![\"\"](\"https:\/\/surveyinsights.org\/wp-content\/uploads\/2026\/02\/Table_1_.png\")
<\/a><\/p>\nTable 2: Break-down of participation in wave 2 (pre-recruited)<\/p>\n
The main purpose of our analyses is to compare second-wave participation between households of Ukrainian and non-Ukrainian refugees in Germany, and to identify relevant determinants of participation in the second wave, as well as compositional and possibly other unexplained differences between the groups regarding their decision whether to continue participating or not. Therefore, our selection basis is participation in the first wave in 2023 and eligibility to participate in 2024. To ensure comparability, we restrict our analyses to households in the M9 refreshment sample of the IAB-BAMF-SOEP Survey of Refugees that participated in the pre-recruitment survey and whose anchor persons completed the household, the personal, and the biography interview for refugees to have all information needed for the analyses. In total, 675 households of Ukrainian refugees and 1,594 households of refugees from other countries were included in the analyses. These restrictions are driven by data requirements and result in a more engaged analytic sample than the full first-wave sample, which should be kept in mind when interpreting the results.<\/p>\n We go beyond a simple treatment-control comparison, i.e. fitting a single model on the pooled sample with a UKR\/REF indicator. Such a specification would implicitly impose common slope coefficients across groups, unless all covariate interactions with group membership are included. In contrast, the decomposition approach estimates separate models for each group, allowing the full vector of coefficients to differ. This enables us to decompose the outcome gap into differences in characteristics (known as the \u201cendowment effect\u201d) and differences in coefficients (known as the \u201ccoefficient effect\u201d). The retention probability of household (1) <\/span> <\/span> where We employ a linear probability model as the decomposition method described below is derived for linear regression models (Blinder, 1973; Kitagawa, 1955; Oaxaca, 1973). Nonlinear decompositions for binary outcomes exist (Fairlie, 2005), but we focus on the linear specification for the ease of interpretation. We use the full model specification for the UKR sample (excluding REF-specific regressors) to conduct a Kitagawa-Blinder-Oaxaca (KBO) decomposition[i]<\/a>,[ii]<\/a><\/sup>. Although this method was originally conceptualised to decompose wage differentials across different population groups, it can be applied to other contexts.<\/p>\n Suppose we want to estimate the retention probability for households in the UKR and the REF samples, respectively. Let (2) <\/span> <\/span> By expanding the second line of equation (2), we get the KBO decomposition into the following terms:<\/p>\n The first term (\u201cendowment effect\u201d <\/em>or effect due to characteristics,<\/em> reported as \u201ccharacteristics\u201d <\/em>in Table 5) describes the part of the differential driven by differences in their compositional characteristics The second term (\u201ccoefficient effect\u201d <\/em>or effect due to varying effects of these characteristics across groups, <\/em>reported as \u201ceffects o<\/em>f characteristics\u201d in Table 5) describes the part of the differential driven by differences in coefficients The third term (\u201cshift coefficient\u201d<\/em>) represents the unexplained portion of the differential that is not captured by differences in covariates.<\/p>\n It is important to note that in our application the KBO decomposition serves as a descriptive accounting tool. The endowment and coefficient components do not represent causal effects, as the compositional characteristics are not exogenous but reflect migration-related selection processes that may have an impact on participation behaviour. The decomposition therefore quantifies how much of the observed retention gap is statistically associated with compositional differences and should be interpreted descriptively rather than causally. It should be noted, however, that in experimental or quasi-experimental settings where group membership or compositional characteristics can be treated as exogenous, a causal effect can be decomposed.<\/p>\n The following three subsections summarise our results. We first report the sample descriptives, followed by the regression results, and conclude with the decomposition analysis.<\/p>\n The description in Table 3 indicates that the samples of Ukrainian and non-Ukrainian refugees differ significantly in socio-demographic and interview-related characteristics. Ukrainian refugees are more likely to be female, have a high formal education, and live in private accommodations, but report poorer health status compared to the non-Ukrainian refugees. The median age of Ukrainian refugees is 43, while it is 35 for refugees from other countries. Additionally, Ukrainian refugees have a lower proportion of households in the third income tercile, single-person households, and households with children under 10 years. Approximately 90% of non-Ukrainian refugees express permanent stay intentions in Germany, whereas only 65% of Ukrainian refugees do, with further 18% undecided. The UKR sample includes only households that arrived in Germany since 2022, while the REF sample shows more variation in stay duration, with 61% having arrived in 2019 or earlier. The REF sample includes refugees from various countries, primarily Syria, Turkey, Iran, and Afghanistan.