Epi Comp Exam

Question 1

Transportability

Answer 1

taking a causal effect from one target population and extrapolating the findings to another non-overlapping target population

Question 2

Transportability Assumptions

Answer 2

• Positivity: each person in the target populations have a non-zero change of receiving each level of treatment, including subgroups (effect modifiers); therefore there must be similar distributions of effect modifiers in the target populations
• Consistency: same definition and distribution of treatment (and therefore expected outcomes) in the target populations
  -No interference (SUTVA): treatments in the original study population do not interfere with each other

Question 3

Generalizability

Answer 3

taking the causal effect from the study population and applying it to the wider target population

Question 4

Generalizability Assumptions

Answer 4

• Sample representativeness/No selection bias: the study sample is representative of the target population in terms of distributions of characteristics
• Random Sampling: proportions of participants in the study sample are not over or under represented or excluded

Question 5

Difference between Generalizability and Transportability

Answer 5

-Both are concepts of external validity where we are generalizing the causal effects from a study population to target populations
-Generalizability implies we are applying the causal effects to completely overlapping populations, where the study sample is a small part of the larger target population
-Transportability we assume that the amount of overlap between the target populations of interest is little to none

Question 6

RCT

Answer 6

experimental study design where participants are randomly assigned to a group receiving a treatment/exposure of interest or a comparison group receiving a placebo, no treatment, or standard treatment

Question 7

EMM in RCT applied to transportability/generalizability

Answer 7

-There must be a similar distribution of effect modifiers in the study sample to generalize the causal effects of the RCT to its target population
-There must be a similar distribution of effect modifiers in the study sample (and therefore the target population) to transport the causal effects of the RCT to a second, non-overlapping target population

Question 8

Selection bias in RCT applied to transportability/generalizability

Answer 8

-if there is type 2 selection bias, causal effects of a study population cannot be generalized to the target population of interest
-Type 2 selection bias: study sample differs from the target population
-This means that the causal effect of the study population (and therefore the target population) would be biased if transporting to a second non-overlapping population due to different effect modifier distributions, can arise from differential loss to follow up

Question 9

External validity in RCT applied to transportability/generalizability

Answer 9

Causal effect of a study population is externally valid if the causal effects of the study population can be generalized to its target population, and transported to a second non-overlapping population

Question 10

Mechanism of action in RCT applied to transportability/generalizability:

Answer 10

-MoA must be the same in the study sample and the target population to be generalizable
-MoA must be the same in the second non-overlapping population to be transportable

Question 11

External validity

Answer 11

applying causal findings of a study population to the target population of interest

Question 12

Mechanism of Action/Mediation

Answer 12

a variable in the causal pathway between an exposure and an outcome; can be thought of as the mechanism of action that the exposure works through to have a causal effect on the outcome

Question 13

Effect measures that can be generalized

Answer 13

-Conditional outcomes (outcomes given specific covariates) and local treatment effects (treatment effects within a specific subgroup) can be generalized
-Measure used must be a collapsible effect measure; effect calculated for the entire population can be expressed as a weighted average of the effect calculated for subgroups
-Ideally with a collapsible effect measure, you can reweigh (using an estimator) the individuals in the trial according to the covariate distribution in the target population to say that the causal effects in the study sample = causal effects in the target population

Question 14

How well an effect measure generalizes depends on

Answer 14

-Its collapsibility (OR is non-collapsible) because it allows you to reweigh the study sample according to the covariate distribution in the target population
-Its ability to disentangle between the causal effects and the baseline risk (RR and RD can do, OR can’t)

Question 15

Internal Validity

Answer 15

-Concerned whether the estimate of causal effect in a sample differs from the true causal effect in that sample
-The extent to which a study accurately measures the causal relationship between variables within its own specific conditions
-Ensures that observed effects are due to the treatment and not confounding factors
-Relies on careful study design and control of variables

Question 16

Target validity

Answer 16

-Comprehensive metric to assess the total validity of a causal effect estimate for a specified target population
-Takes into account internal validity bias and external validity bias due to observed and unobserved factors
-Target bias = 0 means that the estimated causal effect of a study sample perfectly matches the true causal effect in the target population
-High target validity:
§ Representativeness of the study sample relative to the target population
§ Randomization of treatment within the study sample

