Logistic Regression
Now that we’ve used the stan_glm() function, you might be wondering if you can fit other GLMs with a Bayesian model. Yes we can! The stan_glm() function supports every link function that glm() supports. We will fit a logistic regression model in the next example.
Suppose we are interested in how an undergraduate student’s GRE, GPA, and college prestige ranking affect their admission into graduate school. The response variable is whether or not the student was admitted to graduate school.
seed <- 196
admissions <- read.csv("https://stats.idre.ucla.edu/stat/data/binary.csv")
## view the first few rows of the data
head(admissions)## admit gre gpa rank
## 1 0 380 3.6 3
## 2 1 660 3.7 3
## 3 1 800 4.0 1
## 4 1 640 3.2 4
## 5 0 520 2.9 4
## 6 1 760 3.0 2admissions$rank <- factor(admissions$rank)
admissions$admit <- factor(admissions$admit)
admissions$gre <- scale(admissions$gre)
p <- 5
n <- nrow(admissions)Frequentist Approach
##
## Call:
## glm(formula = admit ~ ., family = binomial(), data = admissions)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.627 -0.866 -0.639 1.149 2.079
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.659 1.166 -2.28 0.0225
## gre 0.262 0.126 2.07 0.0385
## gpa 0.804 0.332 2.42 0.0154
## rank2 -0.675 0.316 -2.13 0.0328
## rank3 -1.340 0.345 -3.88 0.0001
## rank4 -1.551 0.418 -3.71 0.0002
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 499.98 on 399 degrees of freedom
## Residual deviance: 458.52 on 394 degrees of freedom
## AIC: 470.5
##
## Number of Fisher Scoring iterations: 4- Interpret the significant coefficients (at the 0.05 significance level).
Weakly Informative Prior: Normal
We have the choice of a logit or probit link. You should already know what probit and logistic regressions are. If you do not, please do a quick review of the frequentist versions. We should cover Bayesian logistic regression next week but we may or may not get to Bayesian probit regression. With stan_glm(), binomial models with a logit link function can typically be fit slightly faster than the identical model with a probit link because of how the two models are implemented in Stan. In the following code, we simply specify the chosen link, and set priors for the intercept and the predictor coefficients.
post1 <- stan_glm(admit ~ ., data = admissions,
family = binomial(link = "logit"),
prior = normal(0,1), prior_intercept = normal(0,1),
seed = seed,
refresh = 0)- What do our choice of priors say about our beliefs? How do we interpret these normal priors?
As always, it is good practice to run diagnostics to check model convergence. Here is a nice function that will allow you to do this without having to call many different functions:
Now we can look at posterior densities and estimates for the coefficients.
## Warning: `expand_scale()` is deprecated; use `expansion()` instead.
## (Intercept) gre gpa rank2 rank3 rank4
## -2.86 0.26 0.81 -0.47 -1.11 -1.26## 2.5% 97.5%
## (Intercept) -5.14 -0.607
## gre 0.02 0.521
## gpa 0.16 1.451
## rank2 -1.04 0.087
## rank3 -1.71 -0.491
## rank4 -1.98 -0.545- How do the estimated coefficients compare in this model to those from the model fit using
glm? - How do the credible intervals and standard errors of the coefficients compare to the confidence intervals and standard errors from the model fit using
glm?
Posterior Predictive Checks
##
## Computed from 4000 by 400 log-likelihood matrix
##
## Estimate SE
## elpd_loo -235.3 8.6
## p_loo 5.7 0.3
## looic 470.6 17.3
## ------
## Monte Carlo SE of elpd_loo is 0.0.
##
## All Pareto k estimates are good (k < 0.5).
## See help('pareto-k-diagnostic') for details.In the code chunk above, we assessed the strength of our model via its posterior predictive LOOCV. However as we know, this accuracy rate is quite meaningless unless we have something to compare it to. So let’s create a baseline model with no predictors to compare to this first model:
post0 <- stan_glm(admit ~ 1, data = admissions,
family = binomial(link = "logit"),
prior = normal(0,1), prior_intercept = normal(0,1),
seed = seed,
refresh = 0)
(loo0 <- loo(post0, save_psis = T))##
## Computed from 4000 by 400 log-likelihood matrix
##
## Estimate SE
## elpd_loo -250.9 7.1
## p_loo 0.9 0.0
## looic 501.8 14.2
## ------
## Monte Carlo SE of elpd_loo is 0.0.
##
## All Pareto k estimates are good (k < 0.5).
## See help('pareto-k-diagnostic') for details.## elpd_diff se_diff
## post1 0.0 0.0
## post0 -15.6 5.7- Which model is better? Why?
Below, we compute posterior predictive probabilities of the linear predictor via the posterior_linpred() function provided in the rstanarm package. This function will extract posterior draws from the linear predictor. If we used a link function, then specifying the transform argument as TRUE will return the predictor as transformed via the inverse-link.
We calculate these posterior predictive probabilities in order to determine the classification accuracy of our model. If the posterior probability of success for an individual is greater or equal to \(0.5\), then we would predict that observation to be a success (and similarly for less than \(0.5\)). For each observation, we can compare the posterior prediction to the actual observed value. The proportion of times we correctly predict an individual (i.e. [prediction = 0 and observation = 0] or [prediction = 1 and observation = 1]) is our classification accuracy.
## [1] 0.7The Horseshoe Prior
In the case when we have more variables than observations, it will be difficult to achieve good estimates of the coefficients. To address this hurdle, we may consider alternative priors on the \(\beta\)s which place higher prior density on 0, effectively saying that those predictors should not be included in our final model. The horseshoe prior (hs()) is one such “shrinkage” prior. We will not spend time on the horseshoe prior beyond this, so consider this a very brief introduction. If you would like to know more, let the instructor know. In this dataset we have many samples and few covariates, so the horseshoe is not necessary. However, we will examine its effect on posterior inference of the \(\beta\) coefficients.
p0 <- 2 # prior guess for the number of relevant variables
tau0 <- p0/(p-p0) * 1/sqrt(n) # recommended by Pilronen and Vehtari (2017)
hs_prior <- hs(df=1, global_df=1, global_scale=tau0)
post2 <- stan_glm(admit ~ ., data = admissions,
family = binomial(link = "logit"),
prior = hs_prior, prior_intercept = normal(0,1),
seed = seed,
refresh = 0)
round(coef(post2), 3)## (Intercept) gre gpa rank2 rank3 rank4
## -2.48 0.24 0.65 -0.20 -0.88 -1.07## 2.5% 97.5%
## (Intercept) -5.096 -0.192
## gre -0.004 0.511
## gpa -0.004 1.392
## rank2 -0.994 0.176
## rank3 -1.708 -0.059
## rank4 -2.041 -0.031## Warning: `expand_scale()` is deprecated; use `expansion()` instead.
- How does posterior inference for the coefficients compare to when we used the weakly informative Normal prior above?
- How do the two models compare in terms of predictive performance? Consider using the
loofunction as we have been doing.
##
## Computed from 4000 by 400 log-likelihood matrix
##
## Estimate SE
## elpd_loo -237.7 8.2
## p_loo 7.0 0.4
## looic 475.4 16.5
## ------
## Monte Carlo SE of elpd_loo is 0.1.
##
## All Pareto k estimates are good (k < 0.5).
## See help('pareto-k-diagnostic') for details.## elpd_diff se_diff
## post1 0.0 0.0
## post2 -2.4 1.0