Design and intervention

This study was a multiarm parallel trial with random assignment of homes with equal probability into two treatment arms and a comparison arm (1:1:1 allocation). The arms were:

1. Full Theory-Driven Education (Full Education)—an educational programme based on the health belief model integrated with hands-on practice based on the observational learning, reinforcement and modelling precepts of social-cognitive theory.18 The educational programme primarily addressed the individual-level factors that influence use of smoke alarms. These are beliefs that smoke alarms are effective in preventing fire injuries, that a fire at home is likely, that a fire in the home can result in injury and that one is capable of maintaining a smoke alarm in the home.15 These factors correspond to the health belief model constructs of perceived benefits, perceived susceptibility, perceived severity and self-efficacy.18 Observational learning based on social-cognitive theory reinforced the critically important self-efficacy construct in the health belief model. The social-cognitive module consisted exclusively of hands-on practice testing the battery, installing a replacement battery and using the hush button. As part of the practice effort, the resident installed batteries in the alarms we installed. We also completed the fire safety checklist that is standard in fire department installations. It requires a home tour and guides identification of home fire hazards, testing if existing alarms are working and decisions about how many alarms to install or replace and where to hang them. We chose the health belief model because it is an individual-level behaviour change theory developed to explain what factors are important in encouraging preventive health actions that many people are not adopting even though these actions have proven effectiveness.18 Observational learning, reinforcement and modelling precepts of social-cognitive theory were included as they explain how a person learns the behaviour and these factors reinforce self-efficacy.18 Online supplementary appendix A provides the practice module's script. The accompanying brochure and all other materials are available from the lead author without charge.

2. Focused Hands-on Practice (Practice)—a compressed model restricted to the hands-on practice module, typical fire department education (described below) and leaving a brochure describing alarm use and explaining the importance of changing batteries.

3. Comparison arm: Typical Fire Department Education (Current Norm)—generic fire department education including completion of a fire safety checklist and leaving a ‘baseball card’ with fire safety educational messages on the back. The messages cover fire escape planning and encourage smoke alarm testing and semiannual battery changes.

Procedures

We advertised ourselves as a home safety project. A health educator and an alarm installer comprised a visit team. Safety protocols dictated that the educator remain in the home while the installer worked.

Two master-level health educators trained on and delivered scripted education. Including research data collection, home visits averaged 45 min for all arms with wide variance.

Education took less time than installation if the resident stayed on topic. Visit duration, therefore, was sensitive to resident characteristics, not the education programme.

Table 1 summarises data collection forms and other activities associated with the visit. The knowledge, attitudes and behaviour (KAB) survey (see online supplementary appendix B) included four general questions about fire risks, 10 about smoke alarms, and 7 (which were not analysed) about fall risks. We designed KAB questions keyed to the stages in the health belief model, with emphasis on the barriers and facilitators, but did not validate them. The intake form had six questions about fire experience and eight about fall experience. We included questions about falls and installed touchlights in dark stairwells because we described the project as a home safety check. Participant behaviour might have changed if they knew our research focus was strictly on smoke alarm functionality. Of importance, this approach presented the comparison arm with more balanced safety messaging. The protocol was to leave existing alarms in place only if they had a hush button and 10-year batteries (zero homes qualified) or were hard-wired (two homes qualified). We created English and Spanish versions of all materials. All forms, procedures and modifications were approved by Pacific Institute for Research and Evaluation's (PIRE's) IRB. No harm resulted from this trial.

Installers used Kidde model 0916 ionisation smoke alarms with hush button. Manufacturer's specifications state that the alarms start chirping when 6% of battery charge remains. The alarms include test buttons that provide a second mode of testing operability. Installing batteries marked with nail polish allowed follow-up to differentiate resident-supplied batteries from our batteries that were still working. We bulk purchased and used Rayovac D1604 heavy duty 9-volt carbon-zinc batteries with an expiration date of April 2010, but installed longer lasting alkaline batteries in all alarms at follow-up.

