Problems with Running a Dichotomous M...
Message/Author
 julie schiro posted on Friday, June 12, 2015 - 10:22 am
My mediation model has the following variables.

y = continuous
x = discrete (contrast coded -1,1)
m = discrete (contrast coded -1,1)

In my output, the indirect path is marginally significant at p = .09 but the 95 % confidence interval for the indirect path estimate *does not* contain 0, which implies significance at alpha = .05. Why do these estimates conflict? The first p-value implies that the mediation is not significant but the CI implies that it is. Which one is correct? Please see my code and output below.

DATA:
FORMAT IS free;
VARIABLE:
NAMES ARE y x m;
USEVARIABLES y x m;
ANALYSIS:
bootstrap = 5000;
MODEL:
m on x; y on m x;
MODEL INDIRECT:
y via m x;
OUTPUT:
cinterval(bcbootstrap);

*****select output*****

"TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND DIRECT EFFECTS"

Specific indirect [est, SE, est SE, Two-tailed p-value)
Y
M
X 0.027 0.016 1.694 0.090

"Effects from X to Y via M"
Est = .027 (same as above)
Lower 2.5% = .003
Upper 2.5% = .07

Thanks you,
Julie
julieschiro@gmail.com
 Bengt O. Muthen posted on Friday, June 12, 2015 - 4:54 pm
You say cinterval(bcbootstrap); which means that you are asking for non-symmetric CIs based on bootstrap. The p-values are for regular z-test, which is the same as a symmetric CIs. You should use the bootstrap CIs because they are more appropriate.
 julie schiro posted on Saturday, June 13, 2015 - 9:49 am