Latent profile analysis - getting graph with predicted means and CIs

[Marcos Almeida]

Dear Statalisters.

I have 5 continuous variables and a binary variable

This is the abridged dataset:

* Example generated by -dataex-. To install: ssc install dataex
clear
input float(binvar var1 var2 var3 var4 var5)
0   50 67.85714 54.16667  83.33334  43.75
1 62.5 39.28571     62.5  58.33333   37.5
0   50 64.28571 54.16667        50  43.75
1 62.5 71.42857       75  91.66666 53.125
1 62.5 53.57143 66.66667  58.33333 34.375
0 37.5 82.14285 54.16667  41.66667     25
0 87.5 89.28572 95.83334        75     75
0 87.5 78.57143 70.83333  83.33334  56.25
1  100 89.28572     87.5  83.33334  81.25
1 62.5 45.83333 58.33333        50   37.5
0   50 67.85714     62.5  66.66667     50
1   75 67.85714 70.83333        75  68.75
1 87.5       75     87.5        75 46.875
0 87.5 89.28572       75  83.33334  68.75
0  100 96.42857 91.66666  83.33334  81.25
0   50 64.28571       75  66.66667  43.75
0   75 57.14286 66.66667  83.33334 59.375
1 62.5 53.57143 54.16667  58.33333  43.75
0 87.5       75     87.5        75     75
0 62.5 64.28571     62.5  66.66667 59.375
1 87.5 71.42857 79.99999  83.33334 53.125
1 37.5 57.14286       75        75   37.5
1 62.5 53.57143 58.33333  83.33334   37.5
0 62.5 64.28571       75  58.33333   37.5
0 87.5       75 79.16666  83.33334  56.25
1   50       75 70.83333  58.33333  56.25
1   75       50     62.5        75 40.625
0 62.5 71.42857       70  58.33333 65.625
1  100 78.57143 91.66666  91.66666 59.375
0   75 92.85715 83.33334  58.33333 53.125
0   75 67.85714 83.33334  83.33334     75
1   75 60.71429 58.33333  83.33334 53.125
0  100 71.42857       75  66.66667   62.5
0 62.5       75 66.66667        75 46.875
0  100 96.42857     87.5       100     50
1 62.5 53.57143 79.16666        75  56.25
1 62.5 53.57143 66.66667        75     50
1   50       50       75        50      .
1 87.5 53.57143     62.5  83.33334 65.625
0 62.5       75 54.16667        75 59.375
0   75 78.57143       75  83.33334 65.625
1 37.5 64.28571 54.16667        75  18.75
0   50 60.71429 66.66667  41.66667   37.5
1 62.5 60.71429 70.83333  83.33334 40.625
0 62.5 96.42857 79.16666       100 78.125
0 62.5 67.85714       75  66.66667 46.875
1 87.5 60.71429 54.16667        50     50
0 62.5 85.71428 54.16667  58.33333 53.125
1 87.5 78.57143 83.33334       100     50
0 87.5 85.71428 70.83333  66.66667     50
0  100 67.85714 79.16666        75 71.875
1 62.5 46.42857 41.66667  41.66667 28.125
1 87.5 67.85714 70.83333  91.66666  56.25
1   75 57.14286     62.5        75  43.75
0 87.5       75       75  58.33333  43.75
0 62.5 60.71429     62.5  66.66667 40.625
1   75 71.42857 58.33333        75     50
1   50 60.71429 54.16667  41.66667  43.75
1   75 60.71429       75        75  43.75
1  100 64.28571 70.83333       100 71.875
1 62.5 82.14285 54.16667        50 53.125
0   75 78.57143       75  91.66666  43.75
1 87.5 71.42857       50       100 71.875
0 87.5       75       70  91.66666   62.5
0 62.5 71.42857 70.83333  58.33333 46.875
0    . 67.85714 58.33333  66.66667   62.5
1   50 64.28571 66.66667        50 40.625
1   75 71.42857     87.5        75     50
0  100       75 70.83333  66.66667 78.125
0   50 64.28571     62.5        75 53.125
0  100      100 91.66666       100 78.125
0   75 82.14285 83.33334  83.33334 65.625
1   75 57.14286       75  41.66667     50
1 62.5       75       75  91.66666  43.75
0   75 64.28571 70.83333        75 53.125
0   75 64.28571 70.83333        50     50
0   75 78.57143     87.5 33.333332  68.75
0 62.5 60.71429 58.33333  83.33334   37.5
0 62.5 60.71429     87.5  83.33334  56.25
0 62.5 67.85714 70.83333  83.33334   37.5
1 62.5 53.57143 54.16667       100 71.875
0 62.5 71.42857     87.5  91.66666 59.375
0    . 91.66666     87.5  83.33334 71.875
1   75       75       75        75     50
0   75       50 58.33333  91.66666 59.375
1 87.5 64.28571 70.83333        75   62.5
0 87.5 89.28572       75       100  56.25
0 62.5 64.28571 70.83333        75  56.25
1   50 67.85714 66.66667  41.66667 40.625
0   75 60.71429 45.83333        75 65.625
0   75 82.14285 70.83333  58.33333 59.375
1   75       75       75  91.66666 40.625
1  100 82.14285       90      87.5  68.75
0 87.5      100 91.66666       100     75
1   75     62.5 58.33333      37.5 40.625
0   75 64.28571       75       100 59.375
1   50 57.14286 54.16667  83.33334 59.375
1    . 60.71429     62.5        75  68.75
0  100 71.42857       75        75   62.5
0   75 71.42857 66.66667  66.66667     50
end

