The study was conducted in ten urban parks and their surrounding communities. Five intervention parks had been scheduled for major improvements with budgets in excess of $1,000,000 after December of 2003, and each intervention park was matched with a similar park (i.e., comparison park) which was not planned to receive upgrades by the city. The selected comparison park had similar size, features, and amenities and it served a population with similar sociodemographic characteristics as its intervention counterpart. (See ).
Park and Community Characteristics (Paired parks are in the same consecutive shading, “a” is a comparison park, “b” intervention) and Respondent Characteristics
Each intervention park scheduled open public meetings to discuss improvements and a Voluntary Oversight Committee was formed with members appointed by local elected officials to ensure community participation. Three parks constructed completely new gymnasiums. Two of the three parks had old gymnasiums: one retained the old gym, so they ended up with two gyms, while the other razed and replaced the one they had. The fourth park had its old gymnasium refurbished and underwent some field improvements in watering and landscaping. The fifth had improvements to picnic areas, upgrades to a walking path, and enhancements to a playground area so that it had rubberized surfacing around the climbing apparatus and stationary horses.
Assessments consisted of direct observations of park use and park characteristics plus intercept and household interviews at pre- and post-construction. Baseline data were collected between Dec 2003 and Nov 2004, and follow-up data between April 2006 and March 2008. To account for seasonal variation, pre- and post-measures were conducted at the same time of year. In addition, follow-up measures were initiated at least 3 months after construction; thus, the onset of post construction observations varied from 3 months to 14 months.
We used the System for Observing Play and Recreation in Communities (SOPARC) protocol that was developed specifically to objectively assess baseline park use and physical activity in the is project. SOPARC has been found to have good reliability6
and has recently been validated as an accurate method for providing estimates for total park use.7
Observations were conducted in all activity areas 7:30–8:30am, 12:30–1:30pm, 3:30–4:30 pm, and 6:30–7:30pm during each of the 7 days of the week. Any observation cancelled because of inclement weather was made up at the same time and on the same day of the following week. All area users are counted by gender (female or male), age group (child, teen, adult, or senior), race/ethnicity (Latino, black, white, or other), and activity level (sedentary, walking, or vigorous). The characteristics of each target area were also recorded (e.g., accessibility, usability, equipped, and whether activity in the area was being organized or supervised).
We also surveyed park users and recruited them systematically from the most and the least busy areas, by gender, and by activity level (i.e., sedentary, physically active). In addition, residents living within a 2-mile radius of the park were surveyed. More specifically, households were classified into four strata (within ¼ mile, from ¼ to ½ mile, from ½ to 1 mile, and from 1 to 2 miles from each park) and sampled approximately equal numbers of households from each stratum. Field staff, trained bilingual promotoras from a community-based organization, administered the interviews in either English or Spanish with the adult at home whose birthday most closely matched the visit date. Interviewers returned to a sampled household up to 5 times to locate residents before selecting an alternate address. Respondents were questioned about their use of the park and their physical activity. The same households were visited at baseline and follow-up, but unique identifying personal information was not collected from respondents. All methods were approved by the RAND IRB.
Propensity score analysis
To assess whether park improvements had an effect on outcomes of interest (such as park use, perceived park safety, physical activity during leisure time and health, and use of other parks), a propensity score analysis was conducted. This analysis included only 8 of the10 study parks. The first pair was eliminated from this analysis because a few key questions had not been included in the initial survey given to residents living near them. Park users and residents within a 2-mile radius were sampled and interviewed both before and after the park improvements (note that two different samples of people were drawn at the two time points). Because this is an observational study, survey respondents are not (and cannot be) randomized to live in a certain area or use a certain park; therefore, differences in the respondents’ characteristics, which in a randomized study would likely be null, might in part explain the observed intervention effect.
Propensity score weighting is an effective way of eliminating the differences in the observed characteristics (such as age, gender, and race) between survey respondents sampled at an intervention park at follow-up and respondents sampled at a control park at follow-up. Regression models rely too heavily on the linear assumption and are highly sensitive to model specification, such as the inclusion of important interaction terms. Propensity score weighting does not make linear assumptions and is more robust to model specification. The propensity score weights were fitted using the R package TWANG8
. Four distinct groups of respondents were compared: those sampled at intervention parks at follow-up (the “treated” group), those sampled at intervention parks at baseline, those sampled at control parks at follow-up, and those sampled at control parks at baseline. Because the respondents of the treated group differed from the respondents of the other three groups with respect to some observed characteristics such as age, race, and gender, three propensity score models were run. The obtained propensity score weights were then used to weight the other three groups of respondents to make them look like the “treated group” with respect the observed characteristics. The following respondent characteristics were included in the propensity score model: age, gender, Latino versus non-Latino, BMI, distance from home to the park, and whether the respondent engages in moderate to vigorous physical activity at work. The propensity score weights eliminated differences with respect to the characteristics between the treated group and the three other groups. A propensity score weighted logistic regression was then run to assess whether the changes in the intervention parks were significantly different from the changes in the control parks over time.