SME: Case-control studies Flashcards
What do you calculate in case-control study in your analysis?
Measure of association ie OR
What can you NOT calculate in case-control study in your analysis?
Incidence rate
Incidence risk
We do not usually know the sampling fraction for either cases (SD) or controls (SH), therefore we cannot calculate absolute risks or rates
Goal of interpretation of case-control studies, ie what conclusion are you trying to draw from your analysis?
Exposure increases/decreases/keeps same the risk of outcome
How do the OR and RR compare in rare diseases?
OR very similar numerically to RR (rare disease assumption)
For common diseases, what do measures of OR estimates depend on?
How controls are sampled
What is the null hypothesis vs H1 in case-control studies?
Null: OR=1
H1: OR NOT equal to 1
What is a key statistical feature of case control studies?
Cases and non-cases have different probabilities of being selected for inclusion i.e. SD ≠ SH
What does a OR vs a p value measure?
OR: magnitude of association between risk factor and outcome
p: strength of evidence against null hypothesis
Pros and Cons of Case Control Studies
Advantage: Efficient, Quick, Cheap (especially for rare diseases, diseases of long latency)
Disadvantages: - Susceptible to selection bias and recall bias
- Can only estimate relative measures of disease frequency (e.g. odds ratio)
Odds ratio for exposure
Odds of exposure in disease/Odds of disease in unexposed
Odds of disease
Risk/1-risk
Which in CC studies is worked out as
number of cases/number of persons without disease
What are the three ways you can sample controls for case-control studies?
- Exclusive sampling: from individuals still at risk at end of study period
odds ratio - Inclusive sampling: from individuals at risk at start of the study period
risk ratio - Concurrent sampling: from individuals still at risk when each case occurs
rate ratio (implies matching on time)
matched analysis
(unless time not associated with exposure)
Odds ratio is gained from what type of sampling from controls
Exclusive sampling
OR can be interpreted as a risk ratio based on what sampling from CC studies?
Inclusive sampling
OR in CC studies equation?
Odds of exposure in cases/odds of exposure in controls
Interpret as what are the odds of outcome amongst those who have been exposed compared to those who haven’t been exposed
How to approach potential confounders?
1) Stratify by the potential confounder
2) Analyse each stratum as a separate table -> calculate stratum-specific ORs
Can then either:
- Pool ORs back together again to obtain summary OR (using MH weighting)
or
- Present stratum specific ORs
Definition of residual confounding? How can it arise?
Some confounding effect of a factor left over after we have tried to control for it
Can arise if within each stratum of the confounder, individuals are not similar with respect to the confounder:
- categories too broad for confounding factor
- underlying confounder has been measured by a proxy factor which is imperfect
- misclassification of information on the confounder
How to obtain summary OR for pooled stratum ORs?
ORmh = w1OR1 + w2OR2 + W3*OR3 / w1 + w2 + w3
where Mantel-Haenszel weight for a given stratum is given by
(number of unexposed cases) x (number of exposed controls) / total no. of cases and controls
Gives point estimate of adjusted OR for association of exposure and outcome having adjusted for confounder
Then need to:
- Calculate CI
- Test null hypothesis eg ORmh = 1
Define interaction
When the association between exposure and outcome genuinely depends on another factor
- can help understand the processes underlying association between exposure and outcome
- if there is evidence of important interaction (after controlling for all confounders), present stratum specific ORs
- the test for interaction have low power -> a high p value does NOT mean an interaction is not present
How to approach ordered categorical variables to assess association between exposure and outcome?
1) Define a baseline level for exposure
2) Calculate OR of outcome for each non-baseline level compared to baseline level
POSSIBLE TESTS
1) Use standard chi-squared test to test if general association between exposure categories and outcome (ignores ordering of categories)
2) Use trend chi-squared statistic to look for systematic increase or decrease (trend or dose-response relationship) in odds of disease for increasing exposure levels
3) Test for departure from trend
Features and meaning of chi-squared test for trend
Xi squared test for trend is looking for straight line relationship
1) assign score to each exposure level eg 1,2,3,4 etc
2) Null hypothesis: no association between outcome and exposure level, alternative hypothesis: linear relationship between log(odds of outcome) and the score of the exposure level
Ie for each unit increase on the exposure score scale is there a constant increase on the log odds scale
In real life that means the odds of outcome is multiplied by constant factor
3) Comparison of mean score between cases and controls - gives X2 statistic to 1 DEGREE OF FREEDOM (is ALWAYS 1 degree)
ISSUE:
- other relationships can also give small p value for trend, ie you can have a curve and it will still output a small p value, so need to check that relationship is approximately linear first, otherwise might want to use the test for departure trend instead