SemiParBIVProbit can be used to fit bivariate binary models where the linear predictors of the two main equations can
be flexibly specified using
parametric and regression spline components. Several bivariate copula distributions are supported. During the model fitting process, the
possible presence of associated error equations, endogeneity, non-random sample selection or partial observability is accounted for. Regression
spline bases are extracted from the package mgcv. Multi-dimensional smooths are available
via the use of penalized thin plate regression splines. Note that, if it makes sense, the dependence parameter of the employed bivariate
distribution can be specified as a function of covariates.
SemiParBIVProbit can also be used to fit bivariate models in which one response is binary and the other is continuous. This case
is relevant for estimating, for instance, the effect that a binary endogenous variable has on a continuous response or viceversa
or when interest is in modelling jointly binary and continuous responses.
There are many continuous distributions and copula functions that can be employed when using SemiParBIVProbit and we plan to include more
options. Other bivariate models can/will be implemented.
In the basic setup this will be a list of two formulas, one for equation 1 and the other for equation 2. s terms
are used to specify smooth smooth functions of
predictors. SemiParBIVProbit supports the use shrinkage smoothers for variable
selection purposes and more. See the examples below and the documentation of mgcv for further
details on formula specifications. Note that
if Model = "BSS" then the first formula MUST refer to the selection equation. Furthermore,
if it makes sense, a third equation for the dependence parameter can be specified (see Example 1 below).
When one outcome is binary and the other continuous then the first equation must refer to the binary outcome whereas
the second to the continuous one. For binary-continuous bivariate models, a maximum of three more equations can be
specified, one modelling
the scale parameter, the other the shape parameter (this would be the case
for "DAGUM" and "SM") and the last for the dependence parameter. The order matters. Examples are provided below.
data
An optional data frame, list or environment containing the variables in the model. If not found in data, the
variables are taken from environment(formula), typically the environment from which SemiParBIVProbit is called.
weights
Optional vector of prior weights to be used in fitting.
subset
Optional vector specifying a subset of observations to be used in the fitting process.
Model
It indicates the type of model to be used in the analysis. Possible values are "B" (bivariate model),
"BSS" (bivariate model with non-random sample selection), "BPO" (bivariate model with partial observability)
and "BPO0" (bivariate model with partial observability and zero correlation).
margins
It indicates the distributions used for the two margins. The first is one of "probit", "logit", "cloglog" which refer
to the link function of the first equation whose response is always assumed to be binary. The response
for the second equation can be binary ("probit"), normal ("N"),
log-normal ("LN"), Gumbel ("GU"), reverse Gumbel ("rGU"),
logistic ("LO"), Weibull ("WEI"), inverse Gaussian ("iG"), gamma ("GA"),
gamma with identity link for the location parameter ("GAi"), Dagum ("DAGUM"),
Singh-Maddala ("SM"), beta ("BE"), Fisk ("FISK", also known as log-logistic distribution).
gamlssfit
In the case of continuous margin for the outcome equation, if gamlssfit = TRUE then
a gamlss if fitted for the outcome equation. This is useful to calculate, for instance, the average
treatment effect from the model which does not account for endogeneity/associated equation errors.
This may also be used for obtaining better calibrated starting values, for instance.
BivD
Type of bivariate error distribution employed. Possible choices are "N", "C0", "C90", "C180", "C270", "J0", "J90", "J180", "J270",
"G0", "G90", "G180", "G270", "F", "AMH", "FGM" which stand for bivariate normal, Clayton, rotated Clayton (90 degrees), survival Clayton,
rotated Clayton (270 degrees), Joe, rotated Joe (90 degrees), survival Joe, rotated Joe (270 degrees),
Gumbel, rotated Gumbel (90 degrees), survival Gumbel, rotated Gumbel (270 degrees), Frank, Ali-Mikhail-Haq and
Farlie-Gumbel-Morgenstern. Note that Clayton, Joe and Gumbel are somewhat similar. Also,
there might be situations in which the use of a specific copula may result in more stable computations.
