Sea denotar la respuesta y el vector predictor (respectivamente) del estudiante i en la escuela j .yij,xijyoj
(1) Para los datos binarios, creo que la forma estándar de hacer descomposiciones de varianza análogas a las realizadas para datos continuos es lo que los autores llaman Método D (comentaré los otros métodos a continuación) en su enlace: imaginando los datos binarios como que surge de una variable continua subyacente que se rige por un modelo lineal y descompone la varianza en esa escala latente. La razón es que los modelos logísticos (y otros GLM) surgen naturalmente de esta manera:
Para ver esto, defina manera que se rija por un modelo mixto lineal:y⋆yo j
y⋆yo j= α + xyo jβ + ηj+ εyo j
donde son coeficientes de regresión, η j ∼ N ( 0 , σ 2 ) es el efecto aleatorio a nivel escolar y ε i j es el término de varianza residual y tiene una distribución logística estándar . Ahora dejaα , βηj∼ N( 0 , σ2)εyo j
yyo j= ⎧⎩⎨⎪⎪10 0si y ⋆yo j≥ 0si y ⋆yo j< 0
vamos ahora, simplemente usando el CDF logístico que tenemospagyo j= P(yij= 1|xij,ηj)
pagyo j= 1 - P( y⋆yo j< 0 | Xyo j, ηj) = exp{ - ( α + xyo jβ + ηj) }1 + exp{ - ( α + xyo jβ + ηj) }
ahora tomando la transformación logit de ambos lados, tienes
log(pij1−pij)=α+xijβ+ηj
que es exactamente el modelo logístico de efectos mixtos. Entonces, el modelo logístico es equivalente al modelo de variable latente especificado anteriormente. Una nota importante:
- La escala de no se identifica ya que, si fuera a reducirla pero a una constante s , simplemente cambiaría lo anterior aεijs
exp{−(α+xijβ+ηj)/s}1+exp{−(α+xijβ+ηj)/s}
por lo tanto, los coeficientes y los efectos aleatorios simplemente se ampliarían en la cantidad correspondiente. Por lo tanto, s = 1 se utiliza, lo que implica v un r ( ε i j ) = π 2 / 3 .
s=1var(εij)=π2/3
Ahora, si usa este modelo y luego la cantidad
σ^2ησ^2η+π2/3
estima la correlación intraclase de las variables latentes subyacentes . Otra nota importante:
- εij
σ^2ησ^2η+1
estimates the tetrachoric correlation between two randomly selected pupils in the same school, which were shown by Pearson (around 1900 I think) to be statistically identified when the underlying continuous data was normally distributed (this work actually showed these correlations were identified beyond the binary case to the multiple category case, where these correlations are termed polychoric correlations). For this reason, it may be preferable (and would be my recommenation) to use a probit model when the primary interest is in estimating the (tetrachoric) intraclass correlation of binary data.
Regarding the other methods mentioned in the paper you linked:
(A) I've never seen the linearization method, but one drawback I can see is that there's no indication of the approximation error incurred by this. In addition, if you're going to linearize the model (through a potentially crude approximation), why not just use a linear model in the first place (e.g. option (C), which I'll get to in a minute)? It would also be more complicated to present since the ICC would depend on xij.
(B) The simulation method is intuitively appealing to a statistician since it would give you an estimated variance decomposition on the original scale of the data but, depending on the audience, it may (i) be complicated to describe this in your "methods" section and (ii) may turn off a reviewer who was looking for something "more standard"
(C) Pretending the data is continuous is probably not a great idea, although it won't perform terribly if most of the probabilities are not too close to 0 or 1. But, doing this would almost certainly raise a red flag to a reviewer so I'd stay away.
Now finally,
(2) If the fixed effects are very different across years, then you're right to think that it could be difficult to compare the random effect variances across years, since they are potentially on different scales (this is related to the non-identifiability of scaling issue mentioned above).
If you want to keep the fixed effects over time (however, if you see them changing a lot over time, you may not want to do that) but look at the change in the random effect variance, you can explore this effect using some random slopes and dummy variables. For example, if you wanted to see if the ICCs were different in different years, you culd let Ik=1 if the observation was made in year k and 0 otherwise and then model your linear predictor as
α+xijβ+η1jI1+ η2 jyo2+ η3 jyo3+ η4 jyo4 4+ η5 jyo5 5+ η6 jyo6 6
esto le dará un ICC diferente cada año pero los mismos efectos fijos. Puede ser tentador usar una pendiente aleatoria en el tiempo, haciendo que su predictor lineal
α + xyo jβ + η1+ η2t
pero no lo recomiendo, ya que eso solo permitirá que sus asociaciones aumenten con el tiempo, no disminuyan .