General Epi Flashcards
What is Berkson’s Bias?
Selection bias in case-control studies conducted within hospitals due to the manner in which risks of hospitalization can combine in patients who have more than one condition.
repeated measures
when subjects are measured at
multiple time points
cluster randomized trials
performed to assign interventions to groups of people rather than to individual subjects (for
example, schools, classrooms, cities, clinics, or communities)
If a statistical test assuming observations are independent is used with correlated observations in analyzing within-subject or within-cluster effects, what happens to the p-values?
overestimation of the P-values - decreasing the statistical power and increasing the type II error rate
If a statistical test assuming observations are independent is used with correlated observations in analyzing between-subject or between-cluster effects, what happens to the p-values?
underestimation of the P-values - increase in statistical power and increase in type I error rate
What test?
Continuous or ordinal, non-normally distributed, independent
Wilcoxon rank sum test
What test?
Continuous or ordinal, non-normally distributed, correlated
Wilcoxon signed rank test
What tests?
Continuous, normally distributed, independent
Two-sample t-test
ANOVA
Linear regression
What tests?
Continuous, normally distributed, correlated
Paired t-test
Repeated-measures ANOVA or
Mixed models or
hierarchical linear models
What tests?
Binary/categorical , independent
X2 Test (Chi-squared) or
Fishers Exact Test or
Logistic Regression
What tests?
Binary/categorical , correlated
McNemar X2 test (for 2x2 data) or
McNemar exact test (for 2x2 data) or
Conditional logistic regression or
generalized estimating equations
what is the logit function?
ln(p/1-p) - aka log odds
another name for log odds
logit
what do you use to hypothesis test in a logistic regression model
Wald and Wald chi-squared for single Betas and
Likelihood Ratio Test or testing ALL Betas
conditional exchangeability
Conditional exchangeability essentially means that, even if there are confounding variables that differ between the treatment and control groups that affect the outcome, if we only look at individuals who take a single value for that confounding variable, then the treatment assignment within each strata is “as if” random.
ITT
intention to treat - ITT analysis includes every subject who is randomized according to randomized treatment assignment. It ignores noncompliance, protocol deviations, withdrawal, and anything that happens after randomization. ITT analysis maintains prognostic balance generated from the original random treatment allocation.
compare and contrast iptw vs propensity score matching
propensity score matching discards some samples while IPTW uses all samples
propensity scores reduce the dimensionality of the data/counfounders into one metric which is beneficial, but limits the ability to analyze the effects of specific confounders
matching may not perfectly balance confounders, as there are many ways to generate similar scores
IPTW - extreme measures have outsized impacts - the tails can bias the results- matching discards these
What are the 8 assumptions for MSM’s?
- time ordering (Exposure precedes the outcome
Confounders precede exposure and outcome) - no interference (An individual’s counterfactual outcome under treatment
does not depend on others’ treatment values) - Consistency Assumption (The observed outcome is one of all possible counterfactual
outcomes) - No unmeasured confounding/conditional
exchangability/Ignorable treatment assignment - Experimental Treatment Assignment/Positivity (you need to observe all levels of treatment within each stratum of
the covariates in the real data) - Correct model specification
- No selection bias (Selection bias limits the ability to make inference to your target
population and may distort your estimates of effect) - No measurement error
How can selection bias your analysis?
This can affect the magnitude and even direction of the
estimate of effect
positivity assumption
determining if for any value of
covariates, the probability of treatment was either 0 or 1
How can you correct for a positivity assumption violation?
Drop persons who violate positivity/restrict to population in which there are both treated/untreated
◦ Limitation is that this may limit generalizability and induce selection bias
Try dropping variables from treatment model to determine if 1 or 2 variables is driving violation
◦ If variable is not a possible confounder, remove from treatment model
◦ If variable is a possible confounder, consider “coarsening” the variable
classification
◦ E.g. transform linear variable into quartiles; collapse categories/levels
of a variable
◦ Limitation is that this will may induce some residual confounding
An alternative is to use a different estimator (g-computation; double robust)
region of common support
in propensity score matching, it’s the region that includes both treated and untreated individuals - members out of this region cannot be matched
What can you do for a linear in the logit violation?
`1. transform the variable (log, square, z-score)
- discretize / categorize the variable
- convert to ordinal - same as above
If your results are not linear in the logit, what does that mean?
It means the relationship between your predictors and outcomes may be quadratic, exponential or some other form
If something is not linear in the logit, and you use the results anyway, what is likely to occur?
The model is looking for the best fit line, so it will give you that, but your confidence intervals will be huge
In logistic regression, if the residuals are non-linear, what does this mean?
Possible that predictors need to be modeled as quadratic or similar.
How do you get an overall p-value for interaction in multivariable logistic regression?
Use the likelihood ratio test
How do you get an overall p-value for interaction in multivariable logistic regression?
Use the likelihood ratio test - compare model that has interactions with a model that doesn’t
What % difference in Beta coefficients does Kristin recommend to identify confounding or interaction?
10% diff in B coefficients
Can p-values be used to determine if something has an interaction or is a confounder?
NO! Only Beta coefficients/OR effect sizes.
What is residual confounding?
Confounding that remains even after adjustment
What’s the general effect range (in OR) that can sometimes be attributed to residual confounding?
.6 to 1.6
How many backdoor paths are there in this picture?
2
What sets can you control on to block the backdoor paths?
V
VZ
ZW
VZW
What is wrong with controlling for Z here?
It’s a collider, so it opens up a path from W to V
Chi squared formula
