research design and experimentation Flashcards
correlation does..
not imply causation
spurious correlation
two variables that have a high correlation by chance but are not related
confounding factor
third variable that influences dependent and independent variables, creating omitted variable bias
research design
strategy of how one addresses questions by integrating all parts of analysis to provide opportunity to deliver answers
A/B experiment
- sample split into controlled and treatment groups
- split randomised
- researcher assigns individuals without bias
- environment controlled as test before entering real world
field experiment
- real world application with one group exposed to treatment
- other factors influencing effect of treatment held fixed
natural experiment
- situation cannot be experimented due to unsafety/ethical issues
- reliable answers from available observational data
- existing data only designs experiment
- data already exists but not in context of design
- not randomised
reverse causality (regression)
eg market with low sales can influence increased advertising, as opposed to advertising increasing sales
selection bias (RCT)
eg group choose to participate in an intervention based on confounding variable
- experiment likely to provide a biased estimate of treatment effect
- random assignment avoids this
systematic correlation
- split into control/treatment can be skewed in a way as this is random, but not perfectly correlated
- this can occur due to a characteristic that determines split that effects outcome; confounding
2 rules of results of RCT
1) characteristics of treatment and control should look the same, differences should not be attributed to treatment
2) only one of the two groups should be treated
- expect 100% of treatment group is treated, or else effect is understated
