t****k
发帖数: 1095 | 1 两个model, 一个的R2提高很多，但是AICc也大不少，这个什么原因呢？变量的VIF都小
于5，大多在4一下。 | y*****w
发帖数: 1350 | 2 This is all about the number of model parameter estimates.
AICc = AIC + 2k(k+1)/(n-k-1) = 2k - 2lnL + 2k(k+1)/(n-k-1) = 2kn/(n-k-1) -
2lnL
Adjusted R-squared = 1 - (RSS/(n-k-1)) / (TSS/(n-1))
For AICc, assuming a fixed maximum likelihood, the higher the number of
model parameters (k), the larger 2kn/(n-k-1), and thus the larger the AICc;
whereas for adjusted R-squared, while the residual sum of squares (RSS)
always decreases as the number of model parameters increases, at the same
time (n-k-1) decreases as well, such that RSS/(n-k-1) may increase or
decrease. | t****k
发帖数: 1095 | 3 谢谢。两个model的independent variables一样多，但是第二个estimate 的parameter
是要多， AICc算出来大了10以上，但是adjusted R2提高了一倍。这种情况到底哪一个
更好呢？看文献说如果AICc差10以上，就不应该选AICc大的model,可是R2差这么多，也
不考虑吗？
;
【在 y*****w 的大作中提到】
: This is all about the number of model parameter estimates.
: AICc = AIC + 2k(k+1)/(n-k-1) = 2k - 2lnL + 2k(k+1)/(n-k-1) = 2kn/(n-k-1) -
: 2lnL
: Adjusted R-squared = 1 - (RSS/(n-k-1)) / (TSS/(n-1))
: For AICc, assuming a fixed maximum likelihood, the higher the number of
: model parameters (k), the larger 2kn/(n-k-1), and thus the larger the AICc;
: whereas for adjusted R-squared, while the residual sum of squares (RSS)
: always decreases as the number of model parameters increases, at the same
: time (n-k-1) decreases as well, such that RSS/(n-k-1) may increase or
: decrease.
| y*****w
发帖数: 1350 | 4 If one model is a full model and the other is a reduced model, have you run
a likelihood ratio test?
parameter
【在 t****k 的大作中提到】
: 谢谢。两个model的independent variables一样多，但是第二个estimate 的parameter
: 是要多， AICc算出来大了10以上，但是adjusted R2提高了一倍。这种情况到底哪一个
: 更好呢？看文献说如果AICc差10以上，就不应该选AICc大的model,可是R2差这么多，也
: 不考虑吗？
:
: ;
| t****k
发帖数: 1095 | 5 第一个model是用spss automated linear regression筛选的10个most significant
variables. 然后第二个model是基于这是个变量建立的geographically weighted
regression.
run
【在 y*****w 的大作中提到】
: If one model is a full model and the other is a reduced model, have you run
: a likelihood ratio test?
:
: parameter
| p***r
发帖数: 920 | 6 Because one of the model you use have missing values in one of the variable,
and records are dropped, rendering smaller AICs |
|
