Hi all,
I am (happily) running an ordinal regression model with brms, but noticed some inconsistencies when comparing the results with the same model fitted with the package “ordinal”. Specifically, the results of the “brms” model fitted with link=“probit” appear broadly in line with the results of the “ordinal” model, but only if the latter is fitted with with link=“logit”. In sum, there does not seem to be a 1:1 correspondence between the results of the models fitted with the same link function.
What am I missing?
Best,
Luca
frequentist logit-link ordinal model
clmm.logit ← clmm(contrast ~ (scale(horn) + scale(mass))*Sex + scale(age) + scale(julian.date) + (1|area),
link = “logit”,
threshold = “equidistant”,
data=data.f)
frequentist probit-link ordinal model
clmm.probit ← clmm(contrast ~ (scale(horn) + scale(mass))*Sex + scale(age) + scale(julian.date) + (1|area),
link = “probit”,
threshold = “equidistant”,
data=data.f)
Bayesian logit-link ordinal model
brm.logit ← brm(contrast ~ (scale(horn) + scale(mass))*Sex + scale(age) + scale(julian.date) + (1|area),
family=cumulative(link=“logit”, threshold = “equidistant”),
warmup = 1000, iter = 5000, chains = 4,
data=data.f)
Bayesian probit-link ordinal model
brm.probit ← brm(contrast ~ (scale(horn) + scale(mass))*Sex + scale(age) + scale(julian.date) + (1|area),
family=cumulative(link=“probit”, threshold = “equidistant”),
warmup = 1000, iter = 5000, chains = 4,
data=data.f)
brm.logit
> Population-Level Effects:
> Estimate Est.Error l-95% CI u-95% CI
> Intercept[1] -11.33 2.99 -18.24 -6.64
> Intercept[2] -8.00 2.22 -13.16 -4.52
> Intercept[3] -4.67 1.51 -8.16 -2.32
> Intercept[4] -1.34 0.97 -3.54 0.29
> scalehorn 0.08 0.64 -1.15 1.36
> scalemass 2.72 1.21 0.69 5.45
> SexMales -2.87 1.35 -5.96 -0.67
> scaleage -1.44 0.55 -2.67 -0.52
> scalejulian.date 0.77 0.41 0.08 1.69
> scalehorn:SexMales 1.23 1.05 -0.64 3.51
> scalemass:SexMales -1.77 1.26 -4.51 0.46
brm.probit
Population-Level Effects: Estimate Est.Error l-95% CI u-95% CI Intercept[1] -6.58 1.57 -10.00 -3.90 Intercept[2] -4.64 1.17 -7.18 -2.66 Intercept[3] -2.70 0.80 -4.46 -1.34 Intercept[4] -0.77 0.53 -1.91 0.17 scalehorn 0.05 0.37 -0.66 0.79 scalemass 1.58 0.65 0.42 2.98 SexMales -1.67 0.74 -3.26 -0.38 scaleage -0.85 0.30 -1.49 -0.31 scalejulian.date 0.46 0.23 0.05 0.97 scalehorn:SexMales 0.73 0.61 -0.39 2.02 scalemass:SexMales -1.02 0.68 -2.45 0.24
car::Confint(clmm.logit)
Estimate 2.5 % 97.5 %threshold.1 -6.87 -9.343 -4.403
spacing 2.10 1.409 2.792
scale(horn) -0.07 -0.915 0.774
scale(mass) 1.51 0.263 2.764
SexMales -1.44 -2.852 -0.029
scale(age) -0.75 -1.342 -0.158
scale(julian.date) 0.41 -0.066 0.878
scale(horn):SexMales 0.62 -0.639 1.886
scale(mass):SexMales -0.86 -2.327 0.608
car::Confint(clmm.probit)
Estimate 2.5 % 97.5 %threshold.1 -4.235 -6.191 -2.278
spacing 1.285 0.731 1.838
scale(horn) -0.008 -0.532 0.516
scale(mass) 0.946 0.108 1.784
SexMales -0.892 -1.875 0.091
scale(age) -0.494 -0.910 -0.078
scale(julian.date) 0.266 -0.045 0.577
scale(horn):SexMales 0.383 -0.438 1.203
scale(mass):SexMales -0.551 -1.499 0.398