Can a measurement effect of survey mode be estimated from an AB/BA crossover test: the issue of non-ignorable non-response in the Time Use test survey?

Because it involves interviewing the same respondent twice, using a different collection method, an AB/BA crossover trial appears to be an appropriate tool for highlighting a measurement effect. However, the presence of a possible endogenous selection (non-ignorable non-response), as is the case for the Time Use test survey, biases the results obtained. It is nevertheless possible to discuss the existence of a measurement effect by comparing the responses given in the first interrogation by similar people surveyed using a different collection method. To do this, it is necessary to carry out a Rosenbaum sensitivity analysis on the conclusions obtained. With regard to the Time Use test survey, we found that (i) Internet data collection had an impact on the declared duration of sleep, meals and leisure activities, (ii) paper data collection had an impact on the declared duration of journeys, but in each case (iii) only for a limited number of respondents. These results, which are robust to a possible endogenous selection, confirm those obtained by the crossover analysis, which initially assumed ignorable nonresponse.