SafeGraph

• Source name: safegraph
• Available for: county, MSA, HRR, state (see geography coding docs)
• Time type: day (see date format docs)
• License: CC BY

This data source uses data reported by SafeGraph using anonymized location data from mobile phones. From June 2020-July 2022, SafeGraph provided several different datasets to eligible researchers. We surface signals from two such datasets.

Table of Contents

1. SafeGraph Social Distancing Metrics
   1. Lag
2. SafeGraph Weekly Patterns
   1. Lag
3. Limitations

SafeGraph Social Distancing Metrics

• Earliest issue available: June 20, 2020
• Number of data revisions since June 23, 2020: 1
• Date of last change: November 3, 2020

This dataset is no longer updated after April 19th, 2021.

Data source based on social distancing metrics. SafeGraph provided this data for individual census block groups, using differential privacy to protect individual people’s data privacy.

Delphi creates features of the SafeGraph data at the census block group level, then aggregates these features to the county and state levels. The aggregated data is freely available through the COVIDcast API.

For precise definitions of the quantities below, consult the SafeGraph social distancing metric documentation.

After computing each metric on the census block group (CBG) level, we aggregate to the county-level by taking the mean over CBGs in a county to obtain the value and taking sd / sqrt(n) for the standard error, where sd is the standard deviation over the metric values and n is the number of CBGs in the county. In doing so, we make the simplifying assumption that each CBG contributes an iid observation to the county-level distribution. n also serves as the sample size. The same method is used for aggregation to states.

SafeGraph’s signals measure mobility each day, which causes strong day-of-week effects: weekends have substantially different values than weekdays. Users interested in long-term trends, rather than mobility on one specific day, may prefer the 7dav signals since averaging over the preceding 7 days removes these day-of-week effects.

Lag

SafeGraph provides this data with a three-day lag, meaning estimates for a specific day are only available three days later. It may take up to an additional day for SafeGraph’s data to be ingested into the COVIDcast API.

SafeGraph Weekly Patterns

• Earliest issue available: November 30, 2020
• Number of data revisions since June 23, 2020: 0
• Date of last change: never

This dataset is no longer updated after July 15th, 2022.

Data source based on Weekly Patterns dataset. SafeGraph provided this data for different points of interest (POIs) considering individual census block groups, using differential privacy to protect individual people’s data privacy.

Delphi gathers the number of daily visits to POIs of certain types (bars, restaurants, etc.) from SafeGraph’s Weekly Patterns data at the 5-digit ZipCode level, then aggregates and reports these features to the county, MSA, HRR, and state levels. The aggregated data is freely available through the COVIDcast API.

For precise definitions of the quantities below, consult the SafeGraph Weekly Patterns documentation.

SafeGraph delivered the number of daily visits to U.S. POIs, the details of which are described in the Places Manual dataset. Delphi aggregates the number of visits to certain types of places, such as bars (places with NAICS code = 722410) and restaurants (places with NAICS code = 722511). For example, Adagio Teas is coded as a bar because it serves alcohol, while Napkin Burger is considered to be a full-service restaurant. More information on NAICS codes is available from the US Census Bureau: North American Industry Classification System.

The number of POIs coded as bars is much smaller than the number of POIs coded as restaurants. SafeGraph’s Weekly Patterns data consistently lacks data on bar visits for Alaska, Delaware, Maine, North Dakota, New Hampshire, South Dakota, Vermont, West Virginia, and Wyoming. For certain dates, bar visits data is also missing for District of Columbia, Idaho and Washington. Restaurant visits data is available for all of the states, as well as the District of Columbia and Puerto Rico.

Lag

SafeGraph provided newly updated data for the previous week every Wednesday, meaning estimates for a specific day are only available 3-9 days later. It may take up to an additional day for SafeGraph’s data to be ingested into the COVIDcast API.

Limitations

SafeGraph’s Social Distancing Metrics and Weekly Patterns are based on mobile devices that are members of SafeGraph panels, which is not necessarily the same thing as measuring the general public. These counts do not represent absolute counts, and only count visits by members of the panel in that region. This can result in several biases:

• Geographic bias. If some regions have a greater density of SafeGraph panel members as a percentage of the population than other regions, comparisons of metrics between regions may be biased. Regions with more SafeGraph panel members will appear to have more visits counted, even if the rate of visits in the general population is the same.
• Demographic bias. SafeGraph panels may not be representative of the local population as a whole. For example, some research suggests that “older and non-white voters are less likely to be captured by mobility data”, so this data will not accurately reflect behavior in those populations. Since population demographics vary across the United States, this can also contribute to geographic biases.

---