Epidemiology Flashcards
Odds ratio < 1
Indicates study favors the tx
Fixed Effect Model
Assumes each study answers same question, has same effect size, so results differ only by chance
Length-time Bias
Slow developing conditions are more likely to be picked up in screening, and screening will MISS many of the fast progressions
Confounding bias/factors
When a factor is related to both the exposure and outcome, but not on the causal pathway - the factor distorts or confuses the effect of exposure on outcome
(E.g. - age, known risk factors, known prognosis factors)
Types of studies for systematic review
Randomized trials
Cohort
Case-control
Diagnostic tests
Meta-Analysis
Use of statistical methods to combine results of individual studies, usually from systematic reviews
- advantages: adequate sample size and power to evaluate small tx effects; good if analysis can be done w/ data from individual pt
- disadvantages: quality is dependent on studies; may be too heterogeneous to combine; pt’s are variable
Random Effect Model
Assumes studies address different but related questions, takes heterogeneity into account, less likely to overestimate precision, wider CI, more realistic
Subgroup analyses (without specifying in advance)
Analyses and outcomes must be specified before study is conducted
Odds ratio = 1
Indicates no association
Validity (accuracy)
Extent that the measurement represents what it’s supposed to
-compromised by systematic error
Odds ratio > 1
Indicates the study favors the placebo/control
Selection Bias
Error in assigning subjects to a study group resulting in an unrepresentative sample
*most commonly a sampling bias
Compliance Bias
Compliant pt tend to have better prognosis regardless of preventative activities
Reliability (consistency; precision)
Extent to which repeated measurements are similar
-compromised by random error
Multiple comparisons (to find something)
It can’t be OK to keep testing one subgroup against another forever until one is significant
Measurement Bias
*AKA Information Bias
Information is gathered in a systematically distorted/inconsistent manner
Non-respondent bias (volunteer effect)
A sampling bias, where the research only includes those who say “yes” and those ppl are different from the ppl that say “no” so that doesnt tell you much about a population
Ascertainment bis
Sampling bias where ppl w/ more severe cases are more likely to be seen so we miss the more subtle cases
Late-look bias
Sampling bias where ppl w/ severe dz are less likely to be included in a study bc they’re hard to access or already dead (so bias is toward less sick cases)
Solution to sampling biases
Random sample (getting random ppl into the study), weigh data so sample matches the population
Selection bias (design bias)
Different ppl in the treatment and control groups (its like comparing apples and oranges)
Solution: random assignment (which part of the study for the participant to be in)
Hawthorn effect
The fact of measurement can change what is measured (act different when you think someone is watching)
Solution: control group
Recall bias
Ppl don’t remember what happened in the past so they make things up
Solution: confirmation
Observer bias
You see what you’re tuned in to see (makes assessment based on prior knowledge or experience)
Solution: multiple observers
Lead-time bias
False estimate of benefits of an intervention (early detection is confused w/ living longer)
Solution: use life expectancy
Expectancy bias
Researcher unintentionally acts to influence behavior of subjects and change results
Solution: double-blind design (if researchers dont have knowledge they can’t influence how they deal w/ subjects)
Proficiency bias
New intervention/tx are not applied w/ equal skill to all research subjects
Solution: tx providers selected @ random
Confounding bias
“Found within” aka things wrapped up together
*an additional variable, not the subject of research interest that produces the observed results (often the “hidden cause” or underlying issue)
Solution: thoughtful research design, multiple studies (meta-analysis)