Dave MacKinnon - Distribution of Products Approaches to Testing the
Significance of Mediation Effects
Overview:
-
Type 1 error rates are inaccurate and power is low for most tests of mediation
-
A new method to test for mediation effects has more accurate type 1 error
rates and power
-
The new method is based on the distribution of Products of Random Variables
-
The statistical background and a computer program to compute cumulative
probabilities is described
Equations used to test for mediation:
3 Methods to Test for Significance of the Mediated Effect
1. Causal Step Methods
-
Judd & Kenny (1981)
-
Baron & Kenny (1986)
2. Difference in Coefficient Methods
-
Variance of
-
Variance of the raw minus partial correlation
3. Product of Coefficient Methods
-
Variance of
- first order
-
Variance of
- second order
-
Distribution of
-
Empirical distribution of
Distribution of Products
(Craig, 1936, Springer, 1978, Aroian & colleagues, 1947, 1977)
-
Normal distribution is well known
-
Product of two regression coefficients,,
is the product of two independent random normal variables
-
Distribution is complicated and not normal
-
When,
the distribution has kurtosis of 6. This is why tests for mediated effects
have low power and incorrect type 1 error rates.
Moments of the distribution:
skewness and kurtosis (mean and variance are similar)
|
|
skew |
empirical skew |
kurtosis |
empirical kurtosis |
0 |
0 |
0 |
-.16 |
6 |
6.97 |
small |
small |
1.15 |
1.15 |
3.50 |
3.32 |
Simulation study of comparison of methods:
Causal step - low Type 1 error
Diff. In Coefficient - low Type 1 error
Product of Coefficient - and
Empirical Distribution - good type 1 error and pretty good power at sample
size of as low as 50 and 100
Conclusions:
-
The new method directly addresses the distribution of the product which
is the mediated effect estimate
-
Accurate type 1 error rates and more statistical power than the other methods
-
Can be used to obtain more accurate asymmetric confidence limits
-
It can be extended for more complicated mediational tests
-
Mathematica program has been developed