[R-sig-ME] LSmean: transformed error bars
ecp52 at cornell.edu
Sat Mar 5 17:20:00 CET 2016
Dear R gurus,
I am using the "lsmeans" function to extract predicted marginal means from
a Poisson family mixed model with a log link function. When I ask for the
"lsmeans" on the scale of the linear predictor, the standard errors and
confidence limits are symmetrical about the mean.
However, I would like to plot the responses on the original scale, so that
I can depict "number of eggs" rather than "ln(number of eggs)".
The function "lsmeans" offers this reverse transformation via the argument
"type='response'". However, I don't understand why there is only one
standard error after the response. When I manually exponentiate the
predicted means and CI's from the scale of the linear predictor to the
scale of the original response variable, my calculations match perfectly.
However, when I exponentiate the "mean + se" and "mean - se" for plotting
on the original scale, I get asymmetric bars.
So in lsmeans, what does the "SE" mean when using "type="response""?
Shouldn't there be separate + and - SE values?
The SE value seems to be very close to (upperCI-lowerCI)/4, but not
exactly. Should I plot mean +/- SE using "type="response"" for symmetric
error bars, or use my manually exponentiated mean +/-SE?
Thanks for your help!
More information about the R-sig-mixed-models