fp
If TRUE then a fully parametric model with unpenalised regression splines if fitted. See the example below.
hess
If FALSE then the expected/Fisher (rather than observed) information matrix is employed. The Fisher information matrix
is not available for cases different from binary treatment and binary outcome.
infl.fac
Inflation factor for the model degrees of freedom in the approximate AIC. Smoother models can be obtained setting
this parameter to a value greater than 1.
rinit
Starting trust region radius. The trust region radius is adjusted as the algorithm proceeds. See the documentation
of trust for further details.
rmax
Maximum allowed trust region radius. This may be set very large. If set small, the algorithm traces a steepest
descent path.
iterlimsp
A positive integer specifying the maximum number of loops to be performed before the smoothing parameter estimation
step is terminated.
tolsp
Tolerance to use in judging convergence of the algorithm when automatic smoothing parameter estimation is used.
gc.l
This is relevant when working with big datasets. If TRUE then the garbage collector is called more often than it is
usually done. This keeps the memory footprint down but it will slow down the routine.
parscale
The algorithm will operate as if optimizing objfun(x / parscale, ...) where parscale is a scalar. If missing then no
rescaling is done. See the
documentation of trust for more details.
extra.regI
If "t" then regularization as from trust is applied to the information matrix if needed.
If different from "t" then extra regularization is applied via the options "pC" (pivoted Choleski - this
will only work when the information matrix is semi-positive or positive definite) and "sED" (symmetric eigen-decomposition).
Details
The bivariate models considered in this package consist of two model equations which depend on flexible linear predictors and
whose association between the responses is modelled through parameter θ of a standardised bivariate normal
distribution or that of a bivariate copula distribution. The linear predictors of the two equations are flexibly specified using
parametric components and smooth functions of covariates. The same can be done for the dependence parameter if it makes sense.
Estimation is achieved within a penalized likelihood framework with integrated automatic multiple smoothing parameter selection. The use of
penalty matrices allows for the suppression of that part of smooth term complexity which has no support
from the data. The trade-off between smoothness
and fitness is controlled by smoothing parameters associated with the penalty matrices. Smoothing parameters are chosen to
minimise an approximate AIC.
Details of the underlying fitting methods are given in Radice, Marra and Wojtys (in press). Releases previous to 3.2-7
were based on the algorithms detailed in Marra and Radice (2011, 2013).
For sample selection models, if there are factors in the model, before fitting, the user has to ensure
that the numbers of factor variables' levels in the selected sample
are the same as those in the complete dataset. Even if a model could be fitted in such a situation,
the model may produce fits which are
not coherent with the nature of the correction sought. As an example consider the
situation in which the complete dataset contains a factor variable with five levels and that only three of them
appear in the selected sample. For the outcome equation (which is the one of interest) only three levels of such variable
exist in the population, but their effects will be corrected for non-random selection using a selection equation
in which five levels exist instead.
Having differing numbers of factors' levels between complete and selected samples will also
make prediction not feasible (an aspect which may be particularly important for selection models);
clearly it is not possible to predict the response of interest for the missing entries using a
dataset that contains all levels of a factor variable but using an outcome model
estimated using a subset of these levels.
Value
The function returns an object of class SemiParBIVProbit as described in SemiParBIVProbitObject.
WARNINGS
Convergence failure may sometimes occur. Convergence can be checked using conv.check which provides some
information about
the score and information matrix associated with the fitted model. The former should be 0 and the latter positive definite.
SemiParBIVProbit() will produce some warnings if there is a convergence issue.
In such a situation, the user may use some extra regularisation (see extra.regI) and/or
rescaling (see parscale). Using gamlssfit = TRUE (this is relevant
when one outcome is not binary) is typically more effective than the first two options as
this will provide better calibrated starting values as compared to those obtained from the default starting value procedure.
