Which factors influence mobility change during COVID‐19 in Germany? Evidence from German county data
2022; Elsevier BV; Volume: 14; Linguagem: Inglês
10.1111/rsp3.12537
ISSN1757-7802
Autores Tópico(s)Urban, Neighborhood, and Segregation Studies
ResumoThis study analyzes the role of regional demographic, socioeconomic, and political factors in mobility changes during the COVID-19 pandemic in Germany. Spatial econometric models are applied using data from the 401 counties in Germany. The model incorporates measures to reduce potential endogeneity effects. Our results show that mobility change shows significant socioeconomic heterogeneity, which could affect future policy measures to contain the pandemic. For example, case numbers and the share of academics are negatively associated with changes in mobility. On the contrary, a region's mean age and rural location have a positive impact. Political and economic implications of the results are discussed. The findings point to a possible reorganization of spatial, economic, and social activities beyond the course of the pandemic. Este estudio analiza el papel de los factores demográficos, socioeconómicos y políticos regionales en los cambios de movilidad durante la pandemia de COVID-19 en Alemania. Se aplicaron modelos econométricos espaciales utilizando datos de los 401 condados de Alemania. El modelo incorpora medidas para reducir los posibles efectos de endogeneidad. Nuestros resultados muestran que el cambio de movilidad presenta una importante heterogeneidad socioeconómica que podría afectar a las futuras medidas políticas para contener la pandemia. Por ejemplo, el número de casos y la proporción de académicos se asocian negativamente con los cambios en la movilidad. Por el contrario, la edad media de una región y su ubicación rural tienen un impacto positivo. Se discuten las implicaciones políticas y económicas de los resultados. Los hallazgos apuntan a una posible reorganización de las actividades espaciales, económicas y sociales más allá del periodo de la pandemia. 本研究では、ドイツにおけるCOVID-19パンデミック時の移動性の変化における、地域の人口統計学的、社会経済的、および政治的要因の役割を分析する。ドイツの401の郡から得たデータを空間計量経済モデルに適用した。このモデルは内生性効果を抑制する方法を取り入れている。結果から、移動性の変化が著明な社会経済的異質性を示されるが、これはパンデミックを封じ込めるための将来の政策手段に影響を与える可能性がある。例えば、症例数と学者の割合は、移動性の変化と負の相関関係にある。逆に、地域の平均年齢と農村部の立地はプラスの影響を与える。結果から得られる、政治的、経済的インプリケーションを考察する。知見から、パンデミックの経過を超えた空間的、経済的、社会的活動が再編成される可能性が指摘される。 The coronavirus disease 2019 (COVID-19) pandemic has been one of the greatest social and economic challenges of recent years, with approximately 427 million infections and 5.9 million deaths worldwide (as of February 25, 2022). The COVID-19 pandemic is still in full swing, although the momentum and lethality has slowed in some countries, in part because of vaccination success. Policymakers have responded with a variety of tools to limit mobility and, thus, contacts and chains of infection. In Germany, measures to restrict contacts were implemented for the first time in March 2020. Examples include direct mobility restrictions, temporary entry restrictions to certain federal states or counties (Kreise), and (nighttime) curfews. Indirect mobility restrictions consist of repeated appeals to the population to avoid private and tourist travel, the closure of restaurants, cafés, and leisure facilities, and self-motivation (caution and insight) to refrain from contacts and travel. From an economic point of view, therefore, a bundle of measures has increased the individual costs of mobility (transaction costs) and, at the same time, reduced its attractiveness (utility). This paper joins a growing body of work examining the impact of the COVID-19 pandemic on economic activity. For example, Ferreira et al. (2021) discuss worldwide interdependencies and supply chain disruptions based on different economic scenarios. They highlight how international trade linkages may be hit by pandemic shocks. Our analysis of mobility change elucidates the empirical impact of policy measures to contain COVID-19 in light of population behavior adjustment varying regionally with respect to socioeconomic characteristics. The results have broad economic, social, and policy implications. While the question of the macroeconomic impact of mobility changes has been discussed in Deb et al. (2021), for example, there are few studies analyzing the relationship between social status and the associated ability to effectively restrict one's mobility (in terms of mobility restriction as a luxury good, see, e.g., Huang et al., 2021). Related to this is the political possibility of deriving, for example, regionally specific and thus more targeted measures than before, if it is known which and how socioeconomic factors limit or promote mobility change in the wake of COVID-19. Note that the analysis of spatial interaction effects (question (iii)) is of central importance also from a statistical point of view, as such interactions can have biasing effects on the answer to the first two research questions. In addition, when making political decisions, it