Question 16

External Validity

Answer 16

-Applying the causal effect of a study population to the target population
-How well results of causal effects generalize to other contexts including different populations, settings, or times

Question 17

Dichotomy between internal and external validity

Answer 17

-General idea and oversimplified idea that there is a trade-off between the two
-When internal validity increases (tightly controlled study), external validity decreases (decreasing generalizability of findings
-Or a study must be completely internally valid before thinking about external validity
-Real world application of study findings can involve numerous factors that are not accounted for in a tightly controlled study that has high internal validity
-Findings of a causal relationship is helpful and important but without considering external validity, it feels like the finding is limited In its applicability

Question 18

Reasons for Mediation Analysis

Answer 18

-Determine how exposure exerts its effect on the outcome by understanding the mechanism of action through the causal pathways with and without mediators
○ Quantify how much of the effect is occurring through the direct and indirect pathways
-Distinguish direct effects from indirect effects of exposure on outcome, which can strengthen causal conclusions (through the interventionist approach)
-Identify potential targets for public health interventions by understanding causal pathways to determine proximal causes that affect the outcome, when manipulation on the exposure/intervention is not feasible/realistic (e.g. race/ethnicity)
-Helps with more accurate generalizability/transportability by understanding the distribution of the mediator and the mechanism of action in other populations and settings

Question 19

NDE (Counterfactual Approach)

Answer 19

The average causal effect of the exposure on the outcome if, for each individual, the mediator is fixed at the value it would have been if the individual had been unexposed

Question 20

Crossworld effect:

Answer 20

-Refers to any causal effect involving the counterfactual outcomes indexed by exposure levels, which cannot occur naturally at the same time in the same individual
-Requires knowledge of the counterfactual value of the mediator for the same individual had they not been exposed

Question 21

CDE (Interventionist Approach)

Answer 21

the average causal effect of the exposure on the outcome if the mediator is fixed at a constant value (controlled)

Question 22

Interventionist critique of NDE

Answer 22

NDE/crossworld effect cannot be enforced experimentally and therefore is not subject to strict scientific scrutiny

Question 23

Interventionist mediation effects identified but not NDE

Answer 23

Ideally, you can intervene on a mediator, therefore intervention mediation effects can be identified when you have a mediator you can act on (to control at a set value)

Question 24

Separable direct effect

Answer 24

decomposing the exposure/treatment into different components, affecting the outcome and affecting the mediator/competing risk , separately

Question 25

SDE Assumptions

Answer 25

-relies on science-fictiony ways to create the components where the identifiability conditions are less likely to hold as derived from the observed data
-Need to measure all common causes of the E-Y and E-M/Competing risk

Question 26

SDE Pros

Answer 26

-Does not require cross-world counterfactuals or conceptual interventions on competing risk
-Can be verified empirically by an RCT if needed
-Useful for obtaining the direct effect if there is a confounder between the competing risk and the outcome that is not impacted by the component of the exposure impacting the outcome (because then you can adjust for the confounder)

Question 27

SDE Cons

Answer 27

-Needs subject matter knowledge to explicitly describe decomposition of the exposure pathway into different components and be able to intervene on each component to estimate the two effects
-Relies on science-fictiony ways to create components where the identifiability conditions are less likely to hold as derived from observational data
-Ignores exposure-mediator interaction

Question 28

Total Indirect Effect

Answer 28

effect of the exposure on the outcome due to the fact that it caused the mediator, including any interaction between the exposure and the mediator

Question 29

NCE

Answer 29

a variable that is not known to be causally related to the outcome of interest but is known to share a potential source of bias

Question 30

Qualities of a good negative control

Answer 30

-Should not be causally associated with the exposure or outcome of interest
-Should reproduce a condition that cannot involve the hypothesized causal mechanism but likely involve the unmeasured causes as the exposure and outcome of interest
-Should be similar in context to the primary exposure and outcome to reflect similar sources of bias

Question 31

NCO

Answer 31

a variable that is known to share a potential source of bias with the outcome of interest but that is known not to be causally associated with the exposure

