Latent Profile Analysis (LPA) is a statistical modeling approach for estimating distinct profiles, or groups, of variables. In the social sciences and in educational research, these profiles could represent, for example, how different youth experience dimensions of being engaged (i.e., cognitively, behaviorally, and affectively) at the same time.
tidyLPA provides the functionality to carry out LPA in R. In particular, tidyLPA provides functionality to specify different models that determine whether and how different parameters (i.e., means, variances, and covariances) are estimated and to specify (and compare solutions for) the number of profiles to estimate.
This introduction to tidyLPA vignette is an overview of LPA and the tidyLPA package. This vignette covers the following topics:
Latent Profile Analysis (LPA) is a statistical modeling approach for estimating distinct profiles of variables. In the social sciences and in educational research, these profiles could represent, for example, how different youth experience dimensions of being engaged (i.e., cognitively, behaviorally, and affectively) at the same time. Note that LPA works best with continuous variables (and, in some cases, ordinal variables), but is not appropriate for dichotomous (binary) variables.
Many analysts have carried out LPA using a latent variable modeling approach. From this approach, different parameters - means, variances, and covariances - are freely estimated across profiles, fixed to be the same across profiles, or constrained to be zero. The MPlus software is commonly used to estimate these models (see here) using the expectation-maximization (EM) algorithm to obtain the maximum likelihood estimates for the parameters.
Different models (or how or whether parameters are estimated) can be specified and estimated. While MPlus is widely-used (and powerful), it is costly, closed-source, and can be difficult to use, particularly with respect to interpreting or using the output of specified models as part of a reproducible workflow.
The goal of tidyLPA is to make it easy to carry out LPA using R. In particular, tidyLPA provides an interface to the powerful and widely-used mclust package for Gaussian Mixture Modeling. This means that tidyLPA does not contain code to carry out LPA directly, but rather provides “wrappers” to mclust functions that make them easier to use. The primary contributions of tidyLPA are to:
data.frame
(specifically its modified version, a tibble
) that can be used to create plots or can be used in subsequent analyses%>%
to compose functionsIn the open-source R software, there is not yet a tool to easily carry out LPA, though there are many tools that one could use to. For example, the R version of OpenMx can be used for this purpose (and to specify almost any model possible to specify within a latent variable modeling approach). However, while OpenMx is very flexible, it can also be challenging to use.
Other tools in R allow for estimating Gaussian mixture models, or models of multivariate Gaussian (or normal) distributions. In this framework, the term “mixture component” has a similar meaning to a profile. While much more constraining than the latent variable modeling framework, the approach is often similar or the same: the EM algorithm is used to (aim to) obtain the maximum likelihood estimates for the parameters being estimated. Like in the latent variable modeling framework, different models can be specified.
In addition to following the same general approach, using tools that are designed for Gaussian mixture modeling have other benefits, some efficiency-related (see RMixMod, which uses compiled C++ code) and others in terms of ease-of-use (i.e., the plot methods built-in to RMixMod, mclust, and other tools). However, they also have some drawbacks, in that it can be difficult to translate between the model specifications, which are often described in terms of the geometric properties of the multivariate distributions being estimated (i.e., “spherical, equal volume”), rather than in terms of whether and how the means, variances, and covariances are estimated. They also may use different default settings (than those encountered in MPlus) in terms of the EM algorithm, which can make comparing results across tools challenging.
This package focuses on models that are commonly specified as part of LPA. Because MPlus is so widely-used, it can be helpful to compare output from other software to MPlus. The functions in tidyLPA that use mclust have been benchmarked to MPlus for a series of simple models (with small datasets and for models with small numbers of profiles. This becnhmarking vignette contains information on how mclust and Mplus compare. As long as you have purchased MPlus (and installed MplusAutomation), this vignette can be used to replicate all of the results for the benchmark. Note that most of the output is identical, though there are some differences in the hundredths decimal places for some. Because of differences in settings for the EM algorithm and particularly for the start values (random starts for MPlus and starting values from hierarchical clustering for mclust), differences may be expected for more complex data and models. An important direction for the development of tidyLPA (the functions that use mclust) is to continue to understand when and why the output differs from MPlus output. Note that tidyLPA also provides functions to interface to MPlus, though these are not the focus of the package, as they require MPlus to be purchased and installed in order to be used.
