NEW
Font size
WorksheetsM235 Module 5 Trivia
Total questions: 10
Worksheet time: 10mins
What can we say about the two key assumptions we make when performing causal inference with observational data?
Overlap is an assumption that we can assess by looking at the data
Unconfoundedness is an assumption that we can test for
Balance of covariates is a key assumption necessary for inference
All of the above
What assumptions does sensitivity analysis attempt to examine when we are performing causal inference with an observational study?
Tests for whether the unconfoundedness assumption is violated or not
Assesses how sensitive the results are in regards to how much overlap there is
Evaluates how sensitive the results are when the unconfoundedness assumption fails in certain ways
Attempts to balance the covariates for more efficient inference
What is not one of the four sensitivity parameters that Rosenbaum and Rubin vary in their model-based sensitivity analysis? Let U = one unmeasured binary confounder.
Effect of U on treatment probability
The probability that U = 1
Effect of U on the outcome
Weights on each observation
How does Rosenbaum’s bounds approach to sensitivity analysis use less assumptions than Rosenbaum and Rubin’s model-based sensitivity analysis?
It avoids making assumptions on how the unmeasured confounder affects the outcome
It avoids looking at the association of the unmeasured confounder and treatment
It uses randomization inference framework to avoid distributional assumptions on the outcome
All the above
Which of the following is not part of the three-factor decomposition in Tukey’s factorization when using Bayesian sensitivity analysis?
The observed data distribution
The propensity score model
The marginal distribution of the observed outcomes
The selection function
What is the E-value?
The minimum strength of association required for a confounder to explain away an observed treatment effect
The probability of the unmeasured binary confounder being one
The amount of bias in our treatment effect that results from having an unmeasured binary confounder
The model-based estimate of how much we want to set the association between the unmeasured confounder and treatment probability
What justification is there for assuming there is only one unmeasured confounder that is binary in model-based sensitivity analyses?
Even if there are more than one unmeasured confounder, you can consider U to be a scalar summary of the other confounders (like propensity scores)
Causal conclusions are more sensitive to unobserved binary covariates than continuous ones
Sensitivity analysis is already a secondary analysis, so overcomplicating it with additional assumptions may distract from the main analysis
All of the above
When using causal trees, which variables should be used for splitting when constructing the tree structure?
Covariates, X
Outcomes, Y
Treatment indicator, W
All of the above
What is not an advantage of causal trees?
Good at filtering out noise, so less prone to overfitting
Can handle non-linearity of covariates
Automatically captures covariate interactions
Can easily separate tree construction from treatment effect estimation
When might you want to use causal forests over causal trees?
When you want to be able to identify clear and interpretable subgroups
When you want to directly estimate CATE as a function of the covariates
When you want to estimate the CATE for specific, pre-specified subgroups
When you have high dimensions and don’t want your CATE estimates to be biased
