Skip to contents

The task of model selection targets the question: If there are several competing models, how do I choose the most appropriate one? This vignette1 outlines the model selection tools implemented in {RprobitB}.

For illustration, we revisit the probit model of travelers deciding between two fictional train route alternatives from the vignette on model fitting:

data("model_train", package = "RprobitB")
model_train
#> Probit model 'choice ~ price + time + change + comfort | 0'.

As a competing model, we consider explaining the choices only by the alternative’s price, i.e. the probit model with the formula choice ~ price | 0:2

model_train_sparse <- nested_model(model_train, form = choice ~ price | 0)

The nested_model() function helps to estimate a new version of model_train with new specifications. Here, only form has been changed.

The model_selection() function

{RprobitB} provides the convenience function model_selection(), which takes an arbitrary number of RprobitB_fit objects and returns a matrix of model selection criteria:

model_selection(model_train, model_train_sparse, 
                criteria = c("npar", "LL", "AIC", "BIC", "WAIC", "MMLL", "BF", "pred_acc"))
#>                          model_train model_train_sparse
#> npar                               4                  1
#> LL                          -1727.74           -1865.86
#> AIC                          3463.48            3733.72
#> BIC                          3487.41            3739.70
#> WAIC                         3463.76            3733.91
#> se(WAIC)                        0.18               0.07
#> pWAIC                           4.32               1.15
#> MMLL                        -1732.14           -1867.48
#> BF(*,model_train)                  1             < 0.01
#> BF(*,model_train_sparse)       > 100                  1
#> pred_acc                      69.55%             63.37%

Specifying criteria is optional. Per default, criteria = c("npar", "LL", "AIC", "BIC").3 The available model selection criteria are explained in the following.

npar

"npar" yields the number of model parameters, which is computed by the npar() method:

npar(model_train, model_train_sparse)
#> [1] 4 1

Here, model_train has 4 parameters (a coefficient for price, time, change, and comfort, respectively), while model_train_sparse has only a single price coefficient.

LL

If "LL" is included in criteria, model_selection() returns the model’s log-likelihood values. They can also be directly accessed via the logLik() method:4

logLik(model_train)
#> [1] -1727.742
logLik(model_train_sparse)
#> [1] -1865.861

AIC

Including "AIC" yields the Akaike’s Information Criterion (Akaike 1974), which is computed as \[-2 \cdot \text{LL} + k \cdot \text{npar},\] where \(\text{LL}\) is the model’s log-likelihood value, \(k\) is the penalty per parameter with \(k = 2\) per default for the classical AIC, and \(\text{npar}\) is the number of parameters in the fitted model.

Alternatively, the AIC() method also returns the AIC values:

AIC(model_train, model_train_sparse, k = 2)
#> [1] 3463.485 3733.723

The AIC quantifies the trade-off between over- and under-fitting, where smaller values are preferred. Here, the increase in goodness of fit justifies the additional 3 parameters of model_train.

BIC

Similar to the AIC, "BIC" yields the Bayesian Information Criterion (Schwarz 1978), which is defined as \[-2 \cdot \text{LL} + \log{(\text{nobs})} \cdot \text{npar},\] where \(\text{LL}\) is the model’s log-likelihood value, \(\text{nobs}\) is the number of data points (which can be accessed via the nobs() method), and \(\text{npar}\) is the number of parameters in the fitted model. The interpretation is analogue to AIC.

{RprobitB} also provided a method for the BIC value:

BIC(model_train, model_train_sparse)
#> [1] 3487.414 3739.705

WAIC (with se(WAIC) and pWAIC)

WAIC is short for Widely Applicable (or Watanabe-Akaike) Information Criterion (Watanabe and Opper 2010). As for AIC and BIC, the smaller the WAIC value the better the model. Including "WAIC" in criteria yields the WAIC value, its standard error se(WAIC), and the effective number of parameters pWAIC, see below.

The WAIC is defined as

\[-2 \cdot \text{lppd} + 2\cdot p_\text{WAIC},\]

where \(\text{lppd}\) stands for log pointwise predictive density and \(p_\text{WAIC}\) is a penalty term proportional to the variance in the posterior distribution that is sometimes called effective number of parameters, see McElreath (2020) p. 220 for a reference.