Question 32

How negative control is used in studies

Answer 32

-Detect confounding, recall bias, selection bias (if an effect is found for a negative control, it is likely due to bias and that bias is affecting the effect estimate of interest)
-Can be used to remove bias from studies under additional assumptions as in proximal interference

Question 33

Proximal Inference Assumptions to remove bias

Answer 33

-U-comparable: requires that unmeasured confounders of the primary association are identical to those affecting the associations between the treatment, outcome, and negative controls
-Monotonicity: treatment and control effects work in the same direction for all individuals, with no reversal effect
-Completeness: requires that negative controls need to be sufficiently informative (high dimensional or high in quantity) to serve as good proxies for unobserved confounding
-Positivity: all combinations of treatment and negative controls are represented across all strata of observed covariates, meaning there are no zero-probability scenarios for any of the treatment-control-outcome combinations

Question 34

Purpose of sensitivity analyses

Answer 34

-Used to assess how robust a study’s conclusions are in the presence of unobserved confounding
-Does not measure the amount of unobserved confounding in a study

Question 35

Logic behind conducting sensitivity analyses

Answer 35

-Analysis adjusts the effect estimate obtained under the assumption of no unobserved confounding
-Provides alternate estimates or bounds under different assumptions about the strength of the unobserved confounding
-By considering a range of assumptions, it shows how the effect estimate would change
○ If effect estimate varies significantly under a plausible range of assumptions, the study’s conclusions are sensitive to unobserved confounding
If the effect estimate remains stable across different assumptions, the conclusions are robust to unobserved confounding

Question 36

Relationship between EMM and MoA

Answer 36

-To incorporate EMM into a DAG (since you can’t show EM in a DAG), portray EM as acting through a MoA
○ Example: if age is a modifier, use metabolism or immune response as a mediator for different age groups
-Makes it easier to visualize if there are different MoA within different strata of EMM that affect the exposure-outcome relationship
Depending on differing MoA, different strata of an EMM may show different magnitudes of effect or even different directions of causal effects of the exposure-outcome relationship

Question 37

Conceptual Constructs

Answer 37

• We use conceptual constructs to represent variables of interest for our study
  Conceptual constructs represent theoretical ideas (e.g. religiosity, health)

Question 38

Operationalizing constructs:

Answer 38

• Operationalizations are specific ways to measure conceptual constructs
  Ex. Church attendance for religiosity, self-reported health status for health score

Question 39

Competing risk

Answer 39

an event that prevents or modifies the primary event/outcome of interest

Question 40

Examples of competing risks

Answer 40

-In a study focusing on Disease A after recruitment as a result of Exposure E, disease causing death D before Disease A serves as a competing risk for Disease A because an individual who dies of disease causing death (D) before disease A cannot subsequently have Disease A
-In a study focusing on prostate cancer mortality, death from CVD (which we assume occurs earlier in life), serves as a competing risk for prostate cancer death because an individual who dies from CVD cannot subsequently die from prostate cancer
-In a study focusing on infant mortality due to maternal smoking, death from birth defects in utero serves as a competing event for infant mortality due to maternal smoking because an infant who dies from birth defects in utero cannot subsequently die from the effects of maternal smoking

Question 41

Limitations when you have competing risks

Answer 41

• Total effect encompasses all causal pathways including those influenced by competing events, therefore you obtain attenuated estimates of the total effect if exposure does cause the outcome and has a competing risk that is not accounted for
  
  • Choosing a causal estimand (theoretical quantity we aim to measure) that does not take into account competing risk
    ○ Survival analysis (collider stratification due to death by a common cause)
    ○ Treating competing risks as censoring events
  • Selection bias by selecting on survivors of an exposure and outcome, or competing risk of an outcome
    Results in differences between the true causal effect (in the population before the selection process) and the observed estimate that is due to the selection of the sample into the study

Question 42

Censoring event:

Answer 42

circumstances where it becomes impossible for the event of interest to occur (death, LTF)

Question 43

Competing risk as a censoring event:

Answer 43

Assumes that treatment or exposure under study prevents the occurrence of the competing event

Question 44

Interpretation of effect estimate when competing risk is treated as a censoring event

Answer 44

Estimating the effect of the exposure on the outcome had the competing risk been completely prevented from occurring

