Epi Final Flashcards
2 types of causal relationships
Necessary
Sufficient
Necessary
without the factor, the disease will not exist
Sufficient
with the factor the disease will exist
3 characteristics of a cause
- Time order (temporality)
- Associtation
- Direction
Time order (temporality)
cause must predate effect in time
Association
The likelihood of the outcome is different under the cause. There is a correlation between the putative cause and effect
Direction
A change in the cause will induce a change in the outcome. If you could change the cause and leave everything else the same, you should still see a change in result. requires counterfactual
schedule of potential outcomes
contains treatment, counterfactual, and difference. But this does not happen IRL because you cannot build a time machine to go back and retest the person
If you randomize enough units (RCTs) to intervention or controls, what factors should be balanced?
all pre-randomization features of the groups should be balanced. Anything else that should cause the outcome should be about the same
After RCT, the only remaining cause of observed differences should be wat 3 things
- the cause
- difference arising after randomization
- randomization failing by change
If you can exclude 2 and 3, can make strong causal inference
3 features of strong RCT
- randomization worked
- Excludability - exclude other factors from explaining outcomes observed
- Non-interference: treatments dont spill over
Randomized trials are the gold standard for causal inference if..
conducted well
2 major benefits of RCTs
- ensure groups are not systematically different when trial begins
- prevent bias in allocation
Randomization does not equal
random sampling
A good randomization approach is truly..
random and cannot be predicted (like time or day of service)
For RCTs, measuring before and after versus only after
before and after: account for differences in baseline. treatment effect estimated as difference between groups in the change in blood pressure
after: difference between groups in endline blood pressure
Methods of randomization
simple
stratified: randomizing for a confounder (age)
Cluster: must have sufficiently large number, and clusters cannot be super different
Crossover design
use each participant as their own control by switching them between placebo and treatment, but these designs only work if the effect of the treatment is temporary
Non-inferiority and equivalency designs
Test hypothesis that:
one treatment produces results approx the same as existing (equivalency), or one treatment produces results that are at least as good than the other (non-inferioirty)
Useful when a new treatment is developed (not more effective, but might be faster or cheaper)
Do RCTs account for problems arising after randomization?
NO
Non-adherence to treatment and solutions
Participants may just stop taking treatment, take wrong one, etc.
Can analyze by intention to treat (misclassification of exposre)
Can analyze by actual treatment (no longer have randomization)
Downsides of RCTs
If you want to describe a health problem or understand how an intervention works, not good
High internal validity often trades off with external valdiitiy (well designed observational studies often generalize better)
Unethical/unacceptable/impossible
Analytical epi
seeks to understand effect of various exposures, characteritsics, or interventions on health status
2 steps of analystical epi
- measured association
- inferences draw from association? is it causal, or by chance/bias?
Studies (like reports or case series) only generate
hypotheses. valid inferences about associtions require hypotheses be rigorously tested
To assert association between exposure and outcomes… 3 needs
- accurately measure exposure
- measure outcomes
- see if outcome is different in presence or absence of exposure
2 major observational studies
cohort studies: group based on presence/absence of exposure, then researcher evaluates outcomes. look at association with incident disease
case control studies: classified by presence or absence of outcoome, then exposures are evaluated. look at association with prevalent disease
Strongest designs to identify association and approximate randomization:
- prevent bias in allocation
- make characteristics of both study groups comparable with respect to everything besides intervention
- Randomized triala - gold standard for causality
Fundamental elements of cohort studies
-2 groups of people identified: one with and one without exposure
-researcher then identifies cases among each group
Selecting a study pop for cohort studies: two options
- select explicity on basis of exposure (good for when rare)
- select defined pop and classify members as exposed ot not (more generalizability, more representative pop.)
