Covid19
Conditions
Brief summary
Facebook ads with physician-delivered videos were shown before the Thanksgiving and Christmas holidays and focused on staying safe during the COVID pandemic by limiting travel and mask-wearing.
Detailed description
The investigators used Facebook ads to show a 20 second video clip recorded by Massachusetts General Hospital, Harvard and Lynn Community health center doctors (6 people in all) to approximately 20,000,000 Facebook users. The ads will be shown before the Thanksgiving and Christmas holiday and will focus on staying safe - limiting travel and mask-wearing. The investigators will randomize exposure to the ad campaign at the ZIP code or county level to ask: Do the videos change mobility and Thanksgiving and Christmas holiday travel? Do they reduce the spread of COVID-19? Are there spillover impacts from the video messages? For example, if individuals decide to stay home, then do the geographical regions that tend to be visited by people from treated areas experience any effects, either through information spillovers or through a reduction in travel?
Interventions
BEHAVIORALFacebook Ads on the importance of staying safe during the Thanksgiving / Christmas holiday
Facebook ads to show a 15 second video clip recorded by MGH, Harvard and Lynn Community health center and other health care professionals. The ads will be shown before the Thanksgiving/Christmas holiday and will focus on staying safe - limiting travel and mask-wearing. The investigators will randomize exposure to the ad campaign at the ZIP code or county level.
Sponsors
Massachusetts Institute of Technology
Harvard University
Massachusetts General Hospital
Yale University
Stanford University
National Bureau of Economic Research, Inc.
Study design
Allocation
RANDOMIZED
Intervention model
SEQUENTIAL
Primary purpose
PREVENTION
Masking
NONE
Intervention model description
A total of 13 states, comprising 829 counties, centrally report COVID-19 cases at the ZCTA level. Two separate experiments were run within the study, one around Thanksgiving and one around Christmas. For each experiment, approximately half of the counties were randomized to high-intensity treatment with the remaining randomized to low-intensity treatment. In high-intensity counties, 3/4 of ZCTAs were treated and 1/4 were not. In low-intensity counties, 1/4 of ZCTAs were treated and 3/4 were not. In treated ZCTAs, Facebook users received ads with messages from physicians and other health personnel urging them to stay home and wear masks to stop the spread of COVID-19. Users in non-treated ZCTAs received no such ads. After exclusions, 820 counties were randomized to high- or low-intensity treatment during Thanksgiving and 767 counties were during Christmas.
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
Yes
Inclusion criteria
\- Individuals viewing the ads must be Facebook users, 18 years and older. Facebook will decide specifically which users receive the messages within a target geographical area (zip or county).
Exclusion criteria
\-
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Facebook Movement Metric | November 26, 2020 (Thanksgiving); December 24-25, 2020 (Christmas); February 2-29, 2020 (Baseline benchmark) | The change in movement metric is the percent change in distance covered in a day compared to the same day of the week in the benchmark period of February 2-29, 2020, by people who started the day in a particular location. We define holiday travel as travel during the three days preceding each holiday, since the available data does not allow us to compute the impact of the intervention on the return travel (after the holiday). The reason is that the mobility data describes the behavior throughout the day, for people who were in each county that morning. Since the campaign was targeted based on home location, we can only capture its impact on travel away from home, not back home. |
| Percentage Leaving Home on Day of Holiday | November 26, 2020 (Thanksgiving); December 24-25, 2020 (Christmas) | The Facebook stay put metric is the percentage of people who stay within a small geographical area (a tile of 600m\*600m in which they started the day). We use it to compute the share of people leaving home on the day of the holiday (i.e. this variable = 1 - stay put on the day of the holiday). |