Question 45

Total effect

Answer 45

all causal pathways from a treatment to an outcome, including those influenced by competing risks

Question 46

Survivor Average Causal Effect (SACE)

Answer 46

• Causal estimand that focuses on quantifying the average causal effect of a treatment on an outcome of interest within a specific subgroup of individuals who are “survivors” of the competing event
  
  • Individuals who are unaffected by the treatment’s impact on reducing the competing event
  • Being more explicit in your research question by choosing a causal estimand that takes into account comepting event instead of just ignoring it

Question 47

Target trial:

Answer 47

using observational data to emulate a hypothetical RCT that mimics the design features of a true experiment when they are not feasible, ethical, or timely

Question 48

Strengths of target trial framework:

Answer 48

• Have a well-defined treatment/exposure
  Ensure that time zero aligns with eligibility criteria, the start of the treatment, and the start of outcome assessment

Question 49

Sources of bias target trials address by following the target trial framework:

Answer 49

• Selection bias: eligibility criteria being assigned after treatment or when prevalent users (those already on the treatment) are included at the start of follow-up
  ○ Target trials address this by ensuring that time zero aligns with meeting eligibility criteria, then treatment assignment and starting of follow-up can occur at the same time
  
  • Immortal time bias: treatment is assigned after time zero and meeting eligibility criteria, meaning that individuals in the study would have had to have survived to treatment assignment to be included in the study, making them “immortal”
    ○ Target trial addresses this by ensuring that time zero aligns not only eligibility criteria but also treatment assignment and start of follow-up

Question 50

Per-protocol:

Answer 50

the causal effect of a treatment on an outcome among those that fully adhered to the treatment protocol (excluding patients who did not comply to the treatment protocol)
Estimates the average treatment effect among the treated

Question 51

Why is it important to estimate PP:

Answer 51

• Provides a true treatment effect using those who completely adhered to the protocol
• More relevant to decision makers because they want to know the effectiveness of a treatment under ideal conditions where it is perfectly adhered to

Question 52

Per Protocol RCT

Answer 52

-Since PP does not maintain the equal distribution of confounders achieved at randomization, must use methods to adjust to account for factors influenced by prior treatment adherence
○ Methods to adjust: matching via PS, stratification, IPTW, standardization, g-methods
-Gives straightforward effect of treatment A on causal contrasts of interest

Question 53

Per Protocol Observational:

Answer 53

-To emulate random treatment assignment, methods to adjust must be made for the prognostic factors that predict treatment adherence at baseline
○ Methods to adjust:
§ Estimators: Matching via PS, stratification, IPTW, standardization, g-methods
§ Use of clones and assign clones to a specific treatment strategy so there are no issues with overlapping treatment assignment
-Align eligibility criteria met, treatment assignment, and start of treatment follow-up at time zero to simulate point of randomization in an RCT

Question 54

Potential outcomes framework:

Answer 54

structured way to conceptualized and estimate causal effects by comparing what actually happened with what would have happed under different circumstances (the counterfactual)

Question 55

Limitations of target trial and potential outcomes framework

Answer 55

• Exposures in marginalized communities often are social determinants of health of involuntary exposures where it would be difficult/unethical to fit into frameworks that typically use medicine or biological foundations
  
  • Interventions typically have to be multifaceted and community based, making it difficult to work with isolating specific exposures which is typical in target trials/PO framework
  • Exposure can change over time
  • Social exposures can be difficult to measure
  • May be a lack of data due to systemic biases in data collection
  • Interventions may need to be tailored to specific context/communities
  • Generalizability and transportability of causal findings is difficult from one community to another due to different exposures, mediators, effect modifiers, confounders
    Marginalized communities face overlapping exposures that interact in complex ways making the traditional framework difficult to use

Question 56

ITS

Answer 56

-quasi-experimental study design that evaluates the impact of an intervention (the interruption) by comparing outcome measures before an intervention and after the intervention
-can be viewed as a special case of regression discontinuity, where the assignment variable is calendar time
-continuity assumption mandates that no other interventions or covariates should change at the threshold (time of policy change)
-Counterfactual scenario is established to represent what would have occurred in the absence of the intervention, based on the pre-existing trend