is important to know whether external effects of own parameters and measures must be considered. This paper is structured as follows. A brief overview of the literature is given in Section 2. The data used and the statistical methodology are covered in Section 3. The results are given in Section 4. Section 5 then discusses these results and relates them to their economic and political implications. There is now an extensive literature on the relationship between changes in population mobility behavior and COVID-19 dynamics. Most of the studies refer to transportation and (cross-national) long-distance travel as outcomes. For example, Linka et al. (2021) investigate the dynamics between mobility and COVID-19 operationalized by global air traffic and local mobility. Their study demonstrates different intensities of disease dynamics by using passenger air travel, cell phone data, and COVID-19 cases. For ten European countries – among others, Austria, Belgium, Denmark, France, Germany, and the United Kingdom – they find a time lag between mobility and disease dynamics of around 14.6 days on average. Moreover, it is discussed how local mobility data can help to identify super-spreading events. Kapitsinis (2020) is an example of a European Union (EU)-wide study at the regional level. Explanatory factors of COVID-19 dynamics are regional (NUTS 2 and NUTS 1) air quality, demographic variables, global interconnectedness, urbanization, and trends in health expenditures. In this study, regions with a high mortality rate are characterized by high shares of old people (65+ years). Regions with a weak health system exhibit higher COVID-19 mortality rates. Cutrini and Salvati (2021) analyze the spatial patterns during the first wave of the COVID-19 pandemic in Italy. They discuss the relevance of multiple factors to explain the forces beyond the spatial dynamics of the pandemic such as airline networks, urbanization, economic sectors, and firm size for North and South Italy. Further, weather conditions have been discussed as a factor underlying the spread of COVID-19, such as by influencing people's behavior of staying outside or inside, (which, in turn, affects the COVID-19 infection rate). Palialol et al. (2020) and Santos et al. (2021), to name a few, also address this question. They find that the exogenous variations of a weather variable reduce the COVID-19 transmission rate by approximately 9%. Credit (2020) brings up underlying racial and ethnic disparities with respect to COVID-19 infection and testing in the US cities Chicago and New York. White-majority neighborhoods are characterized by lower infection rates than other racial groups. The findings suggest that socioeconomic factors may help to explain these differences. Hatayama et al. (2020) discuss that working from home and avoidance of mobility may grow with the level of income. The authors point out that, for example in service occupations, more activities can be expected to be transferred to the home office. In the context of COVID-19, Crowley and Doran (2020) highlight differences in remote working potential by occupation and sector. Some occupations do not allow for remote work, so mobility and social distancing will not increase in certain occupations. However, the spatial distribution of economic activities due to sectoral or industrial clustering has an influence on the course of the pandemic. Where home office was not possible, the results showed a higher impact. For example, the US shutdown policies resulted in substantial increases in unemployment insurance claims (the policies caused a 12.4% increase of unemployment insurance claims according to Kong and Prinz, 2020). Yilmazkuday (2021) discusses the welfare loss of travel restrictions in the United States. The costs are measured by corresponding distances. The results show an estimated 11% loss of welfare of at its peak. On the basis of this result, the paper proposes that the legal and regulatory framework should be aligned with regard to future pandemics. Iacus et al. (2020) study data on global air passenger traffic addressing the impact of travel bans on the air transport sector and its general economic consequences. A more specific strand of literature deals with the associations between (COVID-19-induced) changes in mobility behavior and socioeconomic characteristics. The focus of the present study is to complement the literature in this field. For example, on the basis of a global online survey (approximately 600 participants), Dingil and Esztergár-Kiss (2021) use a multinomial logit model to estimate the impact of sociodemographic and travel characteristics on mobility behavior (before and after COVID-19 awareness, first wave). They find that age, income, travel distance, and mode are important influencing factors. Similar results are found by Czech et al. (2021), who relate the Human Development Index to mobility changes for 124 countries. For a global country dataset, Mendolia et al. (2021) examine how changes in information