Given its longitudinal nature, the YAPS survey faces the inevitable problem of attrition. Of the 2,820 individuals first interviewed in 1999, 1,575 were successfully re-interviewed in 2009. This generated an attrition rate of 44 % over the 10 year period, which is similar to the rates typically observed in longitudinal surveys from other developed countries (Becketti et al. 1988; Abraham et al. 2006). The high non-response in the YAPS gives rise to concerns about the existence of an attrition bias. In what follows, first, the main characteristics at baseline of the people who attrit (are not re-interviewed in 2009) and who do not attrit are compared. Then, two main problems related to attrition are discussed: selection on migration, and selection on unobserved time-varying characteristics related to the changes in the dependent variables of the study.
At baseline, attritors have generally lower income,20
lower economic satisfaction, and less years of education, then the people who are interviewed in both 1999 and 2009. Attritors are also more likely to be male, young, and have Swedish background (Table 7
). The first series of characteristics related to income and education, stands in opposition to what has been observed in previous studies in both developing (Thomas et al. 2001
) and developed countries (Hausman and Wise 1979
; Becketti et al. 1988
), where attrition has been found to have a positive association with income and education levels. This difference is probably due to the specific design of the YAPS survey which targets young adults (ages 22–30 in 1999), and therefore has a high relative proportion of student respondents (characterized by low income) at the time of the first survey. Given that young people are more likely to leave the survey, a higher percentage of attritors would have not achieved their final levels of education in 1999, lowering the average education level of this group, as well as their income and economic satisfaction.
Comparison of the characteristics at baseline (1999) of surveyed people who consequently attrit (not interviewed in 2009) and do not attrit (interviewed in 2009)
The relationship between the birth cohort and attrition is similar to that observed in previous literature, with younger cohorts being more likely to attrit in subsequent interviews. The difference in the attrition rates of people with Swedish and non-Swedish background may be related to previous findings that early life experience and parent characteristics are related to attrition (Thomas et al. 2012). Interestingly, higher levels of attrition are not associated with more hours worked per week, as could be expected if busy people were less likely to be re-interviewed. Previous studies conducted with surveys from the United States have found that non-contact is in fact associated with longer work times, though the same did not hold for refusals, with refusal rates showing no association with work time (Abraham et al. 2006).
Attrition in the YAPS survey could represent a major problem if it was selective on migration given that the main focus of the present study is on comparisons of migrants and non-migrants. Past research has found that attrition in longitudinal surveys may, in fact, be selective on migration. This problem arises especially in the case of surveys performed in developing countries (Thomas et al. 2001, 2012), as in developed countries non-response rates in surveys are mostly associated with refusals as opposed to failure to contact the respondents. Still, Abraham et al. (2006) find that non-contact rates may also be high in developed countries, as documented by their observations about the American Time Use Survey.
The problem of attrition due to migration should be lessened in the YAPS due to the access of the employees of Statistics Sweden, who were in charge of the data collection, to the Swedish Register records. The Register consists of data collected by the Swedish Tax Agency and includes specific information about current place of residence for all individuals. Access to this information should potentially make the task of following migrants considerably easier than in countries with less precise demographic information on their inhabitants.
A comparison of non-contact versus refusal rates in the YAPS could be informative, as non-response associated with non-contact may be more related to trouble finding a person who has moved. Unfortunately, the YAPS survey was performed by mail, and no information of non-contact versus refusal rates was collected. Still, because attrition is generally associated with similar demographic characteristics across different surveys (Zabel 1998), a comparison of the characteristics of attritors in the YAPS to the characteristics of attritors due to non-contact in other surveys could provide insight into this problem. In developed countries such as the United States, non-contact is typically associated with being single, working longer hours, and being a high school graduate (Abraham et al. 2006). In the YAPS, the proportion of people married and the hours worked at baseline are not statistically different for attritors and non-attritors. Moreover, attritors have significantly less years of education, which is the opposite of the association between education and non-contact found by Abraham and co-authors. If the same associations between non-contact and demographic characteristics hold for Sweden as for United States, this could imply that a big proportion of attrition in the YAPS is due to refusal. Still, it is not clear that Swedish attrition should follow the same patterns as those observed in studies from other countries, and so the previous implication may be considered inconclusive.
