Study Design:
A cluster randomized clinical trial (RCT) is planned in 40 (22 originally planned +16 additional) publicly subsidized housing facilities (HUD Section 202) and other low-income housing voucher programs in NE Ohio and will be randomized to 2 arms: Arm 1 (20 sites) - Participants will receive biannual SDF; Arm 2 (20 sites): Participants will receive ART + FV biannual application. All participants will be followed for one year and will receive a dental screening and prophylactic cleaning at all visits.
Study Population and Setting:
The long-term goal is to reduce disparities and improve oral health equity for low-income older adults. The immediate goal is to test interventions to address disparities. Therefore, the population/setting will be low socioeconomic status (SES) adults ≥ 62 years of age in 6 NE Ohio counties (Cuyahoga, Lorain, Summit, Ashtabula, Geauga, and Lake) residing in HUD Section 202 and other publicly subsidized housing facilities. Administrators have provided letters of support confirming the availability of sites and intent to participate. The population is diverse: African American/Black, Caucasian/White including Hispanic/Latino, and Asian. There are 64 HUD housing sites with 3,661 tenant units in the 6 counties. An additional 87 low-income facilities housing elderly are accounted for by regional housing authorities, adding 11,667 units. Therefore, a total of 151 facilities and 15,328 older adults are available.
The sample size goal of 40 (22+16) sites and 590 (550 originally planned + 40 additional) participants will be easily achieved. Medicaid participation among the tenants is 76% among facilities. From a prior study in these facilities it is expected that the mean age is 74 years, 76% female, 51% Caucasian, 45% African-American, 2% Hispanic, and 2% other racial/ethnic groups; 60% of HUD older adults had none/rare dental visits in over 3 years.
Due to COVID-19 restrictions which prevented outside visitors/providers at the housing sites, recruitment will take longer than the anticipated 24 months. Additionally, approximately 40 enrolled individuals did not complete the 26 week visit/second dose of treatment. We therefore increased the total number of enrolled participants by 40 (included in total above). We also added sixteen additional housing facilities to the study to address some delays in recruitment due to the COVID-19 pandemic.
The additional 16 sites will also be randomized to Arm 1 and Arm 2 for a total of 20 sites in each arm
Recruitment of Participants:
Successful strategies from prior studies will be used in recruiting and input will be solicited from the stakeholders. Strategies that will be used are as follows: Service coordinators at all 40 facilities will serve as site liaisons. The study PI/project manager will provide the service coordinators with an introductory letter/flyer containing study information to be given to tenants and to be posted in public areas of the facility. Service coordinators will first arrange an informational meeting (study dentist will give a talk on the interventions as suggested by our stakeholders), and will arrange a second recruitment meeting for study staff to present information regarding the study at scheduled group events (e.g. tenant meetings, health fairs). For planning purposes, service coordinators will have a sign-up sheet for those interested in the sessions. Study staff will schedule those who are interested and meet the inclusion criteria for one-on-one sessions at the housing facility to obtain informed consent and collect baseline survey data. Baseline dental exam and treatment appointments will then be scheduled at each facility according to a designated exam day(s) for each facility, which will occur approximately 1-2 weeks following consent and baseline data collection. Six and twelve month dental exams/treatment and one-on-one interviews for follow-up survey completion for each time point will also occur at the housing facility where participants reside.
Randomization Procedures:
Randomization is at the level of the cluster (housing facility) for logistical efficiency. This will greatly reduce the potential for error that could otherwise occur with people at the same site assigned to different treatments. Furthermore, keeping the same treatment at each site reduces chances of 'contamination' (i.e. participant discussing their treatment with others). Additionally, stratified cluster randomization will be used, i.e. a block (constrained) randomization approach in which balance over treatments is assured for 2 key cluster-level (stratification) variables, namely facility size (\>100 versus ≤100 residents), and geographic location (Cuyahoga County vs other). The project statistician will generate the randomization scheme for the 40 housing facilities.
Analytic plan:
Each primary outcome will be compared between the SDF and ART+FV groups. For tooth pain, a 95% confidence interval (CI) based on a t-test for the difference in mean responses (SDF minus ART+FV) will be computed. If this confidence interval lies within the interval (-∞, 8) we may conclude 'non-inferiority' of SDF relative to ART+FV treatment. The confidence interval may secondarily be examined to assess possible superiority of one intervention over the other. For arrest rate, a 95% CI for the difference in rates (based on a z statistic) will be computed. If this confidence interval lies within the interval (-0.09, 1) we may conclude 'non-inferiority' of the SDF relative to ART+FV treatment. As above, possible superiority of one intervention over the other may also be assessed. For other outcomes, computing 95% CI for differences in means (or proportions for binary outcomes) will also be used. These secondary outcomes will be assessed in an exploratory manner for possibly superiority or inferiority based on appropriate margins.
