Final Flashcards
Define a cohort study: A group of people _______________ is ____________ for a certain period of time to _____________________.
A group of people free of the disease of interest is identified and followed for a certain period of time to ascertain the occurrence of events.
The goal is to compare disease incidence between exposed and unexposed.
What is the difference between RCT’s and cohort studies?
In an RCT, exposure is assigned by the investigator. In a cohort study, groups already exist in the source population and are selected into the study.
What are three strengths of cohort studies?
Ability to study exposures that are difficult to randomize. Allows calculation of incidence. Multiple outcomes/exposures can be assessed.
What is the main advantage of prospective studies, and the main disadvantage of retrospective?
We can see temporal ordering of exposure and disease.
Can only use available information which may be of lower quality or for different purposes.
Studies will often include an __________________ that must occur before beginning to count person-time at risk of exposure effects (varies by subject matter).
induction/latency time
Cohort studies most often deal with acute exposures (T/F).
False, chronic.
____________ refers to how we will conceptualize exposure
based on our study hypotheses and what we currently understand
about the exposure/disease of interest
Dose representation
What do we consider when defining exposure/dose?
Dose represention and causation; do we look for mean exposures, peak, cumulative?
What are the most important time periods of exposure?
Cumulative dose is the sum of _____.
duration x frequency x intensity
If the biological effects of a particular exposures are not well understood, it is a good idea to evaluate only one representations of exposure (T/F).
False, several!
Selection bias is caused by conditioning on a __________________.
common effect of exposure and disease
Exposure and Disease status should be independent of both _________ and _________ in order to avoid selection bias?
selection; retention
As opposed to conditioning on a common effect, confounding is more about common _____.
causes
What are the basic requirements for survival analysis definition?
Clear time scale and origin of time at risk; Precise definition of ‘event’ and how it will be measured; Accounting for censoring in the an analysis
Why wouldn’t we do linear regression instead of survival analysis?
Time to event is usually quite skewed, and this would also ignore censoring.
What are the two types of survival analysis models?
Parametric; assumes that the times follow a distribution.
Semi-parametric; assumes proportional hazards
Right censoring causes an underestimation of the true time to event (T/F).
True
S(t) describes the probability that ________.
an individual survives from the start of follow-up to a specified future time (t).
The hazard function is the ____________ at time (t).
instantaneous event rate (incidence)
The Cox Proportional Hazards Model estimates the effect of covariates on the hazard ratio. What is the hazard ratio?
The ratio of hazards betweenexposed/non-exposed, leaving the baseline hazard rate unspecified. Like an incidence rate raio.
h(t) = h0(t) * exp{b1x1 +b2x2+…+bpxp)
The baseline hazard is the value of h(t) is all covariates __________.
equal zero
What is the key (multiplicative) assumption of the Cox model?
The hazard in either group is a constant multiple of the hazard in any other; the baseline hazards are the same. Thus, the e^betas gives the hazard ratio for a change from x1 to x2. Curves don’t cross!
Does hazard ratio depend on time?
No girl not at all
How can violations of the proportional hazards assumption be addressed?
Including interaction terms between covariates and time, or by using strata variables. If the baseline isn’t the same for everyone, we can have different ones for different strata (stratified proportional hazards model).
+strata(variable)
What is the disadvantage of stratifying baseline hazards?
That covariate will disappear as a predictor, so we won’t see the effects of the stratification we’ll just have controlled for the variable.
In which ways are case-control studies more efficient than cohort studies?
Time (eliminate the need for follow-up)
Personnel + lab analysis
Ideally what proportion of cases in the source population will be included in the study?
All (minimizes standard error)
Sampling must be independent of ______.
exposure
Why should we select incident cases instead of prevalent cases?
Prevalence is related to disease duration / survival time. Sampling must be independent of exposure.
What will an odds ratio approximate in each case?
Risk Ratio when controls are selected at the beginning of follow-up. Rate ratio when they’re selected from members of the population at risk (density / risk-set sampling). Odds ratio when controls are selected from those who are non-cases at the end of follow-up, or risk ratio if the rare disease assumption is kept.
What are the two basic rules of control selection in a case-control study?
They should be selected from the same population as the cases and independently if exposure.
What are some alternatives to source sampling for controls?
Neighbourhood, friend, random digit dialing, hospital-based.
