STAT212 Exam 3

Ch. 9-12

Experiment

Designed to identify causal relationships;
we have the power to assign participants to an intervention

Observational studies

Identify associations that may signal causation;
we can only observe what participants choose to do

Confounder

Explains why two outcomes may correlate, but not be causally linked;
truly causes in/direct change in the response variable;
linked to explanatory variable, but not necessarily causally;
linked to both cause and effect

Mediating variable

Mediator between cause and effect;
links two things

Stratification

Breaking down comparison groups to control for confounders;
if differences still exist after stratification, argument for causality can begin to be built

Cross-sectional studies

Collect response and explanatory variable at a single point in a participant’s life

Cohort studies

Participants have an explanatory outcome, and then response outcomes are collected later;
typically prospective — response variable data is not available until later

Case-control studies

Participants are identified by their response outcomes, then data for explanatory outcomes is collected separately;
typically retrospective — explanatory variable data not available until later

Weakness of experiments, cross-sectional and cohort studies

If studying a rare incidence response outcome, we may need a very large sample size in order to estimate true risk with reasonable precision;
case-control studies can go straight to picking people with the rare outcome, bypassing this weakness

Odds ratio

Used instead of relative risk in cases like case-control studies;
not affected by non-natural incidence sampling;
tends to exaggerate

Blinded

Participants don’t know what group they are in

Double blinding

Neither the participants nor the people administering the intervention know what group they are in

Single group pre-post design

Control/intervention → Pre-measure → Treatment intervention → Post-measure;
Placebo effect and test familiarity are threats

Multiple group design

Separate treatment/control;

Treatment intervention → Response measure OR

Control (intervention) → Response measure

Random assignment w/ blocking

Units are blocked by potential confounders before random assignment into groups;

Helps assure equivalent groups when sample sizes are small

Non-randomized controlled experiment

Uses a non-random assignment method;
may create bias

Randomized controlled experiment with repeated measures

Take multiple measures per person at different time points;

may present a test familiarity threat

Randomized controlled experiments with crossover measures

Special case of repeated measures;

Each group completes both experimental conditions and produce response measures from both;

Need to be wary of lingering measures — may have buffer time between

Causality threats

• Group selection

• Drop out differences

• Test familiarity

• Timing effects

• Setting effects

• Independence

Group selection (threat)

If groups are being compared, they should be similar/equivalent

Drop out differences

Attrition/mortality → dropouts;

Threaten causality argument when drop out rates are different or the reasons for dropping out are different;

When there is some level of attrition due to non-adherence, researchers may make a comparison of all intent-to-treat participants

Setting effects

• Placebo effect

• Researcher effects

• Environment condition differences

Sample biases in random sampling

• Undercoverage bias

• Volunteer bias

Undercoverage bias

Some in the population don’t have an equal chance or have no chance to be selected

Volunteer bias

Once selected, those who chose to participate may be systematically different than those who chose not to participate (self-select) or forgot to/are unable to respond (non-response)

Survivorship bias

Especially important in biomedical studies;

Sampling process results in an inaccurate number of un/successful cases that don’t represent the population

Convenience sampling

Non-random sampling plan

Quota sampling

Identifies key features we want to represent and tries to sample a target number from each subgroup

Weighted sampling

Attempts to use a mathematical model to weigh for subgroups that are under/overrepresented

3 types of uncertainty

• Causality

• Generalizability

• Precision

Statistical power

How much we have reduced our uncertainty in precision

Two ways to improve statistical power

• Increase sample size

• Decrease random sources of variation

• + dividing sample size equally has a small effect

Abstract

150-300 word summary of a research paper

Four parts of research process

• Problem

• Design

• Analysis

• Conclusions