Conference Paper World Association for Public Opinion Research Annual Conference
Measurement Error in Recall Questions: a consolidated model
18 Sep 2003
The literature contains many studies regarding 'memory effects' in surveys. Most of these concern apparent decay in memory over time, as reflected in a reduction in the number of incidents reported, or in the accuracy of information reported, the longer the elapsed time between the event and the survey interview. Other studies look at apparent 'rounding,' 'censoring,' and 'telescoping' effects. A large majority of these studies are empirical in nature, merely documenting the nature of observed effect and, in some cases, discussing implications for survey design. There is little discussion of why these effects manifest themselves in a particular form. Progress with the theory of recall measurement error accelerated only in the 1980s, largely as a result of collaboration between psychologists and survey statisticians.
However, in all of this literature, most writers have addressed only one phenomenon at a time. In reality, it is likely that all of these effects are present, to a greater or lesser extent, in all survey recall data. In this paper, we attempt to summarise what is known about measurement error in recall data and to synthesise this into a single error model. By recall data, we refer to survey questions that ask respondents to recall specific events. This is often done in surveys of consumer behaviour, social interactions and so on. Such behavioural data are often used to help explain opinion data that may be collected in the same interview.
The consolidated theoretical model will be illustrated using data from the British Crime Survey, in which sample members are asked to recall crime incidents that they have experienced over a period of 12 to 20 months prior to the interview. A large-scale split-run experiment was run on the survey in 2001 and this allows key components of recall measurement error to be identified. Specifically, bounding and length-of-recall effects can be separated from other sources of measurement error.
The paper will describe the effect of these measurement errors on estimates of the number of incidents taking place per annum and will discuss implications for survey design. We will investigate the relationship between the salience of an incident and the rapidity of memory decay and also the extent of net forward telescoping due solely to random error in the recall of dates. We will also examine the effect of changing the nature of the reference period bound from a date which forms a natural memory bound for many people (1st January) to a less memorable date.
We will discuss the applicability of the proposed consolidated model of recall effects. The model makes explicit the factors that must be either held constant or estimated if a particular type of effect is to be studied.