My experimental design is GRBD's (generalized randomized block design) with split plot (strips 1&2).
I made my model with the lmerTest package to check the effects of g_diversity and t_diversity on the response variable decomposed_weight:
m3b lsmeans(m3b, list(pairwise ~ g_diversity:t_diversity), adjust = "tukey")
Mix,2 years - No-Mix,2 years 0.8687
Mix,2 years - Mix,4 years 0.0526
Mix,2 years - No-Mix,4 years 0.6410
No-Mix,2 years - Mix,4 years 0.2573
No-Mix,2 years - No-Mix,4 years 0.1999
Mix,4 years - No-Mix,4 years 0.0212
_______________________________________
> difflsmeans(m3b, test.effs = "g_diversity:t_diversity")
g_diversity:t_diversity Mix 2 years - No-Mix 2 years 0.474
g_diversity:t_diversity Mix 2 years - Mix 4 years 0.013 *
g_diversity:t_diversity Mix 2 years - No-Mix 4 years 0.258
g_diversity:t_diversity No-Mix 2 years - Mix 4 years 0.076 .
g_diversity:t_diversity No-Mix 2 years - No-Mix 4 years 0.056 .
g_diversity:t_diversity Mix 4 years - No-Mix 4 years 0.005 **
I would like to know-
Why the tests provides different results?
Which of the tests fits to my situation? (if any)
Thanks!
I agree with Salvatore. The difference is because of different adjustments in the two commands i.e. Tukey in lsmeans but no adjustment in difflsmeans. The Ismeans command results are more reliable because they are adjusted for the risk of false discoveries (Type 1 errors) inherent in multiple comparisons.
Tukey is essentially a Student's t test – it simultaneously compares the mean of every treatment to the mean of every other treatment and identifies any difference between two means that is greater than the expected standard error AND corrects for family-wise error rates (or false discoveries) to counteract the chance of such type 1 errors, which can occur simply due to making multiple comparisons :)
I suspect that if you change the option in lsmeans to adjust="none", that the results will be more similar.
I agree with Salvatore. The difference is because of different adjustments in the two commands i.e. Tukey in lsmeans but no adjustment in difflsmeans. The Ismeans command results are more reliable because they are adjusted for the risk of false discoveries (Type 1 errors) inherent in multiple comparisons.
Tukey is essentially a Student's t test – it simultaneously compares the mean of every treatment to the mean of every other treatment and identifies any difference between two means that is greater than the expected standard error AND corrects for family-wise error rates (or false discoveries) to counteract the chance of such type 1 errors, which can occur simply due to making multiple comparisons :)
You will find it helpful to read the documentation for each of the functions you are considering using. This can be accessed in R for example with library(lmerTest); ?difflsmeans . I don't see that there is an adjust option in difflsmeans.
Personally, I just use the lsmeans package. It works for many different model types and has options for p-value adjustment and contrasts among treatments.
- Badges
- Science topic