1

I made a lmer with an interaction between odour and concentration and random effect of date. I'm getting very high df in the emmeans output below. It's actually higher than the number of observations in my df (954). Why? Should I be concerned?

Also, can I specify "side = '>'" when I do the dunnett trt.vs.ctrl instead of what I am doing here?

summary1 <- emmeans(mod1, specs = trt.vs.ctrl ~ conc|odour)
summary(summary1)
test(summary1, side = ">",adjust = "tukey", type = "response")

$contrasts
odour = A:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.290 0.252 292 1.149   0.3320 
 0.1 - 0     0.594 0.214 933 2.775   0.0084 
 1 - 0       0.506 0.256 876 1.980   0.0702


odour = B:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.770 0.253 379 3.046   0.0037 
 0.1 - 0     0.706 0.259 302 2.721   0.0103 
 1 - 0       0.852 0.265 328 3.214   0.0022


odour = C:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.328 0.184 718 1.782   0.1085 
 0.1 - 0     0.658 0.230 275 2.858   0.0069 
 1 - 0       0.759 0.282 332 2.694   0.0111


odour = D:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.536 0.250 245 2.146   0.0484 
 0.1 - 0     0.659 0.292 381 2.261   0.0361 
 1 - 0       0.284 0.283 283 1.004   0.4034


odour = E:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.157 0.215 910 0.733   0.5470 
 0.1 - 0     0.508 0.256 291 1.988   0.0699 
 1 - 0       0.693 0.361 440 1.918   0.0814


odour = F:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.496 0.220 899 2.251   0.0365 
 0.1 - 0     1.282 0.213 983 6.018   <.0001 
 1 - 0       1.144 0.232 978 4.922   <.0001


odour = G:
 contrast estimate    SE  df t.ratio p.value
 0.01 - 0    0.312 0.242 263 1.288   0.2698 
 0.1 - 0     0.952 0.225 509 4.229   <.0001 
 1 - 0       0.997 0.204 473 4.881   <.0001


Degrees-of-freedom method: kenward-roger 
P value adjustment: sidak method for 3 tests 
P values are right-tailed
AdamO
  • 62,637
Michael
  • 11
  • This comes up a lot. emmeans isn't for the fledgling analyst. The KR df expression is complicated and there's no guarantee it's less than $n$. – AdamO Mar 08 '23 at 20:14
  • Thanks @AdamO I guess I am now concerned that by performing a dunnett comparison within odours could lead to bad results based on the questions asked in the link you sent. Since my model uses all data, but I am testing effect of concentration within levels of odours. Either way, thanks! Would there be a way to do a trt.vs.cntrl using glht so I can see if the results of emmeans and glht post-hoc are similar? Or would that not tell me anything? – Michael Mar 08 '23 at 20:28
  • There are tons of ways to analyze panel data. I don't know what to tell you. If the design is balanced, you can even just use a saturated least squares model (ignore subject, adjust time and test). I'm not a fan of post-hoc testing - it's an oxymoron. In my world, I'd be a "target trial framework" kind of guy, that is if you wanted to estimate X, design the trial with X as the endpoint. – AdamO Mar 08 '23 at 20:57

0 Answers0