| Inverse Hyperbolic Sine of COVID-19 Cases | December 1-14 (Thanksgiving arms) and January 1-14 (Christmas arms) | Inverse hyperbolic sine of COVID-19 cases during a 14 day period starting 5 days after each holiday |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Knowledge of COVID-19 Prevention Message (Recall of ad, Intent to Travel, Mask Wearing and Beliefs About Travel) | up to one month | Facebook will ask a set of users in treated zip codes to answer four survey questions (each user answers only one question each). A small within-zip control group will be held from being treated, enabling comparisons between those who see the videos and those who don't. These questions will be asked a few days after the ad is shown on the User's feed and will include 1) recall of the ad; 2) intention of traveling over holiday; 3) intention of wearing a mask; and 4) beliefs about whether people should travel over holiday. |
| Average Number of Tiles People Occupy (Mobility Measure) | up to one month | this measures each day how much people move around by counting the number of level-16 Bing tiles they are seen in within a day. People seen in more tiles are probably moving around more, while people seen in fewer are probably moving around less. Each day take eligible people in a given region and compute the number of distinct tiles they were seen in. This is aggregated to the county level. |
| Percentage of Eligible Participants Only Observed in a Single Level-16 Bing Tile (no Change in Movement) | up to one month | Facebook Stay-put data: this measures each day the percentage of eligible people who are only observed in a single level-16 Bing tile (600m x 600m) during the course of a day aggregated to they county level. |
Countries
United States
Participant flow
Pre-assignment details
The 13 states which centrally report COVID-19 cases at the ZCTA level contained 829 counties. At Thanksgiving, 9 counties were excluded due to data limitations; at Christmas, 62 counties were excluded due to data limitations and potential negative impacts of treatment in rural, conservative areas resulting from polarization in the wake of the 2020 presidential election.
Participants by arm
| Arm | Count |
|---|---|
| High-Intensity Counties Treatment: Individuals received approximately 3 Facebook ads over a 2-week period. Each ad contained a short video recorded by a physician using a script that discusses the importance of staying safe during Thanksgiving or Christmas by considering not traveling and using a mask when appropriate.
Randomization to treatment: The investigators randomized exposure to the ad campaign as follows. The 13 states which centrally report COVID-19 cases at the ZCTA level contained 829 counties. At Thanksgiving, 9 counties were excluded due to data limitations; at Christmas, 62 counties were excluded due to data limitations and potential negative impacts of treatment in rural, conservative areas resulting from polarization in the wake of the 2020 presidential election. Approximately half of the counties were randomized to high-intensity treatment with the remaining randomized to low-intensity treatment. In high-intensity counties, 3/4 of ZCTAs were treated (i.e., Facebook users in those ZCTAs received ads) and 1/4 were not. In low-intensity counties, 1/4 of ZCTAs were treated and 3/4 were not. | 0 |
| High-Intensity Counties Treatment: Individuals received approximately 3 Facebook ads over a 2-week period. Each ad contained a short video recorded by a physician using a script that discusses the importance of staying safe during Thanksgiving or Christmas by considering not traveling and using a mask when appropriate.
Randomization to treatment: The investigators randomized exposure to the ad campaign as follows. The 13 states which centrally report COVID-19 cases at the ZCTA level contained 829 counties. At Thanksgiving, 9 counties were excluded due to data limitations; at Christmas, 62 counties were excluded due to data limitations and potential negative impacts of treatment in rural, conservative areas resulting from polarization in the wake of the 2020 presidential election. Approximately half of the counties were randomized to high-intensity treatment with the remaining randomized to low-intensity treatment. In high-intensity counties, 3/4 of ZCTAs were treated (i.e., Facebook users in those ZCTAs received ads) and 1/4 were not. In low-intensity counties, 1/4 of ZCTAs were treated and 3/4 were not. | 410 |
| Low-Intensity Counties Control: Individuals did not receive ads containing short physician-recorded videos.
