心理学报
心理學報
심이학보
Acta Psychologica Sinica
2008年
1期
109~118
,共null页
郭庆科 李芳 陈雪霞 王炜丽 孟庆茂
郭慶科 李芳 陳雪霞 王煒麗 孟慶茂
곽경과 리방 진설하 왕위려 맹경무
结构方程模型 模型拟合 2指数策略 2界值策略
結構方程模型 模型擬閤 2指數策略 2界值策略
결구방정모형 모형의합 2지수책략 2계치책략
Structural Equation Modeling, model - data fit, 2 - indite strategy, 2 - cutoff- value strategy.
在本模拟研究中设计了6种样本容量，6种因子载荷，和4种评分等级，并考察了正态和非正态分布两种情况。采用的错误模型为参数误置（真模型中每个因子各由5个题目来测量，错误模型中则是第一个因子由6个题测量，另两个因子各由4个和5个题来测量，即有一个因子载荷被误置）模型。结果发现（1）样本量、载荷量、评分等级数和分布形态都对GOF的取值确有影响。其中分布形态的影响最大。NNFI、IFI在不同条件下的平均值是最稳定的，其次是CFI、RMSEA和SRMR。它们都算是值得推荐的GOF，尤其是NNFI和IFI。（2）在正态分布中，当样本量≥1000时，根据NNFI、IFI、CFI、RMSEA、SRMR对模型是否拟合做出判断时有很低的两类错误率，在样本量〈1000时则不理想。在偏态条件下无论选择哪个GOF两类错误率都很高。（3）采用2指数策略在很多情况下也不能显著降低两类错误率。（4）由于在数据分布非正态，或正态但样本量〈1000时是难判断模型是否拟合的。因此我们提出了2界值策略。即为每个GOF确定上下两个界值。低于下界值时可判断模型是不正确的，而高于上界值时则可判断模型是正确的。GOF取值处于上下界值之间时难以判断模型是否拟合，只能说越高拟合的可能性越大。这时就要通过跨样本验证和增加样本量来确定模型是否正确。
在本模擬研究中設計瞭6種樣本容量，6種因子載荷，和4種評分等級，併攷察瞭正態和非正態分佈兩種情況。採用的錯誤模型為參數誤置（真模型中每箇因子各由5箇題目來測量，錯誤模型中則是第一箇因子由6箇題測量，另兩箇因子各由4箇和5箇題來測量，即有一箇因子載荷被誤置）模型。結果髮現（1）樣本量、載荷量、評分等級數和分佈形態都對GOF的取值確有影響。其中分佈形態的影響最大。NNFI、IFI在不同條件下的平均值是最穩定的，其次是CFI、RMSEA和SRMR。它們都算是值得推薦的GOF，尤其是NNFI和IFI。（2）在正態分佈中，噹樣本量≥1000時，根據NNFI、IFI、CFI、RMSEA、SRMR對模型是否擬閤做齣判斷時有很低的兩類錯誤率，在樣本量〈1000時則不理想。在偏態條件下無論選擇哪箇GOF兩類錯誤率都很高。（3）採用2指數策略在很多情況下也不能顯著降低兩類錯誤率。（4）由于在數據分佈非正態，或正態但樣本量〈1000時是難判斷模型是否擬閤的。因此我們提齣瞭2界值策略。即為每箇GOF確定上下兩箇界值。低于下界值時可判斷模型是不正確的，而高于上界值時則可判斷模型是正確的。GOF取值處于上下界值之間時難以判斷模型是否擬閤，隻能說越高擬閤的可能性越大。這時就要通過跨樣本驗證和增加樣本量來確定模型是否正確。
재본모의연구중설계료6충양본용량，6충인자재하，화4충평분등급，병고찰료정태화비정태분포량충정황。채용적착오모형위삼수오치（진모형중매개인자각유5개제목래측량，착오모형중칙시제일개인자유6개제측량，령량개인자각유4개화5개제래측량，즉유일개인자재하피오치）모형。결과발현（1）양본량、재하량、평분등급수화분포형태도대GOF적취치학유영향。기중분포형태적영향최대。NNFI、IFI재불동조건하적평균치시최은정적，기차시CFI、RMSEA화SRMR。타문도산시치득추천적GOF，우기시NNFI화IFI。（2）재정태분포중，당양본량≥1000시，근거NNFI、IFI、CFI、RMSEA、SRMR대모형시부의합주출판단시유흔저적량류착오솔，재양본량〈1000시칙불이상。재편태조건하무론선택나개GOF량류착오솔도흔고。（3）채용2지수책략재흔다정황하야불능현저강저량류착오솔。（4）유우재수거분포비정태，혹정태단양본량〈1000시시난판단모형시부의합적。인차아문제출료2계치책략。즉위매개GOF학정상하량개계치。저우하계치시가판단모형시불정학적，이고우상계치시칙가판단모형시정학적。GOF취치처우상하계치지간시난이판단모형시부의합，지능설월고의합적가능성월대。저시취요통과과양본험증화증가양본량래학정모형시부정학。
In this simulation study we designed 6 sample - size conditions, 6 factor - loading conditions, 4 rating- category conditions, and 2 distribution conditions. To each data - set in the conditions a correct model and a mis - specified model are fitted. In the correct model there are 15 items and 3 factors, each factor is measured by 5 items. While in the mis - specified model the factors are measured by 6, 4, and 5 items, we call the mis - specified model wrong - parameterized model, which is different from those studied by former researchers.
The results are： 1） Sample size, loading size, rating category and distribution form all have influence on the values of fit indices. And the influence of distribution form is the largest. The values of NNFI, IFI are most stable across all conditions, values of CFI, RMSEA and SRMR are less stable, but their variations are rather small, these 5 indices should be recommended. 2） In normal distribution conditions, when sample size≥1000, model right - wrong judgment based on NNFI, IFI, CFI, RMSEA, SRMR all have low two type error （α＋β） rate. But distribution forms are non- normal and when sample size 〈 1000, α＋β error rate cannot be reduced to satisfactory level. 3） 2-indices strategy recommended by Hu ＆ Bentler（1999） cannot reduce α＋β error rate significantly in many conditions. 4） Since model judgment is difficult when samples are small and distributions are non - normal, we present 2 - cutoff- value strategy. 2-cutoff- value strategy means, when the value of a model is lower than the low bound cut - off of the recommended indices, the model can be judged as wrong when the value of a model is higher than the upper bound cut - off of the recommended indices, the model can be judged as right, when the value of a model fall between lower and upper bound of the recommended indices, the model cannot be judged as right or wrong. When a model cannot be judged as right or wrong larger sample size and cross - validation of the model aJ：e needed before a clear conclusion can be drawn.
