Making Causal Inferences and Ruling out Rival Explanations 29 February
Questions? How do we know that X is causing Y? Did X have any effect on Y? If X had not happened would Y have changed anyway?
Hypothesized relationship: %Women elected in National Parliaments Party rules gender quotas
Questions? How do we know that party quotas causing changes in %women elected? Standard Design Party adopts quotas % women elected X O Where X = treatment and O = observation
Establishing Causation: Co-variation Time – (x occurs before y) Consistent with other evidence Rule out rival explanations Example – spurious relationship
Spurious Relationship a relationship in which two variables that are not causally linked appear to be so because a third variables in influencing both of them
Spurious Relationship Fire damage in $ # of fire engines responding to call + Intensity of fire + (the third variable problem)
Alternative explanations: Electoral System Political Culture %Women elected in National Parliaments Women’s Labor Force Participation Party rules - quotas Access to educational opportunities Women’s Political Resources % of women candidates standing for election
Spurious Relationship %women elected Party quotas + Political culture +
When choosing a research design? When and how to make observations: Internal Validity Ability to establish causality External Validity Ability to generalize
Types of Designs: Experimental designs Quasi-experimental Control groups Quasi-experimental Non-experimental designs Statistical controls
Manipulation of an independent variable Experiments come in a wide variety of apparent types but all share three basic characteristics: Random assignment Manipulation of an independent variable Control over other potential sources of systematic variance X O1 R O2
These basic characteristics effectively solve the two basic problems in nonexperimental (correlational) research: The directionality problem The third variable problem
Random Assignment Random assignment means that assignment to experimental conditions is determined by chance. Participants have a equal probabilities of being assigned to a treatment or control group. This insures that any pre-existing characteristics that participants bring with them to the study are distributed equally among the experimental groups . . . in the long run. Treatment group = (equivalent to) Control group
Think about example of party quotas a % women elected: Randomly assign countries to two groups: treatment and control Theoretically should end up with two groups that have equivalent distributions on all other “third variables” (i.e. culture, % women in labour force, etc.) Have one group adopt quotas Observe % women elected, treatment group expected to have higher average for % women elected.
Problems? Random assignment might be difficult in this case. Turn to quasi-experiments when randomization not possible
To Review - One Group Post-Test Only Design X O The simplest and the weakest possible design: Lack of a pretest prevents assessment of change Lack of a control group prevents threats from being ruled out.
Threats to Internal Validity Selection Threats Maturation History Testing Instrumentation Regression Note: Experimental designs control for these Party adopts quotas % women elected X O
One Group Post-Test Only Design X O Without changing the basic nature of this design, it can be improved considerably by adding additional outcome measures: O1 X O2 O3 Compared to norms or expectations, only O2 should be unusual.
Post-Test Only Design with Nonequivalent Groups X O O Threats: Selection
One-Group Pretest Post-Test Design O X O This very common applied design is susceptible to all threats to within-groups comparisons: History Maturation Testing Regression Instrumentation
One-Group Pretest Post-Test Design O X O One powerful modification is to add pretests: O O O O O X O Maturation threats can now be examined and their influence separated from treatment effects.
O O O O O O O O O X O
Untreated Control Group Design with Pretest and Posttest O1 X O2 O1 O2 Can compare change within groups and across groups Expect change in treatment group to be greater Selection still a threat
Conclusions: Experiments best for internal validity May not be good on external validity In non-experimental designs, use statistical controls (hold constant all possible “third” variables.