I decided to perform a latent profile analysis. Below, command and output :

 gsem (var1 var2 var3 var4 var5 <- _cons), family(gaussian) link(identity) lclass(C 2)

(output omitted, because the summary result is presented below)

 estat lcmean

Latent class marginal means                     Number of obs     =         96

------------------------------------------------------------------------------
             |            Delta-method
             |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
1            |
        var1 |    63.1006   1.727262    36.53   0.000     59.71523    66.48597
        var2 |   63.35653   1.420895    44.59   0.000     60.57162    66.14143
        var3 |   64.09592   1.381402    46.40   0.000     61.38842    66.80342
        var4 |   67.63647    2.22846    30.35   0.000     63.26877    72.00417
        var5 |   47.28173   1.523574    31.03   0.000     44.29558    50.26788
-------------+----------------------------------------------------------------
2            |
        var1 |   85.80579   2.117787    40.52   0.000       81.655    89.95657
        var2 |    78.5297   1.816236    43.24   0.000     74.96994    82.08946
        var3 |   79.42668   1.602816    49.55   0.000     76.28521    82.56814
        var4 |   82.44715   2.532471    32.56   0.000      77.4836    87.41071
        var5 |   62.92604   1.905211    33.03   0.000     59.19189    66.66018

Now, the obstacle I'm facing.

In spite of the presentation of the margins as well as the CIs, I have not found a way to get the marginsplot of var1 to var5, according to the two classes. Ideally, I would like to have a graph for the means + CI of each var, with the 2 classes side by side.

Indeed, I see these values are available in the r(table) matrix:

return list
matrix list r(table)

In spite of being a - gsem - , I believe we can use - margins - as well as - marginsplot -, for a couple of reasons:

There is an example here, provided by StataCorp, albeit the case is slightly different from mine.

Additionally, if I type:

margins, dydx(*)
marginsplot

I get the predictive values for mu overvall.

However, I wish to have the predictive margins according to each class (1 and 2) as presented after typing - estat lcmean - command.

To end, I wonder whether it is possible to get it, maybe:

a) Under -margins + marginsplot
b) Using Mata, after transforming the r(table) in a dataset
c) Some sort of - predict - under "twoway" command, using the mean + rcap.

So far, I tried options a and c, to no avail.

Thank you in advance for the suggestions.

[Marcos Almeida]

I'm still trying to curb the issue.

Among the solutions, quite "paleolithic", so to speak, I copied and pasted the latent marginal means (plus CIs) in a new dataset, and managed to get a graph by using - graph twoway - with - rcap - commands.

Sitll, I wonder whether there is a more elegant solution, for the strategy seems to be "anywhere nearby" (since Stata gives the marginal means and CIs).

So near, yet so far, I mean, in terms of getting the marginsplot for the latent marginal means, according to each class?

[Anton Schwarz]

Helly Marcos Almeida,

I'm facing the same issue. Did you find a (more elegant) solution yet?

Best regards,
Anton

[Marcos Almeida]

Hello Anton,

Actually, I did manage to get the wishful graph, albeit under a dowdy way, you know, as described in #2.

[Marcos Almeida]

This is an update of my own query (also a reply to Anton Schwarz), more than two months later.

Indeed, I couldn't keep this (yet unsurpassed) obstacle out of mind, hence I had to face it anyway.

Basically, we need to fiddle with the - post - option beforehand, then save the columns of the matrix e(b) as variables names in the dataset.

estat lcmean, post
ereturn list
matrix D = e(b)
svmat D, names(var)
matrix list D

After performing ordinary data management operations (rename, generate, reshape, label, etc.), we will find the desired display for a graph bar (mean) command.

That being said, and feeling somewhat relieved, this is just a strategy (a little bit more elegant than typing results directly in a blank spreadsheet, I expect) to provide a similar graph such as the one presented in the link I shared in #1, then with the LCA frame, now with the LPA frame.

But I still prefer the more informative presentation under graph two way (scatter for the mean + rcap for the CIs).

Hopefully that helps!

[Ahmed Bayoumi]

I think the graph you are looking for can be obtained with the following code

estat lcmean, post
margins, dydx(*) noesample predict(outcome(1)) predict(outcome(2))
marginsplot, bydimension(_outcome, allsimplelabels)

Or as a bar graph

marginsplot, bydimension(_outcome, allsimplelabels) recast(bar)

Best,

Ahmed

[Marcos Almeida]

Thank you for the suggestion. The graphs weren't exactly what I wished. That said, just typing - marginsplot - after the command you shared -
margins, dydx(*) noesample predict(outcome(1)) predict(outcome(2)) - gave practically what I desired, apart from the fact that it presents a connected display, instead of just means + CIs ,as we would get with rcap command. Anyway, that was really an elegant rendition. Thanks again.