<\/p>\n Major differences also exist in interview-related characteristics. The REF sample has a higher proportion of partially completed households (25% vs. 18% for UKR sample) and a lower proportion of households interviewed by the same interviewer in both waves (interviewer continuity, 21% vs. 34% for UKR sample). There is a higher proportion of language matches between the anchor person\u2019s native language and the interview language in the UKR sample (48% vs. 34% for REF sample). About 55% of interviewers in the UKR sample are female, compared to 24% in the REF sample. Interviewers in the UKR sample are generally older and less experienced, with only 10% having worked in previous waves of the panel, as Ukrainian refugees were included in the IAB-BAMF-SOEP Survey of Refugees for the first time in 2023.<\/p>\n Table 3: Descriptive statistics: sample composition<\/p>\n The results of the linear probability models are presented in Table 4 (the significance level is set to 5%). The largest coefficients for both samples are attributed to interviewer continuity between the first and second waves. The coefficients remain stable across models, indicating that the regressor is likely uncorrelated with others.<\/p>\n No detectable effects were found for being a pure face-to-face household or the anchor\u2019s personal interview duration. However, a positive effect exists in the UKR sample for the match of the interview language with the anchor person\u2019s native language. For both samples, PUNR negatively affects participation in the second wave. Factors such as municipality size, household income, presence of children under the age of 10, and type of accommodation do not significantly affect participation.<\/p>\n Table 4: Determinants of household participation in the second wave<\/p>\n Regarding individual characteristics of the anchor person, gender, age, and self-reported health status do not correlate with participation in the second wave. In the UKR sample, low education levels are linked to lower participation likelihood, which does not apply to the REF sample. The absence of permanent stay intentions reduces retention probability in the REF sample, but not in the UKR sample. Households whose anchor person arrived in Germany in 2020-2021 are more likely to participate in the second wave than those who arrived in 2019 or earlier. No significant effects of country of origin or legal status on participation were found in the REF sample, providing empirical support for treating non-Ukrainian refugees as a single analytical category: with respect to panel retention, the null results suggest that the shared institutional framework outweighs within-group differences across origin groups.<\/p>\n Interviewer experience positively affects household retention in the REF sample, but this effect is negligible for the UKR sample due to most interviewers being newly recruited. Interviewer gender and age do not significantly influence participation.<\/p>\n The models explain a modest share of the variation in household retention, with R\u00b2 values of 0.172 for the UKR model, 0.069 for the REF model, and 0.078 for the extended REF model. This is not unusual for models predicting survey participation behaviour, where a large share of variation reflects idiosyncratic household-level factors that are difficult to capture with observable characteristics. The decomposition results presented below should be interpreted in this light.<\/p>\n The decomposition results in Table 5 show that the average retention rates for the UKR and REF samples are 76% and 63%, respectively, resulting in a retention differential of 13.4 percentage points. This difference is almost entirely attributed to variations in characteristics (endowment effect, 13.2 percentage points), with almost no difference in how these characteristics influence retention decisions (coefficient effect, 0.2 percentage points). Although the decomposition is performed using the model specification shown in the first two columns of Table 4, Table 5 only reports the covariates that significantly contribute to the retention differential.<\/p>\n Table 5: Decomposition of the retention differential<\/p>\n Source: IAB-BAMF-SOEP Survey of Refugees v41 (DOI: 10.5684\/soep.core.v41eu<\/a>).<\/em> About four percentage points of the retention differential arise from 34% of households in the UKR sample retaining the same interviewer between waves, compared to 21% in the REF sample. This difference is not due to one group valuing interviewer continuity more than the other. Similarly, the lower proportion of PUNR in the UKR sample contributes one percentage point to the retention differential. However, the difference in how the two samples deal with the PUNR shows no significant impact on retention.<\/p>\n About an additional four percentage points of the retention differential arise from differences in education between the two samples. When examining anchor persons with low formal education, a more complex picture emerges. Although the endowment effect for low formal education is positive, its coefficient effect is negative. This indicates that, despite a higher proportion of anchor persons with a low education degree in the REF sample contributing to the retention differential, the households with low educated anchor persons in the UKR sample face a greater risk of dropping out than their REF counterparts. In the UKR sample, anchor persons with low education have a more detrimental impact on retention. However, this effect cancels out in the total contribution to the retention differential.<\/p>\n Lastly, uncertainty about stay intentions in Germany is more pronounced in the UKR sample, which reduces the retention differential. This effect is driven by differences in characteristics rather than by differences due to the effect of this characteristic, as the intention to leave Germany does not significantly affect retention probability for the UKR sample compared to the REF sample.<\/p>\n As the shift coefficient is not statistically significantly different from zero, there is no tangible residual effect of possible unexplained variation.