To corroborate initial results, a generalized estimating equations (GEE) approach will be used. For each outcome, a GEE (marginal) model will be fit that includes a treatment indicator and prognostic variables (including sociodemographic variables, medical conditions, and oral health behaviors). Appropriate link functions (e.g., logit link for binary outcomes and identity link for continuous outcomes) will be specified and an exchangeable working correlation matrix used to allow for correlations within site. The arrest outcome will be analyzed as a binary outcome (as described in the sample size section), and secondarily as the number of arrested lesions assuming an appropriate distribution (e.g., negative binomial) and link function (e.g., log link). Robust t tests with correction for a small number of clusters will be used to test for treatment effects and corresponding 95% confidence intervals computed.
Secondarily, the above GEE approach will be extended to analyze the repeated (baseline, 26 and 52 week) measures for each outcome. The models for each outcome will include the same prognostic variables as before, as well as time and a time by treatment interaction. Correlations among the repeated measures will be allowed, e.g., using a first-order autocorrelation structure. If a substantial within-facility correlation is found it would be necessary to incorporate facility as a second cluster levels (within which person - the first cluster level - is nested). Additionally, estimation and testing (via a robust t-test) the interaction term to compare trends over time for the two interventions will also be employed. If the use of two cluster levels is not found to be feasible in the GEE approach, a generalized mixed effects model approach will be considered.
Causal Inference standards: The use of randomization and adjustment in regression models should be sufficient to provide causally interpretable intervention effect estimates; special causal inference techniques such as propensity score or instrumental variable methods often indicated for observational studies, will therefore not be necessary for the data analysis. Biases will be avoided in causal inferences about the intended interventions by using an intent-to-treat approach, in which individuals are analyzed according to randomized groups without regard to (extent of) treatment actually received.
Sensitivity Analyses: Sensitivity of conclusions to model assumptions will be checked as well as methodological decisions such as outcome definition. In particular, for the arrest outcome, the primary analysis is based on a binary indicator of 'complete' arrest of all lesions for an individual. Results will be corroborated from this approach with an alternative approach analyzing number of lesions arrested. For the GEE analyses, planned for all outcomes, sensitivity will be assessed in part by using alternative working covariance structures and alternative sets of predictors. For some outcomes, alternative distributions will be considered; for example, for number of arrested lesions, possible distributions include Poisson, negative binomial, zero-inflated Poisson (ZIP), and zero-inflated negative binomial (ZINB). Sensitivity to assumptions regarding missing data will also be investigated as described further below.
Missing Data. (a) Methods to prevent and monitor missing data: All questionnaires (using tablets) and dental assessment forms (using paper), will be managed in an electronic database (REDCap) by study staff. Weekly quality control checks will be run for outliers, entry errors, missing data, and potential data anomalies. Statistical analyses, summary and missing data reports will be generated by the study biostatistician monthly during the study. (b) Statistical Methods to Handle Missing Data and Account for Statistical Uncertainty Due to Missingness: The primary analyses, which use GEE based on all available observations, assume outcomes are missing completely at random (MCAR). As an initial assessment of missing data patterns, intervention groups (including by relevant subgroups and time point) will be compared with regard to missing outcome rates. A test for a relationship between missingness and outcomes at early time points will evaluate the plausibility of the MCAR assumption. Multiple imputation methods (an established method) will be used to help assure valid inferences. (d) Plans to Record and Report Dropout and Missing Data: The trial data will be managed using REDCap software, currently running at Case Western Reserve University (CWRU). REDCap is a secure web-based application providing an intuitive interface for validated data entry, audit trails for tracking data manipulation and export procedures, and automated export procedures for seamless data downloads to common statistical packages. Missing data reports will be generated weekly from REDCap for timely resolution and reporting purposes. (e) Plans to Examine Sensitivity of Inferences to Missing Data Methods As noted above, multiple imputation will be used to obtain valid inferences in the presence of data that are not missing completely at random. Predictive mean matching with an appropriate prediction model will be used depending on the outcome (e.g. logistic regression for binary outcomes, linear regression for continuous and count outcomes). Alternative imputation models will be used as part of sensitivity analyses.
