Latent Class Analysis

[Biene Maja]

Hello Everyone,

I've done a Latent Class Analysis with the lclogit command described by Pacifico and Yoo. http://papers.ssrn.com/sol3/papers.c...act_id=2174146
I've estimated the parameters in the different classes of a choice-based conjoint experiment. But I'm not sure, if it's allowed to interpret the numbers if significant of the parameters or only the trend (positive/negative) if significant.

I've written already to Mr. Pacifico and Yoo but have not received an answer yet.

You can compare my results to that of the paper in the link above, page 12 and 13 (after lclogitml command).

I really need help with this question that I'm not wrong with my interpretation.

Thank you very much.

With best regards

[Nathan E. Fosse]

Hi, Biene Maja -

It would very much help if you provided a sample data set and code for examples. The issue with latent class models (LCM) is that there are many different types of parameters and it's unclear from your question to what parameters you are referring.

It's unclear whether you have already compared model fits to determine the number of classes. Assessing the number of latent classes is assessed with a variety of measures, typically using AIC or BIC (the lower, the better). Determining the number of latent classes will necessarily involve theory, however, particularly given your two models of fit (there are many others, but not in the Stata environment). The help file for lclogit provides a clear example of looping through values of numbers of classes to assess model fit.

For parameters with p-values, take a look at lclogit postestimation. There are at least three options available to you, most of which are specific to the needs of your particular analysis.

To be clear, there are other types of latent class models other than discrete choice modeling. The latter is available only using gllamm in the Stata environment; as its developers note, gllamm can be "computationally intensive." For those options, I'd consider specialty software such as Latent GOLD, which provides demo downloads (link).Sometime, I'm going to write a wrapper for Latent GOLD for Stata, given the need and the research possibilities.

Cheers,

- Nate

[Biene Maja]

Hello Nate,

thanks for your answer. I estimate already the AIC, BIC and CAIC as well the Log-Likelihood. The best solution would be a 4-class model. I used data from a choice-based conjoint experiment. Recruiter have choosen the best fictitious resume of a set of three resumes of candidates and did this experiment for 10 times. The CV contains 5 attributes of two or three levels.
Based on the decisions of the choosen CVs I did a latent class analysis to analyse the recruiting preferences. The table shows the results for a solution of 4 classes: My first question was: Does it make a difference that the numbers e.g. for the level „no patents“ differ from -1.9992 to -0.4253, so can you say, that recruiter in class 4 evaluate people with no patent experience more negatively than recruiter in class 1 do, even if individuals in Class 1 estimate that level negative as well?

Is it possible in Stata to estimate the relative importance/part-worths of the attributes in the 4 classes, e.g. Professional background is the first important attribute followed by Patent experience and so on e.g. in class 1. For the relative importance the attributes (Professional Background, Job experience, Specific skills, Motivation, Patent experience) have to count to 100 %.

Thank you very much for your answer.

Best regards

	Class 1	Class 2	Class 3	Class 4
Professional background, reference: Natural sciences
Engineer	0.1301	1.4117***	-0.3486	0.251
Job experience, reference: Specialist
Generalist	1.1229***	-0.4278***	-0.9810***	-0.3136
Specific skills, reference: Creativity
Risk_attitude	-0.8115***	-0.8468***	-3.1593***	-0.1546
Analytical_thinking	0.3239*	0.4328***	-0.3085	0.2663
Motivation, reference: Autonomous working
Technology_orientation	0.3962*	1.2570***	2.8919***	0.4603**
Environment_sustainability	0.5070**	-0.4210**	2.3421***	-0.2736
Patent experience, reference: Patents in other field
No_patents	-0.4253**	-0.5441***	-1.7956***	-1.9992***
Patents_CTorME	0.4316**	0.3228*	0.7874**	1.5189***
Notes: Independent Variable: Choice of fictitious CV (Dummy-Variable);
*** p<0.01, ** p<0.05, * p<0.1

[Biene Maja]

Hello Nate,

I attached also my commands in Stata and the results. The results of the table of my post is in the txt-file in line 285. It was generated with the lclogitml command.

Best regards