Randomization to treatment: The investigators randomized exposure to the ad campaign as follows. The 13 states which centrally report COVID-19 cases at the ZCTA level contained 829 counties. At Thanksgiving, 9 counties were excluded due to data limitations; at Christmas, 62 counties were excluded due to data limitations and potential negative impacts of treatment in rural, conservative areas resulting from polarization in the wake of the 2020 presidential election. Approximately half of the counties were randomized to high-intensity treatment with the remaining randomized to low-intensity treatment. In high-intensity counties, 3/4 of ZCTAs were treated (i.e., Facebook users in those ZCTAs received ads) and 1/4 were not. In low-intensity counties, 1/4 of ZCTAs were treated and 3/4 were not. | 0 |
| Low-Intensity Counties Control: Individuals did not receive ads containing short physician-recorded videos.
Randomization to treatment: The investigators randomized exposure to the ad campaign as follows. The 13 states which centrally report COVID-19 cases at the ZCTA level contained 829 counties. At Thanksgiving, 9 counties were excluded due to data limitations; at Christmas, 62 counties were excluded due to data limitations and potential negative impacts of treatment in rural, conservative areas resulting from polarization in the wake of the 2020 presidential election. Approximately half of the counties were randomized to high-intensity treatment with the remaining randomized to low-intensity treatment. In high-intensity counties, 3/4 of ZCTAs were treated (i.e., Facebook users in those ZCTAs received ads) and 1/4 were not. In low-intensity counties, 1/4 of ZCTAs were treated and 3/4 were not. | 410 |
| Total | 820 |
Baseline characteristics
| Characteristic | High-Intensity Counties | Low-Intensity Counties | Total |
|---|---|---|---|
| Baseline fortnightly cases Christmas (Period 2) | 654.77 Individuals | 598.54 Individuals | 626.84 Individuals |
| Baseline fortnightly cases Thanksgiving (Period 1) | 683.90 Individuals | 496.70 Individuals | 590.30 Individuals |
| Baseline fortnightly deaths Christmas (Period 2) | 5.70 Individuals STANDARD_DEVIATION 23.07 | 5.07 Individuals STANDARD_DEVIATION 11.29 | 5.38 Individuals STANDARD_DEVIATION 18.19 |
| Baseline fortnightly deaths Thanksgiving (Period 1) | 5.51 Individuals STANDARD_DEVIATION 22.35 | 4.64 Individuals STANDARD_DEVIATION 11.08 | 5.07 Individuals STANDARD_DEVIATION 17.63 |
| Baseline leave home Christmas (Period 2) | 82.40 Percentage STANDARD_DEVIATION 2.43 | 82.44 Percentage STANDARD_DEVIATION 2.4 | 82.42 Percentage STANDARD_DEVIATION 2.41 |
| Baseline leave home Thanksgiving (Period 1) | 82.33 Percentage STANDARD_DEVIATION 2.42 | 82.49 Percentage STANDARD_DEVIATION 2.53 | 82.41 Percentage STANDARD_DEVIATION 2.47 |
| Baseline movement metric Christmas (Period 2) | -8.69 Percent change relative to Feb 2020 STANDARD_DEVIATION 6.88 | -9.09 Percent change relative to Feb 2020 STANDARD_DEVIATION 6.56 | -8.89 Percent change relative to Feb 2020 STANDARD_DEVIATION 6.72 |
| Baseline movement metric Thanksgiving (Period 1) | -8.58 Percent change relative to Feb 2020 STANDARD_DEVIATION 7.1 | -8.88 Percent change relative to Feb 2020 STANDARD_DEVIATION 6.42 | -8.73 Percent change relative to Feb 2020 STANDARD_DEVIATION 6.77 |
| Missing baseline Facebook outcomes Christmas (Period 2) | 0.11 Proportion of counties STANDARD_DEVIATION 0.32 | 0.13 Proportion of counties STANDARD_DEVIATION 0.33 | 0.12 Proportion of counties STANDARD_DEVIATION 0.32 |
| Missing baseline Facebook outcomes Thanksgiving (Period 1) | 0.13 Proportion of counties STANDARD_DEVIATION 0.34 | 0.17 Proportion of counties STANDARD_DEVIATION 0.38 | 0.15 Proportion of counties STANDARD_DEVIATION 0.36 |
| Population in 2019 Christmas (Period 2) | 116,787 individuals | 122,875 individuals | 119,811 individuals |
| Population in 2019 Thanksgiving (Period 1) | 122,491 individuals | 102,818 individuals | 112,654 individuals |
| Race/Ethnicity, Customized Christmas (Period 2) | NA Proportion | NA Proportion | NA Proportion |
