Anca, Are you trying to fit a traditional latent class model to
nominal indicators? In Latent Gold, this is the Cluster module. As the modeler, you can try increasing the number of classes to see whether you can obtain a model that fits. The p-value to which you refer is associated with a model with a particular number of classes. It sounds from your reported p-values as if the model does not fit for the given number of classes. You can try increasing the number of classes. As you do so, your fit L-square should improve, at the "cost" of using an increased number of parameters. Some of the available information measures such as BIC or AIC can be used to compare models. Other things equal, you would choose the model that minimizes BIC or AIC (and these might not signal the same model). Or, you can switch modeling frameworks and use the Factor module in the same situation. That is, instead of increasing the number of classes, explore models with 2 or more factors. There are some tutorials on this available at the Latent Gold website. Finally, with summative scores, you might consider the cluster module with the variable type declared continuous, which would specify the normal model. One question is: What does the shape of the summative score look like? If it is roughly symmetric, then the normal model should work OK. However, you lose the L-square and p-value as fit assessment statistics, for you are no longer in the discrete indicators framework. _____ From: Classification, clustering, and phylogeny estimation [mailto:[EMAIL PROTECTED] On Behalf Of a Sent: Tuesday, December 12, 2006 5:28 AM To: [email protected] Subject: when p values significant for all models Dear Dr Babinec sorry to bother you again with questions about cluster analysis in Latent Gold, but what happens if p-value is very significant for all models? (e.g. 9.7E-204 or 2.8E-191) how can we assess the models? Would you treat total scores from different measures as continuous variables(answers to items are scaled from 1-6 or 1-10, but then a total score is calculated for each measure)? Thank you, Anca Anthony Babinec <[EMAIL PROTECTED]> wrote: When you have NOMINAL indicators, for example, your model gives rise to expected counts that can be compared to observed counts. The distribution theory is based on the chi-square statistic (L-squared), which has an associated p-value. When you have CONTINUOUS indicators, your model is based on normal theory. The parameters being estimated are means, variances, and covariances. Since the data are continuous and not discrete, you no longer have a model framework of observed and expected counts. The model is the normal finite mixture model. Classification can work well with a good-fitting model. -----Original Message----- From: Classification, clustering, and phylogeny estimation [mailto:[EMAIL PROTECTED] On Behalf Of SUBSCRIBE CLASS-L Anonymous" Sent: Thursday, December 07, 2006 9:06 AM To: [email protected] Subject: p values in Latent Gold Hello Does anyone know why p-values and chi-squared statistics are not available in Latent Gold summary output for models using continuous variables and what is the statistical explanation behind it? Also, how reliable is the classification with continuous variables in latent gold given the fact that it is based - from my understanding - on means and not on probabilities? Many thanks Anca ---------------------------------------------- CLASS-L list. Instructions: http://www.classification-society.org/csna/lists.html#class-l ---------------------------------------------- CLASS-L list. Instructions: http://www.classification-society.org/csna/lists.html#class-l __________________________________________________ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com ---------------------------------------------- CLASS-L list. Instructions: http://www.classification-society.org/csna/lists.html#class-l ---------------------------------------------- CLASS-L list. Instructions: http://www.classification-society.org/csna/lists.html#class-l