Question 57

ITS estimand

Answer 57

ATT

Question 58

ITS Assumptions

Answer 58

-Continuity (exchangeability): the intervention group’s baseline trend of the outcome would have continued unchanged if the intervention had not occurred

-No anticipation: the intervention had no causal effect on the outcome before it was actually implemented

-No other simultaneous/co-occurring interventions: there should be no other major events or interventions occurring at the same time as the intervention that could also affect the outcome variable

-No model misspecification of the pre-trend: the pre-intervention trend must be completely specified in the model (e.g. linear, quadratic)

Question 59

cITS

Answer 59

-introduce a control group or control outcome unaffected by the intervention for comparison
-Control group is used to address history bias where the intervention group’s pre and post-trends are subject to history bias
-Control group has to be exposed to the same concurrent events (only looking for temporal factors to address history bias)

Question 60

Types of Controls for ITS

Answer 60

location biased, characteristic based, behavior based, historical cohort, synthetic, and control outcomes or time periods

Question 61

cITS Assumptions

Answer 61

Parallel trends:
Control group and treatment group experience similar trends in outcomes before the treatment is applied and would have the same trend in outcomes continue without the treatment
Any difference in the trend between the two groups after treatment can be attributed to the treatment itself rather than pre-existing differences in trends

Question 62

DiD Estimand:

Answer 62

ATT

Question 62

DiD

Answer 62

-evaluates the impact of an intervention by comparing the changes in outcomes over time between a group that receives the intervention and a group that does not
* Controlled before and after (CBA): compares outcomes before and after the intervention for both the treatment group and the control group
-Control group design with pre-test and post-test: group that does not receive the intervention and measure outcomes pre-intervention and post-intervention to determine effect of the intervention

Question 62

DiD Assumptions

Answer 62

• Parallel trends (exchangeability and consistency):
  ○ treatment and control groups must have experienced similar trends in outcomes before the treatment was applied,
  ○ any difference in the trend between the two groups after the treatment can be attributed to the treatment itself rather than pre-existing trends differences between the trends;
  ○ assumption that average outcome among the treated and comparison groups would have followed the same trend in absence of treatment
  
  • No anticipation: treatment had no causal effect before implementation
  • Exogeneity of instrument: timing of treatment should not be influenced by events that also affect the outcome
    -Sampling: data should represent a broad and randomly assigned sample population so that results reflect trends and outcomes that are representative of larger population, not just the specific sample studied

Question 63

How DiD Assumptions are relaxed

Answer 63

• Multiple periods and variation in treatment timing:
  ○ Instead of two groups, two time periods, and dichotomous treatment, use multiple time periods and variation in treatment timing (staggered adoption design)
  § Example: evaluating the impact of a policy change that is implemented at different times in different regions. Instead of comparing just before and after a single time point, the analysis can track outcomes over several periods as the policy is rolled out in different regions
  
  • Conditional parallel trends:
    ○ Situations where trends are only parallel conditional on observed covariates
    ○ This means that trends are parallel when controlling for certain variables that influence the outcome
  • Non-linear transformations:
    ○ Traditional DiD assumes that outcome data can be analyzed using linear models
    ○ Cases where the outcome data requires non-linear transformation (i.e. log transformation), traditional linear DiD models may violate parallel trend assumptions
  • Alternative Sampling assumptions:
    ○ Traditional DiD relies on standard cluster-robust methods to account for the fact that data may be clustered (e.g. individuals within schools or regions)
    Permutation and bootstrapping: resampling techniques can be used instead to provide a way to estimate the distribution of the treatment effect by repeatedly resampling the data

Question 63

Scenario to use cITS

Answer 63

• Time series data available
  
  • Control for history bias
    Looking for changes over time

Question 63

Scenario to use ITS

Answer 63

• Single group over time; no control group
  Will be using population’s pre-trends as the counterfactual

Question 64

Scenario to use DiD:

Answer 64

• Time is not the focus
  
  • Average treatment effect
  • Focus on before and after
    Comparison group is as similar as possible to the intervention group

