Sicherheit älterer Menschen im Wohnquartier (SENSIKO) Projektberichte / Nr. 3 Heleen Janssen & Dominik Gerstner An attrition analysis in the SENSIKO survey (waves 1 and 2) Freiburg 2016 SENSIKO Working Paper / 3 istock.com/chinaface istock.com/chinaface istock.com/chinaface
Berichte aus dem Projekt Sicherheit älterer Menschen im Wohnquartier (SENSIKO). Analysen und Konzeption des Praxismodells Seniorensicherheitskoordination Verbundprojekt gemeinsam mit der Technischen Hochschule Köln, Forschungsschwerpunkt Sozial Raum Management, gefördert 2013-2016 vom Bundesministerium für Bildung und Forschung im Rahmen des Programms "Forschung für die zivile Sicherheit Heleen Janssen & Dominik Gerstner An attrition analysis in the SENSIKO survey (waves 1 and 2) weitere Informationen: https://www.mpicc.de/de/forschung/forschungsarbeit/kriminologie/sensiko.html Max-Planck-Institut für ausländisches und internationales Strafrecht Abteilung Kriminologie Günterstalstrasse 73, D-79100 Freiburg i.br. (Germany) https://www.mpicc.de/
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -1- Table of Contents 1 Introduction... 2 2 Descriptives of response behavior... 2 3 Results of multivariate analyses... 3 Demographics... 3 Socio-economic status... 4 Neighborhood characteristics... 5 Other factors... 7 4 Differences by age... 11 5 Conclusion... 13 6 References... 14
- 2 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO 1 Introduction Attrition the loss of respondents over subsequent waves or measurement points - is one of the major methodological problems in longitudinal studies. Selective attrition of sub-populations can lead to biased estimates of association and reduce the generalizability of findings. Attrition bias can occur when those who drop out of a panel are systematically different from those who continue to participate (e.g. Lynn, 2009). This report details the analysis of socio-demographics and psychological factors as predictors of attrition in the survey Zusammenleben und Sicherheit in Köln [Essen] conducted by the Max Planck Institute in 2014 and 2015. In addition, descriptive statistics describing the T1 and T2 sample are compared. We refer to Gerstner and Oberwittler (2016) for basic documentation on this survey. 2 Descriptives of response behavior In total, 6565 respondents participated in T1. The main reason why respondents did not participate in T2 is because they indicated in the T1 questionnaire that they did not want to be contacted again (n=1438). According to data protection regulation, addresses of respondents could only be kept after T1 with their active consent which we asked for as the final question in T1. Of the 5127 respondents who approved to be contacted again, 385 could not be sent a questionnaire because the respondent died (n=76), was too ill to participate (n=17), moved outside the study area (n=269), or because the questionnaire was undeliverable for unknown reasons (n=23). In 345 cases, a different person from the household returned the questionnaire. For the remaining cases (n=996) we do not know why the respondent did not participate again; a questionnaire was sent, but not returned. Table 1. Response SENSIKO Study n % of T1 % of could Participated in T1 6565 Ready to be contacted again in T2 5127 75 Could participate in T2 4742 72 Participated in T1 & T2 3401 52 72 Table 2. Dropout SENSIKO Study n % of T1 Did not want to be contacted again 1438 22 Could not participate 385 6 Other person in household returned questionnaire 345 5 Reason unknown 996 15
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -3-3 Results of multivariate analyses Logistic regression analyses were performed to assess how variables from the baseline (T1) questionnaire predicted drop-out at T2. In the analyses we compared individuals who did participate at T1 and T2 (n=3401) to individuals who participated at T1, but not in T2 (n=3164) in Model 1, and to individuals who participated at T1, and could but did not in T2 (n=996) in Model 2. Results of Model 1 are described. Demographics The results reported in Table 3 indicate that several socio-demographic variables predict the likelihood of dropping out of the study. There is a substantial effect of age, which is curvilinear and illustrated in Figure 1. It shows clearly that younger persons and older persons are more likely to drop out of the study; the probability of dropping out for the youngest and oldest persons is.70, compared to.40 for individuals 50-65 years of age. This selective dropout by age resulted in a smaller standard deviation and range of age at baseline at T2 compared at T1. Figure 1. Attrition by Age
