This vignette defines the models and historical borrowing metrics
supported in the historicalborrow package.
Each primary model is parameterized thus:
\[ \begin{aligned} E(y) = X_\alpha \alpha + X_\delta \delta + X_\beta \beta \end{aligned} \]
Above, \(X_\alpha\), \(X_\delta\), and \(X_\beta\) are fixed matrices. \(X_\beta\) is a conventional model matrix for the baseline covariates \(\beta\), and the details are explained in the “Baseline covariates” section below. \(X_\alpha\) is a matrix of zeroes and ones. It is constructed such that each scalar component of \(\alpha\) is the mean response of the control group in a particular study. Likewise, \(X_\delta\) is a matrix of zeroes and ones such that each scalar component of \(\delta\) is the mean response of a non-control treatment group in a particular study.
To illustrate, let \(y_{ijk}\) be patient \(i\) in treatment group \(j\) (where \(j = 1\) is the control group) of study \(k\), and let \(\left ( X_\beta \beta \right )_{ijk}\) be the corresponding scalar element of the vector \(X_\beta \beta\). Then,
\[ \begin{aligned} E(y_{ijk}) = I (j = 1) \alpha_{k} + I (j > 1) \delta_{jk} + \left ( X_\beta \beta \right )_{ijk} \end{aligned} \]
This parameterization is represented in the more compact expression \(X_\alpha \alpha + X_\delta \delta + X_\beta \beta\) in the model definitions in this vignette.
The baseline covariates model matrix \(X_\beta\) adjusts for baseline covariates.
It may contain a continuous column for baseline and binary indicator
columns for the levels of user-defined covariates. All these columns are
included if possible, but the method automatically drops baseline
covariate columns to ensure that the combined model matrix \(X_i^* = \left [ {X_\alpha}^* \quad {X_\delta}^*
\quad {X_\beta}^* \right ]_i\) is full rank. (Here, \(X_i^*\) denotes the rows of matrix \(X\) corresponding to study \(i\), with additional rows dropped if the
corresponding elements of \(y\) are
missing. The additional row-dropping based on the missingness of \(y\) ensures identifiability even when the
user supplies complicated many-leveled factors as covariates.) The
choice of columns to drop from \({X_\beta}_i\) is determined by the rank and
pivoting strategy of the QR decomposition of \(X_i\) using the Householder algorithm with
pivoting (base::qr(), LINPACK routine DQRDC).
Separately within each study, each column of \(X_\beta\) is centered to have mean 0, and if possible, scaled to have variance 1. Scaling ensures that the priors on parameters \(\beta\) remain relatively diffuse relative to the input data. Study-level centering ensures that the \(\alpha\) parameters truly act as unconditional study-specific control group means (as opposed to conditional on the subset of patients at the reference level of \(X_\beta\)), and it ensures that borrowing across \(\alpha\) components fully presents as control group borrowing.
The hb_summary() function post-processes the results
from the model. It accepts MCMC samples of parameters and returns
estimated marginal means of the response and treatment effect. To
estimate marginal means of the response, hb_summary() takes
group-level averages of posterior samples of fitted values while
dropping covariate adjustment terms from the model (i.e. \(X_\alpha \alpha + X_\delta \delta\)).
Because the columns of \(X_\beta\) are
centered at their means, this choice is mathematically equivalent to
emmeans::emmeans() with the
weights = "proportional" (Lenth
(2016)).
Functions:
hb_sim_mixture()hb_mcmc_mixture()The mixture model analyzes only the data from the current study, so we use \(X_\alpha^{\text{mixture}}\) instead of \(X_\alpha\). \(X_\alpha^{\text{mixture}}\) is a one-column matrix to indicate which elements of \(y\) are part of the control group of the current study.
The historical studies contribute to the model through hyperparameters \(({m_\omega})_i\) and \((s_\omega)_i\). If study \(i\) is a historical study, \(({m_\omega})_i\) and \((s_\omega)_i\) are the posterior mean and posterior standard deviation, respectively, of the mean control group response estimated from the simple model described later. If study \(i\) is the current study, \(({m_\omega})_i\) and \((s_\omega)_i\) are chosen so the mixture component Normal(\(({m_\omega})_i\), \((s_\omega)_i\)) of study \(i\) is diffuse and non-informative. Variable \(\omega_i\) of study \(i\) is the latent variable of mixture component \(i\), and the index variable \(\pi\) chooses which \(\omega_i\) to use for the current study control group mean \(\alpha\). Hyperparameter \(p_\omega\) is a constant vector of prior mixture proportions of each study. The posterior histogram of \(\pi\) gives the posterior mixture proportions.
