标题: 关于hlm里的离异数问题 [打印本页] 作者: 蛋羹 时间: 2014-8-10 19:02 标题: 关于hlm里的离异数问题 罗老师,您好!之前向你请教了icc相关问题,获益良多。随后进行了跑数据活动,发现一个问题,向你请教请教。我建立的模型使用两个指标作为因变量。一个是roa,一个是eps。但我发现在,roa的离异数(Deviance)出现了负值,不论是零模型还是随机系数模型亦或完整模型。 * q. r3 |' b. W+ I我查询了一些文献:6 P, Y8 {6 `0 R" h, j- A1 Q0 o
同一研究可以使用多个HLM模型,由此产生一个问题,多个HLM模型的优劣如何判断?如果两个模型是嵌套模型,可以通过离异数(-2LogLikelihood,-2LL)来判断,由于-2LL值近似服从χ2分布,如果两个模型的-2LL值之差的χ2检验显著,表明改进的模型更优。——我国近十年来心理学研究中HLM方法的应用述评 方杰 邱皓政 张敏强 方路. A, o) I8 H3 W& J1 x
Importantly,PROCMIXED produces a variety of statistics useful for comparing the goodness-of-fit / q+ y' H" c$ Z& O9 Cofmultiplemodels using the degrees of freedom(df) associated with the number ofmodel differences* }4 A$ P! W P1 T9 C- ] j
between the contrasted models. Of particular relevance are the 2 log likelihood 3 |/ _! }6 t: [- s6 j+ Y- L(2LL, also referred to as the deviance statistic) and Akaike’s information criterion (AIC).8 a. P9 Y" o3 t# U0 a
In both cases,models that fit better produce smaller values. The2LL depicts difference scores* v, l) O6 F$ S, [4 L2 z; ~
between nested models—for example, comparing one model containing n predictors with8 g, V7 c2 Q3 \
another containing n1 predictors—and follows an approximate w2 distribution. Thus, unlike5 e' ], R, X' C0 P0 o
AIC, it offers a w2 significance test. For model comparison purposes and for consistency with . {- R' l* ?- S, x; O8 Cprevious studies, we report the 2LL and evaluate goodness-of-fit using the w2 test. 1 Q0 W$ S* G# p0 K7 ^& y5 O1 ~$ ^, j; r
Modeling Levels and Time in Entrepreneurship Research An Illustration With Growth Strategies and Post-IPO Performance: }) p: M, _5 X' G2 w! w8 q
Tim R. Holcomb James G. Combs David G. Sirmon Jennifer Sexton 4 `3 \7 T1 a7 C3 `5 T' \/ a$ g. b; p+ {( S: Q5 `; D
文献中提到了多个模型的比较,我从温福星老师的书中也发现通过模型的离异数来计算模型优劣程度。并没有提及假如离异数为负值,是哪些方面出了问题?是否在数据方面有异常?作者: Kenneth 时间: 2014-8-10 21:31
蛋羹, -2LL 其实就是 minimum value of the fit function 的一个表现,所以这个卡方其实就代表了模型的拟合。因此比较两个模型的 -2LL 就是比较它们的卡方差。但是这与负的方差没有什么关系。为什么你会把它们连起来说呢?) t* j; R" L, z1 s' @
6 ~8 ^* b" F& @9 ^! m/ p
多层线性模型是有可能出现古怪的估计的。但是我自己没有见过负的方差,也不知道如何理解。对不起。- }0 i% J2 k$ }
ROA 与 EPS 的分布本来就很偏差,你有没有把它们 log?作者: 蛋羹 时间: 2014-8-11 15:50
Kenneth 发表于 2014-8-10 21:31 " x ], F. ]5 G1 E6 b蛋羹, -2LL 其实就是 minimum value of the fit function 的一个表现,所以这个卡方其实就代表了模型的拟合 ...
8 |9 [9 b; ^$ {, b4 o+ pThe value of the log likelihood depends on the scale of the data. It is defined as the product of the probability density functions, evaluated at the estimated parameter values. Although the total area under a probability density function is scaled to be equal to 1, this does not imply that the probability density function evaluated at a certain point in parameter space has to be less than1. The likelihood function can therefore exceed 1, in which case (the deviance) will be negative.1 Q! r! m+ g" D6 K5 o- \ 作者: 蛋羹 时间: 2014-8-12 18:52