top of page
Search
khanspwwx

Lisrel 8.8 Full Version Free 15: Features, Benefits, and Limitations



The aim of the current study was to develop a pregnancy specific psychological functioning scale. Based on the outcome of group interviews with pregnant women, new mothers and clinicians, a 16-item self-rating scale was developed (i.e., the TPDS; for full scale see appendix). Subsequent analyses confirmed that the TPDS has a two factor structure: 'negative affect' and 'partner involvement'. Negative affect, in turn, appeared to have three sub components: negative affect with regard to confinement, postpartum period and general health. Both AMOS and Lisrel showed appropriate structure of the final scale using CFA.


In this seminar, we will understand the concepts of CFA through the lens of a statistical analyst tasked to explore the psychometric properties of a newly proposed 8-item SPSS Anxiety Questionnaire. Due to budget constraints, the lab uses the freely available R statistical programming language, and lavaan as the CFA and structural equation modeling (SEM) package of choice. We will understand concepts such as the factor analysis model, basic lavaan syntax, model parameters, identification and model fit statistics. These concepts are crucial to deciding how many items to use per factor, as well how to successfully fit a one-factor, two-factor and second-order factor analysis. By the end of this training, you should be able to understand enough of these concepts to run your own confirmatory factor analysis in lavaan.




lisrel 8.8 full version free 15




The concept of a fixed or free parameter is essential in CFA. The total number of parameters in a CFA model is determined by the number of known values in your population variance-covariance matrix $\Sigma$, given by the formula $p(p+1)/2$ where $p$ is the number of items in your survey. Suppose the principal investigator thinks that the third, fourth and fifth items of the SAQ are the observed indicators of SPSS Anxiety. To obtain the sample covariance matrix $S=\hat\Sigma$, which is an estimate of the population covariance matrix $\Sigma$, use the column index [,3:5], and the command cov. The function round with the option 2 specifies that we want to round the numbers to the second digit.


The example above is unrealistic because it would be pointless to have all the parameters be fixed. Instead, many models are just-identified or saturated with zero degrees of freedom. This means that the number of free parameters takes up all known values in $\Sigma$. This is commonly seen in linear regression models, and the main drawback is that we cannot assess its model fit because it supposedly is the best we can do. An under-identified model means that the number of known values is less than the number of free parameters, which is undesirable. In CFA, what we really want is an over-identified model where the number of known values is greater than the number of free parameters. Over-identified models allow us to assess model fit (to be discussed later). To summarize


Answer: We start with 10 total parameters in the model-implied covariance matrix. Since we fix one loading, and 3 unique residual covariances, the number of free parameters is $10-(1+3)=6$. Since we have 6 known values, our degrees of freedom is $6-6=0$, which is defined to be saturated. This is known as the marker method.


Thankfully for us, we have just the right amount of items to fit a CFA because a three-item one factor CFA is just-identified, meaning it has zero degrees of freedom. Because this model is on the brink of being under-identified, it is a good model for introducing identification, which is the process of ensuring each free parameter in the CFA has a unique solution and making surer the degrees of freedom is at least zero. There are many rules for proper identification, but for the casual analyst identification helps us avoid the following message in lavaan:


Answer: We start with 10 unique parameters in the model-implied covariance matrix. Since we fix one factor variance, and 3 unique residual covariances, the number of free parameters is $10-(1+3)=6$. Since we have 6 known values, our degrees of freedom is $6-6=0$, which is defined to be saturated. This is known as the variance standardization method.


The first line is the model statement. Recall that = represents the indicator equation where the latent variable is on the left and the indicators (or observed variables) are to the right the symbol. Here we name our factor f (or SPSS Anxiety), which is indicated by q03, q04 and q05 whose names come directly from the dataset. We store the model into object m1a for Model 1A. The second line is where we specify that we want to run a confirmatory factor analysis using the cfa function, which is actually a wrapper for the lavaan function. The model to be estimatd is m1a and the dataset to be used is dat; storing the output into object onefac3items_a. Finally the third line requests textual output for onefac3items_a, listing for example the estimator used, the number of free parameters, the test statistic, estimated means, loadings and variances.