An additional indirect test of selection on attrition used by previous literature consists of comparing characteristics of interest of the observed survey sample to those of a similar sample of the general population (Groves 2006
). Using this method, a test of attrition selective on migration in the YAPS is performed comparing rates of mobility by cohort of survey respondents interviewed in both years to those of the general population of Sweden (Table 8
). For every cohort, the mobility of the general population is slightly above that of the non-attritors from YAPS, with the difference between the two populations being highest for the 1976 cohort. For all cohorts combined, the difference in the migration proportions between the general population and the YAPS is 3 % (44 % for general population and 41 % for YAPS). This difference implies that, though selection on migration might have certainly taken place in the YAPS survey, the magnitude of this selection appears small.
Proportion of mobility by cohort: general population versus YAPS non-attritors
The second way in which attrition could bias the results is through selection on time-varying characteristics associated with either changes in life satisfaction or any of the other dependent variables used. To analyze this issue of selective attrition, the panel structure of the data is used to implement a test described by Wooldridge (2010). Selective attrition may represent a source of bias if it is correlated with the error term conditional on the explanatory variables (including the variable of interest, in this case migration). That is, in the model Yic = β’xi + ηi + εic the condition E(εic|xi, ai, ηi) = 0 (where ai represents attrition) must be satisfied to assure consistency of the parameter estimates (Wooldridge 2010). To test for this Wooldridge suggests adding to the equation a lead attrition indicator (that turns to one in the period before attrition) and testing for its significance using a standard t test. If the attrition indicator turns out to be not significant in this regression, that indicates that attrition should not represent a source of bias.
To implement the above method in the present analysis requires using the intermediate 2003 survey. Since some of the respondents who attrited between 1999 and 2009 still participated in the 2003 survey, using this data allows to estimate the first-difference model with the 99–03 variables adding future attrition (in the period 03–09) into the model. To assure robustness to the possible correlation between attrition and the explanatory variables, this test should control for all explanatory variables included in the original model (including migration). This rises a practical issue, as Register data on municipality of residence (used to determine migration) and on some of the control variables (such as education) is only available in the YAPS survey for the respondents who participated in the 2009 round, and is therefore missing for all attritors. Due to the absence of the Register data, migration may not be directly included into the model when estimating the first-difference regression using the 99–03 variables. Still, in an attempt to control for the correlation between attrition and migration, 1999 variables expected to determine the probability of future migration are included into the model testing for attrition bias.21
Additionally, since education completion is also unavailable (due to absence of Register data on education), a first difference in the student dummy is used to proxy for education completion. Estimating this full model attrition proves not to be a significant determinant of any of the dependent variables used in the analysis (Table 9
). This indicates that attrition is unlikely to be selective on the first difference variables used in the main analysis.
Test for attrition bias: OLS regressions of variables of interest (in 99–03 changes) regressed on attrition in 2009 and control variables
Finally, as a last test for selective attrition, a comparison may be carried out between the changes in a clue variable for the selected sample of respondents interviewed in both 1999 and 2009, and the changes in the same variable for the general population. This comparison is carried out for income changes (Table 10
). There are two main reasons to use income for this test. First, disposable income is readily available from the Statistics Sweden for both, the YAPS sample, and the general population. Second, attrition has been specifically found to be selective on changes in returns to human capital, such as education (Thomas et al. 2012
), which could possibly be reflected in changes in disposable income.
Mean disposable income (in hundreds of SEK) from Register, whole population (1968, 1972 and 1976 cohorts) and YAPS (non-attritors), by migration status, by year
For both migrants and non-migrants observed in the YAPS survey in 1999 and 2009, the changes in disposable income are slightly above those of the general population.22 Because the present study is based on the comparison of migrants versus non-migrants, one may be especially interested in comparing the difference in changes in income for these two groups for the YAPS sample and the general population. For the sample of non-attritors from YAPS, the difference between changes in income for migrants and non-migrants is 21,800 SEK; the difference between the migrant groups for the general population is 26,500 SEK (Table 10). The closeness between these two differences is reassuring.
Because of the high levels of attrition in the YAPS survey, concerns with possible bias may certainly arise. However, given the previous analysis selective attrition on migration, though possible, appears to be generally small in magnitude. The first-difference regression analysis used in the study allows to control for all time invariant unobserved characteristics that could be related to both attrition and the variables of interest. Though the possibility of time varying unobserved characteristics related to attrition remains, the tests performed (using first difference variables over 99–03 and a comparison of the changes in income for migrants and non-migrants for the YAPS sample and the general population) both provide results indicating that the first difference variables do not appear to be selective on attrition. In conclusion, the results of the analysis performed in this section provide reassurance that the possible attirition bias in the survey should not have a strong effect on the main results of the study.