W2: Study Designs Overview Flashcards
Cohort study
Start with source population, select group exposed and unexposed. Follow over time; measure outcomes in each group. Samples based on exposure status and follows prospectively. Investigator does not control exposure.
Experimental (Randomized) trial
Source population –> invite people to participate, then randomly assign to treatment or control group. Follow over time, then measure outcomes. Investigator can control the intervention. Intervention precedes outcome measurement.
Case-control study
People from source population are selected based on disease status (cases are diseased, controls are not). Ideally drawn from same source population. Investigator does not control exposure (measured in the past). Unique d/t sampling based on outcome.
Cross-sectional study
Take sample from source population, measure who is exposed and unexposed. Snapshot – measures at only one point in time; compares outcomes in exposed and unexposed. No information about chronology of exposure and outcome. Can’t show causality.
Ecological study
Samples only group-level data; compares different populations (eg colon CA and egg consumption in different countries). Ecological fallacy: incorrect assumption that association at population level exists at individual level (i.e. the men with cancer may not be the same men eating eggs).
Quasi-experimental
Study in which groups with and without the intervention appear to differ only in intervention status, but were not randomly assigned
Observational
Study in which investigator does not control the intervention
Features of experimental studies
- Better quality comparison group
- Must consider ethical aspects of randomization
- Setting may be less realistic
Features of observational studies
- More natural setting
- Fewer ethical concerns (no randomization)
No individual level data?
Ecological study
Goal OTHER THAN to understand how exposure/intervention affects disease?
Cross-sectional study
Can you control the exposure/intervention?
YES: Experimental study
NO: Observational study
For obs study- Are any of the following true?
Disease is rare
Exposure data difficult/expensive to obtain
Little is known about disease
Disease has long induction or latent period
Underlying population is dynamic
Study needs to be done quickly (eg outbreak setting)
YES: Case-control study
NO: Cohort study
Parallel arm study
Type of experimental study in which investigators randomize one group to intervention, one group to control; measure predictors prior and outcomes post.
Crossover study
Enrollment process same as parallel arm trial; mid-way through investigators measure outcomes, then have washout period and groups change assignments. Outcomes measured again at the end.
Factorial study
Investigators randomize individuals to single or combined treatment (eg drug A &B, drug A & placebo, drug B & placebo, placebos A&B). Interested in single vs combined effects. Assumes intervention effects are independent (if not, larger sample sizes needed)
Cluster/community study
Entire groups are randomized rather than individuals. Cluster can be any group w/shared characteristic that connects them (eg people w/certain healthcare provider, villages, schools). Use when intervention delivered at the group level. Special sample size calculations and analyses needed to account for clustering.
Exchangeable
Control and treatment groups being balanced on measured and unmeasured characteristics
Key requirements for randomization
Truly random (all participants have fixed probability of assignment to each group). Not determined by investigator, clinician, or participants. Not based on predictable pattern (because then investigator could predict who gets intervention vs control).
Truly random randomization techniques
Coin flip, random number table, computer generated random number (most commonly used)
Non-random randomization techniques
MRN, name, DOB, order of enrollment
How do you check that randomization was successful?
Check the balance table (compares characteristics of each group at baseline)
What’s the most common reason randomization fails?
Sample sizes are too small
Stratified randomization
Randomizing groups separately based on certain given characteristic (eg randomize men separately from women, to ensure even balance of treatment assignment by gender)
