Participating trialists worked with the InSPACE team to generate residential environmental measures for each trial participant. This measurement collaboration leveraged the ACMT, an open-source tool for linking environmental data to any point in the United States (eg, participant addresses for InSPACE) that was provided to each trial team. The ACMT includes an RShiny-driven user-friendly interface that does not require prior geospatial software experience to link environmental measures. The ACMT’s geocoding and linkage occur locally such that trial teams can generate and share the measures without sharing personally identifiable participant data with the InSPACE team or a third-party geocoding site [35]. The ACMT is described more fully elsewhere [33].

To run the ACMT, a user must have administrative privileges on their computer. In some cases, our trialist partners were not administrators on their computers or encountered institutional firewalls that prevented them from installing and running the ACMT without IT assistance. Additionally, because geocoding and data linkage involve large geographic files, ACMT-driven linkage often requires considerable disk space and memory resources. While requirements depended on the number of state files that needed to be downloaded, in most cases, 100 GB of available disk space and 8 GB of memory were adequate for the ACMT to function properly on trialists’ computers. Despite challenges in working through firewalls and administrator privileges, none of the recruited trials dropped out of the InSPACE study due to technical issues with the ACMT.

Trialists who had not already used their own geocoding systems to identify participants’ home locations (this was rare) used the geocoder embedded in the ACMT to generate the latitude and longitude values for each participant’s address. Trialists who used the ACMT geocoder were asked to update addresses that received geocode ratings higher than 0 (the PostGIS geocoder embedded in the ACMT uses 0 to indicate the highest level of confidence in the accuracy of the geocode; Corti et al [36]). To improve the geocoding accuracy, trialists used the ACMT to edit and improve the address text, including removing apartment numbers, fixing obvious typos, and spelling out street names and cardinal directions. If these initial edits did not improve the geocode to a rating of 0, the ACMT mapping tool provided the user a way to view geocoded locations on an OpenStreet map (embedded within the ACMT so as not to reveal addresses to third parties) and to document a note as to whether the geocode appeared accurate (mapped to the correct street and city) or not. This geocode checking process helped improve the quality of the geocodes and deterred trialists from using web-based third-party mapping tools to verify addresses and geocodes.

Next, trialists used the ACMT to link environmental data from the datasets listed in the table below (Table 2) to the geocoded home addresses. For most datasets, only 1 year of data was available. Notable exceptions include American Community Survey (ACS) 5-year estimates, for which data are provided in 5-year increments, starting with 2005‐2009. The National Land Cover Database (NLCD) includes data every 2‐3 years from 2004 through 2019, regional price parity was available for all years between 2008 and 2018, and gentrification was available for 2000 (indicating gentrification from 1990 through 2000) and 2010 (indicating gentrification status from 2000 to 2010). For datasets with multiple years of data, we chose 2015 as the a priori estimated midpoint of enrollment years for all recruited trials and linked environmental data for the year of available data closest to 2015 (ie, default year for NLCD data was 2016—the closest to 2015 of the available years of data; for ACS this meant pulling the 5-year span of data with a midpoint of 2015 [2013‐2017]). In addition to the default year(s) of data, trialists were asked to pull environmental data that most closely aligned with the trial’s first year of enrollment. For example, if a trial enrolled participants from 2018to 2019, they would pull data for ACS for both the 5-year span of data with a midpoint of 2015 (ACS 5-year data for 2014‐2018; ie, the default year for environmental data), as well as the 5-year span whose midpoint most closely aligned with their enrollment (ACS 5-year data for 2016‐2020; ie, ACS data with a midpoint of 2018). The same trial would also pull NLCD data for both 2016 and 2019 (data for 2018 are not available for NLCD), CDC (Centers for Disease Control and Prevention) Places data for 2017 (closest available year to the default year of 2015) and 2018, and Regional Price Parity data for 2015 and 2018.

Due to the limited availability of many of the environmental datasets, there is not perfect temporal alignment between the year of participant enrollment and the year of environmental measures. We do not expect that this temporal misalignment will create significant bias, particularly given the evidence that environmental attributes, such as sidewalk availability, greenspace, and gentrification change very slowly [47-49]. Study metadata, including year of enrollment for participants, length of intervention, and year of follow-up, will also be available in the pooled data to allow for more precise temporal alignment so that future sensitivity analyses may assess potential bias due to temporal misalignment.