Neighbourhood and friend may be related to exposure and not in the same source pop. All homes aren’t likely to be recruited in random dialing. For hospital, make sure that controls are selected independently of exposure.
What are the two types of matching in case-control studies and what is the effect?
Individual and frequency (distribution). We match on confounders but matching doesn’t control for confounding, and those variables will have to be included as covariates to adjust for the bias.
Matched analysis will have a slightly wider confidence interval.
Which function do we use to model logistic regression?
logit, to ensure the range of pi is between 0 and 1.
How do we interpret B1 in a logistic regression?
e^B1 is the odds ratio for a unit change in X. (e^B1)^2 is the OR for a two unit change
What are the two opportunities to introduce selection bias into a case control study?
Control selection and case ascertainment
Why is effect modification so important and informative?
Unlike biases which should be reduced, effect modification provides insight into the nature of the exposure-outcome relationship. It should be reported and understood.
Helpful in identifying the groups of individuals that would benefit the most from an intervention.
Effect modification examines whether the strength of an association related to the exposure ______________.
varies across strata of another variable
Interaction refers to a situation where a risk factor and another factor ___________ each other.
strengthen or weaken
What are the two main strategies for assessing and detecting Effect Mod?
Assessment of homogeneity: comparing effects of one exposure over the strata of a third.
Comparing observed effects to expected joint effects under an additive or multiplicative disease model.
Effect Modification could be caused by heterogeneity due to…?
random variability (stratify), confounding (we assume none), and other biases caused by error.
Additive interactions are more relevant to public health because they
reflect___________.
absolute differences in risk (e.g., numbers of additional cases)
What is the difference between prediction and inference?
Prediction uses data to map inputs to outputs, requiring expert knowledge to specify the question, data sources, and variables.
Inference estimates how a feature would change under different circumstances, requiring expert knowledge to define causal structure and identify biases.
Good predictive models tell you about how to intervene to change the outcome (T/F).
False, they don’t necessarily
It is best if predictive models capture ___________________, otherwise predictions may have poor generalizability.
true causal relationships
Health risk scores are useful in stratifying _______________ patients through prediction of health events.
low/high-risk
Why are prediction model coefficients hard to interpret?
These models are typically quite complex so they can flexibly model the relationships of interest. Then, it’s difficult to understand how any individual predictor influences the outcome.
What is the most common metric used to evaluate prediction accuracy?
Mean squared error (MSE) which we want to be small for the test data. But it doesn’t need to be too small (overfitting/noise inclusion).
Cross-validation is often used to evaluate predictive models when ______________ are not available.
external test data
What is a “hold-out” test sample?
A sample from the training set used only for model evaluation. We use the trained model to predict its values multiple times to estimate average performance.
Describe leave-one-out cross validation
Testing on n-1 samples. Each data point is used as a test set once, repeated for every data point. MSE values are averaged and squared errors are calculated for each.
Describe k-fold cross validation.
Fold into subsets by “k” (k=5/10) and use each, one at a time, as a test set. MSE values are averaged and squared errors are calculated for each.
K-fold gives more accurate estimates than LOOCV because there is _______________ in the estimates of LOOCV.
more variance
If the goal is prediction, we must think about confounding, bias, and model interpretation heavily (T/F).
False! This is true for inference.
For prediction, interpretation is less important but the predictors should still capture the causal model in question.
How do we approximate the target tissue dose?
We use proxies like urine and blood measures
What are the steps in planning a study?
Define exposure and outcome, identify and recruit population, etc.
What is an instrumental variable?
A factor which causes the exposure, not directly associated with the outcome or any confounders. A perfect proxy for exposure.
We use it to extract variation in the exposure not influenced by confounders, and use this variation to estimate causal effect.
Any paths between the instrumental variable and the outcome must pass through ______________.
exposure or be closed
When does the IV method lead to incorrect inferences?
When the relationship between the instrumental variable and the outcome is affected by unmeasured confounding
Propensity scores adjust for confounding by mimicking _____________ and making groups __________________.
randomization; as similar as possible with respect to known confounders.
Identifies groups of people equally likely to receive treatment to minimize systematic differences
What, technically, is a propensity score?
A balancing score. The predicted probability of exposure for every individual, using exposure as the dependent variable in a logistic regression model with potential confounders, interactions, and variables associated only with the outcome of the main relationship, regardless of significance.
How are propensity scores used to mimic randomization?
