A loan dataset has missing employment-duration values, and the fact that applicants leave this field blank may itself relate to approval outcomes. Which approach can preserve that potential signal while also supplying a numeric value for a model that requires complete inputs?