about the spread of the COVID-19 pandemic affected community mobility, depending on government policies implemented at the time. They find that human mobility is significantly responsive to information about the spread of the pandemic. Borkowski et al. (2021) study the impact of the pandemic on daily mobility behavior. The analysis is based on a sample of 1,069 people from Poland during March and April 2020. For data analysis, they apply a generalized linear model. In contrast to many publications that focus on long-distance or neighborhood behavior changes, the paper by Borkowski et al. (2021) covers behavioral adaptation in the short distance of daily life. Explanatory factors include homeschooling, quarantine, and level of education. The model confirms the assumption that household composition is crucial for short-distance travel avoidance. Here, the fear of COVID-19 infection was also included as a control variable showing a significantly negative impact on mobility. Further, the study finds no significant effect for age or gender on the change in mobility, but does for occupation or car availability. Unlike our case, however, the study is based on individual data. The study also contains a very comprehensive literature review (also with respect to the selection of covariates), which will not be duplicated here. Using network mobility data for Ontario, Canada, Long and Ren (2022) study the association between three different mobility measures and four socioeconomic indicators during the first and second waves of COVID-19. They find strong associations between mobility and the socioeconomic indicators. They also discuss how the relationships between mobility and other socioeconomic indicators vary over time. Liu et al. (2021) study the extent to which socioeconomic factors are related to the reduction in population mobility for both 358 Chinese cities and 121 countries worldwide. The analysis is based on mobile phone data and Google mobility reports from early 2020. They find that a higher socioeconomic index is significantly associated with a greater reduction in mobility at both city and country levels. Schlosser et al. (2020) investigate the impact of the pandemic on the mobility in Germany using mobile phone data. The authors emphasize that long-distance travel in particular declined sharply. Koenig and Dressler (2021) address a question similar to our study (albeit with different data for Germany). Their mixed-methods analysis assesses mobility changes in a rural region (Altmarkkreis Salzwedel). Their study is based on quantitative household surveys (301 persons) and qualitative telephone interviews (15 persons) on perceived mobility changes. Socioeconomic variables such as age, employment status, and income are included. Among other findings, the study shows that reductions in car trips were significantly associated with household income. Anke et al. (2021) analyze mobility behavior for a survey-based dataset for Germany (about 4,000 participants) and find that curfew measures have little effect on mobility changes. However, neither of the above analyses is based on econometric models. In selecting the possible socioeconomic factors influencing mobility change for our econometric model (see Sections 3 and 4), we largely followed the above literature on the association of mobility change with COVID-19 dynamics. A selection of literature sources and possible hypotheses regarding socioeconomic factors influencing mobility behavior are summarized in Table 1. This is an ecological study based on aggregated data at the level of the 401 counties in Germany, whose populations range from about 34,000 (Zweibrücken) to about 3,664,000 (Berlin). 1 As an outcome variable, the study uses the change in general mobility behavior at the county level in January 2021 (average values are calculated separately for weekdays and weekends for the entire month from January 4 to exclude the influence of the New Year's weekend) compared with the same period of the previous year. To map mobility at the county level, anonymized mobile communications data from the network of the communications provider Telefónica (Germany-wide market share ≥ 30%) are used, which are processed by Teralytics and made available by the Federal Statistical Office (2021). The data provide an overview of the number of mobile devices performing a certain movement. Movements are recorded when a mobile device changes the cell. The target region of a movement is reached when the mobile remains in a cell for at least 30 min. Possible distortions due to regionally varying market shares of Telefónica are compensated for by Teralytics using an algorithm that extrapolates geographically differentiated local market shares to the total German population. 