The IAB-SMART-Mobility Module: An Innovative Research Dataset with Mobility Indicators Based on Raw Geodata

2.1 Verifying Participants

Initially, we ensure that smartphone owners installing and using the IAB-SMART app are PASS participants since the IAB-SMART app data can be linked to the PASS data. After installing the IAB-SMART app, participants answered a short survey within the app containing questions about their demography (Kreuter et al. 2020). To verify that the PASS and the IAB-SMART study participants are identical, we compare the date of birth and gender from this IAB-SMART app survey to PASS answers. If at least one variable does not match, we drop the observation. We can verify that 618 persons from the PASS panel installed the app and participated in IAB-SMART.

Table 1 displays the distribution of the first installation and last signal from the app. In sum, most participants installed the app at the beginning of the survey and sent information until around the survey’s end: The majority of the 618 verified participants installed the app within a month after the beginning of the field time on January 8, 2018. Specifically, 75 % of the verified participants installed the app by February 10, 2018, i.e. within a month after the start of field time. Turning to the latest signals, 50 % of participants had the app installed until the end of field time on August 30, 2018.

2.2 Structuring Data into 30-min intervals

After restricting our sample to verified participants, we structure the raw geodata to harmonize the time of measurements across participants. The app aimed to collect geolocation information every 30 min, yet the measurements were not taken at precisely the same time for each participant. In order to harmonize the measurement times, we structure the measurements into 30-min intervals starting at 0:00 am. Based on the 618 participants, we create a raw data file with 48 intervals for each 30-min interval per field day. We then assign the geolocation measurements to the closest interval before the measurement. For example, if a geocode was measured between 9:00 and 9:29, we assign this measurement to the 9:00 measurement interval (Figure 1). Allocating geocodes into intervals allows us to structure the data and ensures that data is comparable between persons.

Figure 1:

Structuring the data into 30-min-intervals: measurements categorized into.

2.3 Ensuring a Participant’s Consent

In the next step, we focus on geolocation measurements for which participants gave consistently consent daily. Users can grant or deny access to the smartphone location and network connectivity data, which is necessary for the geolocation data collection by the SMART app. We remove 51 participants from the sample because they did not consent to collect geodata on a single day.

2.4 Cleaning Geolocation Data

We remove problematic geolocation measurements for the remaining 567 participants who consented to data collection. There are three types of measurement problems in the raw geolocation records. First, some measurements have failed and recorded a longitude and latitude position of 0. Missing geolocation measurements might occur because of a participant’s phone running out of battery, getting switched off, or losing network service (see Bähr et al. 2022). We remove these measurements of 0/0, which reduces the sample to 543 participants with 1,465,541 measurement points.

A second problem is imprecise geolocation records. Geolocation measurements are not precise points but come with a radius of accuracy (e.g. 350 m) describing the area where the geolocation might be. Geolocations were recorded using different collection providers associated with different levels of measurement accuracy. Table 2 displays the measurement accuracy for all geolocation records by type of measurement provider. GPS localization is substantially more precise than geolocation based on Google’s Fused Location Provider: the median accuracy is 13.0 m for GPS, while measurement using the Fused Location Provider is 4,898.4 m are in line with previous studies (e.g. Zandbergen 2009). Following the literature (e.g. Jongs et al. 2020), we focus on geolocation data with an accuracy of no more than 350 m. Since we restrict our sample to geolocation points with sufficient accuracy, we do not exclude specific providers.

A third measurement problem concerns the collection of multiple geolocation measurements per 30-min interval for approximately 25,000 intervals, which is less than 0.02 % of the intervals. The duplicates were an exact match in approximately 75 % of the cases. In the remaining cases, we keep the measurement with higher accuracy. In sum, this yields a data set of 1,288,050 geolocation measurements from 543 participants that meet the minimum requirements regarding individual measurement quality.

2.5 Ensuring Participants’ Sufficient Observation Periods

In addition to including only high-quality records at the measurement level, we also ensure a minimum number of measurements at the participant level. Specifically, we restrict the data to participants with at least seven ‘valid’ days, i.e. days with at least 8 h with sufficient geolocation records. Following the example provided by Müller et al. 2022, we first check whether at least one geolocation measurement was recorded in a given hour. We restrict the dataset to valid days with at least 8 h and at least one geolocation record. We focus on participants with a minimum of seven valid days. Furthermore, we exclude days when app usage data reveals that participants used geolocation spoofing apps (e.g. Fake-GPS). This step leaves us with 1,218,715 geolocation measurements from 398 participants.

2.6 Summary of Data Preparation and Sample Description

In summary, our data preparation yields a final data set of 398 out of the initial 618 verified participants. On the level of individual geolocation measurements, 1,218,715 out of 1,465,541 usable measurements remain (Table 3). Our final sample comprises 64 % of the verified participants and 83 % of the geolocation measurements. The discrepancy in the number of removed participants and geolocations illustrates that we remove primarily participants with few geolocation measurements. Thus, we only remove participants who had very few geolocation measurements, for whom mobility indicators are of little use due to gaps in the data. See Appendix A for a detailed overview of the data collection (following Müller et al. 2022).