Generate a new dataset of matched subjects based on similar propensity scores and then include the scores as a covariate and then to stratify. No guarantee of similarities between unmeasured confounders, though. It can also restrict sample size.
1:1 nearest-neighbour matching. Matched pair analysis.
What does propensity score stratification do?
Creates strata based on individual score, then takes an average of the effect estimates across the different strata. Can reveal potential heterogeneity.
Is it useful to use propensity score as a covariate?
Yes, if we include with other covariates it’s a doubly robust method.
PS can be used to adjust for confounding in observational studies, but
unlike randomization, cannot account for __________________.
unknown or unmeasured confounding
Name 3 advantages and 3 disadvantage of propensity scores.
Can adjust for confounding, adjust for more covariates than normal, and maximize comparability.
Prediction is limited, can’t handle time-varying exposures, can lower generalizability.
Machine learning models train _______ to make predictions.
algorithms
What is the main limitation of linear regression for prediction?
The assumed linear relationship between predictors and the outcome.
Generalized additive models estimate the weight for each basis function to ___________________________________________.
minimize the residual square error of the model
caputures non-linear relationships while still being interpretable.
Name two assumptions of GAMs.
independent observations and normal distribution of residuals
What are hyper-parameters in machine learning?
Parameters set before training that are tuned and influence the speed and accuracy of learning.
Random forest (number of trees, node size, etc). XGBoost (number of sequential trees to build, depth of each tree). Dense neural net (learning rate, number of hidden layers).
Name two methods for small samples in machine learning.
Transfer learning updates weights from a previously trained model with a new dataset.
Data augmentation can increase the numbers of sample by randomly flipping images.
Define training set, tuning set, and test set.
Primary dataset used to train the model.
Subset used to tune the hyperparameters to optimize model performance.
External dataset not used for model development.
The quality of reference standard for model performance should be based on ______________________.
subject matter expertise
What is overfitting in the context of machine learning?
Model performs really well on the training set but not so much on the test set; can happen if data in the two sets overlap. Pay attention to your data splits!
Machine learning _____________ should be reproducible.
predictions
The ___________ design is for answering causal questions at the group level.
difference-in-differences
time and group are both confounders
In occupational exposure studies, 8-hour ____________ are often used for dose representation.
Time-Weighted Average Exposures
Describe what happens in a difference in differences study.
The first difference isolates the change in outcome over time within groups, controlling for group differences.
The second difference compares the within group differences between the groups; how much change in the outcome we would expect in the treated group if the treatment had not occurred.
What do we assume in a DiD study?
Exchangeability and the Parallel Trends Assumption. If no treatment had occurred, the difference between the treated group and the untreated group would have stayed the same post-treatment
period as it was in the pre-treatment period. So the groups should have similar trajectories before treatment.
How do we evaluate the parallel trends assumption?
Compare prior trends between groups, placebo test to estimate DiD at earlier time periods (there should be no difference).
What is the structure of a DiD interpretation?
In [Year], compared with [Year], __________ experienced 7.7 additional ______________ for AMI per 1000 Medicare beneficiaries than would have been expected in the absence of ____________.
What is a meta-analysis?
Quantitative approach for systematically combining results of previous research to arrive at conclusions about the body of research.
Effect size is the measure of the ____________ and ________ of a
relationship between variables
magnitude; direction
What are the three types of heterogeneity?
Clinical, methodological, and statistical.
The Cochrane Q statistic looks for heterogeneity by calculating ______.
the weighted sum of squared differences between individual study effects and the pooled effect across studies
The I^2 statistic is the percentage of variation across studies that is due to _____________ rather than chance.
heterogeneity
What is the difference between a fixed and random effects model in meta-analysis?
A fixed effects model assumes low heterogeneity (same effect for every study, variations due to random error) whereas a random effects model assumes the true estimate for each study varies and follow a normal distribution, with a higher CI.
The random model allows for inter-study variability, although with wider confidence intervals, more equal weighting, and less stability.
An ____________ funnel plot indicates clinical and methodological heterogeneity.
asymmetrical
Name three ways to deal with publication bias?
Combine the results of the larger studies, file drawer method to estimate how many null studies it would take to make the results insignificant, trim and fill to duplicate and center the data and make new estimations.
How do we find Simpson’s paradox in meta-analysis.
Effect size is based on the comparison of a group with its own control group – avoids
Simpson’s paradox. Could appear if we pool.
