4.2 Attrition

All surveys have non-response; repeated surveys have attrition: later non-response of earlier respondents. Cross-sectional surveys have very little information on who their non-responders are; panel surveys have lots of information on their attriters. We can model the attrition process.

Attrition is a binary outcome, therefore logistic regression is a suitable method. Focussing on a single pair of waves, attriters are defined as those present in the first and not in the second. A simple way to generate the data set for analysis is to match the waves on PID and use the merge variables (/in=... in SPSS, _merge in Stata) to define attrition (present last wave, absent this wave).

A more sophisticated approach would match XWAVEID to pINDRESP and use the value of qIVFIO, the interview outcome variable for the next wave (death, and movement out of scope have different process from refusal or non-tracing). However the outcome is defined, the independent variables must be drawn from pINDRESP:

answers at the second wave cannot be causes of the participation
we don't have second wave info for the attriters

Given a variable GONE indicating attrition, syntax like this will fit a logistic regression in SPSS:

logistic gone              /* identify outcome var     */
  /cat = ajbstat asex      /* Declare categorical vars */
  /method = enter ajbstat asex.

Subsections