Primary outcome measures were an increase in (1) number of homes with at least one working smoke alarm and (2) average number of smoke alarms working in each home.

Recruitment, randomisation, and time schedule

The trials were conducted in Maryland, USA. Installation began in September 2005 and ended in May 2008. Our initial power calculation targeted 1000 installation homes with follow-up per arm, which would support detection at the 95% confidence level of a 4–6% change in operability at follow-up assuming half of the comparison arm had an operable alarm.

Typically most residents canvassed by smoke alarm installers do not open the door.13 The objective was to learn how to increase alarm maintenance by those who accept an installation, not how to increase access to homes. Therefore, entry into our sample began when a resident age 18 or over consented to alarm installation, even if education was refused (three homes).

Initially recruitment used door-to-door weekday and weekend daylight canvassing in four low income neighbourhoods of suburban Prince Georges’ County. Using census data, we grouped the neighbourhoods into contiguous target areas of roughly 60 occupied dwelling units, then used computer-generated random numbers to select a simple random sample of target areas to canvas.

A change in county fire chief prior to the start of canvassing forced us to install without fire engine assistance, greatly reducing our ability to get doors answered. Randomising to practice (arm 2) stopped in February 2007 to increase the likelihood that the other two arms would achieve an adequate sample size. To inform that decision, without examining arm assignment, we conducted an interim analysis of the mean and variance of number of alarms functioning at enrolment, finding that two-thirds of the comparison arm had working alarms preintervention with an SD of 0.4 so adequate power to detect a 10% differential postintervention required only 100–200 installations in each arm. Because of a continuing sample shortfall, in January 2008, we added two Baltimore city recruitment streams: marketing at church services and canvassing in neighbourhoods where presence of a fire truck with flashing lights facilitated access to low -income homes in blocks the fire department targeted.

To simplify the health educators’ task, we assigned all of a day's participants to one arm that was randomly selected by rolling a die each day when the team assembled.

Using pilot-tested procedures and materials, staff knocked on doors and sent letters to schedule follow-up visits to check alarm operability and to ask questions about alarm testing, fire experience, and fall experience. Although follow-up was not formally blinded, it generally was performed by the installers; they were unaware of arm assignment.

Ideally follow-up would not occur until all batteries we initially installed had died. Otherwise we could not distinguish cases where residents were maintaining alarms that did not yet need new batteries from cases where the alarms would have become inoperable when initial batteries died. Since no data existed on carbon-zinc battery life in smoke alarms, we hung 100 alarms at PIRE's offices. Their battery failure rate led us to wait at least 500 days before attempting to return to a home to check operability. However, we returned up to 4 months early to the last 40 homes with installed alarms because follow-up (and the grant) ended in September 2009.

Statistical analysis

Analytically, we compared outcomes in treatment versus Current Norm homes using ordinary least squares linear regression to analyse whether the number of working alarms grew significantly from baseline to follow-up and logistic regression to analyse whether presence of at least one working alarm at follow-up increased significantly. All analyses were conducted in STATA V.11 from Statacorp, College Station, Texas, USA during 2009–2011.

Before starting data analysis, we decided to include data from pilot test homes because neither methods nor materials changed after pilot-testing. A dummy variable in the regressions tested whether including pilot test data affected outcomes (which it did not).

Because outcomes for the two theory-based health education arms proved extremely similar and sample size was modest, most analyses combined the two arms. To handle missing income data, we coded missing income as 0 and added a 0–1 variable valued at 1 for those who refused the income question. Online supplementary appendix C describes all other model covariates and shows the covariate-adjusted regressions with each arm separate. Regression analyses included preintervention alarm status (whether any alarms were operable or number of operable alarms) as a predictor to adjust for any bias associated with differences in functioning between arms at enrolment. The analysis tested whether time to follow-up affected outcomes in homes where batteries we initially installed were not still working. We initially tested models with all demographic variables and all KAB fire and smoke alarm item variables. We excluded most variables lacking statistical significance from final models to minimise cases excluded due to missing data and avoid multicollinearity.