The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.

Bias reduction in generalized linear models

Ioannis Kosmidis, Euloge Clovis Kenne Pagui

29 April 2017

The brglm2 package

brglm2 provides tools for the estimation and inference from generalized linear models using various methods for bias reduction or maximum penalized likelihood with powers of the Jeffreys prior as penalty. Reduction of estimation bias is achieved either through the mean-bias reducing adjusted score equations in Firth (1993) and I. Kosmidis and Firth (2009), or through the direct subtraction of an estimate of the bias of the maximum likelihood estimator from the maximum likelihood estimates as prescribed in Cordeiro and McCullagh (1991), or through the median-bias reducing adjusted score equations in Kenne Pagui, Salvan, and Sartori (2017).

In the special case of generalized linear models for binomial, Poisson and multinomial responses (both nominal and ordinal), mean and median bias reduction and maximum penalized likelihood return estimates with improved frequentist properties, that are also always finite, even in cases where the maximum likelihood estimates are infinite, like in complete and quasi-complete separation as defined in Albert and Anderson (1984).

The workhorse function is brglmFit(), which can be passed directly to the method argument of the glm function. brglmFit implements a quasi-Fisher scoring procedure, whose special cases result in various explicit and implicit bias reduction methods for generalized linear models (the classification of bias reduction methods into explicit and implicit is given in I. Kosmidis 2014).

This vignette

This vignette

Other resources

The bias-reducing quasi Fisher scoring iteration is also described in detail in the bias vignette of the enrichwith R package. I. Kosmidis and Firth (2010) describe a parallel quasi Newton-Raphson procedure.

Most of the material in this vignette comes from a presentation by Ioannis Kosmidis at the useR! 2016 international conference at University of Stanford on 16 June 2016. The presentation was titled “Reduced-bias inference in generalized linear models” and can be watched online at this link.

Generalized linear models

Model

Suppose that \(y_1, \ldots, y_n\) are observations on independent random variables \(Y_1, \ldots, Y_n\), each with probability density/mass function of the form \[ f_{Y_i}(y) = \exp\left\{\frac{y \theta_i - b(\theta_i) - c_1(y)}{\phi/m_i} - \frac{1}{2}a\left(-\frac{m_i}{\phi}\right) + c_2(y) \right\} \] for some sufficiently smooth functions \(b(.)\), \(c_1(.)\), \(a(.)\) and \(c_2(.)\), and fixed observation weights \(m_1, \ldots, m_n\). The expected value and the variance of \(Y_i\) are then \[\begin{align*} E(Y_i) & = \mu_i = b'(\theta_i) \\ Var(Y_i) & = \frac{\phi}{m_i}b''(\theta_i) = \frac{\phi}{m_i}V(\mu_i) \end{align*}\] Hence, in this parameterization, \(\phi\) is a dispersion parameter.

A generalized linear model links the mean \(\mu_i\) to a linear predictor \(\eta_i\) as \[ g(\mu_i) = \eta_i = \sum_{t=1}^p \beta_t x_{it} \] where \(g(.)\) is a monotone, sufficiently smooth link function, taking values on \(\Re\), \(x_{it}\) is the \((i,t)th\) component of a model matrix \(X\), and \(\beta = (\beta_1, \ldots, \beta_p)^\top\).

Score functions and information matrix

Suppressing the dependence of the various quantities on the model parameters and the data, the derivatives of the log-likelihood about \(\beta\) and \(\phi\) (score functions) are \[\begin{align*} s_\beta & = \frac{1}{\phi}X^TWD^{-1}(y - \mu) \\ s_\phi & = \frac{1}{2\phi^2}\sum_{i = 1}^n (q_i - \rho_i) \end{align*}\] with \(y = (y_1, \ldots, y_n)^\top\), \(\mu = (\mu_1, \ldots, \mu_n)^\top\), \(W = {\rm diag}\left\{w_1, \ldots, w_n\right\}\) and \(D = {\rm diag}\left\{d_1, \ldots, d_n\right\}\), where \(w_i = m_i d_i^2/v_i\) is the \(i\)th working weight, with \(d_i = d\mu_i/d\eta_i\) and \(v_i = V(\mu_i)\). Furthermore, \(q_i = -2 m_i \{y_i\theta_i - b(\theta_i) - c_1(y_i)\}\) and \(\rho_i = m_i a'_i\) with \(a'_i = a'(-m_i/\phi)\). The expected information matrix about \(\beta\) and \(\phi\) is \[ i = \left[ \begin{array}{cc} i_{\beta\beta} & 0_p \\ 0_p^\top & i_{\phi\phi} \end{array} \right] = \left[ \begin{array}{cc} \frac{1}{\phi} X^\top W X & 0_p \\ 0_p^\top & \frac{1}{2\phi^4}\sum_{i = 1}^n m_i^2 a''_i \end{array} \right]\,, \] where \(0_p\) is a \(p\)-vector of zeros, and \(a''_i = a''(-m_i/\phi)\).