You can install tidyLPA from CRAN with:
install.packages("tidyLPA")
You can also install the development version of tidyLPA from GitHub with:
install.packages("devtools")
::install_github("data-edu/tidyLPA") devtools
Here is a brief example using the built-in pisaUSA15
data set and variables for broad interest, enjoyment, and self-efficacy. Note that we first type the name of the data frame, followed by the unquoted names of the variables used to create the profiles. We also specify the number of profiles and the model. See ?estimate_profiles
for more details.
In these examples, we pass the results of one function to the next by piping (using the %>%
operator, loaded from the dplyr
package). We pass the data to a function that selects relevant variables, and then to estimate_profiles
:
library(tidyLPA)
library(dplyr)
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
single_imputation() %>%
estimate_profiles(3)
#> tidyLPA analysis using mclust:
#>
#> Model Classes AIC BIC Entropy prob_min prob_max n_min n_max BLRT_p
#> 1 3 639.57 676.04 0.71 0.67 0.91 0.11 0.60 0.06
We can use Mplus simply by changing the package argument for estimate_profiles()
(not run):
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
single_imputation() %>%
estimate_profiles(3, package = "MplusAutomation")
A simple summary of the analysis is printed to the console (and its posterior probability). The resulting object can be further passed down a pipeline to other functions, such as plot
, compare_solutions
, get_data
, get_fit
, etc. This is the “tidy” part, in that the function can be embedded in a tidy analysis pipeline.
If you have Mplus installed, you can call the version of this function that uses MPlus in the same way, by adding the argument package = "MplusAutomation
.
We can plot the profiles by piping the output to plot_profiles()
:
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
single_imputation() %>%
scale() %>%
estimate_profiles(3) %>%
plot_profiles()
The function compare_solutions()
compares the fit of several estimated models, with varying numbers of profiles and model specifications:
1:100, ] %>%
pisaUSA15[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"))
#> Warning in (function (data, modelName = NULL, nboot = 999, level = 0.05, : some
#> model(s) could not be fitted!
#> Compare tidyLPA solutions:
#>
#> Model Classes AIC BIC
#> 1 1 675.131 690.762
#> 1 2 637.105 663.157
#> 1 3 625.807 662.279
#> 6 1 616.765 640.212
#> 6 2 624.221 673.719
#> 6 3 628.881 704.431
#>
#> Best model according to AIC is Model 6 with 1 classes.
#> Best model according to BIC is Model 6 with 1 classes.
#>
#> An analytic hierarchy process, based on the fit indices AIC, AWE, BIC, CLC, and KIC (Akogul & Erisoglu, 2017), suggests the best solution is Model 6 with 1 classes.
Additional arguments can be passed as follows.
For MPlus (here, the additional argument is ANALYSIS
; see:
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
single_imputation() %>%
estimate_profiles(3,
package = "mplus",
ANALYSIS = "starts = 100, 20;")
For mclust (here, the additional argument is to prior
; see ?mclust::mclust
for other options):
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
single_imputation() %>%
estimate_profiles(3,
prior = priorControl())
In addition to the number of profiles (specified with the n_profiles
argument), the model can be specified in terms of whether and how the variable variances and covariances are estimated.
The models are specified by passing arguments to the variance
and covariance
arguments. The possible values for these arguments are:
variances
: “equal” and “varying”covariances
: “varying”, “equal”, and “zero”If no values are specified for these, then the variances are constrained to be equal across classes, and covariances are fixed to 0 (conditional independence of the indicators).
These arguments allow for four models to be specified:
Two additional models (Models 4 and 5) can be fit using MPlus. More information on the models can be found in the vignette.
