檢驗與擴展社會網絡分析在結構方程模型的應用：蒙地卡羅模擬研究

Abstract

心理學家主張個人是鑲崁在社會網絡之中，個人的潛在特質、狀態、與外顯行為受到社會網絡的影響，並建議研究者應在社會網絡脈絡中探討個體的行為表現。同時考量社會網絡與潛在變項資料至單一分析模型是一個重要的議題，然有關此議題之方法學相關研究仍有不足，因此本論文藉由執行兩個模擬研究，以探討此議題。研究一設計不同程度的因素負荷量、跨因素負荷量、樣本數、因素規模、與模型設定，以檢驗結合潛在因素資料與社會網絡資料之二階段最大概似估計程序（two-stage maximum likelihood estimation procedure）（Liu et al., 2018）的影響因子及其適用情境。整體而言，外顯預測變項之迴歸係數在所有模擬情境中皆有良好的表現。而潛在預測變項之迴歸係數及其估計標準誤的相對偏差程度隨跨負荷量越高，或因素負荷量越低而越為明顯。模擬研究二設計多元的網絡型態、節點連結程度、網絡規模、與模型設定，以評估該分析程序的實徵表現，以檢驗將社會網絡資料整合至SEM中的分析程序的實徵表現。研究結果顯示，此分析程序在所有模擬情境中皆表現良好，具有良好的模型適配度，以及可接受的參數估計及其標準誤的相對偏差。透過研究一與二的發現，除了可增加SNA與SEM在方法學知識的累積，填補文獻上的不足，並且藉由考量模擬近似於現實場域所面臨到的情境，可提供實徵研究者在實徵應用上的參考訊息。

Psychologists claim that individuals are embedded in social networks, and their implicit traits and explicit behaviors are affected by social networks. They suggest that researchers should explore individual behaviors in the context of social networks. It is important to consider social network and latent variable data in a single statistical model. However, there still needs to be more methodological research on this issue. Therefore, the dissertation investigates this issue by performing two simulation studies. In study 1, various levels of factor loadings, cross-factor loadings, sample sizes, factor sizes, and model specifications were designed to test the impact factors and applicable conditions of the two-stage maximum likelihood estimation procedure that combines latent variable and social network data (Liu et al., 2018). Overall, the regression coefficients of observed predictors performed well in all simulated conditions. The degree of relative bias of the parameter estimates and their standard error are elevated as cross-factor loadings increase, or factor loadings decrease. In study 2, multiple network types, connected level between nodes, network sizes, and model specifications were designed to test the empirical performance of the analytic procedure proposed by this dissertation that integrates social network data into a structural equation model. The results showed that this analytical procedure performed well in all simulated conditions, with a good model fit and acceptable relative bias of parameter estimates and their standard errors. This dissertation not only improves the accumulation of methodological knowledge on social network analysis and structural equation modeling but also provides valuable suggestions for empirical researchers.