- 4 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO There was no difference in the likelihood of dropping out between male and female respondents. Individuals with a migrant background, or a missing value on this variable, were more likely to dropout compared to native Germans. This selective drop-out resulted in a smaller share of respondents with a migration background at T2 (16% compared to 22% at T1). Homeownership did not predict dropout, although individuals with a missing value on this variable were more likely to dropout. Compared to individuals who reported to be single, married and widowed individuals were more likely to drop out. The reason why married individuals are more likely to drop out compared to single individuals is probably due to the fact that only persons who are not living alone could end up in the dropout category in which another person from the household returned the questionnaire. This explanation is supported by the non-significant effect of being married in Model 2, in which persons whose questionnaire is returned by another person from the households are excluded. Again, individuals with a missing value on marital status were also more likely to drop out. Part-time employed individuals and pensioners were less likely to drop out compared to fulltime employed individuals. This selective dropout by occupational status, however, did not result in differences in the composition of the sample between T1 and T2. Socio-economic status In order to capture the socio-economic position of respondents we included a poverty measure, educational level, and occupational prestige. The results indicate that individuals with a lower socio-economic status were more likely to drop out of the study. The odds of dropping out are 1.5 larger for every unit increase in poverty. The mean of poverty is slightly lower in the T2 sample (M=.04) compared to the T1 sample (M=-.05). Compared to individuals with a university degree, individuals without any education, with lower or general secondary degree were more likely to drop out. This resulted in a slightly lower share of individuals with no education (1% compared to 3% at T1) and lower secondary education (32% compared to 34% at T1) and a larger share of individuals with a university degree (27% compared to 23% at T1) in the T2 sample. Occupational prestige also predicted dropout, with individuals with lower levels of occupational prestige being more likely to drop out. As a result, the share of individuals with a very low occupational prestige is lower in the T2 sample (6% compared to 9% at T1), and the share of individuals with a high level of occupational prestige is higher in the T2 sample (42% compared to 36% at T1).
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -5- Yet again, the missing categories of educational level and occupational classification significantly predicted higher chances of dropping out. This consistent finding of the missing category predicting dropout is most likely a reflection of individuals who were already less motivated to fill out the questionnaire at T1. Neighborhood characteristics In a model without any individual characteristics as predictors, the percentage of immigrants (OR 2.194, p=0.015) and the percentage of welfare dependents (OR=1.009, p=.033) in the neighborhood predicted dropout. In the full model, however, controlling for all individual characteristics, the percentage of immigrants and welfare dependents in the neighborhood did not predict dropping out of the study. Individuals from more ethnic diverse and poorer neighborhoods were likely more likely to drop out, but this could be completely explained by their individual characteristics. The mean percentage of immigrants in the neighborhood (32% compared to 34% at T1) and the mean percentage of welfare dependents in the neighborhood are lower in the T2 sample (17% compared to 18% at T1)
- 6 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO Table 3. Logistic regression analyses of dropout on socio-demographics at baseline Model 1 All who responded T1, n=6,554 Model 2 All who could have responded T2, n=4,397 OR SE OR SE Age 0.860*** 0.011 0.862*** 0.015 Age 2 1.001*** 0.000 1.001*** 0.000 Male (ref.=female) 0.957 0.056 0.929 0.078 Migrant background (ref.= native German) 1.568*** 0.110 1.486*** 0.141 Missing 2.209*** 0.454 1.173 0.382 Renter (ref.=home owner) 1.123 0.067 1.001 0.087 Missing 1.610* 0.384 0.753 0.320 Marital status (ref.=single) Married 1.343*** 0.117 1.233 0.147 Divorced 1.139 0.132 1.125 0.183 Widowed 1.346* 0.162 1.390 0.241 Missing 2.756*** 0.832 2.358* 0.959 Occupational status (ref.=full time) Part-time employed 0.611*** 0.061 0.616*** 0.084 Unemployed 0.738 0.115 0.833 0.170 Home keeper 0.791 0.108 0.882 0.162 Pensioner 0.646*** 0.062 0.620*** 0.086 Other or missing 0.877 0.112 0.987 0.163 Poverty 1.493*** 0.087 1.455*** 0.120 Educational level (ref.=university degree) No education 2.510*** 0.482 1.875* 0.499 Lower secondary 1.425*** 0.120 1.298* 0.158 General secondary 1.222* 0.103 1.188 0.141 Higher (pre-university) secondary 1.198 0.111 1.172 0.146 Other degree 1.298 0.178 0.791 0.178 Missing 2.958*** 0.684 2.578** 0.794 Occupational prestige (ref.=very high) Very low 1.813*** 0.232 1.364 0.245 Low 1.369** 0.141 1.085 0.159 Moderate 1.300* 0.147 0.980 0.159 High 1.086 0.095 0.932 0.113 Missing 1.866*** 0.209 1.092 0.179 % immigrants in neighborhood 1.131 0.390 1.057 0.513 % welfare dependents in neighborhood 1.004 0.005 1.005 0.007 Note: * p<0.05, ** p<0.01, *** p<0.001.