Steps of prospective cohort
- select pop
- identify who smokes at start/follow nd periodiclly assess who begins smoking
- identify who develops lung cancer
Strengths/weaknesses of cohort
strength: you know exposure predated outcome
weakness: long, expensive, people are dying, diseases have long latent periods
Steps of retrospective cohort
- identifiy a pop for which there is past assessment of smoking
- ascertain smoking based on past records
- identify who has developed lung cancer
Strengths/weaknesses of retrospective cohort
strengths: faster and cheaper than prospective
weaknesses: follow up difficult, med hisotries don’t always exist, accuracy of exposure assessment
Combined prospective and retrospective cohort studies steps
enroll patients who’ve gone to same clinic for 10 years for some reason other than cancer
1. look at med records to see if they smoked previously
2. continue to assess smoking
3. follow into future to see if they develop lung cancer
Association exists if…
outcomes are coorelated with exposure: a difference in exposure is associated with a difference in outcomes
Absolute difference vs relative difference
absolute: this much hihger
relative: this % higher
Relative risk
proportion who develop disease among exposed/proportion who develop disease among unexposed
Incidence rate ratio
of cases of disease per person-year among expose / # of cases of disease per person year among unexposed
IRR should be used when participants in study for varying amounts of time
Interpret relative risk:
> 1 = means exposure associated with greater risk of outcome. RR=1.6, risk of outcome is 60% greater among exposed than not
<1 = exposure is associated with reduced risk of outcome. RR=.6, risk of outcome 40% less among exposed than not
Equal to 1= no association, same proportion developed outcome
bias
systematic errors in design and analysis skewing observed RR away from true
Common biases in cohort studies
bias in assessing outcomes (preconcieved notions) - blind assessor to exposure
Info bias - two groups have systematic differences in data available
non-response/LFU
analytic bias
selection bias - exposed grouop may come from different pop. than non-exposed group
confounding - other fundamental differences between exposure gorups
Hypothesis testing: how do we decide whether to conclude H0 from a 1 pt difference from testing
Through P-value: probability of observing an association if null is trye
We reject the null if p value is sufficiently low
Then problem witht he p value
P value tells us probability of observing dta if H0 is true (assuming no other sources of error). But we actually want probability H0 is true if we observe out data.
3 options of what to do with p-vaue alternatly
- Interpret P(D/H0) as if it is the P(H0/D) we want. BAD OPTION
- Use bayesian appraoches, estimating likelihood of H0 and Ha before data collection
- Use p values more informally as one inductive tool among others (disclaims that hypothesis testing is fully deductive)
Deducitive vs inductive
Inductive reasoning involves starting from specific premises and forming a general conclusion, while deductive reasoning involves using general premises to form a specific conclusion
Concluding that there is a difference between treatment groups (choosing H over Ho) does not mean
that pop. difference = observed difference
Rather it means that if we were to reassign treatment vs control over and over, 95% of CIs would include the true pop. association (if only random change in assignment is at play)
A 95% CI interval not including H0 will have a p-value less than
.05
Clinical significance
With large enough samples, very small associations can still be statistically signifcaint, but difference may not be large enough to matter
basic design of case control sudy
start by identifying people w disease and those without (cases and controls)
Work backwards to determine past exposure
If exposure is associated w/disease, we expect exposure to more common among cases than controls
Selection of cases for case-control
Often want prevalent cases (or can use people as cases as they diagnosed)
Prevalent cases are often easier to idenitfy, but using incident cases reduces biases
prevalent case - factors/exposures associted with being a cause might actually be associated with development of disease or survival
SURVIVAL BIAS IMPORTANT FOR CASE CONTROL
Where do case control cases come from?
- particular medical institution (can be problematic because hospitalized cases may be poorly representative, isntitiions often serve particular sub-pops. of society)
- critical to use a precise case defintiion
Selection of controls
Must be from same pop. as cases (or similar). If controls and cases come from different ref. pops., an exposure that appears associate with diseease may just be associate with pop.
Selection bias exists in case control studies esepcially, but is also a major threat for
cohort studies
we want exposed and unexposed grouop to be comparable with respect to causes of outcome of interest other than exposure of interest
Sources of controls
Hospital: similar to cases/good quality of info, but controls may be from a differenet reference pop. than cases and sick controls may have particular exposure contributing to illnesses
Dead controls
Best friend/neighbor
Pop. controls
Matching importance
critical to have cases and controls as similar as possible on other characteristics
matching addresses this by selecting controls identical on relevant charactertistics like cases
Group vs individual amtching
group - if 25% of cases are male, select control so that 25% are also male. requires cases are identiifed before controls can be selected
Individual matching - controlling for relevant characterstics like age, sex, race, can be difficult to do
Inadvertant matching/accidental matching
Strengths/weaknesses of amtching
strengths: easy to do, some factors must be amtched
weaknesses: matching on multiple factors may be impossible
can decrease extent to whichs tudy pop. represent pop as a whole
matching cannot be undone
Recall Errors
People may incorrectly
remember whether they were
exposed to something.
* Especially when the exposure is
subtle or long ago.
* Inaccurate recall can result in
random error.
