A penalized likelihood approach for dealing with separation in count data regression model