Introduction to Epidemiology Flashcards
Matching?
restricts enrollment within comparison group: subjects in comparison are chosen so they have same distributiaton of matching factor in index group
een van drie manieren voor voorkomen confounding (randomization, restriciton).
implication matching :
1) cohort oz
2) CC oz
1) unexposed have same distribution as exposed (matching factor geen cf-er meer).
2) controls same distribution of maching factor as cases). NB!! matching alleen niet genoeg, ook matching in analyse toepassen.
doel matching
1) avoid confouding
2) improve efficiency of analyses (minder pt nodig).
matching in cohort oz op geslacht
1) for elke E+ man, zoek een E- man
2) for elke E+ Vrouw zoek een E- vrouw.
3) result: distributie gelijk in E+ en E- groep
4) . kan individueel (one-to-one) als frequency matching.
RR crude = RR adjusted!
matching CC op geslacht
1: for elke Case man, vind controle die man is.
2: for each case that is female, find controle who is female
3: same sex distribution in case and control
matching in CC oz: hoe kan na matching sex nogsteeds cf zijn?
controls not sampled independently of exposure (selection bias).
Dus: corrigeren in analyse op matching factor.
ratio matchen in CC onderzoek:
improves efficiency of stratified analyses (je heb tiig zowel cases als controls within strata of matching facotr).
vb: prostaat kanker oz, zorgt age matching dat je voldoende controls hebt, omdat cases waarschijnlijk ouder zijn
Overmatching
kan bias veroorzaken of study minder efficient.
Defintion effect modification (ookwel INTERACTIE in statistiek)
Exposure has differente effect on outcome in differen tgroups of subjection
effect mod & effect measure
Afhankelijk welke effect measure je neemt is er wel of geen effect modification (bv risk difference vs risk ratio).
detection of effect mod: (3)
- most studies underpowered
- must distuingish from other reasons (bias, cf by other factors, chance)
- need for clinical or biologic rationale
Ruling out chance variation 2:
(1) test of homogeneity (breslow-day test, regressie)
(2) Examine stratum-specific confidence intervals:
Rapporteren bij effect modificatie:
- per stratum specific result (best option
- weighted average of stratum specific estimates (Mantel-Haensze; or regression
- standadization (average effect of exposure in a standard population.
Standardiseren in 3 stappen
1: select standard population
2: stratifcy by confounder (age)
3: calculate risks of outcome within levels of confounder for EXPOSED AND UNEXPOSED
Standardization: wat voor populatie?
1: internanl population (full dataset)
2: external population: take common set of rates and apply to popultion to be standardized.
Redenen voor regressie in epidemiologie?
To control for confounding
to predict outcomes
Motivatie propensity score analysis
1: control for large number of confounder (combine individual confounders ito a summary score (propensity score)
propensity score =
probability of receiving the treatment as a function of the confoudners.
3 steps in propensity score analysis:
Estimate propensity score using logistic regression (P(prob o receiving treatment)
2: use PS to balance observed covariates
3: estimate differnece in outcomes between treatment groups.
Sensitiviteit
Specificiteit
sens: P(Test+|D+) true positive rate : a/ a+c (verticaal)
spec: P(T-/D-) = 1 - false positive rate
b / b+d
Pos pred val:
Neg pred val:
P(Disease+| T +)
a/ a+b (horizontaal)
P(D- | T -): d / a+b
Wat is afhankelijk van prevalentie van ziekte?
PREDICTIVE VALUES!
prevalentie = D+ / Dtot
diagnostic tests vs clincal predication rules:
Diagnostic tests predict: currect state of health
Clinical predication rules predict FUTUre state of health.
Performance measures: discrimination
Disc: ability to separaate subjects in different outcome states
ROC CURVES / classification tables:
ROC curve
ideal curve
C statistic?
Plots sens (true positve rate) vs 1-spec (false postive rate).
ideaal: punt in li boven hoek (area close to 1.0
AUC (C-statistic) = probability that prediciteve vlaue for a subject with outcome is > than predicited value for a subject without outcome.
Callibration:
Pertains to agreement between est risk and actual outcome.
Method of assessment: resubstitution
testing set
resampling methods:
use same data to build model and evaluate its performance ( cave: overtraining!)
2: splitsen training en validatie set
bootstrapping / crsoov validation
bootstrapping:
resampling with replacement as many times as individuals in original data.
Build model on bootstrap sample (training set).
evaluate performance onf original set (testing set)
cross validation
Divide orginal data in 10 disjoint parts
Each part used as testing set for model for remaining 90%
Average performance of 10 models on 10 testing sets:
u
Use this average as estimate of performance