Model Comparison

Mauricio Garnier-Villarreal

Introduction

The traditional method for model comparison in frequentist SEM (fSEM) is the $\chi^2$ (Likelihood Ratio Test) and its variations. But for BSEM, we would take the Bayesian model comparison methods, and apply them to SEM.

Specifically, we will focus on two information criteria, (1) Widely Applicable Information Criterion (WAIC), and (2) Leave-One-Out cross-validation (LOO).

These methods intend to evaluate the out-of-sample predictive accuracy of the models, and compare that performance. This is the ability to predict a datapoint that hasn’t been used in the training model (McElreath 2020)

For this example we will use the Industrialization and Political Democracy example (Bollen 1989).

model <- '
  # latent variable definitions
     ind60 =~ x1 + x2 + x3
     dem60 =~ a*y1 + b*y2 + c*y3 + d*y4
     dem65 =~ a*y5 + b*y6 + c*y7 + d*y8

  # regressions
    dem60 ~ ind60
    dem65 ~ ind60 + dem60

  # residual correlations
    y1 ~~ y5
    y2 ~~ y4 + y6
    y3 ~~ y7
    y4 ~~ y8
    y6 ~~ y8
'

fit1 <- bsem(model, data=PoliticalDemocracy,
            std.lv=T, meanstructure=T, n.chains=3,
            burnin=500, sample=1000)

Widely Applicable Information Criterion

WAIC (Watanabe 2010) can be seen as a fully Bayesian generalization of the Akaike Information Criteria (AIC), where we have a measure of uncertainty/information of the model prediction for each row in the data across all posterior draws. This is the Log-Pointwise-Predictive-Density (lppd). The WAIC is defined as

$$\begin{equation} WAIC= -2lppd + 2efp_{WAIC}, \end{equation}$$ The first term involves the log-likelihoods of observed data (marginal over latent variables) and the second term is the effective number of parameters. The first term, $lppd$, is estimated as:

$$\begin{equation} \widehat{lppd} = \sum^{n}_{i = 1} log \Bigg(\frac{1}{S}\sum^{S}_{S=1}f(y_{i}|\theta^{S}) \Bigg) \end{equation}$$

where $S$ is the number of posterior draws and $f(y_{i}|\theta^{S})$ is the density of observation $i$ with respect to the parameter sampled at iteration $s$.

The effective number of parameter ($efp_{WAIC}$) is calculated as:

$$\begin{equation}\label{eq:efpWAIC} efp_{WAIC} = \sum^n_{i=1}var_{s}(logf(y_{i}|\theta)) \end{equation}$$

A separate variance is estimated for each observation $i$ across the $S$ posterior draws.

Leave-One-Out cross-validation

The LOO measures the predictive density of each observation holding out one observation at the time and use the rest of the observations to update the prior. This estimation is calculated via (Vehtari, Gelman, and Gabry 2017):

$$\begin{equation} LOO = -2\sum_{i=1}^{n} log \Bigg(\frac{\sum^{S}_{s =1} w^{s}_{i}f(y_{i}|\theta^{s})}{\sum^{s}_{s=1} w^{s}_{i}}\Bigg) \end{equation}$$

Where the $w^s_{i}$ are Pareto-smoothed sampling weights based on the relative magnitude of individual $i$ density function across the $S$ posterior samples.

The LOO effective number of parameters involves the $lppd$ term from WAIC:

$$\begin{equation} efp_{LOO} = lppd + LOO/2 \end{equation}$$

Model comparison

As both WAIC and LOO approximate the models’ performance across posterior draws, we are able to calculate a standard error for them and for model comparisons involving them.

The model differences estimate the differences across the Expected Log-Pointwise-Predictive-Density (elpd), and the standard error of the respective difference.

There are no clear cutoff rules on how to interpret and present these comparisons, and the researchers need to use their expert knowledge as part of the decision process. The best recommendation is to present the differences in elpd $\Delta elpd$, the standard error, and the ratio between them. If the ratio is at least 2 can be consider evidence of differences between the models, and a ratio of 4 would be considered stronger evidence.