For most environmental data linkages, radial buffers were used to define a participant’s neighborhood, and environmental measures were computed using area-weighted averaging for the buffered area. For example, to calculate the walkability index for a buffer around a participant’s house that was comprised 50% of Block group A, 20% of Block group B, and 30% of Block group C, the ACMT calculates the weighted mean of the walkability score for Block groups A, B, and C using the proportion of the buffer that is comprised each Block group as that tract’s weight.

Across the lifespan, a substantial amount of physical activity occurs or originates near or around one’s home or neighborhood, with walking in particular most commonly occurring within 500 to 1000 m of the home [50,51]. Larger buffers may be necessary to more accurately capture neighborhood attributes in more suburban or rural areas, particularly for attributes such as green space and park access [52-54]. No buffer size fully captures the bandwidth at which neighborhood factors might influence PA; the most appropriate buffer to use may vary by age, urbanicity, and by the neighborhood environment factor [24,55-57]. Because all linkage was performed at one time, but future analyses might want to consider different buffer sizes, we selected 3 sizes for all area-buffered environmental data: 500, 1000, and 5000 m. Notable exceptions to the buffering approach include gentrification and Regional Price Parity (Table 2).

Trialists provided individual-level participant demographic data. While we anticipated that the available demographic variables would differ across trials, we requested that trials share baseline demographic data, including gender or sex, age, race, ethnicity, household size, or other variables that could help us determine household size, any measure or indicator of income and socioeconomic status, including income, educational level, insurance status, etc, and a variable to indicate which group the participant was randomized to. For trials where more than one household member was included (ie, parent-child dyads), we requested that a household ID be included.

A set of principles was developed to inform the harmonization of race and ethnic identity categories across trials. Race harmonization categories were based in part on review of trialists’ collection of race and ethnicity data and were further informed by racial harmonization methods used in prior research [58]. Race values were not assumed or assigned unless a participant was specifically asked and self-selected or affirmed in the race or ethnicity category. For example, in cases where a trial asked about race and ethnicity in a single question, participants’ ethnicity was marked as missing if they selected a race or ethnicity category other than Hispanic, and participants’ race was marked as missing if they selected only Hispanic. Participants who selected more than one race category were categorized as multiracial, and participants who selected “Other” for their race but did not specify any specific races or specified a race other than the designated categories were marked as Other for their race. Ultimately, the harmonized racial categories included: American Indian or Alaska Native, Asian, Black, Native Hawaiian or Other Pacific Islander, White, Multiracial, Other, or Missing (Table 3). Ethnicity categories were Hispanic or Latino or not Hispanic or Latino. In trials where a specific demographic population was recruited, and only individuals from that specific demographic group were eligible for the study, trialists typically did not record this information in their demographic dataset. For such cases, we imputed the relevant data (eg, for a trial whose inclusion criteria required participants to self-identify as Black women, we imputed their race as Black and their gender as female).

Education harmonization was informed by the education categories used by trials in the pool. Education level for adults was combined and collapsed as needed to create six categories: (1) no schooling; (2) less than high school degree; (3) high school degree or General Educational Development; (4) some college, including technical school or associate’s degree; (5) bachelor’s degree; (6) some postgraduate education or more (Table 4). For trials that recruited children and collected data on parent education, we harmonized parent education according to these same 6 categories.

Harmonizing participant-reported income data was challenging. Preharmonization tables provided no clear options for harmonizing income by collapsing or combining categories (as having categories was the common practice, such as <US $20,000, US $20,001-US $40,000, etc) due to vastly different ranges of income categories across trials. It was particularly problematic to determine how to harmonize across the highest income category option, given the variability across trials and the open-ended nature of the highest category (eg, >US $75,000 vs >US $300,000). Furthermore, because the purchasing power of income varies across regions, the same income level might represent substantially different economic resources for participants in trials in different geographic regions. Due to these complexities, we chose to use not to harmonize income at all for primary aims analyses, but rather to use a nonharmonized within-trial categorical income variable for confounding adjustment in analyses.

