建立隐性变量模型(英文)
2023-01-13
来源:意榕旅游网
上海精神医学2012年第24卷第2期 observed and latent variables is fully specified according to a simple linear regression model: y {+∈ where the outcome variable y is an observed variable and there may be i=1….,m Of them in any factor analysis.The predictor f is a Iatent variable(a common factor),and∈,may be regarded as a disturbance term 《a unique factor).The regression coe仟icient^f is called the factor Ioading of variable y on common factor£ representing the strength of association between the observed variable and the latent cornm0n factor.There is one unique factor per observed variable,and typically they are assureed to be normaIIy distributed with zero means,uncorrelated with ,and with unique variance .Thus,the conditional distribution of y,given is norma1 with mean A f and varianee ,iust as in a Iinear regression mode1.This conditional distribution is a measurement modeIin the sense that it provides the necessary linkage between the observed variables and the latent va riabIes.1t directly incorporates the component Of measurement error into the obseryed va riabIe.As such,the Iatent va riabIes in a Properly specified measurement modeI can be thought of as having been purged of measurement erron In factor anaIysis.because the common factor is unobserved,a(prior)distribution is imposed on£which is typicaIlY taken to be standa rd normaI.This simpie distribution of f is a structuraI modeI for the Iatent variable.When there are more than one in the mode1. the structu raI modeI can describe the relationshiP among the latent variables.This wilI subsequently be important for structuraI equation modeIing.In generaI, the P rior distribution of Iatent variabIes tyPicaIIY stems not fr0m statistical cOnsideratiOns.but frOm the substantive needs of the research questions that the latent variable modeling is attempting to address.on a case—bv-case basis.1s the question a taxometric one?0r might there be a continuum of underlying dimensions? 0r both?ShouId the structu raI modeI onIY incIude Iatent va riabIes7 0 r Pe rhaps obse rved exogenous covariates are required to explain the heter0geneitV?Or both?0ften the right answer is the more complex one since phenomena studied in mentaI health research are usually quite complex. SuPpose the number of observed va riabIes m is equaI to 4.There exists another equivaIent way of representing the factor analysis model,using a path diagram.Figure 1 is Ia rgeIY based on JOreskog s exampie for sets of congeneric tests.The directionaI arrows represent regression paths.The rectangular nodes a re observed va riabIes and ci rcles a re Iatent variables.The bidirectionaI curved a rrows rePresent variances fwhen the arrow heads point tO the same variable1 or covariances(when the arrow heads point to two diiferent variables).The path diagram makes it clear that the factor loadings and the unique variances ・119・ are the key parameters to estimate.Once their values are known,one can use the model to provide optimal P redictions of the latent va riable scores based on observed variable values. Figure 1.A Path Diagram Example Fu rthermore.the path diagram reP resentation oPens the door to mo re comPIex Iatent variabIe structuraI modeling aIong the Iines of path analysis.【q Indeed.with JOreskog s factor analytic simultaneous equations model and the advent of the LISREL software Program.one may specity simuItaneous regresslon equations for the latent variables.and use maximum IikeIihood or other methods to fit the model directly to a sample of data.For example,one may consider simultaneous regression equations of the type ,7=8,7+厂 + (2) where仃is a vector of endogenous Iatent variables.f js a vector of exogenous latent variables,B and r contain the regression coefficients,and is a vector of equation distu rbance te rms.The simuItaneous regression equations permit the direct estimati0n and testing 0f substantiveIV impOrtant c0nceptual m0dels cOntalning mediati0n e仟ects,that is,variable X causing Z,which in tu rn causes y.BOllen‘口 c0ntains an auth0ritative treatment Of the main tOpics in structuraI equatiOn mOdeling. Eauati0n(1)c0nnects the 0bserved and Iatent variabIes using a linear mOde1.This is pOssibIe because the outcome(observed)variable is assumed t0 be c0ntinu0us.The applicati0n of concepts deveIoped in generaIized Iinear mOdeIs¨ tO Iatent variab Je mOdeling (e.g.,link functions)has Ied t0 a unified treatment Of latent va riable mOdeIs fOr categorical Observed variabIes. The sO.caIIed twO—parameter IOgistic item respOnse theOrV mOdeI is arguably the mOst widely recOgnized member Of the famiIy Of m0dels for discrete ・120・ observed data.Mathematically,this model relates the probability of endorsing a dichotomously scored variable to the underlying latent variable using a logistic function: P(yi=l1日= 1 (1+exp[一( — 日】) where is stilI the Iatent va riabIe,y}the observed va riabIe,and o【i and ,are the intercept and sIape parameters of this Iogistic mode1.Using the Ianguage of generalized Iinear models,equation(3)diifers from equation(2)in that a Iogit Iink function is used instead of an identity link function.With the avaiIabiIity of mode rn item response modeling frameworks and softwa re. item response theo ry has become a standard tooIin psychologicaI assessment and health— related outcomes research.