For the first example, we will compare the standard political democracy model, with a model where all factor regressions are fixed to 0.

model <- '
  # latent variable definitions
     ind60 =~ x1 + x2 + x3
     dem60 =~ a*y1 + b*y2 + c*y3 + d*y4
     dem65 =~ a*y5 + b*y6 + c*y7 + d*y8

  # regressions
    dem60 ~ 0*ind60
    dem65 ~ 0*ind60 + 0*dem60

  # residual correlations
    y1 ~~ y5
    y2 ~~ y4 + y6
    y3 ~~ y7
    y4 ~~ y8
    y6 ~~ y8
'

fit2 <- bsem(model, data=PoliticalDemocracy,
            std.lv=T, meanstructure=T, n.chains=3,
            burnin=500, sample=1000)

Once we have the 2 models, we can compare them with the blavCompare

bc12 <- blavCompare(fit1, fit2)

By looking into this comparison object, you can see the WAIC, LOO, estimates, and the respective differences between them. As these are information criteria, the best model is the one with the lowest value

bc12

## $bf
##   bf mll1 mll2 
##   NA   NA   NA 
## 
## $loo
## $loo[[1]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##          Estimate   SE
## elpd_loo  -1606.3 19.5
## p_loo        37.8  2.9
## looic      3212.6 39.1
## ------
## MCSE of elpd_loo is 0.2.
## MCSE and ESS estimates assume MCMC draws (r_eff in [0.6, 1.4]).
## 
## All Pareto k estimates are good (k < 0.7).
## See help('pareto-k-diagnostic') for details.
## 
## $loo[[2]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##          Estimate   SE
## elpd_loo  -1646.9 18.9
## p_loo        34.6  2.7
## looic      3293.8 37.8
## ------
## MCSE of elpd_loo is 0.2.
## MCSE and ESS estimates assume MCMC draws (r_eff in [0.6, 1.2]).
## 
## All Pareto k estimates are good (k < 0.7).
## See help('pareto-k-diagnostic') for details.
## 
## 
## $diff_loo
##        elpd_diff se_diff
## model1   0.0       0.0  
## model2 -40.6       7.9  
## 
## $waic
## $waic[[1]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##           Estimate   SE
## elpd_waic  -1606.0 19.5
## p_waic        37.5  2.9
## waic        3211.9 39.0
## 
## 38 (50.7%) p_waic estimates greater than 0.4. We recommend trying loo instead. 
## 
## $waic[[2]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##           Estimate   SE
## elpd_waic  -1646.7 18.9
## p_waic        34.4  2.7
## waic        3293.4 37.7
## 
## 33 (44.0%) p_waic estimates greater than 0.4. We recommend trying loo instead. 
## 
## 
## $diff_waic
##        elpd_diff se_diff
## model1   0.0       0.0  
## model2 -40.7       7.9

In this case we can see that model 1 has lower LOOIC, and the ratio shows that the LOO differences is 5 SE of magnitude. This indicates that the model with the estimated regressions is better

abs(bc12$diff_loo[,"elpd_diff"] / bc12$diff_loo[,"se_diff"])

##   model1   model2 
##      NaN 5.119273

Now, lets look at an example with a smaller difference between models, where only the smallest regression (dem65~ind60) is fixed to 0.

model <- '
  # latent variable definitions
     ind60 =~ x1 + x2 + x3
     dem60 =~ a*y1 + b*y2 + c*y3 + d*y4
     dem65 =~ a*y5 + b*y6 + c*y7 + d*y8

  # regressions
    dem60 ~ ind60
    dem65 ~ 0*ind60 + dem60

  # residual correlations
    y1 ~~ y5
    y2 ~~ y4 + y6
    y3 ~~ y7
    y4 ~~ y8
    y6 ~~ y8
'

fit3 <- bsem(model, data=PoliticalDemocracy,
            std.lv=T, meanstructure=T, n.chains=3,
            burnin=500, sample=1000)
bc13 <- blavCompare(fit1, fit3)