Maximum likelihood estimation

The maximum likelihood estimators \(\hat\beta\) and \(\hat\phi\) of \(\beta\) and \(\phi\), respectively, can be found by the solution of the score equations \(s_\beta = 0_p\) and \(s_\phi = 0\).

Mean bias-reducing adjusted score functions

Let \(A_\beta = -i_{\beta\beta} b_\beta\) and \(A_\phi = -i_{\phi\phi} b_\phi\), where \(b_\beta\) and \(b_\phi\) are the first terms in the expansion of the mean bias of the maximum likelihood estimator of the regression parameters \(\beta\) and dispersion \(\phi\), respectively. The results in Firth (1993) can be used to show that the solution of the adjusted score equations \[\begin{align*} s_\beta + A_\beta & = 0_p \\ s_\phi + A_\phi & = 0 \end{align*}\] results in estimators \(\tilde\beta\) and \(\tilde\phi\) with bias of smaller asymptotic order than the maximum likelihood estimator.

The results in either I. Kosmidis and Firth (2009) or Cordeiro and McCullagh (1991) can then be used to re-express the adjustments in forms that are convenient for implementation. In particular, and after some algebra the bias-reducing adjustments for generalized linear models are \[\begin{align*} A_\beta & = X^\top W \xi \,, \\ A_\phi & = \frac{(p - 2)}{2\phi} + \frac{\sum_{i = 1}^n m_i^3 a'''_i}{2\phi^2\sum_{i = 1}^n m_i^2 a''_i} %A_\phi & = \frac{(p - 2)\phi\sum_{i = 1}^n m_i^2 % a''(-m_i/\phi) + \sum_{i = 1}^n m_i^3 % a'''(-m_i/\phi))}{2\phi^2\sum_{i = 1}^n m_i^2 % a''(-m_i/\phi)} \end{align*}\] where \(\xi = (\xi_1, \ldots, \xi_n)^T\) with \(\xi_i = h_id_i'/(2d_iw_i)\), \(d_i' = d^2\mu_i/d\eta_i^2\), \(a''_i = a''(-m_i/\phi)\), \(a'''_i = a'''(-m_i/\phi)\), and \(h_i\) is the “hat” value for the \(i\)th observation (see, e.g. ?hatvalues).

Median bias-reducing adjusted score functions

The results in Kenne Pagui, Salvan, and Sartori (2017) can be used to show that if \[\begin{align*} A_\beta & = X^\top W (\xi + X u) \\ A_\phi & = \frac{p}{2\phi}+\frac{ \sum_{i = 1}^n m_i^3 a'''_i}{6\phi^2\sum_{i = 1}^n m_i^2 a''_i} \, , \end{align*}\] then the solution of the adjusted score equations \(s_\beta + A_\beta = 0_p\) and \(s_\phi + A_\phi = 0\) results in estimators \(\tilde\beta\) and \(\tilde\phi\) with median bias of smaller asymptotic order than the maximum likelihood estimator. In the above expression, \(u = (u_1, \ldots, u_p)^\top\) with \[\begin{align*} u_j = [(X^\top W X)^{-1}]_{j}^\top X^\top \left[ \begin{array}{c} \tilde{h}_{j,1} \left\{d_1 v'_1 / (6 v_1) - d'_1/(2 d_1)\right\} \\ \vdots \\ \tilde{h}_{j,n} \left\{d_n v'_n / (6 v_n) - d'_n/(2 d_n)\right\} \end{array} \right] \end{align*}\] where \([A]_j\) denotes the \(j\)th row of matrix \(A\) as a column vector, \(v'_i = V'(\mu_i)\), and \(\tilde{h}_{j,i}\) is the \(i\)th diagonal element of \(X K_j X^T W\), with \(K_j = [(X^\top W X)^{-1}]_{j} [(X^\top W X)^{-1}]_{j}^\top / [(X^\top W X)^{-1}]_{jj}\).