Recall Bias
If accuracy of recall is associated with outcome or another potentially causative facot,r can create a false association
Rumination bias
For example, patients with cancer often seek a cause for the
disease and may spend much more time thinking about past
exposures than health controls. What appears to be an
association with cancer may be merely an association with
recall
can address by asking about non-associate exposure s to see if thhey are more frequent among cases than controls
Multiple controls
Increases statistical power, can gain more info or assess quality of a control group
Strengths/weaknesses of case-control studies
strengths: cheap, fst, can examine multiple potench exposure, better for rare diseases
Weaknesses: frequently cannot determine conclusively that exposure predated illness, esposure may be associated with survival and not incidence, susceptiblet o inaccurate recall, comparison groups harder to find
Measure of association for case control
odd ratios
could use relative prevalence of exposure, but doesn’t work well if you want to control for multiple confounding variables
odds ratios can
approximates relative risk under rarirty condition. can be used for cohort, but relative risk preferred for ease of interetation
Odds are a way of assessing
likelihood of an outcome to likelihood of that outcome not occurring
odds ratio of exposure
odds of exposure among cases divided by odds of exposure among control (uased in case-control)
OR of disease is algebraically equal to the OR of exposure
odds ratio calculate
ad/bc
if it goes
ab
cd
Interpreting odds ratio
> 1 = exposure associated w more disease
<1 = exposure associated w less disease
equal to 1 = no association between disease and exposure
odds ratios can bea. good estimate of relative risk/relative prevalence if
- cases are representative of all ppl with disease in pop
- control are representative
- disease being studied doesn’t occur frequently among people who are either exposed or not
Unless we are carefully using incident cases or can excluse survivor bias, an odds ratio still only approximates
relative prevalence, not a relative risk
Odds raito with amtched pairs
We can’t learn anything from pairs that are both exposed or unexposed. ratio is b/c
cross sectional studies are usually absolute ______ studies
prevalence
Two steps for cross-secitonal study
- identify pop. to study
- simultatenously identify both present exposure and present disease status
Disadvantages of cross-sectional study
Survivor bias
impossible (often) to know if exposure preceded outcome
Advantages of cross-sectional studies
if well-designed, can be pop.representatntive (rare for case control and cohort)
repeated cross-sections can gain info about changes in pop.level ooutcomes related to an exposure in some instance
Causal inference relies on
theory
Measures of association for cross sectional
Odds ratios are the tradition, but can easily use a relative prevalence or difference in prevalence
Surgeon general guidelines for causality
Temporal relationship
Strength of association
dose-response relationshion
replication of findings
biologic plausibility
consideration of alternate explanation
cessation of exposure
knowledge consistency
most common ways to measure strength of association
relative risk, odds ratio, risk difference
- temporality
one approach to falsifying. causal claim is to see if ssociatin exists at a time point when it should not
- strenght of association
how closely correlated are cause and effect.
strong association guards against confounding or bias, but sometimes large effects are due to bias or confounding. and well-executed studies can still identify causal small effects
- dose-response relationship
Increasing disease as exposure increases (strong evidence of causality, although absence doesn’t imply none)
Replication of findings
will the causal relationship be consistently found
Plausibility
findings should be reliable knowing the biology of the disease
Consideration of alternate explanations
rule out plausible competing hypotheses, including confounding and bias
studies not considering alternate explanations usually bad
Cessation fo exposure
one would expect to see cessationf of exposure lead to reduced disease
Consistency with other knowledge
what it sounds like
specificity of association
Not a particularly persuasive criterion.
* Some argue that a particular exposure should be
associated with a particular disease only, and vice
versa—otherwise it may be a signal of confounding.
Coufounding variable 3 characteristics
- causes outcome of interest
- associated w exposure of interest
- Is not in thd direct causal pathway between exposure of interest and the outcome
An association you observe due to confounding is a real association, but not a causal one. It isn’t
spurious
How to address confoudning
- study design (matching/randomization)
- Analysis (stratification and adjustment0
- Only randomization can remove the effect of an unknown confounder
How does stratification work for confounding variables?
When we stratify, we examine the association between our exposure and outcome at each value of a confounder (age adjustmne,t smokers v nonsmokers)
What is adjustment?
