
2x2 of Karaoke job Decision latitude
low demand, low decision-passive
low demand, high decision-low -strain (scientist)
high demand, low decision-high- strain
high demand, high decision-active (doctor)
Latency
Period between exposure and disease
Cohort study main question
Asking if incidence of an outcome in a group who were exposed different (greater/less) compared to incidence among similar group who were unexposed?
Cohort study result
Get incidence rate of exposed
Get risk ratio, relative risk/difference when compare
Picking a cohort
Should not have outcome when picked
All should be at risk for outcome
Should be observed over natural history of disease
Observe over entire time of disease
How to do cohort study
Find population where everyone at risk for something
Divide people into two groups depending on exposure to risk factor
Cohort study other names
incidence study longitundinal study prospective study retrospective historical
Retrospective vs Prospective Cohort
Prospective takes much longer time, assemble cohorts in present, choose which risk factors/confounder to measure, chose how to measure
Retrospective may not have all data you need, cohorts assembled in past (by medical records), outcome accessed at later date
Relative Risk
a/(a+b)/c/(c+d) draw
Type 1/2 and false … error
- false postive
2. false negative
Causes of error
Chance-nondifferential-random error-type 2 error
Bias-can be differential or non differential-type 1 error
Confouding
Differential vs Nondifferntial bias regarding direction
Differential-towards one direction or another
Nondifferential-towards norm
Cross sectional study
Measure disease and time at same time
Confounding Variable
Associated with DV and IV but not in pathway
Can cause t1 or t2 error
Confounding by indication (and error associated with it)
Sicker patients are more likely to be treated and to have worse outcomes
Increase T1 error
ex. use of drug for really sick people associated with increased mortality because people are really sick (even if it helps)
Decrease confoundng
Randomization-distribute potential confounders between groups
Restriction-restrict a confounding variable during study duration (lower sample size and power though)
Matching-match with people of similar characteristics
Stratifcation-data separated by potential confounder-if confoudner present-risk ratios lower than in unstratified data (risk due to confounding no difference between strata)
MV adjustment-control effects of many variables simulataneously
Effect modification
effect mods-variables that change effect of exposure of interest on risk of disease
AKA- interaction
One exposure effects other exposure
Selection Bias
Selective differences between comparison groups that impact relationship between exposure and outcome
Selection bias examples (4)
healthy worker effect
Self selection bias
Withdrawal bias (primarily cohort studies)
Information bias-Investigators who know exposure status may be more or less likely to ascertain the outcome (diagnostic bias)
Fix for info bias
BLINDING of all personnel, investigators, and subjects to the exposure status of the subjects
CAN ONLY FIX BIAS IN DESIGN STAGE
or after with mv but then must measure bias
Cohort Studies Adv (4)
good when exposure is rare,
can minimize selection and measurement bias,
can directly determine incidence rate and risk,
can look at multiple outcomes from single exposure
Cohort studies Disadv (5)
inefficient for rare outcomes Needs large sample long time to complete Loses to follow up Expensive potential ethical issues
What test should i do for prevalence?
Cross sectional
What test should i do for risk of harm
cohort, cross sectional
tWhat test should i do for treatment or prevention
RCT, Cohort, case-control
prognosis
cohort
screening
rct, cohort, case control