Mixed adjustments

The results in Ioannis Kosmidis, Kenne Pagui, and Sartori (2020) can be used to show that if \[\begin{align*} A_\beta & = X^\top W \xi \,, \\ A_\phi & = \frac{p}{2\phi}+\frac{ \sum_{i = 1}^n m_i^3 a'''_i}{6\phi^2\sum_{i = 1}^n m_i^2 a''_i} \, , \end{align*}\] then the solution of the adjusted score equations \(s_\beta + A_\beta = 0_p\) and \(s_\phi + A_\phi = 0\) results in estimators \(\tilde\beta\) with mean bias of small asymptotic order than the maximum likelihood estimator and \(\tilde\phi\) with median bias of smaller asymptotic order than the maximum likelihood estimator.

Maximum penalized likelihood with powers of Jeffreys prior as penalty

The likelihood penalized by a power of the Jeffreys prior \[ |i_{\beta\beta}|^a |i_{\phi\phi}|^a \quad a > 0 \] can be maximized by solving the adjusted score equations \(s_\beta + A_\beta = 0_p\) and \(s_\phi + A_\phi = 0\) with \[\begin{align*} A_\beta & = X^\top W \rho \,, \\ A_\phi & = -\frac{p + 4}{2\phi}+\frac{ \sum_{i = 1}^n m_i^3 a'''_i}{6\phi^2\sum_{i = 1}^n m_i^2 a''_i} \, , \end{align*}\] where \(\rho = (\rho_1, \ldots, \rho_n)^T\) with \(\rho_i = h_i \{2 d_i'/(d_i w_i) - v_i' d_i/(v_i w_i)\}\).

Fitting algorithm in brglmFit

brglmFit() implements a quasi Fisher scoring procedure for solving the adjusted score equations \(s_\beta + A_\beta = 0_p\) and \(s_\phi + A_\phi = 0\). The iteration consists of an outer loop and an inner loop that implements step-halving. The algorithm is as follows:

Input

Output

Iteration

Initialize outer loop

  1. \(k \leftarrow 0\)

  2. \(\upsilon_\beta^{(0)} \leftarrow \left\{i_{\beta\beta}\left(\beta^{(0)}, \phi^{(0)}\right)\right\}^{-1} \left\{s_\beta\left(\beta^{(0)}, \phi^{(0)}\right) + A_\beta\left(\beta^{(0)}, \phi^{(0)}\right)\right\}\)

  3. \(\upsilon_\phi^{(0)} \leftarrow \left\{i_{\phi\phi}\left(\beta^{(0)}, \phi^{(0)}\right)\right\}^{-1} \left\{s_\phi\left(\beta^{(0)}, \phi^{(0)}\right) + A_\phi\left(\beta^{(0)}, \phi^{(0)}\right)\right\}\)

Initialize inner loop

  1. \(m \leftarrow 0\)

  2. \(b^{(m)} \leftarrow \beta^{(k)}\)

  3. \(f^{(m)} \leftarrow \phi^{(k)}\)

  4. \(v_\beta^{(m)} \leftarrow \upsilon_\beta^{(k)}\)

  5. \(v_\phi^{(m)} \leftarrow \upsilon_\phi^{(k)}\)

  6. \(d \leftarrow \left|\left|(v_\beta^{(m)}, v_\phi^{(m)})\right|\right|_\infty\)

Update parameters

  1. \(b^{(m + 1)} \leftarrow b^{(m)} + 2^{-m} v_\beta^{(m)}\)

  2. \(f^{(m + 1)} \leftarrow f^{(m)} + 2^{-m} v_\phi^{(m)}\)

