在生醫與其他領域中,常會面臨配對二元計數資料(paired bivariate count data)之分析問題。例如,病患於不同醫院之診斷結果、不同類型就診次數, 或同一觀測單位下兩個計數反應,皆可能具有相關性,且有時呈現負相關。為 描述此類資料,Gonçalves et al.(2025)提出二元混合卜瓦松模型,透過共同隨 機效應與相關性參數,同時處理過度離散、未觀測個體差異與正、負相關結 構。 本研究比較二元負二項分佈(Bivariate Negative Binomial, BNB)、伽瑪共 同隨機效應下之二元混合卜瓦松模型 BMPG ,以及逆高斯(Inverse Gaussian)共 同隨機效應下之二元混合卜瓦松模型 BMPIG 三類實作模型在二元相關計數資料 分析中的表現,重點包括均值估計、變異數估計、假設檢定與模型診斷。模擬 與實例分析結果顯示,BNB 模型雖然相關結構較受限制,但形式簡單,且對均 值參數之估計具有穩定性。搭配 Royall 與 Tsou(2003)之強韌化概似方法 後,BNB 模型在模型可能錯誤指定時,其推論表現優於傳統非強韌推論方法。;Paired bivariate count data are common in biomedical and related fields, such as medical visit counts or two count responses from the same observational unit. Such data may be correlated and may sometimes exhibit negative correlation. To describe these data, Gonçalves et al. (2025) proposed bivariate mixed Poisson models, which use a shared random effect and a correlation parameter to account for overdispersion, unobserved heterogeneity, and dependence. This study compares three working models: BNB, BMPG ,and BMPIG . Through simulations and an empirical analysis, we examine their performance in mean estimation, variance estimation, hypothesis testing, and model diagnostics. The results show that the BNB model, although limited in its correlation structure, is simple and provides stable estimation of mean parameters. With the robust likelihood method of Royall and Tsou (2003), BNB provides more stable inference than conventional nonrobust methods under possible misspecification