<\/p>\n The first goal of this paper is to provide field-based insights into panel retention between the first and second waves of data collection in a probability-based household panel of refugees in Germany, approached from a comparative perspective. To achieve this goal, we compared the participation of Ukrainian refugees and refugees from other countries within the same survey over two waves and analysed the determinants of panel retention in these groups.<\/p>\n In contrast to our first hypothesis (H1), our analyses indicate that, except for education in the Ukrainian sample, socio-demographic characteristics (e.g. age, gender, income) are not the primary factors influencing continued participation. Instead, other individual factors, such as long-term stay intentions in the host country, along with interview-related characteristics, were found to be more relevant. Although the survey adhered to the same fieldwork procedures, interviewer training, and questionnaires, the interview-related characteristics did vary between the two groups. In a self-completion setting, these differences may diminish, resulting in a smaller panel retention gap.<\/p>\n Our second hypothesis (H2) is confirmed for the REF sample. An absence of permanent stay intentions in Germany is negatively associated with participation in the second wave. In line with the leverage-salience theory of survey participation (Groves et al., 2000), the perceived reward of participation might be greater for households intending to stay in Germany, as the study’s outcomes hold more significance for them. To improve coverage of households with temporary stay intentions, advance letters should emphasise the importance of continued participation regardless of stay intentions. Related to this, it should be noted that the situation of Ukrainian refugees differs from that of other refugees, as they receive more generous protection status as compared to refugees from other countries (Mickelsson, 2025). In line with the social exchange theory of survey participation (Dillman et al., 2014), the higher participation rates in both waves among Ukrainian refugees may reflect acts of reciprocity and gratitude. However, given the current developments and changing policies regarding Ukrainian and other refugees in Germany, further research is needed to analyse how these changes might affect panel retention in the future.<\/p>\n Our third hypothesis (H3) posited that a higher respondent burden, measured by language mismatch and lengthy interview, would decrease participation in the second wave. This hypothesis is partly confirmed for the Ukrainian sample, where alignment between the anchor person\u2019s native language and the interview language positively affects participation. In contrast, this effect is absent in the REF sample, likely due to limited language options to cover a wide range of languages and the fact that participants have generally spent more time in Germany, allowing them to learn German. Offering survey materials and conducting interviews in various languages is crucial for recently arrived refugees and migrants. Although this approach incurs additional costs, promising advancements in automated translation (e.g. Zavala-Rojas et al., 2024) may provide cost-effective alternatives, especially for self-completion modes.<\/p>\n Our fourth hypothesis (H4) is confirmed for both samples. Consistent with previous research, interviewer continuity positively affects participation in the second wave. Interviewers with prior experience in the survey were associated with higher participation rates in the REF sample. In our analysis, interviewer continuity exhibits the strongest effect on panel retention across both samples. It should be noted, however, that interviewer continuity is not randomly distributed and may be correlated with unobserved household characteristics that independently influence retention. Consequently, the estimated effect of interviewer continuity may partly reflect household-level selection rather than a genuine interviewer effect and should therefore be interpreted with caution. Retaining the same interviewers across waves is advisable, though it is not always feasible. Given the recent challenge of interviewer shortages and the shift to online data collection, the impact of this transition on retention in previously face-to-face panels remains uncertain. For instance, an analysis of a CAWI follow-up for soft refusals and non-contacts in the IAB-BAMF-SOEP Survey of Refugees (B\u00fcchner et al., 2025) demonstrates that offering an online mode of data collection was especially effective among households that experienced an interviewer change between two waves. Regarding PUNR, both samples indicate that incomplete households, where not all eligible members participated in the first wave, are more likely to drop out in the second wave, confirming our fifth hypothesis (H5). Our findings suggest that PUNR negatively impacts household participation, highlighting the need for strategies to minimise it, such as personalised outreach to eligible members and additional incentives for complete data. Future research should address and explore the mechanisms underlying PUNR in household surveys and its impact on panel retention.<\/p>\n The second goal of our paper is to extend the standard regression analysis of panel retention predictors by integrating a decomposition method to enable a more nuanced analysis of the retention gap. Our analyses began by observing differing panel retention rates in the Ukrainian and non-Ukrainian samples of refugees. We anticipated that the differences in panel retention primarily result from variations in sample composition. Applying the KBO decomposition allowed us to disentangle the sources of group differences in panel retention and to distinguish between differences attributable to differences in observable sample composition characteristics, differential effects of these characteristics, and unobserved factors on panel retention. The results of the decomposition support our main hypothesis (H6), confirming that the differences in retention arise almost entirely from variations in sample characteristics rather than from differing effects of these characteristics or other unexplained factors. Consistent with H6, the higher rate of interviewer continuity and the lower proportion of incomplete multi-person households in the UKR sample emerged as the primary contributors to the endowment component. Tertiary education also contributed positively to the endowment component as expected, though the decomposition revealed an additional complexity: low-educated anchor persons in the UKR sample face a substantially higher dropout risk than their REF counterparts, a pattern that partially offsets the compositional advantage. In this sense, the method provided important analytical clarity: it confirmed that the retention gap largely reflects structural differences in sample composition, which is consistent with the substantial heterogeneity between the two groups.