Participants in our final sample differ in socio-demographic characteristics from the general population. For instance, our sample overrepresents 17-39-year-old participants compared to the German population (Destatis 2023). 34 % of participants in our sample are located in new federal states, i.e. eastern Germany, which is higher than the share of Germans living in eastern Germany (20 %) (Kreuter et al. 2020). The overrepresentation of younger participants and participants in eastern Germany is similar to the PASS (Kreuter et al. 2020). Unemployed and inactive persons are overrepresented in our sample (36 %) in comparison to Germany in 2017 (6 %) (Kreuter et al. 2020), which can be attributed to PASS’s oversampling of unemployed persons. In summary, the sample similarly deviates from the German population, given that IAB-SMART participants are randomly drawn from the PASS, yet the deviations for IAB-SMART are larger. Scholars need to take these deviations into account when analyzing the IAB-SMART-Mobility module.

3.1 Visited Unique Locations

The first indicator captures the number of unique locations visited per participant. Locations are identified following the clustering of individual geolocation points using the density-based spatial clustering of applications with noise method (DBSCAN) (Hahsler et al. 2019; Jongs et al. 2020; Müller et al. 2022). For the DBSCAN clustering, we chose a range of 600 m, similar to Jongs et al. (2020), since we also follow the accuracy restriction of 350 m recommended by Jongs et al. (2020). The clustering range needs to be higher than the accuracy to ensure that clusters are not biased by inaccurate measurements. Second, we chose two as the minimum geolocations per cluster. A location represents a place participants visit and spend time.

Since geolocation was measured in 30-min intervals, two geolocation measurements from the same cluster correspond to a stay of 30–60 min or at least two visits to that geolocation at different times. Participants visited 25.4 unique locations within the study period (Table 4). The average number of visited places is slightly higher on weekdays than on weekends (Table 4).

3.2 Traveled Distance

We calculate the distances between two consecutive points to measure the traveled distance. Given that we are using an extensive study region without any limits (participants cross states and even travel abroad), we use the distHarversine function from the geosphere package (Hijmans et al. 2017) to account for the curvature of the Earth when calculating distances (Müller et al. 2022). The Haversine distance calculates the shortest distance between two points (i.e. the ‘great circle distance’ or’ as the crow flies’), assuming a spherical earth, ignoring ellipsoidal effects. Based on the clustering, distances are set to zero for consecutively measured points in the same visited unique location. Consequently, the total distances capture the movement between unique locations as well as solitary geolocations. We aggregate the traveled distances between two locations to calculate the traveled distance on a median weekday and weekend, as well as the total traveled distance. Data users can calculate the mean distance traveled per day from the total distance and number of days with measurement.

3.3 Share Measurements at Home during Daytime

We identify homes following the strategy employed by Müller et al. (2022), who use two conditions: First, the home is the location with the most geodata points at night, i.e. between 0 am and 6 am. Second, to qualify as an identified home cluster, the respective location cluster must be the modal geolocation cluster for at least three nights (Müller et al. 2022, p. 5). After identifying the home cluster, we calculate the share of geolocations from the home cluster during the daytime, between 6 am and 0 am. Since geolocations at night were used to classify homes, we exclude geolocations at night for the share of geodata at home classification, that is, geolocations between 0 am and 6 am. For 23 participants, no home could be identified due to a lack of measurements at night. On average, participants are located in 82.2 % of their geodata points at home during a weekday and 91.8 % during the weekend (Table 4).

3.4 Location Variance

The location variance is calculated using the variance of a person’s latitudinal and longitudinal geolocation records. This variance represents the variability in a participant’s geolocation (Saeb et al. 2015). A higher variance indicates that the measured geodata records are further apart, indicating a higher radius of activities. Since the variance of geolocation points is usually below 1, the log of location variance is negative. A higher value, i.e. closer to 0, indicates a higher variance in geodata records. Similar to the previous indicators, we calculate the location variance for the entire study period and separately for weekdays and weekends.

3.5 Data-Quality Indicators

To provide data users with measures of the underlying data quality for each participant, we calculate three data-quality indicators for the total observation period, weekdays, and the weekend: First, the number of geolocation points summarizes the total number of geolocations that we use for calculating the mobility indicators for each participant. On average, 3,070 data points were used for each participant.

Second, we provide the number of valid days per person. As described in the section Ensuring Participants’ Sufficient Observation Periods, we consider days valid if at least 8 h of geolocation data are available. On average, participants provided sufficient geolocation data for 106.5 days (Table 5).

Third, we provide the average share of geodata records collected per day. Here, we calculate the total collected records and divide by 48, the theoretical maximum of collected points, to calculate this indicator. On average, 56.3 percent of the theoretical maximum is recorded (Table 5). Users can use these quality indicators to restrict the sample to their research needs. For example, the total distance and number of locations correlate with the number of days and the total number of measurement points.