By default, lavaan chooses the marker method (Option 1) if nothing else is specified. In order to free a parameter, put NA* in front of the parameter to be freed, to fix a parameter to 1, put 1* in front of the parameter to be fixed. The syntax NA*q03 frees the loading of the first item because by default marker method fixes it to one, and f 1*f means to fix the variance of the factor to one.


You will notice that the implied variance-covariance matrix is the same as observed covariance matrix. This is because we have a perfectly identified model (with no degrees of freedom) which means that we have perfectly reproduced the observed covariance matrix (although this does not necessarily indicate perfect fit). Taking the implied variance of Item 3, 1.155, obtain the standard deviation by square rooting $\sqrt1.155=1.075$ we can divide the Std.lv loading of Item 3, 0.583 by 1.075 which equals 0.542 matching our results for Std.all given rounding error.


Traditionally, CFA was only concerned with the covariance matrix and only the summary statistic in the form of the covariance matrix was supplied as the raw data due to computer memory constraints. In modern CFA and structural equation modeling (SEM) however, the full data is often available and easily stored in memory, and as a byproduct, the intercepts or means are can be estimated in what is known as Full Information Maximum Likelihood. With the full data, the total number of parameters is calculated accordingly:


Technically a three item CFA is the minimum number of items for a one factor CFA as this results in a saturated model where the number of free parameters equals to number of elements in the variance-covariance matrix (i.e., the degrees of freedom is zero). Suppose that one of the data collectors accidentally lost part of the survey and we are left with only Items 4 and 5 from the SAQ-8. When there are only two items, you have $2(3)/2=3$ elements in the variance covariance matrix. As you can see in the path diagram below, there are in fact five free parameters: two residual variances $\theta_1, \theta_2$, two loadings $\lambda_1, \lambda_2$ and a factor variance $\psi_11$. Even if we used the marker method, which the default, that leaves us with one less parameter, $\lambda_1$ resulting in four free parameters when we only have three to work with.


The puzzle is to somehow fit a model that uses only three free parameters. One solution is to use the variance standardization method, which fixes the variance of the factor to one, and equate the second loading to equal the first loading. The gives us two residual variances $\theta_1, \theta_2$, and one loading to estimate $\lambda_1$. Can you think of other ways?


You can see from the output that although the total number of free parameters is four (two residual variances, two loadings), the degrees of freedom is zero because we have one equality constraint ($\lambda_2 = \lambda_1$). Note the(a) in front of the q04 Estimate means that we have attached a parameter label, and the additional (a) in front of the q05 means we have equated the two loadings, namely 0.605. The limitation of doing this is that there is no way to assess the fit of this model. For example, suppose we have the following hypothetical model where the true $\lambda_1=0.8$ and the true $\lambda_2=0.2$. If we fix $\lambda_1 = \lambda_2$, we would be able obtain a solution, not knowing that the model is a complete false representation of the truth since we cannot assess the fit of the model. It is always better to fit a CFA with more than three items and assess the fit of the model unless cost or theoretical limitations prevent you from doing otherwise.


The benefit of performing a one-factor CFA with more than three items is that a) your model is automatically identified because there will be more than 6 free parameters, and b) you model will not be saturated meaning you will have degrees of freedom left over to assess model fit.


The model chi-square is defined as either $nF_ML$ or $(n-1)(F_ML)$ depending on the statistical package where $n$ is the sample size and $F_ML$ is the fit function from maximum likelihood, which is a statistical method used to estimate the parameters in your model. The model chi-square is a meaningful test only when you have an over-identified model (i.e., there are still degrees of freedom left over after accounting for all the free parameters in your model).


First calculate the number of total parameters, which are 8 loadings $\lambda_1, \cdots, \lambda_8$, 8 residual variances $\theta_1, \cdots, \theta_8$ and 1 variance of the factor $\psi_11$. By the variance standardization method, we have fixed 1 parameter, namely $\psi_11=1$. The number of free parameters is then:


Comparing the Model Test User Model for the eight-item (over-identified) model to the the three-item (saturated) model, we see that the Test Statistic degrees of freedom is zero for the three-item one factor CFA model indicating a saturated model, whereas the eight-item model has a positive degrees of freedom indicating an over-identified model. The Test Statistic is relatively large (554.191) and there is an additional row with P-value (Chi-square) indicating that we reject the null hypothesis. 2ff7e9595c


0 views0 comments

Recent Posts

See All

Comments


bottom of page