Midterm Flashcards
Lectures 1-5
What is the goal of the counterfactual model?
To estimate a true causal contrast?
Define confounding, selection bias, and information bias.
Confounding is a non-causal association
that is observed between a given exposure and an
outcome owing to a third variable. Lowers exchageability.
Selection bias is
Information bias is
DAGs can be used to communicate hypothesized relationships and ________.
identify and understand potential sources of measured/unmeasured selection bias or confounding.
How is the line of best fit obtained?
By minimizing the sum of squared errors (i.e. the vertical distance between each point
and line)
The _______ is the
proportion of the total
variability explained by the model that includes the covariate x.
coefficient of determination (R^2)
1-SSy-SSxy
Name the four assumptions of linearity.
Constant variance, normality of errors, X-Y relationship is linear, and errors are all independent.
What do we call the difference between the measured population mean of X (μx) and the true population mean of T (μT)? How is this different from precision?
Bias. Precision is about low variance of the measurement error itself.
When is measurement error non-differential?
When the sensitivity and
specificity of exposure assessment is equal for both groups
Which type of measurement error moves estimates closer to the null?
Non differential.
Differential moves estimates toward or away or change directions.
If exposure measurement error is non-differential, the bias
in the effect estimate is a function of ________.
precision (random error)
How does measurement error on exposure differ from outcome?
Outcome measurement, because it’s the dependent variable, won’t change the slope estimate but will increase its standard error and widen the confidence interval. Exposure will bias toward the null.
The sample size needed under non-differential measurement error is proportional to ______________, which is the assumed common exposure variance among cases and controls.
standard deviation
When everyone is assigned the same exposure, true exposures vary normally around ____.
Group values
Berkson-type random exposure measurement error is not expected to bias effect estimates (i.e. slopes in regression models) but there is still a loss of ______ (i.e. wider confidence intervals) and __________.
precision; reduced power
Describe the Berkson model of bias.
Random error is attached to the true exposure value, independent of the observed; lowering precision.
Increasing sample size can minimize the impact of measurement
error in continuous outcome variables (T/F).
True
How does error in outcome measurement change categorical/continuous variables?
Categorical outcome measurement error will produce a bias toward the null. Continuous outcome error has no effect, if you recall, on slope and just reduces precision.
Even when the
expected direction of bias is toward the null (because of non-differential exposure misclassification) bias away from the null can occur because ______.
Any study is just one realization, one sample which could be affected by random error in large/small ways each time and only evens out with multiple trials. So what we expect doesn’t always happen.
Misclassification and mixing levels leads to bias toward the null (T/F)
False, it can happen away!
Confounders are typically only CAUSALLY associated with the exposure (T/F).
False, the outcome!
Subject matter expertise is the best way to identify potential confounders (T/F).
True
How do we adjust for confounding through stratification?
Separate confounder into strata and calculate stratum-specific associations, then pool if homogenous.
Name two limits of stratification as a method for controlling confounding.
Leaves room for residual confounding in continuous variables and adjusting for multiple variable requires specific estimates for every different combination.
The adjusted regression coefficient is the expected change in the mean value of ___ per unit change in X keeping _______.
Y; all other variables constant
Residual confounding occurs when categories are _______, or when confounders are measured with _____ or ______.
too broad; error; unmeasured / left out
The distribution of a confounder has to be different across groups (e.g. case/controls) to
cause confounding (T/F).
False, especially when it’s associated with the outcome.
The overall impact of including multiple parameters in the
model depends on their _____ with X.
correlation
Variance of B1 is based on variance in X, error in Y, sample size, and the correlation between parameters.
What are 3 methods for modelling nonlinear relationships?
Dummy variables, quadratic terms, and splines.
Selection bias occurs when exposure and ______ both
affect inclusion in the analysis.
outcome
What does conditioning on a collider do?
Induces a spurious relationship between the variables leading into it.
Confounding is the presence of common effects while selection bias is conditioning on common causes (T/F).
False, the opposite!
Effect modification refers to the situation where the
strength of association between exposure and
outcome differs across ____________.
levels of a third variable
Explain the difference between additive and multiplicative interaction.
Additive interaction is when the absolute risk changes across levels of a third variable, multiplicative interaction is present when the RR varies. You can have additive without multiplicative if the ratios are all the same.
