Hi all
Can anyone recommend a user-friendly and not-too-expensive software for performing LPA other than R? I am routinely using SPSS for my statistical analyses; however, SPSS apparently cannot perform LPA. Am I wrong?
Thanks
Jörg
Hi Jorg
I think the M-Plus software can performing the LPA, but unfortunately I can not guide you, because I am not specialist in this software.
Hi “R” statistical software could be used for LPA. It is free and open sourced.
The `mclust` package in R is really good, an overview can be found here: https://cran.r-project.org/web/packages/mclust/vignettes/mclust.html.
There is also the `tidyLPA` package which I believe is using `mclust` in the backend, or alternatively MPlus. `mclust` is free though and in my opinion has a heap more flexibility than MPlus for LCA/LPA.
Hi anyone please, how we understnd how LPA works with discrete latent variables?
Kalpana Sweety Simon Durai you could estimate your latent variables then use mclust as I recommend above. If they are discrete variables then it would be a LCA rather than LPA though.
My only concern without thinking about it too much would be that the uncertainty in your latent variable classification wouldnt propagate through to the LCA... If I were a reviewer I would let it slide if there was good logic behind the latent variables and they seemed reasonably robust.
Hi Taylor Winter, many thanks for the explanation; So did u mean that irrespective of the latent variable being a discrete/ continuous latent variable and if there was good logic behind using the specific type of analysis we would proceed with LPA !@Taylor Winter
Kalpana Sweety Simon Durai - So LPA is for continuous variables and LCA is for discrete variables. You can often use the same package to implement both and it will operate differently under the hood.
Hi anyone, How can models be specified for the R script using TidyLPA which runs like lpa
Kalpana Sweety Simon Durai - I don't understand.
The code above will take a dataframe with your factors in it and try form three profiles. It then assigns the resulting object to `lpa`.
You want to actually specify a model rather than have it determined ?
Yep taylor, but it gives an default output with model 1, how do i add the model arguements in the script? For ex: we have the 4 models: equal variances and covariances fixed to 0 likewise is one model.
How do I specify these models in the script to comapare the solutions for various models ans different latent classes?
Kalpana Sweety Simon Durai - I think I get you now. You want to compare a variety of model types with different numbers of profiles or clusters.
Thats pretty straight forward in mclust which is the backend of tidyLPA. See a basic example here; https://cran.r-project.org/web/packages/mclust/vignettes/mclust.html
Sorry I couldnt pick out the approach in the tidyLPA vignette here; https://cran.r-project.org/web/packages/tidyLPA/vignettes/Introduction_to_tidyLPA.html
pisaUSA15[1:100, ] %>% select(broad_interest, enjoyment, self_efficacy) %>% single_imputation() %>% estimate_profiles(1:3, variances = c("equal", "varying"), covariances = c("zero", "varying")) %>% compare_solutions(statistics = c("AIC", "BIC")) @Taylor Yep exactly as you have said that’s my question. But how do I use it with my first script- Tis one lpa
Hi everyone,
In a LPA result of 3 groups where one of the group has 2 cases, we have included it in our study and conducted further group differences theough ANOVA- is it appropriate to conduct ANOVA with unequal sample size and very minimal sample size?!
Kalpana Sweety Simon Durai - Sorry I missed your previous message. No it is not okay to do that analysis with only two cases. A general, albeit rough, rule of thumb when doing LPA is that you don't create groups with less than 5% of your sample size. Im sorry to say two cases means it is probably just overfit and highly unlikely to be an actual profile. A reviewer would not let that fly for a second.
You should create fit statistics for different numbers of profiles, such as AIC/BIC or log-likelihood ratios. What does that data tell you ? I suspect it would not advise a three profile classification.
@taylor Yes excatly as you have said its the rule of thumb to not retain a class with less than 5% of cases but we have in my thesis doc owing to the theory behind it as it will be meaningful to have that group (which is similar to a subgroup reported in previous researches) and I have talked about it in the discussion.
Kalpana Sweety Simon Durai - Okay but then it sounds like your results would be impossible to replicate with such a small profile. It could be quite possible the previous research is just overfitting and you are guilty of the same thing. Like I said, you will have a very hard time getting this through review if the reviewer is familiar with LPA.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Software