Here is an example of specifying a model with varying variances and covariances (Model 6):
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
estimate_profiles(3,
variances = "varying",
covariances = "varying")
#> Warning in estimate_profiles_mclust(df_full, n_profiles, model_numbers, : The mclust algorithm does not allow for missing data. Some rows were omitted from analysis. Consider using Mplus, which accounts for cases with partially missing data, or use a non-parametric single imputation technique prior to analysis, such as the R-package 'missForest'.
#> tidyLPA analysis using mclust:
#>
#> Model Classes AIC BIC Entropy prob_min prob_max n_min n_max BLRT_p
#> 6 3 583.13 656.89 0.86 0.87 0.98 0.16 0.62 0.04
In general, the approach to choosing the model is similar to choosing the number of profiles, requiring deciding on the basis of evidence from multiple sources, including information criteria, statistical tests, and concerns of interpretability and parsimony. The article by Pastor and colleagues (2007) has helpful information on the model specifications. Here, the six models that are possible to specify in LPA are described in terms of how the variables used to create the profiles are estimated.
Note that p represents different profiles and each parameterization is represented by a 4 x 4 covariance matrix and therefore would represent the parameterization for a four-profile solution. In all of the models, the means are estimated freely in the different profiles. Imagine that each row and column represents a different variable, i.e., the first row (and column) represents broad interest, the second enjoyment, the third self-efficacy, and the fourth another variable, i.e., future goals and plans.
In this model, which corresponds to the mclust model wit the name “EEI”, the variances are estimated to be equal across profiles, indicated by the absence of a p subscript for any of the diagonal elements of the matrix. The covariances are constrained to be zero, as indicated by the 0’s between every combination of the variables.
It is specified with variances = "equal"
and covariances = "zero"
.
This model is highly constrained but also parsimonious: the profiles are estimated in such a way that the variables’ variances are identical for each of the profiles, and the relationships between the variables are not estimated. In this way, less degrees of freedom are taken used to explain the observations that make up the data. However, estimating more parameters–as in the other models–may better explain the data, justifying the addition in complexity that their addition involves (and their reduction in degrees of freedom). This model is sometimes referred to as a class-invariant parameterization.
\[ \left[ \begin{matrix} { \sigma }_{ 1 }^{ 2 } & 0 & 0 & 0 \\ 0 & { \sigma }_{ 2 }^{ 2 } & 0 & 0 \\ 0 & 0 & { \sigma }_{ 3 }^{ 2 } & 0 \\ 0 & 0 & 0 & { \sigma }_{ 4 }^{ 2 } \end{matrix} \right] \]
This model corresponds to the mclust model “VVI” and allows for the variances to be freely estimated across profiles. The covariances are constrained to zero.
It is specified with variances = "varying"
and covariances = "zero"
.
Thus, it is more flexible (and less parsimonious) than model 1, but in terms of the covariances, is more constrained than model 2. This model is sometimes referred to as a class-varying diagonal parameterization.
\[ \left[ \begin{matrix} { \sigma }_{ 1p }^{ 2 } & 0 & 0 & 0 \\ 0 & { \sigma }_{ 2p }^{ 2 } & 0 & 0 \\ 0 & 0 & { \sigma }_{ 3p }^{ 2 } & 0 \\ 0 & 0 & 0 & { \sigma }_{ 4p }^{ 2 } \end{matrix} \right] \]
This model corresponds to the mclust model “EEE”. In this model, the variances are still constrained to be the same across the profiles, although now the covariances are estimated (but like the variances, are constrained to be the same across profiles).
It is specified with variances = "equal"
and covariances = "equal"
.