When we see the LOOIC, we see that the difference between the two models is minimal, and the ratio is 0.21. This indicates that the models are functionally equivalent. In a case like this, it is up to the researchers to decide which model is a better representation, and theoretically stronger.

bc13

## $bf
##   bf mll1 mll2 
##   NA   NA   NA 
## 
## $loo
## $loo[[1]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##          Estimate   SE
## elpd_loo  -1606.3 19.5
## p_loo        37.8  2.9
## looic      3212.6 39.1
## ------
## MCSE of elpd_loo is 0.2.
## MCSE and ESS estimates assume MCMC draws (r_eff in [0.6, 1.4]).
## 
## All Pareto k estimates are good (k < 0.7).
## See help('pareto-k-diagnostic') for details.
## 
## $loo[[2]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##          Estimate   SE
## elpd_loo  -1606.7 19.4
## p_loo        37.6  2.9
## looic      3213.5 38.9
## ------
## MCSE of elpd_loo is 0.2.
## MCSE and ESS estimates assume MCMC draws (r_eff in [0.5, 1.3]).
## 
## All Pareto k estimates are good (k < 0.7).
## See help('pareto-k-diagnostic') for details.
## 
## 
## $diff_loo
##        elpd_diff se_diff
## model1  0.0       0.0   
## model2 -0.4       0.9   
## 
## $waic
## $waic[[1]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##           Estimate   SE
## elpd_waic  -1606.0 19.5
## p_waic        37.5  2.9
## waic        3211.9 39.0
## 
## 38 (50.7%) p_waic estimates greater than 0.4. We recommend trying loo instead. 
## 
## $waic[[2]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##           Estimate   SE
## elpd_waic  -1606.5 19.4
## p_waic        37.3  2.9
## waic        3212.9 38.8
## 
## 39 (52.0%) p_waic estimates greater than 0.4. We recommend trying loo instead. 
## 
## 
## $diff_waic
##        elpd_diff se_diff
## model1  0.0       0.0   
## model2 -0.5       0.9

abs(bc13$diff_loo[,"elpd_diff"] / bc13$diff_loo[,"se_diff"])

##    model1    model2 
##       NaN 0.4547628

Lets do one last model, where only the largest regression (dem65~dem60) is fixed to 0.

model <- '
  # latent variable definitions
     ind60 =~ x1 + x2 + x3
     dem60 =~ a*y1 + b*y2 + c*y3 + d*y4
     dem65 =~ a*y5 + b*y6 + c*y7 + d*y8

  # regressions
    dem60 ~ ind60
    dem65 ~ ind60 + 0*dem60

  # residual correlations
    y1 ~~ y5
    y2 ~~ y4 + y6
    y3 ~~ y7
    y4 ~~ y8
    y6 ~~ y8
'

fit4 <- bsem(model, data=PoliticalDemocracy,
            std.lv=T, meanstructure=T, n.chains=3,
            burnin=500, sample=1000)
bc14 <- blavCompare(fit1, fit4)

In this case, by looking at the LOOIC, we see that model one is better (lower value), and the ratio of the difference shows that the model is 5 SE in magnitude. Indicating that there is evidence of model predictive differences