\[ \begin{aligned} & y_{ij} \stackrel{\text{ind}}{\sim} \text{N} \left ( \left (X_\alpha^{\text{mixture}} \alpha + X_\delta \delta + X_\beta \beta \right )_{ij}, \ \sigma^2 \right) \\ & \qquad \alpha = \omega_{\pi} \\ & \qquad \qquad \omega_i \stackrel{\text{ind}}{\sim} \text{Normal}(({m_\omega})_i, (s_\omega)_i^2) \\ & \qquad \qquad \pi \sim \text{Categorical}(p_\omega) \\ & \qquad \delta_d \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\delta^2) \\ & \qquad \beta_b \sim \text{Normal} (0, s_\beta^2) \\ & \qquad \sigma \sim \text{Uniform}(0, s_\sigma) \end{aligned} \]
Functions:
hb_sim_hierarchical()hb_mcmc_hierarchical()hb_s_tau()The hierarchical model is equivalent to the meta-analytic combined (MAC) approach analyzes the data from all studies and shrinks the control group means \(\alpha_i\) towards a common normal distribution with mean \(\mu\) and variance \(\tau^2\).
\[ \begin{aligned} & y_{ij} \sim \text{N} \left( \left (X_\alpha \alpha + X_\delta \delta + X_\beta \beta \right )_{ij} , \ \sigma_i^2 \right )\\ & \qquad \alpha_i \stackrel{\text{ind}}{\sim} \text{Normal}(\mu, \tau^2) \\ & \qquad \qquad \mu \sim \text{Normal}(0, s_\mu^2) \\ & \qquad \qquad \tau \sim f_\tau \\ & \qquad \delta_d \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\delta^2) \\ & \qquad \beta_b \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\beta^2) \\ & \qquad \sigma_i \stackrel{\text{ind}}{\sim} \text{Uniform}(0, s_\sigma) \end{aligned} \]
The prior \(f_\tau\) on \(\tau\) is critically important because:
\(f_\tau\) can either be a flexible half-Student-t distribution with \(d_\tau\) degrees of freedom and scale parameter \(s_\tau\):
\[ f_\tau = \text{Student-t}(0, s_\tau, d_\tau)^+ \] or a uniform distribution with lower bound 0 and upper bound \(s_\tau\):
\[ f_\tau = \text{Uniform}(0, s_\tau) \]
Following the recommendation of Gelman (2006), please use half-Student-t if the number of historical studies is small and consider uniform for large numbers of historical studies.
For the half-Student-t distribution, the role of the \(s_\tau\) parameter is equivalent to the \(\sigma\) parameter from the Student-t parameterization in the Stan user manual.
Functions:
hb_sim_independent()hb_mcmc_independent()The independent model is the same as the hierarchical model, but with independent control group parameters \(\alpha\). We use it as a no-borrowing benchmark to quantify the borrowing strength of the hierarchical model and the mixture model.
\[ \begin{aligned} & y_{ij} \sim \text{N} \left( \left (X_\alpha \alpha + X_\delta \delta + X_\beta \beta \right )_{ij} , \ \sigma_i^2 \right )\\ & \qquad \alpha_i \stackrel{\text{ind}}{\sim} \text{Normal}(0, s_\alpha^2) \\ & \qquad \delta_d \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\delta^2) \\ & \qquad \beta_b \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\beta^2) \\ & \qquad \sigma_i \stackrel{\text{ind}}{\sim} \text{Uniform}(0, s_\sigma) \end{aligned} \]
Functions:
hb_sim_pool()hb_mcmc_pool()Like the independent model, the pooled model is a benchmark to quantify the borrowing strength of the hierarchical model and the mixture model. But instead of the no-borrowing independent model, the pooled model represents maximum borrowing. Instead of \(X_\alpha\), below, we use \(X_\alpha^{\text{pool}}\), which has only one column to indicate which observations belong to any control group. In other words, the \(\alpha\) parameters are pooled, and \(\alpha\) itself is a scalar.
\[ \begin{aligned} & y_{ij} \sim \text{N} \left( \left ( X_\alpha^{\text{pool}} \alpha + X_\delta \delta + X_\beta \beta \right )_{ij} , \ \sigma_i^2 \right )\\ & \qquad \alpha \sim \text{Normal}(0, s_\alpha^2) \\ & \qquad \delta_d \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\delta^2) \\ & \qquad \beta_b \stackrel{\text{ind}}{\sim} \text{Normal} (0, s_\beta^2) \\ & \qquad \sigma_i \stackrel{\text{ind}}{\sim} \text{Uniform}(0, s_\sigma) \end{aligned} \]
Functions:
hb_mcmc_mixture_hyperparameters()The mixture model hyperparameters \((m_\omega)_i\) and \((s_\omega)_i\) of study \(i\) are obtained by analyzing the control group data of study \(i\) with the simple model below. \((m_\omega)_i\) and \((s_\omega)_i\) are taken to be the estimated posterior mean and posterior standard deviation, respectively, of \(\mu\) from this model.
\[ \begin{aligned} &y \sim \text{Normal}(\mu, \sigma^2) \\ & \qquad \mu \sim \text{Normal}(0, s_\mu^2) \\ & \qquad \sigma \sim \text{Uniform}(0, s_\sigma) \end{aligned} \]
The package supports the following metrics to quantify borrowing.