Thus, this model is the first to estimate the covariance (or correlations) of the variables used to create the profiles, thus adding more information that can be used to better understand the characteristics of the profiles (and, potentially, better explain the data). This model is sometimes referred to as a class-invariant unrestricted parameterization.
\[ \left[ \begin{matrix} { \sigma }_{ 1 }^{ 2 } & { \sigma }_{ 21 } & { \sigma }_{ 31 } & { \sigma }_{ 41 } \\ { \sigma }_{ 12 } & { \sigma }_{ 2 }^{ 2 } & { \sigma }_{ 23 } & { \sigma }_{ 24 } \\ { \sigma }_{ 13 } & { \sigma }_{ 12 } & { \sigma }_{ 3 }^{ 2 } & { \sigma }_{ 33 } \\ { \sigma }_{ 14 } & { \sigma }_{ 12 } & { \sigma }_{ 12 } & { \sigma }_{ 4 }^{ 2 } \end{matrix} \right] \]
This model, which specifies for the variances to be freely estimated across the profiles and for the covariances to be estimated to be equal across profiles, extends model 3.
It is specified with variances = "varying"
and covariances = "equal"
.
Unfortunately, this model cannot be specified with mclust, though it can be with MPlus; this model can be used with the functions to interface to MPlus described below.
\[ \left[ \begin{matrix} { \sigma }_{ 1p }^{ 2 } & { \sigma }_{ 21 } & { \sigma }_{ 31 } & { \sigma }_{ 41 } \\ { \sigma }_{ 12 } & { \sigma }_{ 2p }^{ 2 } & { \sigma }_{ 23 } & { \sigma }_{ 24 } \\ { \sigma }_{ 13 } & { \sigma }_{ 12 } & { \sigma }_{ 3p }^{ 2 } & { \sigma }_{ 33 } \\ { \sigma }_{ 14 } & { \sigma }_{ 12 } & { \sigma }_{ 12 } & { \sigma }_{ 4p }^{ 2 } \end{matrix} \right] \]
This model specifies the variances to be equal across the profiles, but allows the covariances to be freely estimated across the profiles.
It is specified with variances = "equal"
and covariances = "varying"
.
Like model 4, this model cannot be specified with mclust, though it can be with MPlus. Again, this model can be used with the functions to interface to MPlus described below.
\[ \left[ \begin{matrix} { \sigma }_{ 1 }^{ 2 } & { \sigma }_{ 21p } & { \sigma }_{ 31p } & { \sigma }_{ 41p } \\ { \sigma }_{ 12p } & { \sigma }_{ 2 }^{ 2 } & { \sigma }_{ 23p } & { \sigma }_{ 24p } \\ { \sigma }_{ 13p } & { \sigma }_{ 12p } & { \sigma }_{ 3 }^{ 2 } & { \sigma }_{ 33p } \\ { \sigma }_{ 14p } & { \sigma }_{ 12p } & { \sigma }_{ 12p } & { \sigma }_{ 4 }^{ 2 } \end{matrix} \right] \quad \]
This model corresponds to the mclust model “VVV”. It allows the variances and the covariances to be freely estimated across profiles.
It is specified with variances = "varying"
and covariances = "varying"
.
Thus, it is the most complex model, with the potential to allow for understanding many aspects of the variables that are used to estimate the profiles and how they are related. However, it is less parsimonious than all of the other models, and the added parameters should be considered in light of how preferred this model is relative to those with more simple specifications. This model is sometimes referred to as a class-varying unrestricted parameterization.
\[ \left[ \begin{matrix} { \sigma }_{ 1p }^{ 2 } & { \sigma }_{ 21p } & { \sigma }_{ 31p } & { \sigma }_{ 41p } \\ { \sigma }_{ 12p } & { \sigma }_{ 2p }^{ 2 } & { \sigma }_{ 23p } & { \sigma }_{ 24p } \\ { \sigma }_{ 13p } & { \sigma }_{ 12p } & { \sigma }_{ 3p }^{ 2 } & { \sigma }_{ 33p } \\ { \sigma }_{ 14p } & { \sigma }_{ 12p } & { \sigma }_{ 12p } & { \sigma }_{ 4p }^{ 2 } \end{matrix} \right] \]
There is a lot of output that is possible to obtain from the estimate_profiles()
function - much more than a tidy data frame, which is the default. The easiest way to access it is by using the get_estimates()
function.