2 Empirically, a decline in mobility was observed in most of the regions (with few exceptions, mainly in the eastern part of Germany) (Figure 1). The month of January 2021 was chosen as the core period of the second COVID-19 wave in Germany. The same month of the previous year, January 2020, was the month before the pandemic hit Germany. Source: infas 360 (2021) In addition to the COVID-19 case numbers, which can be considered an indirect (deterrent) factor in relation to mobility (see, e.g., Liu et al., 2021; Mendolia et al., 2021), a selection of covariates (measured also at the county level) was discussed in Section 2, where we largely followed the existing literature dealing with regional effects of COVID-19. These factors may serve as possible explanatory variables for mobility change in our model (Section 3.2). An empirical overview of the dataset is provided in Table 2 (see the appendix for further details). Our econometric approach is a cross-sectional ecological model based on data from the 401 German counties. The outcome variable used is the change in mobility between January 2020 and January 2021. In addition to the number of COVID-19 cases, the factors influencing the change in mobility discussed in Sections 2 and 3.1 serve as covariates. It should be noted that there are currently no publicly available longitudinal data at the county level for the selected variables, and therefore the methods of time series analysis cannot be applied. 3 Three methodological challenges in the identification of influencing factors (at the ecological level) on mobility behavior in the course of COVID-19 will be discussed in this section. These include, first, the problem of historical control, which results from the fact that all regions were hit simultaneously by the COVID-19 shock. Second, the possible endogeneity of COVID-19 case numbers in their influence on mobility change must be considered, and third, possible endogeneity due to the spatial feedback effects of mobility change. Since this research is a historically controlled study, the first question of changed concomitant circumstances that would have varied on an annual basis even in the absence of the COVID-19 pandemic and thus cannot be attributed to the pandemic must be carefully considered. Bias will be only partially avoidable from a statistical perspective but will be mitigated by the comprehensive covariates (variance in area, i.e., their possible influence on the changed accompanying and living circumstances). Conceivable biasing factors include general trends in regional mobility, holiday effects, and the influence of weather on mobility (e.g., for excursions). The last point is considered by the separate analysis of weekdays and weekends (cf. Table 3), as well as the inclusion of differences in sunshine duration and temperature between January 2021 and 2020 (whereby mobility in the month of January is certainly less influenced in this respect than in the summer months, so that some general robustness can be assumed for the study period). Furthermore, holiday effects are already considered in the data preparation by the Federal Statistical Office (2021). General trends in mobility over time are accounted for, at least indirectly, by its relation to socioeconomic variables and, of course, by the inclusion of a constant term (as a quasi-linear time trend) as well as federal state dummies. The second challenge relates to the fact that case numbers cannot be considered an exogenous variable but are themselves very likely to be influenced by the dependent variable in feedback loops (classical econometric endogeneity problem). Plainly, it must be assumed that the change in mobility measured over a period of 1 year will have an effect on the number of cases in that period (this is, after all, the political rationale for inducing changes in mobility in the pandemic). The exact mechanism of action is of course unclear; refer to Gargoum and Gargoum (2021), and Krenz and Strulik (2021), who examine this influence. On the other hand, caseloads are a compelling part of the mobility change model, as they will in turn have an important indirect influence on mobility change through an information and deterrence effect. Taking the above considerations into account, our identification strategy works as follows. We assume that the mobility change is driven, on the one hand, directly by contact and mobility restriction policies such as contact restrictions in public space, wholesale and retail restrictions, restrictions in the tourism sector, and curfews. Our corresponding variables (Table 2) reflect whether these restrictions were in effect on a given day in January 2021. On the other hand, we assume that mobility change is driven indirectly by the COVID-19 case numbers (e.g., people adjust their lifestyle and mobility behavior in response to high COVID-19 case numbers after being repeatedly encouraged to do so by policymakers) in addition to the above policy measures. See also Mendolia et al. (2021), who find that human mobility does respond in a significant way to information on the spread of the pandemic. We may thus assume that regional heterogeneity of case numbers serves as a central COVID-19-related parameter influencing mobility. To reduce