The \(\text{lppd}\) is approximated as follows. Let \[p_{si} = \Pr(y_i\mid \theta_s)\] be the probability of observation \(y_i\) (here the single choices) given the \(s\)-th set \(\theta_s\) of parameter samples from the posterior. Then

\[\text{lppd} = \sum_i \log \left( S^{-1} \sum_s p_{si} \right).\] The penalty term is computed as the sum over the variances in log-probability for each observation: \[p_\text{WAIC} = \sum_i \mathbb{V}_{\theta} \log (p_{si}) . \] The \(\text{WAIC}\) has a standard error of \[\sqrt{n \cdot \mathbb{V}_i \left[-2 \left(\text{lppd} - \mathbb{V}_{\theta} \log (p_{si}) \right)\right]},\] where \(n\) is the number of choices.

Before computing the WAIC of an object, the probabilities \(p_{si}\) must be computed via the compute_p_si() function:5

model_train <- compute_p_si(model_train)

Afterwards, the WAIC can be accessed as follows, where the number in brackets is the standard error:

WAIC(model_train)
#> 3463.76 (0.18)
WAIC(model_train_sparse)
#> 3733.91 (0.07)

You can visualize the convergence of the WAIC as follows:

plot(WAIC(model_train))

plot(WAIC(model_train_sparse))

Here, both approximations look satisfactory. If the WAIC value does not seem to have converged, use more Gibbs samples by increasing R in mcmc() or decreasing B or Q via transform(), see the vignette on model fitting.

MMLL

"MMLL" in criteria stands for marginal model log-likelihood. The model’s marginal likelihood \(\Pr(y\mid M)\) for a model \(M\) and data \(y\) is required for the computation of Bayes factors, see below. In general, the term has no closed form and must be approximated numerically.

{RprobitB} uses the posterior Gibbs samples derived from the mcmc() function to approximate the model’s marginal likelihood via the posterior harmonic mean estimator (Newton and Raftery 1994): Let \(S\) denote the number of available posterior samples \(\theta_1,\dots,\theta_S\). Then, \[\Pr(y\mid M) = \left(\mathbb{E}_\text{posterior} 1/\Pr(y\mid \theta,M) \right)^{-1} \approx \left( \frac{1}{S} \sum_s 1/\Pr(y\mid \theta_s,M) \right) ^{-1} = \tilde{\Pr}(y\mid M).\]

By the law of large numbers, \(\tilde{\Pr}(y\mid M) \to \Pr(y\mid M)\) almost surely as \(S \to \infty\).

As for the WAIC, computing the MMLL relies on the probabilities \(p_{si} = \Pr(y_i\mid \theta_s)\), which must first be computed via the compute_p_si() function. Afterwards, the mml() function can be called with an RprobitB_fit object as input. The function returns the RprobitB_fit object, where the marginal likelihood value is saved as the entry "mml" and the marginal log-likelihood value as the attribute "mmll":6

model_train <- mml(model_train)
model_train$mml
#> 3.54e-117 * exp(-1464)
attr(model_train$mml, "mmll")
#> [1] -1732.138

Analogue to the WAIC value, the computation of the MMLL is an approximation that improves with rising (posterior) sample size. The convergence can again be verified visually via the plot() method:7

plot(model_train$mml, log = TRUE)

There are two options for improving the approximation: As for the WAIC, you can use more posterior samples. Alternatively, you can combine the posterior harmonic mean estimate with the prior arithmetic mean estimator (Hammersley and Handscomb 1964): For this approach, \(S\) samples \(\theta_1,\dots,\theta_S\) are drawn from the model’s prior distribution. Then,

\[\Pr(y\mid M) = \mathbb{E}_\text{prior} \Pr(y\mid \theta,M) \approx \frac{1}{S} \sum_s \Pr(y\mid \theta_s,M) = \tilde{\Pr}(y\mid M).\]

Again, it hols by the law of large numbers, that \(\tilde{\Pr}(y\mid M) \to \Pr(y\mid M)\) almost surely as \(S \to \infty\). The final approximation of the model’s marginal likelihood than is a weighted sum of the posterior harmonic mean estimate and the prior arithmetic mean estimate, where the weights are determined by the sample sizes.