Technique by which stratified results are recombined to eliminate the effect of the confounding variable. create a weighted average across strata of a confounder
Two major types of bias
Selection
Information
Seleciton bias
occurs when there is a systematic error in selection of study participants distorting the observed association between exposure and outcome
Occurs when selections predispose to an association
Selection bias often occurs when
populations chosen are fundamentally different (primarily a problem with case-control studies)
Nonresponse bias
In many studies, grouops with particular illnesses or exposures may be more likely to participate. in cohort studies, different exposure groups may have differnent rates of LTFU
Information bias
Occurs when means of collecting info about subjects is inadequate, resulting in incorrect info about exposures or outcomes
Recall bias
participants inaccurately remember past exposures
rumination bias
ill people spend more time thinking about potench exposures than health people (a type of recall bias)
reporting bias
one group may be less likely to report exposures because of attitudes or beliefs
wish bias
participants seek affirmation that disease was not their fault
interviewer bias
person ascertaining exposure may questions more carefully cases than controls
bias in exposure identification typesr
recall bias
rumination bias
reporting bias
wish bias
interviewer bias
bias in outcome ID
observer bias - classifying exposed as cases more often than non-exposed
respondent bias - participants errs in whether they have disease
detection bias - certain exposures result in more common med visits - better detection
incidence/prevalence bias (survivor) - use of prevalent cases as an outcome may result in associations that increase survivial, not development
temporal bias - factor appearing to cause disease may actually result from it
lead time bias - participnts in a screening study may have just caught the disease earlier
Most bias cannot be addressed in analysis
Thus must be addressed in study design
Why not always draw SRSs?
Require a listing of all units
samples units
Could result in subgroup samples too small to say anything useful/miss groups entirely
Logistically inefficient
How to ensure a min sample of subgroup emembers or ensure you don’t miss some subgroup by cahnce?
stratify by subgroups
How does stratifying impact estimates prescision
makes them more precise
To reduce the risk of a bad sample drawn by change…
random sample proportional to strata sizes (making sure you have populations that don’t overlap)
Benefits of stratifying
Make sure that characteritsics are represented in the sample at the same fraction as in the pop.
Doesn’t require sampling weights
Might get small improvement in statistical precision
If we were to take an SRS of 100 within each stratum (with different pops)…
our sample estimate may be biased
How do we reduce bias in a stratified sample?
Re-weight the samples so that each observation is multiplied by the number of units its represents in the populations
What is the point of weighting data?
Accounts for unequal chances of selection (so when observations represent an unequal number of people in the pop.)
Failing to correctly weight your data will cause the point estimates to be wrong (and probably the SE too)
Remember
Stratification that self weights (proportional to strata size) affects precision how
alsmost never harms
Stratification to overssample some subpops affects precision how?
harms overall precision
Why does stratification to oversample some subpops harm overall precision?
You gain some precision within the smallest stratum, but lose overall precision because a larger part of the estimate is dtermined by a smaller part of the sample under the SRS
What are the best characteristics on which to straitfy?
ones that explain as much variation in our outcome of interest as possible. We get max reduction in variance when stratification results in units that are essentially the same within strata and strata are quite different from one another
PSUs and SSUs
primary sampling units (clusters)
secondary selection units
Cluster sampling
Often only feasible approach
Reduces logistics and time
allows you to sample without needing a listing of everyone eligible for sampling
downsides to cluster sampling
If very different from one another and members are very similar within each community, sample has much worse statistical precision than SRS of same size
Draw a self-weighting possible when possible!!! where all particpants have an equal probability of selection. give examples:
simple/systematic random samples
stratified samples with strata sample proporiton to their size in the pop
All things being equal, weighting will
cause a loss of precision
and complicate analysis
2 options for self-weighting cluster samples
- Draw SRS of clusters and then a constant % of units within each cluster
- Draw a sample of clusters where likelihood of selection is proportional to size of the cluster. Then draw a constant number of unis within each cluster (AKA PPS: probability proportionate to size)
Cluster sampling is unbiased…
but less precise
Why do we lose precision?
Depends on how different your clusters are from each other/how many there are/how homogenous they are inside
What improves precision
stratificatoin
sampling high fraction of pop. effect negligible unless sampling more than 20%
Worsens precision
cluster sampling
relatively small number of clusters
weighting - especially when weights vary substantially
unequal cluster sizes
disproportionate sampling across strata
Basic model of response
comprehension
retrieval
judgement/estimation
reporting
comprehension
do people understand the survey instructions and the specific item
retrieval
how do people recall info from long term mem
estimation and judgment
converting recalled occurrences/beliefs into an estimate/assessment that can be an answer
reporting
selecting a communicating an answer
problems of encoding
- people often don’t encode all occurrences
- surveyors usually cannot affect encoding
- you can sometimes provide cues
Why does misinterpreting Qs occur?
