変量効果なしの負の二項モデルにおける−2残差対数疑似尤度のモデル比較について

[#1]

変量効果を指定しない負の二項・対数リンクのモデルで出力される「−2残差対数疑似尤度」は、説明変数を入れ替えたモデル、または交互作用の構成が異なるモデルの適合比較に使用できますか。使用可能な条件と計算方法を教えてください。

• JMP Student Edition 19.1.1　を使ってます。
• 「一般化線形混合モデル」を使用し、変量効果・offsetは指定していません。
• 応答は5株の1株当たりのある種の葉数の平均値で、日別の按分による非整数値を含みます
• 負の二項分布・対数リンク
• 比較したいのは、気象変数を入れ替えたモデルと、交互作用の構成が異なるモデル
• レポート画像の一つを貼り付けます。

初めてここに質問（投稿）します、ちょっと様子がよくわかりませんが、よろしくお願いします。

[#2]

I would like to clarify how JMP handles this specific case.

I am using JMP Student Edition 19.1.1 and fitting a negative binomial model with a log link in the Generalized Linear Mixed Model personality, without any random effects or offset. My response is the daily mean number of newly diseased leaflets per plant across five plants, including noninteger values resulting from averaging and allocation of counts between survey dates.

I understand that, without random effects, the model has the structure of a generalized linear model. However, JMP reports “−2 Residual Log Pseudo-Likelihood.”

Could you please clarify the following?

1. In this specific setting, how are the negative binomial dispersion parameter and the reported −2 Residual Log Pseudo-Likelihood calculated? Is this statistic based on the actual negative binomial likelihood or on the residual likelihood of a linearized model?
2. Can this statistic be compared between models fitted to exactly the same response values and rows, but using different weather predictors or different interaction terms? Does having the same number of parameters make such comparisons valid?
3. If these comparisons are not valid, should I refit the models using Negative Binomial and Maximum Likelihood in the Generalized Regression personality and compare AICc?

A reference to the relevant JMP documentation or calculation formula would be very helpful. Thank you.