PUBPOL 479 RCT Quiz

bias in rcts

partial/non compliance, balance, differential attrition, spillover/contamination

partial/non-compliance

• when the treated group doesn’t get the treatment

  • Some people defy the treatment (if it makes them respond diff than without)

  • Matters because it can reduce the difference between treatment and control

  • If we look at those  who took the treatment anyways, the T & C groups may no longer be comparable

  • Compliance becomes less of an issue for easy (to get or administer) treatments

balance

• the idea that the T & C look the same on average

  • Matters because the control group is meant to be a counterfactual of the experiment 

  • Ensuring balance addresses selection issues

differential attrition

• attrition (effect to balance) that occurs more/only occurs in one group → treatment and control dont look the same

  • Non-differential attrition is mostly okay, may reduce power

spillover/contamination

• indirect effects of the treatment that change the control group

  • If the treatment group tells control about its intervention, or treating a disease in one community increases herd immunity and protects another

  • Can reduce by randomizing at a higher levels (i.e. randomize at the school, rather than classroom, level)

intent to treat

• addresses partial compliance by averaging treatment on both groups, regardless of whether they were actually treated 

  • CTC ex: not everyone can receive the CTC from treatment group, but ITT assumes they did anyways

treatment-on-the-treated

• addresses partial compliance by analysing treatment effect based only on those who actually got the treatment

  • CTC ex: can be estimated if you know who actually got the credit, but can overestimate the effects if those who got it are differe

Generalizeability issues in RCTs

inclusion/exclusion criteria, differential response rates, who receives the intervention, size/number of participants (statistical power), differences in analytical methods by group, and ethics

methods of controlling for bias

patient act as his/her own control, cross-over designs, blinding of patients to intervention, blinding of investigators, blinding data analysts

research designs

RCT pre/post test w/ control, post test only w/ control, delayed intervention design

RCT Pre/post w/ control

• randomized groups, then both control and treatment observed before and after; only treatment group gets intervention in the middle

post test only w/ control

•  randomized, treatment given intervention, and both treatment and control are observed after T gets intervention

delayed intervention design

• control & treatment both randomized, T given intervention, both observed after, and then after second observation the control gets the treatment

  • Ethically superior

R O X O
R O O X

cross-over design

R O X O (X) O X O

R O (X) O X O (X) O

Compares two or more treatments by having the same participants receive treatments/interventions in a predetermined sequence

• in this case, (X) is the placebo