[Rd] The boot Package with bca Intervals and Inf and NaN Values in an Automated Function [mediation()]
Ken Kelley
xpxixpy at gmail.com
Tue Sep 20 16:28:48 CEST 2011
Hi everyone,
I use the boot package within the MBESS package to automate much of
the hard part for folks interested in performing a (simple) mediation
model. The function mediation() works well, except for a (probably)
unrealistic artificial data set.
When I apply the bootstrap I sometimes get:
Error in t.star[r, ] <- res[[r]]
which is (I think) due to Inf and -Inf and NaN values produced for
some of the estimates in the bootstrap resamples. Here are my warnings
> warnings()
Warning messages:
1: In sqrt((1 - R2.Y_XM) * ((N - 1) * Cov.Matrix[3, 3]/(N - ... : NaNs produced
2: In max(abs(b.contained)) : no non-missing arguments to max; returning -Inf
3: In max(abs(From.b)) : no non-missing arguments to max; returning -Inf
4: In max(abs(From.a)) : no non-missing arguments to max; returning -Inf
5: In sqrt((1 - R2.Y_XM) * ((N - 1) * Cov.Matrix[3, 3]/(N - ... : NaNs produced
6: In max(abs(b.contained)) : no non-missing arguments to max; returning -Inf
7: In max(abs(From.b)) : no non-missing arguments to max; returning -Inf
8: In max(abs(From.a)) : no non-missing arguments to max; returning -Inf
9: In max(abs(b.contained)) : no non-missing arguments to max; returning -Inf
10: In max(abs(From.b)) : no non-missing arguments to max; returning -Inf
11: In max(abs(From.a)) : no non-missing arguments to max; returning -Inf
and here is what boot.out looks like:
Bootstrap Statistics :
original bias std. error
t1* 0 0.01784141 0.030664717
t2* 0 0.00522446 0.009407206
t3* 0 0.01530606 0.026118581
t4* 0 0.10293090 0.132874564
t5* 0 Inf NaN
t6* 0 Inf NaN
t7* 0 0.06200062 0.069837762
t8* NaN NaN 0.165104486
t9* 0 0.01550669 0.026480612
t10* 0 0.01653285 0.029406830
t11* Inf -Inf 10.806967364
t12* 0 0.05329459 0.074168364
t13* 0 0.06153211 0.081591660
t14* 0 0.10558389 0.136462038
The original estimates of 0, NaN or Inf are correct for this
particular set of data (which is artificial for testing purposes).
I'm wondering if there is an easy workaround (e.g., to tell boot to
ignore NaN, etc.). Alternatively, I could make modifications to the
mediation() function in the MBESS package. The idea of the function is
to automate much of the hard part for folks interested in a simple
mediation model (i.e., so that non R users can easily use the
function). However, I'm not sure exactly what I should have
mediation() do so that an extreme data set does not produce the
problem when coupled with boot. I thought about removing the rows
from boot.out$t with NaN and then changing boot.out$R to reflect the
reduced number of rows, but I wasn't sure that was such a good idea.
The NaNs, by the way, come from dividing 0 by 0.
I didn't provide example code because the mediation() function I'm
using is not the released version. However, I can send the updated
mediation() function to whomever if interested (email me off list for
the updated function).
Any thoughts on this would be most welcomed!
Thanks,
Ken
More information about the R-devel
mailing list