The interaction term is the excess change in the outcome not explained by the _____.
Sum of the individual effects of two independent predictor variables.
What are two things to be aware of when interpreting interaction terms?
Heterogeneity due to small sample size and confounding/error across strata of the effect modifier.
Case-crossover studies are used to examine the acute
health effects of ____________.
intermittent exposures with short induction times
What kind of confounding is not present in case-crossover studies?
Confounding due to variables which do not change within single individuals across the reference/case periods.
Transient co-exposures may cause confounding, though.
What are the three ways you can have selection bias in a case crossover study?
nonrepresentative case selection, differential case survival (exposure influences survival), control time not independent of exposure.
Exposures during case and reference periods must be _______, to avoid ______________.
independent of each other; carryover effects
We use a time-stratified design if experiencing the event doesn’t impact the likelihood of ___________, so control periods can
be selected before and after the event.
subsequent exposures.
What analysis do we perform for case-crossover studies?
Logistic regression to get Odds Ratios
Case-crossovers look at the ____ in exposure between two periods, rather than the absolute level.
change
What’s the major con of case-crossover studies?
They only use information from a single point in time, so they capture prevalence rather than incidence. Also cannot determine relationships.
Ecological studies are good for studying variability __________ and for prevention at a _____________.
between homogenous populations; population level
Name the three types of ecological measures.
Aggregate measures (characteristics within
the group)
Environmental measures (physical
characteristics of a location)
Global measures (group characteristics not
reducible to characteristics at the individual level)
What do you call it when within-group correlations are different from between-group and you try to conflate the two?
Cross-level inference (ecological bias)
What three steps have to be at the group level in order to make an ecological study?
measurement, analysis, and inference
What are the pros and cons of ecological studies?
Less expensive (good starting point + existing measurements) and increased generalizability potential.
Hard to control confounding, identify temporal relationships, and threat of ecological fallacy.
The Poisson distribution is used for counts—if events
happen at a constant rate over time, the Poisson distribution
gives the probability of ___________.
X number of events occurring in time T
Count frequencies are often positively skewed with ________________.
most values being low and relatively few high values
In Poisson distribution, the mean event rate is supposed to be equal to the ______.
variance
In Poisson Regression the dependent variable is the log of ________.
the expected count
How does logE[Yi] respond with each parameter in the model?
linear increase!
exp(B1) in a Poisson model approximates to?
Rate Ratio for a one unit increase
What is the offset parameter?
logTi (the
time that unit “i/p” was at risk)
The offset parameter is used to account for differences in exposure time or population size when modeling count data.
The effect of over-dispersion is that the precision of effect
estimates is not correct, leading to ______.
too-narrow confidence intervals
θ is the _____ over-dispersion parameter
Quasi-Poisson (more flexible variable function)
Splines fit a number of different ______ (usually
cubic) over the range of the data and are joined smoothly at
_______ to examine nonlinear forms.
polynomial curves; knots
Splines can be used to flexibly adjust for ______.
confounders
What are two ways to optimize time-series models?
lag models to correct for delayed effects and sensitivity analyses related to spline knots. Autocorrelation (nonindependent days) is unclear though.
_____ analysis is useful when a report makes policy recommendations/actions and draws inferences about
causality
Bias
Non-participation always leads to selection biased results (T/F).
False, only if participation depends on exposure AND outcome (confounding)
We can know this by asking nonparticipants to fill out a short questionnaire
Where do we get information on an unmeasured confounder?
A sub-sample of the study population OR educated guesses based on other studies
______ can be conducted to collect
measurements using both the “gold standard” method and the
less accurate/reliable method
Internal validation studies
What does a simple bias analysis leave out of the picture, other than the other types of biases?
uncertainty in the bias parameter itself
What are the parameters for simple bias analysis for selection bias vs. confounding vs. measurement error?
Contingency table –> multiply OR by selection OR.
Specify association between confounder and the outcome for the unexposed, as well as the prevalence in both groups
Reshuffle contingency table based on sensitivity and specificity values for each cell.
By repeatedly sampling from bias parameter distributions,
the output out probabilistic analysis is ________ which can be interpreted as a point estimate
and frequentist confidence interval.
a distribution of adjusted estimates of
association
Multidimensional bias analysis is an extension of simple bias
analysis that assigns probability distributions to ________ at the same time
multiple bias analysis