<- pisaUSA15[1:100, ] %>%
m3 select(broad_interest, enjoyment, self_efficacy) %>%
estimate_profiles(3)
#> Warning in estimate_profiles_mclust(df_full, n_profiles, model_numbers, : The mclust algorithm does not allow for missing data. Some rows were omitted from analysis. Consider using Mplus, which accounts for cases with partially missing data, or use a non-parametric single imputation technique prior to analysis, such as the R-package 'missForest'.
get_estimates(m3)
#> # A tibble: 18 × 8
#> Category Parameter Estimate se p Class Model Classes
#> <chr> <chr> <dbl> <dbl> <dbl> <int> <dbl> <dbl>
#> 1 Means broad_interest 3.09 0.206 4.74e- 51 1 1 3
#> 2 Means enjoyment 3.51 0.176 4.80e- 89 1 1 3
#> 3 Means self_efficacy 1.58 0.184 7.74e- 18 1 1 3
#> 4 Variances broad_interest 0.489 0.118 3.36e- 5 1 1 3
#> 5 Variances enjoyment 0.230 0.0541 2.01e- 5 1 1 3
#> 6 Variances self_efficacy 0.246 0.0491 5.62e- 7 1 1 3
#> 7 Means broad_interest 1.14 0.459 1.29e- 2 2 1 3
#> 8 Means enjoyment 1.14 0.354 1.26e- 3 2 1 3
#> 9 Means self_efficacy 3.39 0.469 4.33e- 13 2 1 3
#> 10 Variances broad_interest 0.489 0.118 3.36e- 5 2 1 3
#> 11 Variances enjoyment 0.230 0.0541 2.01e- 5 2 1 3
#> 12 Variances self_efficacy 0.246 0.0491 5.62e- 7 2 1 3
#> 13 Means broad_interest 2.27 0.223 2.90e- 24 3 1 3
#> 14 Means enjoyment 2.49 0.178 2.03e- 44 3 1 3
#> 15 Means self_efficacy 2.26 0.100 5.49e-113 3 1 3
#> 16 Variances broad_interest 0.489 0.118 3.36e- 5 3 1 3
#> 17 Variances enjoyment 0.230 0.0541 2.01e- 5 3 1 3
#> 18 Variances self_efficacy 0.246 0.0491 5.62e- 7 3 1 3
Other options include how the raw data is processed.
We can center or scale the data before estimating the profiles with the scale()
or poms()
functions:
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
scale() %>%
estimate_profiles(4) %>%
plot_profiles()
#> Warning in estimate_profiles_mclust(df_full, n_profiles, model_numbers, : The mclust algorithm does not allow for missing data. Some rows were omitted from analysis. Consider using Mplus, which accounts for cases with partially missing data, or use a non-parametric single imputation technique prior to analysis, such as the R-package 'missForest'.
1:100, ] %>%
pisaUSA15[select(broad_interest, enjoyment, self_efficacy) %>%
poms() %>%
estimate_profiles(4) %>%
plot_profiles()
#> Warning in estimate_profiles_mclust(df_full, n_profiles, model_numbers, : The mclust algorithm does not allow for missing data. Some rows were omitted from analysis. Consider using Mplus, which accounts for cases with partially missing data, or use a non-parametric single imputation technique prior to analysis, such as the R-package 'missForest'.
