Section 7.2: How to Test Whether a Treatment Works

1
Section 7.2 How to Test Whether a Treatment
Works
2
Highlights

• Three types of correlation that we will study (or
  have already studied)
• Categorical vs. Categorical
• Categorical vs. Numerical
• Numerical vs. Numerical

3
Categorical vs. Categorical

• Example Is there correlation between year in
  school and whether or not a person smokes?
• Appropriate test Chi-squared

4
Categorical vs. Numerical

• Example is there correlation between yearly
  income and whether or not a person voted in the
  last election?
• Appropriate test ANOVA

5
Numerical vs. Numerical

• Example Is there correlation between college
  GPA and starting salary in your first job after
  college?
• Appropriate test Linear Regression

6
Correlation Study vs. Experiment

• With a correlation study, the researcher simply
  gathers data passive activity
• An experiment is a study in which the objects are
  randomly assigned to conditions.

7
Important Notes

• Correlation does not necessarily imply causation.
  Finding correlation may only improve your
  ability to predict.
• To establish causation you need to run an
  experiment rather than a correlation study.

8
Logic of Experiments

• Randomly assign subjects to treatment groups.
  This ensures that the two groups look the same
  before the experiment begins.
• Assign a different treatment to each group.
• Determine if there is any significant difference
  between the groups. If there is then its likely
  the treatment made a difference. (Why?)

9
Why experiments reveal causation?

• Random assignment of treatments creates groups
  that are not significantly different in almost
  all ways. The only significant difference
  between the groups is how they are treated. If,
  at the end of the study, the groups differ
  significantly, the only respectable theory of
  cause is that the difference in treatments led to
  a difference in the groups.