Multinomial Regression Hyperbolic Tangent (Tanh) Estimator Gauss-Newton Optimization
mGNtanh uses Gauss-Newton optimization to compute the
hyperbolic tangent (tanh) estimator for the overdispersed multinomial
regression model for grouped count data. This function is not meant
to be called directly by the user. It is called by
multinomRob, which constructs the various arguments.
mGNtanh(bstart, sigma2, resstart, Y, Ypos, Xarray, xvec, tvec, jacstack, itmax = 100, print.level = 0)
- Vector of starting values for the coefficient parameters.
- Value of the dispersion parameter (variance). The estimator does not update this value.
- Array of initial orthogonalized (but not standardized) residuals.
Matrix (observations by alternatives) of outcome counts.
Values must be nonnegative. Missing data (
NAvalues) are not allowed.
- Matrix indicating which elements of Y are counts to be analyzed (TRUE) and which are values to be skipped (FALSE). This allows the set of outcome alternatives to vary over observations.
- Array of regressors. dim(Xarray) = c(observations, parameters, alternatives).
- Matrix (parameters by alternatives) that represents the model structure. It has a 1 for an estimated parameter, an integer greater than 1 for an estimated parameter constrained equal to another estimated parameter (all parameters constrained to be equal to one another have the same integer value in xvec) and a 0 otherwise.
- Starting values for the regression coefficient parameters, as a matrix (parameters by alternatives). Parameters that are involved in equality constraints are repeated in tvec.
- Array of regressors used to facilitate computing the gradient and the Hessian matrix. dim(jacstack) = c(observations, unique parameters, alternatives).
- Maximum number of Gauss-Newton stages. Each stage does at most 100 Gauss-Newton steps.
- Specify 0 for minimal printing (error messages only) or 2 to print details about the tanh computations.
The tanh estimator is a redescending M-estimator. Given an estimate of the scale of the overdispersion, the tanh estimator estimates the coefficient parameters of the linear predictors of the multinomial regression model.
mGNtanh returns a list of 16 objects. The returned objects are:
The tanh coefficient estimates in matrix format. The matrix has one
column for each outcome alternative. The label for each row of the matrix
gives the names of the regressors to which the coefficient values in the row
apply. The regressor names in each label are separated by a forward
slash (/), and
NAis used to denote that no regressor is associated with the corresponding value in the matrix. The value 0 is used in the matrix to fill in for values that do not correspond to a regressor.
- A vector containing the tanh coefficient estimates.
Value of the dispersion parameter (variance). This is the value specified
in the argument
sigma2in the call to the function.
Vector of weights based on the tanh estimator's
psifunction for each observation.
Vector of values of the tanh estimator's
psifunction for each observation.
- The outer product of the gradient (expected information) divided by the moment estimate of the dispersion.
- The inverse of the Hessian matrix (observed formation).
- Sandwich estimate of the asymptotic covariance of the tanh coefficient estimates.
- Number of Gauss-Newton iterations.
0, no errors;
sum(w) < nobs*(ncats-1)/2(weights are too small); 32, Hessian not positive definite in the final Newton step.
List reporting final results of the Gauss-Newton optimization. Elements:
coefficients, vector of coefficient parameters (same as
coeffvecvalue in list returned by mGNtanh);
tvec, matrix of coefficient parameters (same as
coefficientsvalue in list returned by mGNtanh);
formation, inverse Hessian matrix;
score, score (or gradient element) matrix;
LLvals, list containing weighted (
LLvals$LL) and unweighted (
LLvals$LLu) log-likelihood values;
convflag, TRUE/FALSE convergence flag;
iters, number of iterations done in final Gauss-Newton stage;
posdef, TRUE if Hessian is positive definite.
- The tanh overdispersion parameter estimate, which is a weighted moment estimate of the dispersion: weighted mean sum of squared orthogonalized residuals (adjusted for effective sample size after weighting and degrees of freedom lost to estimated coefficients).
Ymatrix that was supplied as input, except modified by having done
Y[!Ypos] <- 0.
Yposmatrix that was supplied as input.
- The matrix of predicted probabilities for each category for each observation based on the coefficient estimates.
jacstackthat was supplied as an input argument.
Xarraythat was supplied as an input argument.
Walter R. Mebane, Jr. and Jasjeet Singh Sekhon. 2004. ``Robust Estimation and Outlier Detection for Overdispersed Multinomial Models of Count Data.'' American Journal of Political Science 48 (April): 391--410. http://sekhon.berkeley.edu/multinom.pdf
For additional documentation please visit http://sekhon.berkeley.edu/robust/.