.- help for ^bicdrop1^ - 1.0 - 6 Mar 2005 .-Estimate the probability a model is more likely without each explanatory variab > le .-

^bicdrop1^ [^, h^ighlight^(name)^ ]

Description -----------

^bicdrop1^ is a post-estimation command that uses the Bayesian Information Crit > erion (BIC) to estimate the probability that the model would be more likely after dro > pping one of the explanatory variables. The BIC was first proposed by Schwarz (1978) > and further developed by Raftery (1995).

It works after the following estimation commands: ^regress^, ^logistic^, ^logit^, ^ologit^, ^oprobit^, ^mlogit^, ^poisson^, ^nbre > g^.

It also reports Akaike's AIC, an earlier measure of model likelihood, and BIC' > (BIC prime), an alternative measure proposed by Raftery for model comparison.

The command drops each explanatory variable from the model and reports the AIC, > BIC and BIC' associated with the resulting nested model and uses the differences be > tween the BIC for the reduced model and the full (original) model to calculate a probability that the model is less likely if that explanatory (or independent) variable is removed.

Note that the BIC difference is not a traditional hypothesis test, but a compar > ison between the likelihood of two models: the original model and the model without > one of the variables. Nevertheless, the BIC difference is a more rigourous test of > whether the "true" model (i.e. the most likely model, given the data and the likelihood form) contains the variable in question, especially where the tradit > ional significance tests are weak: where N is large; or where there are a lot of explanatory variables (where k is large).

Acknowledgements ---------------- This program was based on the approach taken by the command lrdrop1 (developed > by Z. Wang) and was suggested by Richard Williams of the University of Notre Dame > . The author is grateful for assistance and encouragement from Richard Williams o > f the University of Notre Dame in the testing of this routine.

Options ------- ^h^ighlight(^"colour"^) highlights variables which are likely to be part of t > he model. "colour" can be any of "^w^hite" "^g^reen" "^y^ellow" "^r^ed" or "^s^peci > al"

Examples -------- . ^regress cbecs09 cprcs03 cprcs05 ^ . ^bicdrop1^

Author: Paul Millar www.ucalgary.ca/~pemillar/stata.htm pemillar@@ucalgary.ca See also: --------- Online: help for @lrdrop1@, @fitstat@ (if installed)

References: Raftery, Adrian (1995) Bayesian Model Selection in Social Research Sociological Methodology, Vol. 25, pp. 111- > 163. Schwarz, Gideon (1978) Estimating the Dimension of a Model The Annals of Statistics vol. 6, pp.461-64.