Update direction

  1. \(v_\beta^{(m + 1)} \leftarrow \left\{i_{\beta\beta}\left(b^{(m + 1)}, f^{(m + 1)}\right)\right\}^{-1} \left\{s_\beta\left(b^{(m + 1)}, f^{(m + 1)}\right) + A_\beta\left(b^{(m + 1)}, f^{(m + 1)}\right)\right\}\)

  2. \(v_\phi^{(m + 1)} \leftarrow \left\{i_{\phi\phi}\left(b^{(m + 1)}, f^{(m + 1)}\right)\right\}^{-1} \left\{s_\phi\left(b^{(m + 1)}, f^{(m + 1)}\right) + A_\phi\left(b^{(m + 1)}, f^{(m + 1)}\right)\right\}\)

Continue or break halving within inner loop

  1. if \(m + 1 < M\) and \(\left|\left|(v_\beta^{(m + 1)}, v_\phi^{(m + 1)})\right|\right|_\infty > d\)

    14.1. \(m \leftarrow m + 1\)

    14.2. GO TO 10

  2. else

    15.1. \(\beta^{(k + 1)} \leftarrow b^{(m + 1)}\)

    15.2. \(\phi^{(k + 1)} \leftarrow f^{(m + 1)}\)

    15.3. \(\upsilon_\beta^{(k + 1)} \leftarrow v_b^{(m + 1)}\)

    15.4. \(\upsilon_\phi^{(k + 1)} \leftarrow v_f^{(m + 1)}\)

Continue or break outer loop

  1. if \(k + 1 < K\) and \(\left|\left|(\upsilon_\beta^{(k + 1)}, \upsilon_\phi^{(k + 1)})\right|\right|_\infty > \epsilon\)

    16.1 \(k \leftarrow k + 1\)

    16.2. GO TO 4

  2. else

    17.1. \(\tilde\beta \leftarrow \beta^{(k + 1)}\)

    17.2. \(\tilde\phi \leftarrow \phi^{(k + 1)}\)

    17.3. STOP

Notes

Contributions to this vignette

The first version of the vignette has been written by Ioannis Kosmidis. Eugene Clovis Kenne Pagui and Nicola Sartori contributed the first version of the section “Median bias-reducing adjusted score functions”, and Ioannis Kosmidis brought the expressions for the median bias-reducing adjustments in the reduced form that is shown above and is implemented in brglmFit().

Ioannis Kosmidis, Kenne Pagui, and Sartori (2020) provides more details about mean and median bias reduction in generalized linear models.

Citation

If you found this vignette or brglm2, in general, useful, please consider citing brglm2 and the associated paper. You can find information on how to do this by typing citation("brglm2").

References

Albert, A., and J. A. Anderson. 1984. “On the Existence of Maximum Likelihood Estimates in Logistic Regression Models.” Biometrika 71 (1): 1–10.
Cordeiro, G. M., and P. McCullagh. 1991. “Bias Correction in Generalized Linear Models.” Journal of the Royal Statistical Society, Series B: Methodological 53 (3): 629–43.
Firth, D. 1993. “Bias Reduction of Maximum Likelihood Estimates.” Biometrika 80 (1): 27–38.
Kenne Pagui, E. C., A. Salvan, and N. Sartori. 2017. “Median Bias Reduction of Maximum Likelihood Estimates.” Biometrika 104: 923–38. https://doi.org/10.1093/biomet/asx046.
Kosmidis, I. 2014. “Bias in Parametric Estimation: Reduction and Useful Side-Effects.” Wiley Interdisciplinary Reviews: Computational Statistics 6 (3): 185–96. https://doi.org/10.1002/wics.1296.
Kosmidis, I., and D. Firth. 2009. “Bias Reduction in Exponential Family Nonlinear Models.” Biometrika 96 (4): 793–804. https://doi.org/10.1093/biomet/asp055.
———. 2010. “A Generic Algorithm for Reducing Bias in Parametric Estimation.” Electronic Journal of Statistics 4: 1097–1112. https://doi.org/10.1214/10-EJS579.
Kosmidis, Ioannis, Euloge Clovis Kenne Pagui, and Nicola Sartori. 2020. “Mean and Median Bias Reduction in Generalized Linear Models.” Statistics and Computing 30: 43–59. https://doi.org/10.1007/s11222-019-09860-6.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.