<\/p>\n Against the criterion we set out, namely whether the KBO decomposition adds analytical clarity beyond a group-indicator regression, the results are positive: the method quantified the characteristics and effects of characteristics split, confirmed the primary source of the retention gap, and surfaced the education asymmetry that would have remained invisible in a pooled model. That said, the analytical value of the KBO decomposition for the analysis of survey participation and panel retention may be particularly pronounced in settings where groups are more similar in their observable characteristics but still exhibit participation differences. In such cases, the decomposition could provide deeper insights into sources of panel retention going beyond compositional differences. In our case, the strong compositional divergence between samples limits the scope for identifying such differences; nevertheless, the method strengthens the robustness and interpretability of our findings by confirming the primary source of the gap observed.<\/p>\n <\/p>\n<\/a>
\nSource: IAB-BAMF-SOEP Survey of Refugees v41 (DOI: 10.5684\/soep.core.v41eu<\/a>).<\/em>
\nNote: Percentages are shares of column totals (gross sample wave 2 (pre-recruited)) and do not follow AAPOR response rate definitions. The number reported under “Total” is not identical to the number of households that participated in wave 1 in Table 1, as some households may drop out of the panel between waves due to eligibility criteria, while new split-off households may be added in subsequent waves.<\/em><\/p>\nMethods<\/h1>\n
interviewed by interviewer in the first wave is estimated using a linear probability model. All covariates in the model are first-wave information, except for the interviewer continuity. The dependent variable \u00a0equals one if the household has participated in 2024, i.e. at least the household questionnaire was completed, and zero otherwise. The model is specified as<\/p>\n<\/p>\n is the constant, is a row vector containing the determinants presented below, is a column vector of coefficients, is the coefficient for interviewer continuity, and is the disturbance term with . The indicator function signals stability in the assignment of the household interviewer between survey waves. Robust standard errors are clustered at the interviewer level, as an interviewer may interview more than one household. The determinants of participation in the second wave were derived from previous research summarised in the second section of this paper. The first (1 UKR) and the second (2 REF) model include characteristics applicable for both samples. The third model (3 REF ext.) includes additional individual characteristics applicable only to the REF sample, as there is no variation in these characteristics for the UKR sample.<\/p>\n and \u00a0be the column vectors of coefficient estimates fitted by equation (1) (including \u00a0for simplicity). Further, let and be the row vectors of the regressor sample means (including the mean of the indicator variable for change of household interviewer, i.e. the within-group rate of keeping the same interviewer). Let and be the estimated constants.<\/p>\n<\/p>\n\u00a0evaluated at UKR\u2019s levels of returns to these characteristics. A positive (negative) effect means a positive (negative) impact on the difference in average retention between the UKR sample and the REF sample due to the given characteristics of households, which are captured by the regressors.<\/p>\n\u00a0evaluated based on REF\u2019s characteristics. This presents a hypothetical scenario where a UKR household, given REF\u2019s average characteristics, decides whether to continue participation. A positive (negative) effect indicates a corresponding positive (negative) impact on average retention across the two samples.<\/p>\nResults<\/h1>\n
Sample descriptives<\/h2>\n
<\/a><\/p>\n<\/a>
\nSource: IAB-BAMF-SOEP Survey of Refugees v41 (DOI: 10.5684\/soep.core.v41eu<\/a>).<\/em>
\nNote: *p<0.05. Precision of differences in means are obtained by Welch\u2019s t-tests.<\/em><\/p>\nRegression results<\/h2>\n
<\/a><\/p>\n<\/a>
\nSource: IAB-BAMF-SOEP Survey of Refugees v41 (DOI: 10.5684\/soep.core.v41eu<\/a>).<\/em>
\nNote: *p<0.05. A Chow test is performed by fitting an LPM on the pooled sample of UKR and REF using the covariates from model (1 UKR), excluding the 37 REF households with missing information on education. A separate model for REF is fitted after excluding these cases.<\/em><\/p>\nDecomposition<\/strong><\/h2>\n
<\/a><\/p>\n
\nNote: *p<0.05. 37 REF households with unknown employment status of the anchor person are dropped from the sample. The overall retention rates are estimated as group means. The values in the \u201cAmount attributable to \/ TOTAL\u201d column are the sum of the \u201ccharacteristics\u201d and \u201ceffects of characteristics\u201d components shown in the two columns to the right. Consequently, they are not independent point estimates and do not have their own standard errors. The same reasoning applies to the \u201cTOTAL\u201d rows.<\/em><\/p>\nDiscussion and conclusion<\/h1>\n