Since we often wish to use the estimated profiles in subsequent analyses, we may want the original data.frame
, with variables that are predictors or outcomes of the profiles, included. Here, we created profiles with just two of the three variables, to demonstrate how the third variable is still returned in the output. We can return this data.frame
, and not just one with the variables used to create the profiles and the profile assignments (and posterior probabilities), using the function get_data()
:
get_data(m3)
#> # A tibble: 100 × 9
#> model_number classes_number broad_interest enjoyment self_efficacy CPROB1
#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 3 3.8 4 1 1.00e+0
#> 2 1 3 3 3 2.75 8.29e-2
#> 3 1 3 1.8 2.8 3.38 8.73e-4
#> 4 1 3 1.4 1 2.75 8.31e-8
#> 5 1 3 1.8 2.2 2 2.69e-3
#> 6 1 3 1.6 1.6 1.88 1.88e-4
#> 7 1 3 3 3.8 2.25 9.27e-1
#> 8 1 3 2.6 2.2 2 1.03e-2
#> 9 1 3 1 2.8 2.62 1.81e-3
#> 10 1 3 2.2 2 1.75 4.32e-3
#> # … with 90 more rows, and 3 more variables: CPROB2 <dbl>, CPROB3 <dbl>,
#> # Class <dbl>
We note that if more than one model is fit at once, then the data is returned not in the (wide) format above, but in long form, e.g.:
<- pisaUSA15[1:100, ] %>%
m4 select(broad_interest, enjoyment, self_efficacy) %>%
single_imputation() %>%
estimate_profiles(c(3, 4))
get_data(m4)
#> # A tibble: 700 × 9
#> model_number classes_number broad_interest enjoyment self_efficacy Class
#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 3 3.8 4 1 1
#> 2 1 3 3 3 2.75 3
#> 3 1 3 1.8 2.8 3.38 3
#> 4 1 3 1.4 1 2.75 2
#> 5 1 3 1.8 2.2 2 3
#> 6 1 3 1.6 1.6 1.88 3
#> 7 1 3 3 3.8 2.25 1
#> 8 1 3 2.6 2.2 2 3
#> 9 1 3 1 2.8 2.62 3
#> 10 1 3 2.2 2 1.75 3
#> # … with 690 more rows, and 3 more variables: Class_prob <int>,
#> # Probability <dbl>, id <int>
If we wish to work with the fit statistics more programatically, we can use the get_fit()
function:
get_fit(m4)
#> # A tibble: 2 × 18
#> Model Classes LogLik AIC AWE BIC CAIC CLC KIC SABIC ICL Entropy
#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 3 -303. 635. 776. 671. 685. 608. 652. 627. -700. 0.758
#> 2 1 4 -292. 620. 802. 667. 685. 586. 641. 610. -681. 0.898
#> # … with 6 more variables: prob_min <dbl>, prob_max <dbl>, n_min <dbl>,
#> # n_max <dbl>, BLRT_val <dbl>, BLRT_p <dbl>
Descriptions of these fit indices follows:
Notes
This is related to prcr, for use of two-step cluster analysis to carry out person-oriented analyses.
To contribute, file issues via GitHub here or get in touch via email or Twitter.
References
Pastor, D. A., Barron, K. E., Miller, B. J., & Davis, S. L. (2007). A latent profile analysis of college students’ achievement goal orientation. Contemporary Educational Psychology, 32(1), 8-47. (https://www.sciencedirect.com/science/article/pii/S0361476X06000543
Helpful resources
Hennig et al’s (2015) handbook for an overview of mixture models, of which LPA is often considered an instance of.
Collins and Lanza (2013) for a book on the related approach (for use with dichotomous, rather than continuous variables used to create the profiles) Latent Class Analysis (LCA)
How to cite tidyLPA
Rosenberg, J. M., Beymer, P. N., Anderson, D. J., Van Lissa, C. J., & Schmidt, J. A. (2018). tidyLPA: An R Package to Easily Carry Out Latent Profile Analysis (LPA) Using Open-Source or Commercial Software. Journal of Open Source Software, 3(30), 978, https://doi.org/10.21105/joss.00978
You can also cite the most latest version with the following citation:
Rosenberg, J. M., van Lissa, C. J., Beymer, P. N., Anderson, D. J., Schell, M. J. & Schmidt, J. A. (2019). tidyLPA: Easily carry out Latent Profile Analysis (LPA) using open-source or commercial software [R package]. https://data-edu.github.io/tidyLPA/
This material is based upon work supported by the National Science Foundation under Grant No.: DRL#1661064. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not reflect the views of the National Science Foundation.