Sensitivity bounds for use with flawed data

Data do not adhere to the usual statistical assumptions and applied work is made difficult by the need to cope with the myriad of problems that arise from flawed data. Prominent examples are data that contain missing values o r where variables are only available in categorised form. Standard solution s include omitting incomplete records and replacing missing or categorised observations by some representative values. In these and other cases of fla wed data, there is some information available on the potential range into w hich any particular observation could feasibly fall. We contend that there is considerable diagnostic value in exploiting this information to compute bounds for the coefficient estimates and related statistics such as t-ratio s. While one of the standard solutions may ultimately be used to produce a set of estimates, the bounds provide an indication of how sensitive these r esults are to the particular solution chosen. This approach is developed an d illustrated by way of several examples. "If the data were perfect, collected from well designed randomised experime nts, there would be hardly room for a separate field of econometrics". (C) 1999 IMACS/Elsevier Science B.V. All rights reserved.