the endogeneity problem discussed above with the available cross-sectional data (and to minimize possible distortions), we use an instrument instead of the case numbers of the respective county, namely the weighted average of the case numbers of the surrounding counties. Here, we choose as surrounding counties those that are at least 20 km and at most 150 km away from the respective (own) county. The reason for choosing the lower limit (20 km) is to reduce the presumed correlation with the error term as much as possible, since the influence of one's own mobility change on the number of cases in counties further away is greatly reduced. (This is still present through commuting linkages, but these are themselves included in the model as control variables.) At the same time, an upper distance limit must be found for a valid instrument, since an influence of case numbers by a deterrent effect should still plausibly exist. The exact limit is of course open to discussion, but for regions with a distance of more than 150 km it can be assumed that the deterrent effect of the case numbers on own mobility strongly decreases. Third, in contrast to a standard linear model (OLS), our analysis takes into account the spatial distance of the observation units (counties) from each other. The reason for this is that the spillover effect caused by feedback effects of mobility from neighboring regions on one's own mobility (endogenous spatial interaction) should also be taken into account in the model, as otherwise a bias of the coefficient estimators may result (see, e.g., Elhorst (2014) for a discussion). Spatial statistical models (see below) reflect the fact that outcomes in one region may be influenced by outcomes and/or covariates in neighboring regions (spatial spillover effects) and/or a spatial autocorrelation of the residuals. This proposition can be explained, for example, via learning effects from neighboring regions or via spatial substitution of mobility. In the latter case, reduced mobility in one region is quasi-substituted by increased mobility in neighboring regions. Spatial models can reduce potential bias of OLS estimation in the case of spatial spillovers and/or increase estimation efficiency. See, for example, Tokey (2021) for an application of such models with regard to COVID-19 and mobility change. However, this general model is only weakly identifiable and is therefore rarely used in practice. It is therefore necessary to focus on a specific subclass of models. A common approach to model selection is to introduce at least one constraint of the form ρ = 0 , θ = 0 , or λ = 0 and/or to base model selection on theory. Since we are particularly interested here in endogenous interactions (i.e., whether mobility change in neighboring regions exerts a spillover or learning effect on one's own region), we focus on the restriction θ = 0, which corresponds to a spatial autoregressive confused (SAC) model and includes as special cases the widely used spatial autoregressive (SAR) and spatial error (SEM) models (see Elhorst (2014) for a discussion). SAR models are also proposed by Santos et al. (2021) in the context of COVID-19, albeit the outcome variable there is incidence. Tokey (2021) uses an SEM model to analyze regional mobility data in the United States. Technically, the spatial statistical models capture the neighborhood relationships using a so-called spatial weighting matrix (i.e., a symmetric N × N matrix). This is based here on the geocodes (longitude and latitude of the circle centers) provided by the provider Opendatasoft (under the Creative Commons license). Specifically, the spmatrix command in Stata/MP 16.1 was used to create an inverse distance matrix from the coordinates, in which regions closer to each other are given a higher weight. The technical details of the spatial statistical models shall be omitted here with reference to the detailed discussion in Elhorst (2014). Data analysis was performed using the spregress command in Stata/MP 16.1 (which effectively reduces to OLS when no spatial interactions are present). Table 3 presents the results for the SAC model, which we focus on here because it includes the spatial autoregressive and spatial error models (as well as OLS) as special cases. Before discussing the individual coefficient estimators, we briefly note that neither the likelihood ratio (LR) test (versus the OLS model, see at the end of the table) nor the estimated coefficients for spatial lag and spatial error turn out to be significant. Thus, the SAC model is not supported and the null hypothesis of a simple linear model cannot be rejected. However, we do not see this as a shortcoming of our modeling, but rather as an empirical result (unknown ex ante) with respect to a model, which allowed flexibility and thus a test with respect to spatial spillover effects, since these can be plausibly motivated from an economic point of view (cf. Section 3). Consequently, in Table 3 we report the results based on the spatial model, since