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -7- Other factors In addition to the socio-demographic characteristics, we examined several other factors, including psychological variables, which might explain why certain people are more likely to drop out of the study. The results of this analysis are reported in Table 4. All socio-demographic variables as presented in Table 3 were also included in this model but not reported. Individuals who rated their health as bad or moderate were more likely to drop out of the study compared to individuals who reported their health as good. The share of individuals reported bad health (5% compared to 7% at T1) or moderate health (25% compared to 28% at T1) was smaller, whereas the share of individuals who reported good (43% compared to 40% at T1) or very good (21% compared to 19% at T1) health was larger at T2. Respondents who reported to feel more unsafe in their neighborhood were more likely to drop out of the study. Consequently, the mean level of neighborhood unsafety is slightly lower at T2 (M=.86 compared to M=.93 at T1). Individuals who reported to have been a victim of a property or violent crime in the two years before baseline were not more likely to dropout. Individuals with higher levels of trust are more likely to continue participation in the study, whereas individuals who distrust other individuals to a greater extent are more likely to dropout. The mean level of general trust at T2 is slightly higher (M=5.59 compared to M=5.32 at T1). In an additional model which is not reported we also analyzed internal and external locus of control, but these variables did not predict dropout. Pseudo R 2 is reported in Table 5. The pseudo R 2 of the full model, including demographics, socio-economic status, neighborhood characteristics and all other variables, indicated that almost 10 percent of the variance is explained.
- 8 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO Table 4. Logistic regression analysis of dropout on psychological characteristics Self-rated health (ref.=good) All who responded T1, n=6,475 All who could have responded T2, n=4,365 OR SE OR SE Bad 1.565*** 0.185 1.532** 0.254 Moderate 1.192** 0.081 1.104 0.111 Very good 0.947 0.075 1.100 0.118 Excellent 1.105 0.142 1.123 0.193 Missing 1.592 0.527 1.076 0.642 Neighborhood Unsafety 1.148** 0.056 1.126 0.078 Victim of property crime 0.916 0.052 0.981 0.080 Missing 1.478 0.486 1.225 0.721 Victim of violent crime 0.929 0.062 1.034 0.094 Missing General trust 0.943*** 0.013 0.939*** 0.018 Note: Estimates are adjusted for all individual and neighborhood characteristics as reported in Table 2;* p<0.05, ** p<0.01, *** p<0.001. Table 5. Pseudo R 2 from logistic regression models Pseudo R 2 All who responded Socio-demographics.089.068 Socio-demographics & other.096.074 T1 All who could have responded T2
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -9- Table 6. descriptives for T1 and T2 T1 T2 Variable T1 N Mean(SD) / % N Mean(SD) / % Age 6,555 60.62(16.21) 3,401 61.01(14.84) Male 6,565 45 3,401 46 Migrant background 6,565 3,401 Native German 75 82 Migrant background 22 16 Missing 3 1 Homeownership 6,565 3,401 Homeowner 39 45 Renter 59 54 Missing 2 1 Marital status 6,565 3,401 Single 16 16 Married 61 64 Divorced 10 10 Widowed 11 10 Missing 2 1 Occupational status 6,565 3,401 Fulltime employed 26 26 Part-time employed 12 13 Unemployed 4 3 Home keeper 5 4 Pensioner 47 48 Other or missing 7 5 Poverty 6,565 0.04(.54) 3,401-0.05(.52) Educational level 6,565 3,401 No education 3 1 Lower secondary 34 32 General secondary 19 21 Higher (pre-university) secondary 13 13 University degree 23 27 Other 5 4 Missing 3 1 Occupational classification 6,565 3,401 Very low 9 6 Low 17 16 Moderate 10 10 High 36 42 Very high 17 17 Missing 15 9 % immigrants in neighborhood 6,563 34 3,401 32 % welfare dependents in neighborhood 6,563 18 3,401 17
- 10 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO Table 6 continued T1 T2 Variable T1 N Mean(SD) / % N Mean(SD) / % Self-rated health 6,565 3,401 Bad 7 5 Moderate 28 25 Good 40 43 Very good 19 21 Excellent 5 5 Missing 1 1 Neighborhood Unsafety 6,543 0.93(.65) 3,396 0.86 Victim of property crime 6,565 56 3,401 57 Missing 1 0.6 Victim of violent crime 6,565 25 3,401 25 Missing 1 0.6 General trust 6,494 5.32(2.12) 3,377 5.59(2)
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -11-4 Differences by age As it can be assumed that the young and the old dropped out for different reasons. Therefore, we ran the models separately for 25-45 and for 65-90 year olds in order to see whether some predictors are more relevant for younger or older individuals. The results of these analyses are reported in Table 7-9. Four notable differences were uncovered. First, the age effect was stronger in the older sample compared to the young sample. Second, socio-economic status seemed to predict drop-out in the older sample slightly stronger than in the younger sample. Especially elderly with no education and very low occupational prestige were more likely to drop out. Third, bad or moderate health at T1 was stronger related to dropout in the older sample compared to the younger sample. Last, the pseudo R 2 is slightly higher for the older sample, indicating that we could predict the dropout for the older respondents slightly better than for the younger respondents. This indicates that the variables that we included might be more relevant for older individuals as a reason to drop out of the study as compared to younger individuals.