| Race/Ethnicity, Customized Thanksgiving (Period 1) | NA Proportion | NA Proportion | NA Proportion |
| Share Democrats Christmas (Period 2) | 0.37 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.37 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.37 Proportion of county voters STANDARD_DEVIATION 0.15 |
| Share Democrats Thanksgiving (Period 1) | 0.36 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.35 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.36 Proportion of county voters STANDARD_DEVIATION 0.15 |
| Share Republicans Christmas (Period 2) | 0.61 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.61 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.61 Proportion of county voters STANDARD_DEVIATION 0.15 |
| Share Republicans Thanksgiving (Period 1) | 0.62 Proportion of county voters STANDARD_DEVIATION 0.16 | 0.63 Proportion of county voters STANDARD_DEVIATION 0.15 | 0.62 Proportion of county voters STANDARD_DEVIATION 0.15 |
| Share urban Christmas (Period 2) | 0.48 Proportion of county ZCTAs STANDARD_DEVIATION 0.33 | 0.50 Proportion of county ZCTAs STANDARD_DEVIATION 0.33 | 0.49 Proportion of county ZCTAs STANDARD_DEVIATION 0.33 |
| Share urban Thanksgiving (Period 1) | 0.47 Proportion of county ZCTAs STANDARD_DEVIATION 0.34 | 0.44 Proportion of county ZCTAs STANDARD_DEVIATION 0.34 | 0.46 Proportion of county ZCTAs STANDARD_DEVIATION 0.34 |
Adverse events
| Event type | EG000 affected / at risk | EG001 affected / at risk |
|---|---|---|
| deaths Total, all-cause mortality | 0 / 0 | 0 / 0 |
| other Total, other adverse events | 0 / 0 | 0 / 0 |
| serious Total, serious adverse events | 0 / 0 | 0 / 0 |
Outcome results
Primary
Facebook Movement Metric
The change in movement metric is the percent change in distance covered in a day compared to the same day of the week in the benchmark period of February 2-29, 2020, by people who started the day in a particular location. We define holiday travel as travel during the three days preceding each holiday, since the available data does not allow us to compute the impact of the intervention on the return travel (after the holiday). The reason is that the mobility data describes the behavior throughout the day, for people who were in each county that morning. Since the campaign was targeted based on home location, we can only capture its impact on travel away from home, not back home.
Time frame: November 26, 2020 (Thanksgiving); December 24-25, 2020 (Christmas); February 2-29, 2020 (Baseline benchmark)
Population: Rural, conservative counties were excluded during Christmas due to potential negative impacts of treatment resulting from polarization in the wake of the 2020 presidential election. The two periods have been entered as rows to avoid double-counting counties.
| Arm | Measure | Group | Value (MEAN) |
|---|---|---|---|
| High-Intensity Counties | Facebook Movement Metric | Thanksgiving (Period 1) | -6.082 Percent change relative to Feb 2020 |
| High-Intensity Counties | Facebook Movement Metric | Christmas (Period 2) | -2.603 Percent change relative to Feb 2020 |
| Low-Intensity Counties | Facebook Movement Metric | Thanksgiving (Period 1) | -5.320 Percent change relative to Feb 2020 |
| Low-Intensity Counties | Facebook Movement Metric | Christmas (Period 2) | -1.823 Percent change relative to Feb 2020 |
Primary
Inverse Hyperbolic Sine of COVID-19 Cases
Inverse hyperbolic sine of COVID-19 cases during a 14 day period starting 5 days after each holiday
Time frame: December 1-14 (Thanksgiving arms) and January 1-14 (Christmas arms)
Population: ZCTAs in rural, conservative counties were excluded during Christmas due to potential negative impacts of treatment resulting from polarization in the wake of the 2020 presidential election. The two periods have been entered as rows to avoid double-counting ZCTAs. Note that there are more ZCTAs in the Christmas control than Thanksgiving control because the two periods were randomized separately and the number of ZCTAs varies by county.