it is the more flexible model and includes OLS as a special case. 4 To classify this result in terms of its actual meaning, the dynamics of possible spatial effects of mobility change should not be confused with the spillover effects of case numbers per se, which are more prominent in the public perception. This is because the latter are based on a quasi-epidemiological process ('viruses need mobility and proximity to spread'), whereas the dynamics of mobility are transmitted only indirectly (e.g., through mobility spillovers and learning and deterrence effects). In this context, note that the latter linkages are also controlled for directly in the model, such as by the commuter balance variable, so this may reduce the significance of the spillover effects according to the SAC model. This is also shown by the fact that, in a sensitivity analysis with a reduced model without covariates, significant spatial spillover effects do indeed occur. However, we do not consider this model without covariates to be appropriate in terms of its economic meaning and therefore refrain from discussing its results further. The obtained coefficient estimators allow some interesting conclusions with respect to overall interpretation: Table 3 presents results for the change in mobility on weekdays (January 2021 versus January 2020) in the left-hand three-column block and results for weekends (in January 2021 versus January 2020, excluding the New Year's weekend) in the right-hand block. In each case, the coefficient estimators, standard errors, and p-values are given. Dummies for the 16 German states were included (not shown) in the estimation to account for the influences of state-specific COVID-19 measures. Note that, as discussed in Section 3, the case numbers for December 2020 and January 2021 refer to the surrounding counties (20–150 km) to mitigate the endogeneity problem. This endogeneity problem does not exist for the case numbers of the first wave in 2020, so that the respective counties' own case numbers were used here. Death counts were not considered in addition to the case counts to avoid a multicollinearity problem (the death counts follow the case counts almost deterministically until January 2021 except for a constant factor). Starting with the estimation results for weekdays, the significant negative association between the average (population-standardized) case numbers in surrounding counties in January 2021 and the mobility change can be noted first, which is in line with the results discussed, for example, in Liu et al. (2021) for cities in China. A possible deterrent effect of high case numbers does not seem to last long, as the previous month's caseload shows no significant association with changes in mobility behavior. The share of employed academics and the share of service providers clearly show a significant negative association with changes in mobility. This also confirms results by Liu et al. (2021) and can be explained in particular by the higher home office rates in academic and service occupations. The average age at the county level is significantly positively related to the mobility change, which may be explained, among other things, by the lack of influence of home office and homeschooling for older persons. Again, similar results are reported by Liu et al. (2021), where the proportion of people over 60 years shows a significant positive influence on intra-city mobility change. Note, however, that age was found to be insignificant in the study by Borkowski et al. (2021) on individual survey data from Poland. The regional share of women shows a significant negative association with mobility trends. One plausible reason may be the higher share of home offices in the service professions, the majority of which are held by women. Note, however, that gender is reported to be insignificant by Borkowski et al. (2021), but significant (at the 10% level) in a model of mobility change proposed by Dingil and Esztergár-Kiss (2021). Among the health variables, only chronic obstructive pulmonary disease (COPD) proportion seems to have a significant (negative) influence on the mobility change. Since people suffering from COPD are a high-risk group in connection with a potential COVID-19 infection, personal precautionary motives may serve as a plausible explanation. This result is also supported by Borkowski et al. (2021), where a variable termed 'being afraid of infection' is found to be significantly associated with a reduction in travel time during the pandemic. In contrast, the number of persons in need of care has a significant positive coefficient. Since the data do not differentiate between institutional and home care, this result may simply reflect the lack of opportunity to substitute mobile outpatient care. Car density exhibits a significantly negative coefficient. Although this may seem counterintuitive at first (see, e.g., Eisenmann et al., 2021), a high car density can also be seen as an indicator for a high potential of mobilit
Referência(s)