Discussion Paper, School of Business & Economics: Economics 2015/25
Nonresponse in surveys may result in a distortion of the distribution of interest. In a panel survey the participation behavior in later waves is different from the participation behavior at the start. With register data that cover also the information for non-respondents one can observe a fade away of the distributional differences between the distribution of the full sample, including nonresponders, and the respondent sample, without the nonrespondents. The mechanics of this effect may be explained by a Markov chain model. Under suitable regularity conditions the distribution on the state space converges to the steady state distribution of the chain, which is independent from the starting distribution of the chain. Therefore the fade-away effect is considered here as the swing-in into the steady state distribution. An essential condition for the fade-away effect assumes the same tran- sition law for the responders and the nonresponders. Such a hypothesis is investigated here for the Finnish subsample of EU-SILC for the equival- ized household net-income. The income is grouped into income brackets which divides the starting sample into quintiles. This analysis is based on register information. For this analysis the null-hypothesis of equal transition behavior between income quintiles for responders and nonre- sponders cannot be rejected. This finding restates a result for Finland for the ECHP (European Community Household Panel). A second condition concerns the selectivity of panel attrition after wave one. Here panel attrition must not depend on the income state of the previous panel wave. The velocity of the swing-in into the steady state distribution depends on the stability to stay in the same income state. The stability may vary among the European countries. Therefore we investigated the transition matrices for 25 EU-SILC countries. We simulated 6 different pattern of nonresponse bias and investigated the fade-away effect across the waves 2006 to 2009. We found remarkable differences between these 25 coun- tries. Expressed by the relative bias, i.e. bias in 2009 divided by bias at start in 2006, we found a reduction down to 26 percent of the initial bias for Bulgaria (foremost reduction) up to 61 percent for Finland (least reduction). Our results vote for longer observation periods in rotation panels like EU-SILC.
Panel surveys nonresponse panel attrition Markov chains income mobility EU-SILC