Response Propensities in the Norwegian Labour Force Survey

With response rates decreasing in the Norwegian Labour Force Survey (LFS) it is important for Statistics Norway to investigate ways to optimize response within a constraint budget. This paper addresses optimal times and days for respondents and looks at how response propensities and response rates change throughout the data collection period. Additionally we observe the progression of R-indicators and a key variable: employment. The data comes from quarters 2, 3 and 4 in 2013, including over 66 000 cases.

The average (mean) case length was around 5 contact attempts while the median was only 2 indicating a skew in the case length data; a few people are followed-up many times. We observed that response propensities show a bi-modal distribution and drop during the data collection period; those that have not been able to be contacted are increasingly harder to reach. The last timeslot (19:00 to 21:30) appears to be the best time for initial contact attempts in order to achieve an interview.

Our study provides some basic ground work on when an appropriate time to swap to a new mode may be. Two of the ways we may consider the best point to change are in terms of phase capacity and for maximizing survey representativity. We see that response rates are not greatly improved once 15 contact attempts have been made. However, our key variable: employment, and the R-indicators and partial R-indicators stabilise before this; already after around 5 contact attempts. We believe this gives a good indication that a switch to a cheaper follow-up mode may be warranted already at this point.

One option could be to base the timing of the swap on a responsive design, whereby certain groups swap to the final cheaper mode when the response rates within the group are above a certain level. Similarly, groups could be routed to the more expensive (but more efficient at recruiting subjects to the survey) telephone follow-up sooner, if initial response rates within the group are below a certain level. Simple control charts for this could be developed with basis in our figures in chapter 6, preferably where the response or outcome influence on the estimate is shown. This since we through this basic analysis have shown that some background variables get more skewed after a certain number of contact attempts.

We would also like to stress that a measure or treatment in responsive design is not only about offering the respondent or groups of respondents different modes; it may well consist of e.g. different contact strategies, differentiated information about the survey and use of incentives.

We suggest taking the necessary steps to implement an adaptive design/responsive design for the Norwegian LFS in the very near future to make the best use of resources. Additionally, a future experiment on the initial day and time of first contact would allow us to investigate how to optimise response rates, and improve resource allocation.