See the hb_ess() function for an implementation.
Neuenschwander et al. (2006) posit a
prior effective sample size metric for meta-analytic predictive (MAP)
priors. In the original paper, the underlying hierarchical model only
uses historical controls, and the hypothetical new study is the current
study of interest. In historicalborrow, we adapt this
metric to a hierarchical model which also includes both control and
treatment data from the current study. We still define \(N\) below to be the number of (non-missing)
historical control patients so we can still interpret ESS on the same
scale as in the paper.
For the pooled model, define \(V_0\) to be the posterior predictive variance of the control mean \(\alpha^*\) of a hypothetical new unobserved study. According to Neuenschwander et al. (2006), it can be derived as an average of study-specific variances. In practice, we estimate \(V_0\) using the average of MCMC samples of \(\frac{1}{\sum \sigma_i^{-2}}\).
\[ V_0 := \text{Var}(\alpha^* | y, \tau = 0) = \frac{1}{\sum \sigma_i^{-2}} \]
For the hierarchical model, we define the analogous posterior predictive variance \(V_\tau\) using the prior distribution.
\[ V_\tau := \text{Var}(\alpha^* | y) = \int E[(\alpha^* - E(\alpha^*|y))^2 | y] \cdot p(\alpha^* | \mu, \tau) \cdot p(\mu, \tau | y) d\mu d\tau \]
The above integral implies a straightforward method of estimating \(V_\tau\) using MCMC samples:
Next, define \(N\) as the number of non-missing control patients from the historical studies only. Given \(N\), \(V_0\), and \(V_\tau\), define the effective sample size as:
\[ \text{ESS} := N \frac{V_0}{V_\tau} \]
\(\frac{V_0}{V_\tau}\) is a weight which quantifies the fraction of historical information that the hierarchical model leverages for borrowing. Notably, the weight should be 1 if the hierarchical and pooled model exhibit the same strength of borrowing. Multiplied by \(N\), the quantity becomes a heuristic for the strength of borrowing of the hierarchical model, measured in terms of the number of historical patients.
The precision ratio is an experimental ad hoc metric and should be
used with caution. It is implemented in the hb_summary()
function for the hierarchical model.
The precision ratio compares the prior precision of a control mean response (an \(\alpha\) component, numerator) to the analogous precision of the full conditional distribution (denominator). The former is \(\frac{1}{\tau^2}\), and the latter is \(\frac{1}{\tau^2} + \frac{n}{\sigma^2}\). Here, \(n\) is the number of non-missing patients in the current study, \(\sigma^2\) is the residual variance, and \(\tau^2\) is the variance of study-specific control means (components of \(\alpha\)). The full precision ratio is:
\[ \begin{aligned} \frac{\frac{1}{\tau^2}}{\frac{1}{\tau^2} + \frac{n}{\sigma^2}} \end{aligned} \]
The precision ratio comes from the conditional distribution of \(\alpha_k\) in the hierarchical model given the other parameters and the data. More precisely, in this conditional distribution, the mean is a weighted average between the prior mean and data mean, and the precision ratio is the weight on the prior mean. This can be seen in a simpler case with a Bayesian model with a normal data model, a normal prior on the mean, and known constant variance. For details, see Chapter 2 of Gelman et al. (2020).
The variance shift ratio is an experimental ad hoc metric and should
be used with caution. It is implemented in the legacy
hb_metrics() function.
Let \(V_m\) be the estimated posterior variance of \(\alpha_I\) (current study control group response mean) estimated by model \(m\). The variance shift ratio is:
\[ \begin{aligned} \frac{V_{m*} - V_{\text{independent}}}{V_{\text{pool}} - V_{\text{independent}}} \end{aligned} \]
where \(m*\) is a historical borrowing model like the mixture model or hierarchical model.
The mean shift ratio is not recommended to measure the strength of
borrowing. Rather, it is an informal ad hoc measure of the lack of
commensurability between the current and historical data sources. It is
implemented in the legacy hb_metrics() function.
To define the mean shift ratio, let \(\theta_m\) be the posterior mean control group response estimated by model \(m\). The mean shift ratio is:
\[ \begin{aligned} \frac{\theta_{m*} - \theta_{\text{independent}}}{\theta_{\text{pool}} - \theta_{\text{independent}}} \end{aligned} \]
where \(m*\) is a historical borrowing model like the mixture model or hierarchical model.
The posterior mixture proportion of study \(i\) is \(P(\pi = i)\), and it is obtained by averaging posterior samples of \(\pi\).