Accuracy and Response Bias in Survey Research
An important kind of survey question attempts, for each respondent, to discriminate between two factual alternatives, often between the occurrence or nonoccurrence of a given event or between the existence or nonexistence of a given condition. For example, the Health Interview Survey may ask whether or not the respondent made a visit to a doctor in the past six months and whether or not the respondent has a certain chronic illness.
Such survey questions may be seen as members of a broad class of "diagnostic" devices that attempt to discern, though admittedly in an imperfect manner, some underlying truth about binary events or conditions. Other devices in this class attempt to reveal whether or not there is a lesion in a patient, a flaw in a metal structure, a malfunction in a processing plant, a lie in a suspect's statements, or a relevant document in a library. Some diagnostic devices look to the future and attempt to predict whether or not rain will fall, an applicant will perform satisfactorily, the market will go up, or a tax return should be audited. In the most general sense, these devices attempt to detect signals (e.g., a lesion, a malfunction, a relevant document) in a background of random events that mimic signals (e.g., a fuzzy radiograph, the multiple processes of an operating manufacturing plant, the many potentially relevant documents in the library) or to discriminate between two confusable alternatives that may be termed '"noise- alone" and "signal-plus-noise." Modern signal detection theory (SDT) has been advanced as the best way of analyzing and assessing their performance ( Swets and Pickett, 1982).
Specifically, SDT provides an analytical technique that separates two fundamental and independent aspects of performance in tasks of the sort exemplified and quantifies each aspect in a numerical index. The two aspects are: (1) discrimination capacity or accuracy