Each InSPACE trial and the associated intervention took place over its own distinct amount of time (eg, some trials lasted weeks and some months or years). Trialists provided individual-level accelerometry files for each data collection period. We defined baseline accelerometry data to include all accelerometry data collected prior to the start of the intervention. We defined the postintervention accelerometry data to be the first tranche of data collected after the end of the intervention period. Some trialists shared additional follow-up accelerometry data (eg, after noncontact follow-up periods), which we harmonized for use in future analyses. We processed the data using established definitions of valid accelerometry data, based on existing literature. Our primary definition of valid accelerometry data requires 8 hours of wear time per day for a given day to be valid, and 3 days of valid wear time for a given time period of data collection to be considered valid [59]. Baseline, postintervention, and any additional follow-up after postintervention accelerometry data were processed using R (R Core Team) as detailed below.

We requested accelerometry data in the rawest form available. For example, in most cases, Actigraph accelerometer (most common device used in pooled trials) files were shared in ActiLife’s native file format (agd), or as a csv file aggregated at no more than the 1-minute epoch level to ensure that the InSPACE team could implement consistent wear time and conversion to physical activity minutes processing protocols across all participants from different trials. Most (31/39, 79%) of trialists used hip-worn Actigraph accelerometers (models 7146, GT3X, or GT9X). However, data from wrist-worn Actigraph, thigh-worn activPAL3, and wrist-worn Actical and GENEactiv accelerometers were also included in the pooled accelerometry data (see below for differences in processing dependent on wear site).

While trials with only day-level accelerometry data were not considered eligible during screening, there were 2 trials that were recruited and entered into a contract believing that they had minute-level accelerometry data but ultimately were only able to locate day-level accelerometry data. In these cases, we confirmed with the trialists’ study team the specific wear time algorithm used as well as the activity count thresholds to determine whether these could rigorously be combined with the other harmonized and processed accelerometry data.

Due to varying accelerometer types, wear locations, and file types, we developed a process to derive minute- and day-level accelerometry summaries for each participant in each study at baseline and postintervention. First, we summarized axis counts for each individual at each data collection time point by 1-minute epochs and derived a wear-time indicator variable using a previously validated wear-time algorithm [60]. Next, an indicator variable was derived that categorized each minute as moderate-to-vigorous physical activity (MVPA) based on the axis count threshold that aligned with the accelerometer type, wear location, and age of the participant [61-66]. Because we did not consistently have triaxial data across all trials, we chose to use only the vertical axis counts to determine wear time and MVPA.

One trial used wrist-worn GENEActiv devices and provided the raw accelerometry data in gravitational units. For this study, wear time was determined using an SD-based approach developed by van Hees et al [67,68]. This approach uses a 60-minute time window with 15-minute overlapping increments. Physical activity thresholds for this study were based on the mean acceleration per epoch rather than counts per minute to be consistent with available validation studies for GENEActiv accelerometry data [69].

To establish MVPA thresholds used for examining physical activity, we conducted a literature review of MVPA threshold validation studies for all types of accelerometers across all participant ages represented in participating trials, as well as thresholds used by participating trialists. Using this information, we chose appropriate MVPA thresholds that (1) considered vertical axis counts only (as opposed to a triaxial vector magnitude measure), (2) were most frequently used by InSPACE trialists, and (3) were well established and commonly used in other physical activity studies (Table 5). Once a threshold was chosen for each accelerometer type, wear location, and age group combination represented the pooled dataset, we applied the chosen thresholds to the pooled accelerometry data to create harmonized PA measures across trials. In cases where age was missing for an individual, study population information was used to categorize the individual as adult, child, or older adult to choose an appropriate MVPA cut point. The current process has documented the accelerometer types, wear location, and age group combinations that are present across all currently recruited trials. If additional trials were to be recruited, the InSPACE investigator team would conduct a similar process for establishing the appropriate MVPA threshold, using existing validation studies that are relevant for the accelerometer type, wear location, and age group as well as the MVPA threshold used by the trial to inform the decision-making process. Raw accelerometry data for all studies will also be available to researchers to allow for different processing and harmonization approaches or to conduct sensitivity analyses.

Day-level summaries for each participant provided the total wear minutes per day, which were used to determine whether a given day was valid. Accelerometry days were considered valid if the day included 8 or more hours of wear time [59]. Of the included trials, 95% (n=37) trials requested that participants wear the accelerometer for 7 days at each assessment time point. One trial had participants wear accelerometers continuously throughout the intervention and into the postintervention period, so we extracted 7 days postintervention from the continuous data collection. Participants with fewer than 3 days of valid accelerometry data at baseline or postintervention were excluded. An exception was made for data from one trial that collected only 2 days of accelerometry data at each time point. For this trial, any participant with at least 2 valid days with 8 or more hours of wear time for baseline and postintervention was included. A final wide-format InSPACE accelerometry dataset was created that included a variable for the mean minutes of MVPA per valid baseline day, mean minutes of MVPA per valid postintervention day, the type of accelerometer and the specific threshold used to define MVPA, total valid days at baseline, and total valid days at postintervention. Figure 1 shows the full cycle of trial recruitment and data collection and processing.