Question 65

RD

Answer 65

assignment to a treatment is based on a threshold rule on a continuous variable; takes advantage of the arbitrary nature of threshold decision (on a micro scale); those just above and just below the threshold are essentially identical on all characteristics (baseline covariate distributions), both observed and unobserved

Question 66

Sharp

Answer 66

treatment assignment is strictly determined by whether the assignment variable crosses a specific threshold; there is universal compliance with the decision rule

Question 67

Sharp Estimand

Answer 67

LATE (local average treatment effect) pertaining to those near the threshold since results may or may not be generalizable to those far from the threshold

Question 68

Fuzzy

Answer 68

treatment assignment is not strictly determined by the assignment variable crossing the threshold; there are individuals above and below the threshold who do or do not receive treatment assignment

Question 69

Fuzzy Estimand

Answer 69

CACE (Complier average treatment effect) pertaining to those people near the threshold since results may or may not be generalizable to those far from the threshold

Question 70

RD assumptions

Answer 70

• Continuity (exchangeability): potential outcomes must be continuous at the threshold; in absence of treatment, there should not be a jump in the potential outcomes at the cutoff point (we want the jump to be due to treatment)
  ○ Participants should be exchangeable across the threshold
  ○ If continuity assumption holds, participants below the threshold can serve as valid counterfactuals for those above the threshold
  
  • No manipulation: if individuals can manipulate their assignment to fall just above or below the threshold, it violates the continuity assumption; there should be no bunching near the threshold
  • Monotonicity: treatment effect moves in the same direction
    No defiers who do the opposite of what they are told

Question 71

Fuzzy RD assumptions

Answer 71

• Exclusion restriction: threshold assignment only affects the outcome through the treatment and not any other pathway
  
  • Independence: threshold assignment is independent of the potential outcomes and any confounders
    Relevance: assignment variable must affect the likelihood of receiving treatment around the cutoff

Question 72

Ways to test RD assumptions:

Answer 72

• Test for continuity: compare covariate distributions just above and below the threshold to ensure that covariates don’t show discontinuities (similar to balancing covariates in an RCT)
  Test for manipulation: assess the density of the assignment variable value at the threshold and check for precise manipulation; if there is a spike around the threshold, it might indicate manipulation

Question 73

IV

Answer 73

use of a specific variable (instrument) that affects a behavior (treatment or exposure) but does not affect the outcome expect through its impact on the exposure

Question 74

IV assumptions

Answer 74

• Relevance: instrument must have a significant effect on the behavior (exposure)
  
  • Exclusion restriction: the intervention should affect the outcome only through the behavior change and not through any other pathway
  • Independence (exogeneity): instrument is independent of the outcome, given the exposure and other covariates, the instrument does not share any common causes with the outcome
  • Monotonicity: all individuals are affected the same way by the instrument
    No selection bias/collider bias

Question 75

Ways to test assumptions

Answer 75

• Test relevance: assess the strength of the instrument using the F-statistic from the first stage regression of the exposure on the instrument; an f-statistic >10 is typically considered a strong instrument
  
  • Test independence (exogeneity): examining the correlation between the instrument and observed covariates that might affect the outcome, ensuring that the instrument is not correlated with the covariates (but this is just based on plausibility, logic, and content knowledge not an actual test)

Question 76

Dynamic treatment strategy in situations with time-varying treatment:

Answer 76

treatment decisions at each time point being influenced on evolving patient conditions and other time-varying covariates

Ex. Chemo regimen regularly monitored by blood tests, imaging, side effects

Question 77

Exchangeability:

Answer 77

allows for fair comparisons between exposure groups by ensuring that any observed differences in outcomes can be attributed to the treatment rather than confounding factors (no unmeasured confounders)

Question 78

Consistency:

Answer 78

potential outcomes are the same as the observed outcomes under the same treatment condition by well defined intervention and no interference

Question 79

Positivity:

Answer 79

every combination of treatment and covariates has a non-zero probability of occurring

Question 80

Confounding Bias

Answer 80

non-causal association introduced into estimate due to uncontrolled/unadjusted common causes of the exposure and the outcome

Question 81

Selection Bias

Answer 81

systematic error when the sample selected into the study differs from its target population, leading to discrepancies between the true causal estimate and the observed estimate