Testes estatísticos em regressão logística sob a condição de separabilidade / Statistical tests in logistic regression under separability condition

AUTOR(ES)
DATA DE PUBLICAÇÃO

2010

RESUMO

Logistic regression is the statistical method of analysis used when the objective is to verify the relationship between one dichotomic response variable and explicative variables of interest. Usually, the model parameters are estimated through the genuine maximum likelihood method, and tests about these parameters are built assuming approximated distributions for the estimators. This means that large samples become necessary to obtain trustable results. In studies involving binary data is common the occurrence of one response variable whose success has low probability, in other words, a rare event that can generate a sparse data sample. In such cases, the data are under separability condition, and this situation is frequently associated to the presence of one categorical co-variable, what means that the maximum likelihood estimators do not exist to one parameter at least. In the separability condition it is recommended to use the Penalized Maximum Likelihood method, proposed by Firth (1993). The main objective of this study was to verify the powers of the Likelihood Ratio Test (LRT) and Wald Test obtained through PML under separability condition by Monte Carlo simulation. The presented methodology has been applied to two real data sets. Monte Carlo simulation with one explicative variable in the model made possible to obtain indicatives that the LRT is most powerful than the Wald test.

ASSUNTO(S)

wald test máxima verossimilhança penalizada simulation simulação ciencias agrarias teste de wald likelihood ratio test teste da razão de verossimilhança penalizmmed maximum likelihood

ACESSO AO ARTIGO

Documentos Relacionados