The default option is, however, gamlssfit = FALSE only because it tends to be computationally cheaper and because the
default starting value procedure has typically been found to do a satisfactory job in most cases.
(The results obtained when using
gamlssfit = FALSE and gamlssfit = TRUE could also be compared to check if starting values make any difference.)
The above suggestions may help, especially the latter option. However, the user should also consider
re-specifying the model and/or using a diferrent dependence structure and/or checking that the chosen marginal
distribution fit the responses well.
In our experience, we found that convergence failure typically occurs
when the model has been misspecified and/or the sample size (and/or number of selected observations in selection models) is low
compared to the complexity of the model. Examples
of misspecification include using a Clayton copula rotated by 90 degrees when a positive
association between the margins is present instead, using marginal distributions that do not fit
the responses (again, this is a bit more relevant when one of the two responses is continuous), and
employing a copula which does not accommodate the type and/or strength of
the dependence between the margins (e.g., using AMH when the association between the margins is strong).
It is also worth bearing in mind that the use of a three parameter marginal distribution requires the data
to be more informative than a situation in which a two parameter distribution is used instead.
When comparing competing models (for instance, by keeping the linear predictor specifications
fixed and changing the copula), if the computing time for a set of alternatives
is considerably higher than that of another set then it may mean that
the more computationally demanding models are not able to fit the data very well (as a higher number of
iterations is required to reach convergence). As a practical check, this may be verified by
fitting all competing models and, provided convergence is achieved, comparing their respective AIC and BICs, for instance.
In the contexts of endogeneity and non-random sample selection, extra attention is required when specifying
the dependence parameter as a function of covariates. This is because in these situations the dependence parameter mainly models the
association between the unobserved confounders in the two equations. Therefore, this option would make sense when it
is believed that the
strength of the association between the unobservables in the two equations varies based on some grouping factor or across geographical
areas, for instance.
Marra G. and Radice R. (2011), Estimation of a Semiparametric Recursive Bivariate Probit in the Presence of Endogeneity. Canadian
Journal of Statistics, 39(2), 259-279.
Marra G. and Radice R. (2013), A Penalized Likelihood Estimation Approach to Semiparametric Sample Selection Binary Response Modeling. Electronic Journal of Statistics, 7, 1432-1455.
Marra G., Radice R. and Missiroli S. (2014), Testing the Hypothesis of Absence of Unobserved Confounding in Semiparametric Bivariate Probit Models. Computational Statistics, 29(3-4), 715-741.
Marra G., Radice R. and Filippou P. (in press), Regression Spline Bivariate Probit Models: A Practical Approach to Testing for Exogeneity. Communications in Statistics - Simulation and Computation.
Marra G., Radice R., Barnighausen T., Wood S.N. and McGovern M.E. (submitted). A Simultaneous Equation Approach to Estimating HIV Prevalence with Non-Ignorable Missing Responses.
McGovern M.E., Barnighausen T., Marra G. and Radice R. (2015), On the Assumption of Joint Normality in Selection Models: A Copula Approach Applied to Estimating HIV Prevalence. Epidemiology, 26(2), 229-237.
Radice R., Marra G. and M. Wojtys (in press), Copula Regression Spline Models for Binary Outcomes. Statistics and Computing.
Marra G. and Wood S.N. (2011), Practical Variable Selection for Generalized Additive Models. Computational Statistics and Data Analysis, 55(7), 2372-2387.
Marra G. and Wood S.N. (2012), Coverage Properties of Confidence Intervals for Generalized Additive Model Components. Scandinavian Journal of Statistics, 39(1), 53-74.
Poirier D.J. (1980), Partial Observability in Bivariate Probit Models. Journal of Econometrics, 12, 209-217.
Wood S.N. (2004), Stable and efficient multiple smoothing parameter estimation for generalized additive models. Journal of the American Statistical Association, 99(467), 673-686.