【 】 More recently,generaI frameworks implemented in software packages such as MplusLl allow the structural modeling of mixtures of discrete and continuous latent variables,for example,regressing a Iatent classification va riabIe on a set of continuous Iatent va riable predictors,further extending the flexibility of structuraI equation modeIing.NonIinear reIationships among latent variables(e…g moderation or interaction effects) can also be assessed with the advent of Bayesian computational methods.u J Finally,structuraI equation modeIing Provides a camprehensive set of tooIs for the anaIysis 0f longitudinal or repeated measures data.through the Iatent curve modeling framework. Here the connection between latent variable modeIs and muItilevel(random coefficient)models becomes transparent.For Iarge subclasses of Iatent curve models. one can find equivaIent mu…t IeveI formuIations. In sum,after more than a centu rv of development, Iatent variable modeling encompasses a broad range of statisticaI techniques that may be usefuI for modeling mental health data. Funding Part of this research is supported by the Institute of Education Sciences(R305B080016 and R305D10OO39) and the National Institute on Drug Abuse fR01DA026943 and R01DA03O4661.The views expressed here belong to the author and do not reflect the views or policies of the funding agencies. References 1. Spearman C.General intelligence objectively determined and measured.Am J Psychol 1904;15:201—293. 2. Bartholomew DJ Knott M.Latent variable models and{Qctor analysis.2nd ed.London,UK:Arnold,1999. Shanghai Archives of Psychiatry,2012,Vo1.24,No.2 3. JOreskog K G.Statistical analysis of sets of congeneric tests Psychometrika 1971;326:109-133. 4. Wright SS.Correlation and causation.J Agric Res 1921;20557— 585. 5. J6reskog KG.A general method for analysis of covar Jance structures.Biometrika 1970;57:239-251. Bollen KA.Structuia/equations with latent variables.New York: John Wiley&Sons.1989. 7. McCullagh P.Nelder JA.Generalized linear models.2nd ed London:Chapman&Hal1.1989. 8. Moustaki I.Factor analysis and latent structure of categorical and metric data.In R.Cudeck&R.C MacCallum(Eds. Factor analysis at 100:Historical developments and future directions. Mahwah,NJ:Laurence Erlbaum Associates.2007. 9. van der Linden WJ,Hambleton RK.Handbook of modern item response theory.New York:Springer Verlag.1997. 10 Cai L,Thissen DJdu Toit SHC.IRTPRO:Flexible,multidimensional, multiple categorical IRTmodeling【Computer software].Chicago, IL:Scientific Software International,Inc.2011. 11 Reeve BB,Hays RD,Bjorner JB,Cook KF,Crane PK,Teresi JA,et a1.Psychometric evaluation and calibration of health・related quality of life items banks:Plans for the patient—reported outcome measurement information system(PROMIS).Medical Care 2007;45(5 suppl 1):S22・31. 12 Muth ̄n,Muth ̄n.Mplus(Version 5.DJ【Computer software].Los Angeles,CA:Author.2008. 13 Arminger G,Muth ̄n BO.A Bayesian approach to nonlinear latent va riahie models using the Gibbs sampler and the Metropolis-Hastings algorithm.Psychometrika,1998;63:271‘ 300. 14 Lee SY,Zhu HT.StatisticaI analysis of nonIinea r structu ral equation models with continuous and polytomous data.Br J Math Stat Psycho/2000;53(pt 2):209-232. 15.Bollen KA.CurranPJ.Latent curve mode ̄:A structuraf equation perspective.Hoboken.NJ:John Wiley&Sons.2006 16.Bauer D J.Estimating multilevel linear models as structura equation models.J Educ Behav Stot 2003;28:135-167. Li Cai is an associate professor of education and psychology at the University of California at Los Angeles (UCLA),where he also serves as Co-Director of the ~afional Canter for Research on Evaluation,Standards, and Student Testing(CRESST}.His methodoloqico| research agenda involves the development,fntegration. and evaluation of innovative latent variable models that have wide—ranfling applications in educational, psvch0|0qico|.and health—related domains of stud ̄A key component on this agenda is statistical computing, particularly as related to mu/tidimensional item response theory(JR 7-J and multilevel modeling.He has also co||oboroted with researchers at UCLA and elsewhere on projects examining measurement ̄sues in mentaf health,substance abuse treatment,and patient- reported outcomes research.