| Arm | Measure | Group | Value (MEAN) |
|---|---|---|---|
| High-Intensity Counties | Inverse Hyperbolic Sine of COVID-19 Cases | Thanksgiving (Period 1) | 4.333 Inverse hyperbolic sine of individuals |
| High-Intensity Counties | Inverse Hyperbolic Sine of COVID-19 Cases | Christmas (Period 2) | 4.368 Inverse hyperbolic sine of individuals |
| Low-Intensity Counties | Inverse Hyperbolic Sine of COVID-19 Cases | Thanksgiving (Period 1) | 4.298 Inverse hyperbolic sine of individuals |
| Low-Intensity Counties | Inverse Hyperbolic Sine of COVID-19 Cases | Christmas (Period 2) | 4.442 Inverse hyperbolic sine of individuals |
Primary
Percentage Leaving Home on Day of Holiday
The Facebook stay put metric is the percentage of people who stay within a small geographical area (a tile of 600m\*600m in which they started the day). We use it to compute the share of people leaving home on the day of the holiday (i.e. this variable = 1 - stay put on the day of the holiday).
Time frame: November 26, 2020 (Thanksgiving); December 24-25, 2020 (Christmas)
Population: Rural, conservative counties were excluded during Christmas due to potential negative impacts of treatment resulting from polarization in the wake of the 2020 presidential election. The two periods have been entered as rows to avoid double-counting counties.
| Arm | Measure | Group | Value (MEAN) |
|---|---|---|---|
| High-Intensity Counties | Percentage Leaving Home on Day of Holiday | Thanksgiving (Period 1) | 71.308 Percentage |
| High-Intensity Counties | Percentage Leaving Home on Day of Holiday | Christmas (Period 2) | 72.859 Percentage |
| Low-Intensity Counties | Percentage Leaving Home on Day of Holiday | Thanksgiving (Period 1) | 71.468 Percentage |
| Low-Intensity Counties | Percentage Leaving Home on Day of Holiday | Christmas (Period 2) | 72.852 Percentage |
Secondary
Average Number of Tiles People Occupy (Mobility Measure)
this measures each day how much people move around by counting the number of level-16 Bing tiles they are seen in within a day. People seen in more tiles are probably moving around more, while people seen in fewer are probably moving around less. Each day take eligible people in a given region and compute the number of distinct tiles they were seen in. This is aggregated to the county level.
Time frame: up to one month
Secondary
Knowledge of COVID-19 Prevention Message (Recall of ad, Intent to Travel, Mask Wearing and Beliefs About Travel)
Facebook will ask a set of users in treated zip codes to answer four survey questions (each user answers only one question each). A small within-zip control group will be held from being treated, enabling comparisons between those who see the videos and those who don't. These questions will be asked a few days after the ad is shown on the User's feed and will include 1) recall of the ad; 2) intention of traveling over holiday; 3) intention of wearing a mask; and 4) beliefs about whether people should travel over holiday.
Time frame: up to one month
Secondary
Percentage of Eligible Participants Only Observed in a Single Level-16 Bing Tile (no Change in Movement)
Facebook Stay-put data: this measures each day the percentage of eligible people who are only observed in a single level-16 Bing tile (600m x 600m) during the course of a day aggregated to they county level.
Time frame: up to one month