Controlled Experiments vs. Observational studiesControlled ExperimentsObservational StudiesAnalyzing Observational StudiesAssociation, Causation & ConfoundingConfoundingCorrecting for confoundingExamplesExample 1: PellagraExample 2: Smoking and Lung CancerExample 3: Death PenaltySimpson's ParadoxSummaryObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryStat 220, Part IDesign of Experiments and StudiesLecture 4Chapter 2: Observational studiesObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryControlled experiments (review)Optimal design: double-blind randomized controlled study•take a representative sample of subjects (also: study units)•randomly divide the subjects into a treatment and acontrol group (randomized controlled)•use placebo and do not let the subjects know in whichgroup they are; do not let evaluators know who getstreatment/placebo (double-blind)•evaluate the response in b oth groups and compareWhy is this a good design?•due to randomization, the two groups will be statisticallyequivalent except for the treatment•hence, if there is a difference in the responses, then this islikely caused by the treatment (or by random chance; butwe’ll deal with that much later).ObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryControlled experiments,observational studiesControlled experimentsInvestigators (randomly)assign subjects totreatment/control groups.Observational studiesSubjects “assign themselves”to treatment/control groups.Investigators just observewhat happens.Why do observational studies at all? In many cases, controlledexperiments can be hard or even impossible to conduct.Examples:•Does smoking cause cancer?•Does carbon emission cause global warming?Problem: often the groups are not comparable. Or there isonly one group!ObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryAnalyzing observational studiesJust like with controlled experiments, we use the method ofcomparison: we compare a treatment (or exposed) group to acontrol group.Examples:1 Studying the effect of a new surgery:•Treatment group: patients who get new surgery•Control group: patients who get standard surgery2 Studying the effect of smoking:•‘Treatment’ group: smokers•Control group: non-smokersSince we could not randomize the two groups, we cannot besure they are similar. Actually, in general they will be differentin additional ways, on top of treatment vs. no treatment.ObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryAssociationThis is why we use a different term.DefinitionTwo variables are associated, if knowing the value of onevariable gives you information on the other.Examples:Weight and height of people are associated:•If you know that somebody is tall, it is more likely that theperson is heavy.•If you know that somebody is heavy, it is more likely that theperson is tall.Smoking habits and the height of adults are not assoc iated:•If you know that somebody is a smoker, that gives you noinformation about her/his height.•If you know somebody’s height, that gives you no informationon whether he/she is a smoker.ObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryAssociation vs. causationA good r andomized controlled experiment can establishcausation.•Example: vaccine decreases risk of getting polio.An observational study can establish association betweenvariables.•Example: smokers tend to have a higher rate of livercancer.What does association mean here?If we compare smokers and non-smokers, there is a higher rateof liver cancer among the smokers. So if you smoke, you aremore likely to get liver cancer.This does not imply that smoking causes liver cancer!ObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryConfoundingWhat could explain the fact that smokers have a higher rate ofliver cancer?•Smokers tend to drink more alcohol than non-smokers•Excessive alcohol consumption causes liver cancerThe effect of smoking is confounded (mixed-up) with the effectof alcohol consumption. Alcohol is a confounding factor.DefinitionConfounding means that the treatment and control group differby some factor (other than the treatment) that influences theresponse/outcome that is studied.ObservationalStudiesControlledExperimentsvs. Observa-tionalstudiesControlledExperimentsObservationalStudiesAnalyzingObserva-tionalStudiesAssociation,Causation &ConfoundingConfoundingCorrecting forconfoundingExamplesExample 1:PellagraExample 2:Smoking andLung CancerExample 3:DeathPenaltySimpson’sParadoxSummaryConfoundingIf the two groups are similar in all aspects that may affect theoutcome, then there is no confounding. In this case differentoutcomes for the two groups are likely caused by the