Message/Author |
|
|
I have egocentric social network data for 835 survey respondents and up to 5 of their social network members. In long form, the total number of ego-alter ties is 2961. I have 8 questions as indicators of social support (binary, yes or no) for each of the ego-alter ties. I theorize that there are latent constructs of social support underlying these observed indicators and believe that LCA would be an appropriate approach. How can I account for the correlation of the data, that ego-alter ties are essentially nested under 835 egos (participants)? I've run LCA with analysis = mixture and analysis = complex mixture with cluster = id. Model fit and probabilities are the same for both. Here's an example: VARIABLE: Names = id ss1argu45 ss2talk5 ss3corn5 ss4pitc5 ss5loan5 ss6stay5 ss7advi5 ss8unde5; Categorical = ss1argu45 ss2talk5 ss3corn5 ss4pitc5 ss5loan5 ss6stay5 ss7advi5 ss8unde5; Missing are all (-9999) ; Classes = c(3) ; Cluster = id ; ANALYSIS: Type = complex mixture ; Starts = 100 20 ; Output: tech10 tech11 |
|
|
A key question is if the latent class variable is a within- or a between-level variable. If the classes represent subjects, the latent class variable should be on the Between = list in a Type=Twolevel analysis. The User's Guide shows several such examples. |
|
|
Thank you Dr. Muthen. I've attempted the Type=Twolevel analysis, but am unable to identify the latent classes in the output. If it helps, I am only looking at LCA to determine classes and class membership in the measurement model at this point. |
|
|
You need to use Type = Twolevel Mixture. Look for UG examples in that chapter. |
|
|
This worked. Thanks. I now would like to construct a model using multinomial logistic regression to identify associations with class membership. I've tried doing both the LCA and model in the same input file, but get different diagnostics and probabilities for the LCA section. Is there a way to perform both the the LCA and multinomial regression in the same input file? |
|
|
You can do it in the same input file; see UG ex7.12. |
|
|
Thank you Dr. Muthen. In the tech10 output of the twolevel mixture analysis, the response patterns and frequencies are reported. I am trying to assess the response patterns by each latent class. Is it possible to see the tech10 output organized by class? |
|
|
No. |
|
|
I renewed my license going from version 7 to 8. Now, when running my twolevel mixture LCA I get the following error message. "Warning: One or more individual-level variables have no variation within a cluster for the following clusters" Is this related to the new version? My data/input file have not changed. |
|
|
See our FAQ WARNING One or more individual-level variables have no within-cluster variance |
|
|
Dr. Muthen, I've decided that EFA might be better suited to my data and conceptualization of social support as a continuous variable. I'm struggling coding to specify that the latent factor(s) are only at the between level, while the indicators are only at the within level. I've been studying examples 4.5 and 4.6, but can't figure out the code for a between level only factor(s) and within only indicators. VARIABLE: NAMES ARE id ss2talk ss3corn ss4pitc ss5loan ss6stay ss7advi ss8unde ; CLUSTER = id ; WITHIN = ss2talk ss3corn ss4pitc ss5loan ss6stay ss7advi ss8unde ; ANALYSIS: TYPE = TWOLEVEL EFA 1 3 UB ; OUTPUT: MODINDICES ; |
|
|
When you put a variable on the Within list, you cannot estimate a variance for it on Between - that is, you are saying that it has zero variation across clusters. Typically, a variable is observed for (say) a student in (say) a school and therefore has variation on both levels. In a factor analysis setting, the factor indicators on between are the latent, random intercepts of the observed variables measured on the student. With this as background, it sounds like you don't want to put your variables on the between list, you want to have factors on between, and you want to freely correlate the variables on within. So maybe you want to say something like Type = twolevel efa 1 1 uw 1 2; |
|
|
Thank you. 1. Ordinal data and EFA: Is there a particular approach to EFA given the 7 indicators of social support from each network member are measured on a Likert scale from: 1) never; 2) a little; 3) sometimes; 4) frequently; 5) always? 2. Frequency of responses: I understand ordinal data is not normally distributed, but is it problematic that the frequency of '5) always' responses is high across all 7 indicators compared to 1) through 4)? Generating a binary variable doesn't seem meaningful given Always compared to Never to Frequently doesn't make sense. 3. The eigenvalues Within: 3.68, 0.79, 0.692, 0.559, 0.498, 0.418 and 0.361. Between: 4.547, 0.606, 0.604, 0.375, 0.336, 0.297 and 0.234. This would indicate only one factor. Is factor analysis useful with only one factor? I'm thinking it would provide a factor "score" for each individual (between level) indicating the degree to which they receive this latent construct of social support. I've read that using factor scores in regression can lead to inconsistent coefficient estimators and I'm unsure how to address this if indeed factor analysis is more appropriate than latent class analysis (generating binary indicators from the Likert scale) |
|
|
1) See our Topic 2 short course video and handout. 2) No (see Topic 2). 3) It is good to have statistical evidence that you have one factor only. I would not use regression based on factor scores if I didn't have to - this is why structural equation modeling was invented. |
|
Back to top |