To use the prior arithmetic mean estimator, call the mml() function with a specification of the number of prior draws S and set recompute = TRUE:8

model_train <- mml(model_train, S = 1000, recompute = TRUE)

Note that the prior arithmetic mean estimator works well if the prior and posterior distribution have a similar shape and strong overlap, as Gronau et al. (2017) points out. Otherwise, most of the sampled prior values result in a likelihood value close to zero, thereby contributing only marginally to the approximation. In this case, a very large number S of prior samples is required.

BF

The Bayes factor is an index of relative posterior model plausibility of one model over another (Marin and Robert 2014). Given data \(\texttt{y}\) and two models \(\texttt{mod0}\) and \(\texttt{mod1}\), it is defined as

\[ BF(\texttt{mod0},\texttt{mod1}) = \frac{\Pr(\texttt{mod0} \mid \texttt{y})}{\Pr(\texttt{mod1} \mid \texttt{y})} = \frac{\Pr(\texttt{y} \mid \texttt{mod0} )}{\Pr(\texttt{y} \mid \texttt{mod1})} / \frac{\Pr(\texttt{mod0})}{\Pr(\texttt{mod1})}. \]

The ratio \(\Pr(\texttt{mod0}) / \Pr(\texttt{mod1})\) expresses the factor by which \(\texttt{mod0}\) a priori is assumed to be the correct model. Per default, {RprobitB} sets \(\Pr(\texttt{mod0}) = \Pr(\texttt{mod1}) = 0.5\). The front part \(\Pr(\texttt{y} \mid \texttt{mod0} ) / \Pr(\texttt{y} \mid \texttt{mod1})\) is the ratio of the marginal model likelihoods. A value of \(BF(\texttt{mod0},\texttt{mod1}) > 1\) means that the model \(\texttt{mod0}\) is more strongly supported by the data under consideration than \(\texttt{mod1}\).

Adding "BF" to the criteria argument of model_selection yields the Bayes factors. We see a decisive evidence (Jeffreys 1998) in favor of model_train.

model_selection(model_train, model_train_sparse, criteria = c("BF"))
#>                          model_train model_train_sparse
#> BF(*,model_train)                  1             < 0.01
#> BF(*,model_train_sparse)       > 100                  1

pred_acc

Finally, adding "pred_acc" to the criteria argument for the model_selection() function returns the share of correctly predicted choices. From the output of model_selection() above (or alternatively the one in the following) we deduce that model_train correctly predicts about 6% of the choices more than model_train_sparse:9

pred_acc(model_train, model_train_sparse)
#> [1] 0.6954592 0.6336634

References

Akaike, H. 1974. “A New Look at the Statistical Model Identification” 19.

Gronau, Q. F., A. Sarafoglou, D. Matzke, A. Ly, U. Boehm, M. Marsman, D. S. Leslie, J. J. Forster, E. Wagenmakers, and H. Steingroever. 2017. “A Tutorial on Bridge Sampling.” Journal of Mathematical Psychology 81: 80–97.

Hammersley, J. M., and D. C. Handscomb. 1964. “General Principles of the Monte Carlo Method,” 50–75.

Jeffreys, Harold. 1998. The Theory of Probability. OUP Oxford.

Marin, J., and C. Robert. 2014. Bayesian essentials with R. Springer Textbooks in Statistics. Springer Verlag, New York.

McElreath, R. 2020. Statistical Rethinking: A Bayesian Course with Examples in R and Stan. 2. ed. Chapman; Hall/CRC.

Newton, M. A., and A. E. Raftery. 1994. “Approximate Bayesian Inference with the Weighted Likelihood Bootstrap.” Journal of the Royal Statistical Society: Series B (Methodological) 56 (1): 3–26.

Schwarz, G. 1978. “Estimating the Dimension of a Model.” The Annals of Statistics 6.

Watanabe, S., and M. Opper. 2010. “Asymptotic Equivalence of Bayes Cross Validation and Widely Applicable Information Criterion in Singular Learning Theory.” Journal of Machine Learning Research 11 (12).