Data-informed priors: unit-information, empirical Bayes and power priors
Source:R/empirical_bayes.R
empirical_bayes_priors.RdDerives priors for the intercept and regression coefficients of a generalised linear model from data, with one of three published methods. Each avoids "double-dipping": centring a tight prior on estimates from the same data that will be fitted counts the data twice and makes the posterior overconfident.
Arguments
- data
A data frame. For
method = "power", the historical data. Rows with missing values in the variables offormulaare dropped (with a message); other columns are ignored.- formula
Model formula, identical to the one you will fit.
- family
Model family (object, function or name), identical to the one you will fit.
- method
One of
"unit_information","eb_shrinkage"or"power".- type
Prior family for the output:
"normal","student_t"or"cauchy".- a0
Power-prior discount in
(0, 1](only formethod = "power").- df
Degrees of freedom when
type = "student_t".- weights
Optional weights for the maximum-likelihood fit: a numeric vector, or the name of a column of
data(unquoted, as instats::glm()). Use the same weights as in the model you will fit.- dist_types
Deprecated (bmbeR 1.x). rstanarm requires a single prior family for all coefficients; use
type.
Value
A prior_config() object (class bmb_prior_config) that can be
passed to fit_model_with_prior(), with attributes table (a data frame
of estimates, standard errors and prior parameters), method, terms,
n (the sample size used by the unit-information prior), for
"eb_shrinkage" tau, and for "power" a0.
Details
A maximum-likelihood GLM with the same formula and family is fitted first.
Coefficients are on the link scale of family, so the priors are directly
usable in fit_model_with_prior(). The intercept prior is placed on the
intercept at the predictor means, which is how rstanarm parameterises the
model.
method = "unit_information"(default; Kass & Wasserman, 1995): each prior is centred at the estimate with variancentimes the sampling variance, i.e. the prior carries the information of a single observation. It is the prior implicit in BIC and is suitable when the same data are used to fit the model, because it adds only1/nof the data's information.ncounts Bernoulli trials for a binomial response given ascbind(successes, failures), and is the sum ofweightswhen weights are given (they are treated as frequency weights, or as the number of trials for proportions).method = "eb_shrinkage"(parametric empirical Bayes; Efron & Morris, 1973; Morris, 1983): coefficients are placed on a common scale by multiplying by the predictor standard deviations, and a shared prior standard deviationtauis estimated by maximising the marginal likelihood ofb_j ~ N(0, tau^2 + se_j^2). Slopes then receive zero-centred priors with scaletau / sd(x_j), which shrink noisy estimates towards zero. Requires at least three slopes. When the estimate oftaufalls below the median standard error it is floored at that value (with a warning), because a boundary estimate of zero would give a degenerate prior. The intercept uses the unit-information prior.method = "power"(Ibrahim & Chen, 2000):datamust be historical or external data, not the data you will analyse. The prior is the normal approximation to the historical likelihood raised to the powera0: centred at the historical estimates, with variances inflated by1 / a0.a0 = 1borrows the historical data at full weight; smaller values discount it. The historical predictor means usually differ from those of the analysed data, sofit_model_with_prior()andprior_predictive_check()re-express the intercept prior at the analysed data's predictor means before fitting.
With type = "student_t" or "cauchy", the same locations and scales are
used with heavier tails, which lets the data override the prior if the two
conflict (O'Hagan & Pericchi, 2012).
References
Kass, R. E., & Wasserman, L. (1995). A reference Bayesian test for nested hypotheses and its relationship to the Schwarz criterion. JASA, 90(431), 928–934. doi:10.1080/01621459.1995.10476592
Efron, B., & Morris, C. (1973). Stein's estimation rule and its competitors—an empirical Bayes approach. JASA, 68(341), 117–130. doi:10.1080/01621459.1973.10481350
Morris, C. N. (1983). Parametric empirical Bayes inference: Theory and applications. JASA, 78(381), 47–55. doi:10.1080/01621459.1983.10477920
Ibrahim, J. G., & Chen, M.-H. (2000). Power prior distributions for regression models. Statistical Science, 15(1), 46–60. doi:10.1214/ss/1009212673
O'Hagan, A., & Pericchi, L. (2012). Bayesian heavy-tailed models and conflict resolution: A review. Brazilian Journal of Probability and Statistics, 26(4), 372–401. doi:10.1214/11-BJPS164
Examples
data(kidiq, package = "rstanarm")
empirical_bayes_priors(kidiq, kid_score ~ mom_iq + mom_hs)
#> <bmb_prior_config> data-informed priors
#> method : unit-information prior (Kass & Wasserman, 1995)
#> family : gaussian (link scale); n = 434
#> type : normal
#>
#> term estimate std_error prior_location prior_scale
#> (Intercept) [centred] 86.800 0.8710 86.800 18.10
#> mom_iq 0.564 0.0606 0.564 1.26
#> mom_hs 5.950 2.2100 5.950 46.10
# Power prior from a (here simulated) historical study, discounted by half
historical <- kidiq[sample(nrow(kidiq), 200), ]
empirical_bayes_priors(historical, kid_score ~ mom_iq + mom_hs,
method = "power", a0 = 0.5)
#> method = "power": `data` is treated as historical data. Do not fit the resulting priors to the same data.
#> <bmb_prior_config> data-informed priors
#> method : power prior from historical data, a0 = 0.5 (Ibrahim & Chen, 2000)
#> family : gaussian (link scale); n = 200
#> type : normal
#>
#> term estimate std_error prior_location prior_scale
#> (Intercept) [centred] 85.600 1.29 85.600 1.820
#> mom_iq 0.614 0.09 0.614 0.127
#> mom_hs 5.460 3.18 5.460 4.500