grammar
complexity
vagueness
unfamiliar terms
Reducing vagueness often increases…
complexity
To know what a Q means to respondents
use cognitive interviewing to understand what a question means to respondents
adapt/borrow questions that have been fully validated
another language won’t always apply
Memory problems
we incoporate subsequent info, older events/blank spaces tend to be answered as a generic event
how much time has passed is faulty
Judgement influenced by waht 3 things
wording
context (framing)
Placement (framing)
tradeoffs
open-ended vs closed questions
with ordinal scales -label every point
for numerical scale - center around expected average, make equal categories
When do respondents tend to be less likely to answer correctly?
when disclosing something illegal or risky
when perceiving that the surveyor or one’s peers expect a favored answer
Guidelines for sensitive questions
- Have survey be self-administered whenever possible
- Open questions for frequency or behavior rather than closed
- Use familiar terminology
- Longer questions to for more time to recall and retrieval
- Deliberately load the question
- position sensitive questions after participants have answered others
- ask about sensitive behaviors from the past
- diary
- assess sensitivity of items by asking questions
- validate data
Social desirability bias
Presence of Interviewer can induce
observable traits can change willingness to report attitudes related to trait
Logic Models / Logical Frameworks 1
Inputs
Activities
Outputs
Short-term outcomes
Long-term outcomes
Another logic model
structures
processes
outcomes
Indicators
Can wrap around each step in logic model
Monitor points where program most likely to break down
Assess end of cascade
Process evaluation
seeing if program is producing what it’s meant to and operating reasonably well in accordance with logic model theory
When measuring programs impact, ideally want to run RCT.. but
unethical
politicall untenable
complicated
Differences between traditional research adn eval
Research: starts with a hypothesis and is fairly unconstrained as to research design
Evaluation: starts with what is believed to be best implementation approach under circumstances
The fundamental challenge of both traditional research and eval
assessing differences compared to an unobservable counterfactual
Difference in analytics/data collection between research and eval
analyticl tools often the same
data collection is usually for routine program purposes so quality usually poorer than for reserach
sample size not determined ahead of time, power can be a challenge
program intervention can chnge over time
can’t really blind evaluators
randomization impossible
4 general approaches for outcome evals
- Only sample is program recipients after program has commenced
- Only have program recipients but under varying program conditions
- Can sample non-recipients and recipients, but only after the program
- can sample non-recipients and recipients, both before and after program
- Post implementation Program Recipients only
challenging for good outcome eval because no approximation of counterfactural
Solutions
Compare to objective outcome becnhmakrs
qualitative investgation of recipient’s perceptions of program
2.Only program recipients under diff. program conditions
You don’t have data from people who didn’t recieve program, but do have data from before, after, or different versions
Options for when you have only program recipients under different program conditions
- before after designs
- placebo outcomes
- ITS
Before after studies
if data existed on outcomes from before an intervention, can be compared to outcmes after intervention
(biases include info bias, selection bias, confounding by time, confounding by other elements of a program)
Interrupted time series
need measurement at lots of time points before and after the intervention/change
lets you assess/exclude histoircal trends that might cause misinterpretation in a simple before-after
you would know about simultaneous other big changes
Major challenge for before-after design
may be confounded by simultaenous changes or selection problems
Isolate effect of intervention by comparing improvements in target pop between…
outcomes the evidence base/theory suggests should result from intervention
outcomes that should not result from intervention but are subject to same confounding or bias
- Non-recipients also, but only at follow up
Counterfactual supported by comparable participants who do not recieve program
challenge is ensuring pops. are comparable on dimensions other than receipt of the program
basically cross-sectional studies where exposure is your program
- Combining control groups with before-after measurement
measuring before-to-after changes in an intervention group compared to comparison produces better counterfactual estimates. would only worry about time-based confounding
Common eval strategies for combining control groups w/before-after measurement
difference-in-differences
interrupted time series with comparison group
step-wedge
Difference in differences
Repeated cross sections among receipeints and non
use before-to-after change among comparable non-recipients to rep what you would have seen among recipients without intervention
What can D-I-D be potench confounded by
different pre treatment trends
simultaneous interventions
changes in pop. over time
Step wedge designs
Extension of DID
implement same program in diff locations over time
evaluate for similar trends over time, with an improvement in outcomes at the time of implementation
should see improvements over time as you finetune your methods