In some cases, trialists recontacted individuals whose initial accelerometry data did not provide adequate data (ie, not enough valid wear days) and asked participants to wear the accelerometer again. For these “rewear” scenarios, we applied the following logic: if valid rewear data were available and initial wear data did not have any valid days, only the rewear data were used. If both the initial wear data and rewear data contained valid wear days, we used the initial wear data up to the first valid rewear day and combined the initial wear data with all valid rewear days.

Several quality assurance reports were generated to help us validate accelerometry data. We did not have complete data from trialists about when and if accelerometers were delivered or returned via mail. Some trials gave accelerometers to participants in person, and some mailed them, and at least 1 trial removed delivery days prior to sharing their accelerometry with the InSPACE team. While previous studies have used machine learning to develop algorithms for identifying delivery days (Moore et al [72]), the inconsistencies in expected delivery days and the lack of triaxial data across all InSPACE trials required us to use a combination of algorithm and manual review to identify potential delivery days.

In reviewing the pooled accelerometry data, we were primarily concerned with days that were marked as a valid wear day (ie, 480 or more minutes of wear time) but were capturing movement of the accelerometer in transit. Days that were already marked as invalid due to insufficient wear time were not reviewed as potential delivery days, given that they were already excluded from analysis. In a manual review of data from 2 trials, we identified patterns that resembled potential delivery days. Participants were expected to wear devices for 7 consecutive days, so we identified deviations from this pattern, such as 1 or 2 days marked as a valid wear day, followed by or preceded by 2 or more invalid days (Figure 2), potentially indicating the device was in transit. When this pattern was identified for individuals who had 4 or more consecutive days of valid wear time, we flagged these individuals for manual review. Flagged days were visually inspected using a plot of accelerometer counts (refer to Figure 2 for a visual of marked possible delivery days) and confirmed as likely delivery (nonwear) days.

Additional QA reports were created to examine the time elapsed between the baseline accelerometry data collection and postintervention data collection. Time between data collection periods was compared to expected time elapsed based on trial metadata (eg, the planned length of time of the intervention) to ensure that accelerometry data files were properly sorted into baseline and postintervention files, and that there were no accelerometry data that were potentially overlapping in time with the trial intervention period for the participant, particularly for interventions that included an intervention team-facilitated physical activity component. In cases where baseline and postintervention accelerometry data overlapped, accelerometry start dates were compared to expected start dates for accelerometry data collection for baseline and postintervention, and only one set of accelerometry data was retained, while the other was omitted due to overlapping data collection period. If no other accelerometry data were available for the individual for the missing data collection period, the individual was omitted from the analysis.

The InSPACE project was approved by Seattle Children’s IRB (STUDY00002955). Because of the nature of InSPACE and the fact that the trials whose data were contributed to the data pool were already covered under the IRB of the trialist’s institution, this study was determined not to require a single-site IRB. The InSPACE study team provided boilerplate IRB language to study teams and offered individual consultation regarding approaches for seeking IRB approval at the trial site. These template documents were updated throughout the recruitment period in response to patterns and common questions that emerged from working with many teams over time. Given that all recruited trials chose to have their own trial team geocode and generate environmental measures, only deidentified data were shared with the InSPACE team, allowing for a much simpler IRB process at the pooled level.

IRB review and approval for participating trials generally followed one of four patterns: (1) Most commonly, the site IRB considered the InSPACE work exempt, and no modifications were required. (2) The trial had existing data sharing language in its protocol or consent documents and had to create a modification and get IRB approval for the new use of the data as a part of InSPACE. (3) The site did not have data sharing language already in their protocol or consent documents and had to create and get IRB approval for a modification requesting a waiver of reconsent and adding data sharing to the protocol. (4) The IRB did not approve the use of the participants’ home address data. This fourth scenario arose most often when the trial’s investigator had changed institutions and was not allowed to transfer identifiable data as part of the move. In some cases, trials were closed with their IRB and needed to be reopened before modification requests and approvals could be made.