- 12 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO Table 7. Logistic regression analyses of dropout on socio-demographics at baseline by age Age 25-45 n=1364 Age 65-90 n=3143 OR SE OR SE Age 0.955 0.135 0.479*** 0.069 Age 2 1.000 0.002 1.005*** 0.001 Male (ref.=female) 0.930 0.120 0.967 0.082 Migrant background (ref.= native German) 1.410** 0.184 1.529*** 0.182 Missing 0.390 0.279 2.661*** 0.750 Renter (ref.=home owner) 1.263 0.185 1.051 0.090 Missing 3.900 3.100 1.596 0.513 Marital status (ref.=single) Married 1.237 0.178 1.360 0.244 Divorced 1.196 0.320 1.109 0.242 Widowed 1.290 0.249 Missing 11.763 15.907 1.750 0.736 Occupational status (ref.=full time) Part-time employed 0.629** 0.106 0.570 0.209 Unemployed 0.537* 0.144 0.590 0.478 Home keeper 0.848 0.205 0.561 0.210 Pensioner 0.772 0.548 0.612 0.167 Other or missing 0.672 0.152 1.083 0.374 Poverty 1.273 0.162 1.564*** 0.137 Educational level (ref.=university degree) No education 2.331 1.101 3.469*** 1.052 Lower secondary 1.660* 0.424 1.625*** 0.202 General secondary 1.402* 0.242 1.484** 0.208 Higher (pre-university) secondary 1.391* 0.208 1.058 0.194 Other degree 1.706 0.576 1.322 0.262 Missing 2.680 2.560 3.683*** 1.099 Occupational prestige (ref.=very high) Very low 1.460 0.436 2.063*** 0.390 Low 1.290 0.307 1.640*** 0.247 Moderate 1.053 0.267 1.344 0.221 High 1.131 0.207 1.211 0.163 Missing 1.926* 0.534 2.230*** 0.343 % immigrants in neighborhood 0.793 0.586 1.528 0.784 % welfare dependents in neighborhood 1.018 0.011 0.997 0.007 Note: * p<0.05, ** p<0.01, *** p<0.001.
An analysis of attrition in the SENSIKO survey (waves 1 and 2) -13- Table 8. Logistic regression analysis of dropout on psychological characteristics Self-rated health (ref.=good) Age 25-45 n=1350 Age 65-90 n=3090 OR SE OR SE Bad 1.520 0.744 1.759*** 0.274 Moderate 0.856 0.184 1.240* 0.111 Very good 0.890 0.125 0.814 0.127 Excellent 1.182 0.235 0.631 0.202 Missing 0.969 0.986 1.691 0.703 Neighborhood Unsafety 1.339* 0.155 1.118 0.080 Victim of property crime 1.093 0.139 0.913 0.075 Missing 1.618 0.671 Victim of violent crime 0.988 0.129 0.890 0.095 Missing General trust 0.983 0.030 0.929*** 0.019 Note: Estimates are adjusted for all individual and neighborhood characteristics as reported in Table 2;* p<0.05, ** p<0.01, *** p<0.001. Table 9 Pseudo R2 by Age Pseudo R 2 Age 25-45 Age 65-90 Socio-demographics.082.101 Socio-demographics & other.088.115 5 Conclusion Attrition is a problem in our study, with an overall loss of 48% of the respondents between the waves. Main correlates of dropping out are age, migration background, socioeconomic status and health. Age, socio-economic status and health, however, seem to better explain why older respondents than why younger respondents dropped out of the study. It can be assumed that younger people more often dropped out for reasons not measured in the survey, i.e. a lower commitment to participation and more distractions from filling in a questionnaire generally. A consequence of this selective attrition, particularly for the older age group, is that we lose the more vulnerable individuals and those who are more affected by fear. This could possibly lead to an underestimation of true effects. These findings should be considered when interpreting results.
- 14 - Max-Planck-Institut Freiburg i.br. Projekt SENSIKO 6 References Gerstner, D., & Oberwittler, D. (2016). Bevölkerungsbefragung Zusammenleben und Sicherheit in Köln/Essen Methodenbericht (Berichte aus dem Projekt SENSIKO / 1). Freiburg: Max-Planck-Institut für ausländisches und internationales Strafrecht. Lynn, P. (2009). Methodology of longitudinal surveys: John Wiley & Sons.