bc14

## $bf
##   bf mll1 mll2 
##   NA   NA   NA 
## 
## $loo
## $loo[[1]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##          Estimate   SE
## elpd_loo  -1606.3 19.5
## p_loo        37.8  2.9
## looic      3212.6 39.1
## ------
## MCSE of elpd_loo is 0.2.
## MCSE and ESS estimates assume MCMC draws (r_eff in [0.6, 1.4]).
## 
## All Pareto k estimates are good (k < 0.7).
## See help('pareto-k-diagnostic') for details.
## 
## $loo[[2]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##          Estimate   SE
## elpd_loo  -1630.1 19.9
## p_loo        38.3  3.0
## looic      3260.1 39.7
## ------
## MCSE of elpd_loo is 0.3.
## MCSE and ESS estimates assume MCMC draws (r_eff in [0.2, 1.3]).
## 
## All Pareto k estimates are good (k < 0.7).
## See help('pareto-k-diagnostic') for details.
## 
## 
## $diff_loo
##        elpd_diff se_diff
## model1   0.0       0.0  
## model2 -23.8       4.0  
## 
## $waic
## $waic[[1]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##           Estimate   SE
## elpd_waic  -1606.0 19.5
## p_waic        37.5  2.9
## waic        3211.9 39.0
## 
## 38 (50.7%) p_waic estimates greater than 0.4. We recommend trying loo instead. 
## 
## $waic[[2]]
## 
## Computed from 3000 by 75 log-likelihood matrix.
## 
##           Estimate   SE
## elpd_waic  -1629.7 19.8
## p_waic        38.0  3.0
## waic        3259.4 39.6
## 
## 39 (52.0%) p_waic estimates greater than 0.4. We recommend trying loo instead. 
## 
## 
## $diff_waic
##        elpd_diff se_diff
## model1   0.0       0.0  
## model2 -23.7       4.0

abs(bc14$diff_loo[,"elpd_diff"] / bc14$diff_loo[,"se_diff"])

##   model1   model2 
##      NaN 5.957015

Bayes factor

In the Bayesian literature you will make use of the Bayes factor (BF) to compare models. There are a number of criticisms related to the use of the BF in BSEM, including (1) the BF is unstable for large models (like most SEMs), (2) it is highly sensitive to model priors, (3) it requires strong priors to have stable estimation of it, (4) it can require large number of posterior draws, (5) the estimation using the marginal likelihood ignores a lot of information from the posterior distributions. For more details on this discussion please see Tendeiro and Kiers (2019) and Schad et al. (2022). These criticisms lead us to recommend against use of the BF in everyday BSEM estimation. For researchers who commit to their prior distributions and who commit to exploring the noise in their computations, the BF can used to describe the relative odds of one model over another, which is more intuitive than some other model comparison metrics.

Summary

We recommend the use of LOO or WAIC as general model comparison metrics for BSEM. They allow us to estimate the models’ out-of-sample predictive accuracy, and the respective differences across posterior draws. They also provide us uncertainty estimates in the comparison.

In most cases LOO and WAIC will lead to similar results, and LOO is recommended as the most stable metric (Vehtari, Gelman, and Gabry 2017). In general, a $\Delta elpd$ of at least 2 standard errors and preferably 4 standard errors can be interpreted as evidence of differential predictive accuracy.

References

Bollen, Kenneth A. 1989. Structural Equations with Latent Variables. Wiley Series in Probability and Mathematical Statistics. John Wiley & Sons, Inc.

McElreath, Richard. 2020. Statistical Rethinking: A Bayesian Course with Examples in R and Stan. 2nd ed. CRC Texts in Statistical Science. Boca Raton: Taylor; Francis, CRC Press.

Schad, Daniel J., Bruno Nicenboim, Paul-Christian Bürkner, Michael Betancourt, and Shravan Vasishth. 2022. “Workflow Techniques for the Robust Use of Bayes Factors.” Psychological Methods, March. https://doi.org/10.1037/met0000472.

Tendeiro, Jorge N., and Henk A. L. Kiers. 2019. “A Review of Issues about Null Hypothesis Bayesian Testing.” Psychological Methods 24 (6): 774–95. https://doi.org/10.1037/met0000221.

Vehtari, Aki, Andrew Gelman, and Jonah Gabry. 2017. “Practical Bayesian Model Evaluation Using Leave-One-Out Cross-Validation and WAIC.” Statistics and Computing 27 (5): 1413–32. https://doi.org/10.1007/s11222-016-9696-4.

Watanabe, Sumio. 2010. “Asymptotic Equivalence of Bayes Cross Validation and Widely Applicable Information Criterion in Singular Learning Theory.” Journal of Machine Learning Research 11: 3571–94.