anna.bucharova
2009-Jul-13 09:37 UTC
[R] zero cells in one variable in logistic regression
Dear all. I am sort of beginner with R. I do logistic regression with binomial response variable and several continuous and categorical variables. In one categorical variable, zero cell occures (2x2 table looks like 7 - 0 23 - 25 This leads to overestimating of odds ratio and inflated confidence interval for odds for given variable. The variable is significant in univariate test. I do not necessarilly need odd ratio, but I need the explained deviance by this variable and I really want to keep this variable in the model. It probably matters for explained deviance. How to treat this problem? Thanks for help, Anna Bucharova -- View this message in context: http://www.nabble.com/zero-cells-in-one-variable-in-logistic-regression-tp24458629p24458629.html Sent from the R help mailing list archive at Nabble.com.
David Winsemius
2009-Jul-13 12:50 UTC
[R] zero cells in one variable in logistic regression
On Jul 13, 2009, at 5:37 AM, anna.bucharova wrote:> > Dear all. > I am sort of beginner with R. I do logistic regression with binomial > response variable and several continuous and categorical variables. > In one > categorical variable, zero cell occures (2x2 table looks like > 7 - 0 > 23 - 25 > This leads to overestimating of odds ratio and inflated confidence > interval > for odds for given variable. The variable is significant in > univariate test. > I do not necessarilly need odd ratio, but I need the explained > deviance by > this variable and I really want to keep this variable in the model. It > probably matters for explained deviance. How to treat this problem? > Thanks for help, Anna Bucharova > --You might consider glmrob in package:robustbase. See http://www.jstatsoft.org/v10/i04 David Winsemius, MD Heritage Laboratories West Hartford, CT
At 10:37 13/07/2009, anna.bucharova wrote:>Dear all. >I am sort of beginner with R. I do logistic regression with binomial >response variable and several continuous and categorical variables. In one >categorical variable, zero cell occures (2x2 table looks like >7 - 0 >23 - 25 >This leads to overestimating of odds ratio and inflated confidence interval >for odds for given variable. The variable is significant in univariate test. >I do not necessarilly need odd ratio, but I need the explained deviance by >this variable and I really want to keep this variable in the model. It >probably matters for explained deviance. How to treat this problem?Anna, you could consider brglm (from CRAN) which does give you finite estimates for the parameter. @ARTICLE{firth93, author = {Firth, D}, year = 1993, title = {Bias reduction of maximum likelihood estimates}, journal = {Biometrika}, volume = 80, pages = {27--38}, keywords = {glm} } gives the theory>Thanks for help, Anna Bucharova >-- >View this message in context: >http://www.nabble.com/zero-cells-in-one-variable-in-logistic-regression-tp24458629p24458629.html >Sent from the R help mailing list archive at Nabble.com.Michael Dewey http://www.aghmed.fsnet.co.uk
At 10:37 13/07/2009, anna.bucharova wrote:>Dear all. >I am sort of beginner with R. I do logistic regression with binomial >response variable and several continuous and categorical variables. In one >categorical variable, zero cell occures (2x2 table looks like >7 - 0 >23 - 25 >This leads to overestimating of odds ratio and inflated confidence interval >for odds for given variable. The variable is significant in univariate test. >I do not necessarilly need odd ratio, but I need the explained deviance by >this variable and I really want to keep this variable in the model. It >probably matters for explained deviance. How to treat this problem?Anna, you could consider brglm (from CRAN) which does give you finite estimates for the parameter. @ARTICLE{firth93, author = {Firth, D}, year = 1993, title = {Bias reduction of maximum likelihood estimates}, journal = {Biometrika}, volume = 80, pages = {27--38}, keywords = {glm} } gives the theory>Thanks for help, Anna Bucharova >-- >View this message in context: >http://www.nabble.com/zero-cells-in-one-variable-in-logistic-regression-tp24458629p24458629.html >Sent from the R help mailing list archive at Nabble.com.Michael Dewey http://www.aghmed.fsnet.co.uk