Primary Outcomes
The primary outcomes are 5 Sleep disturbance, Pain, Anxiety, Depression, and Energy (SPADE) scores and 1 physical function score (“SPPADE”) [2,3,4]. Patients will complete the 6 SPPADE numerical rating scales (NRSs) prior to medical oncology clinic appointments at all stages of the trial, i.e., prior to and after E2C2 initiation. These electronic patient-reported outcomes (ePROs) will be more frequently administered to symptomatic patients outside of their appointment-linked assessments as part of the intervention. To avoid potential bias from oversampling of symptomatic patients, only NRSs collected in association with clinic appointments will be included in the primary analysis. SPPADE scores are self-reported measures using 0-10 Likert scales to rate the degree of each symptom. The physical function score is also a 11 point scale. The 6 SPPADE scores will be our primary outcomes.
Secondary Outcomes
In addition to our primary outcomes, we will assess secondary outcomes including a composite SPPADE score (the average of all 6 SPPADE NRSs), depression, anxiety, pain, and additional measures of physical function using NRSs. We will also assess use of healthcare, including hospitalizations, visits to the ED, visits to the outpatient clinic, and calls to the oncology care team. Vital status will also be assessed.
RANDOMIZATION
In a stepped-wedge cluster randomized study, clusters are randomized to sequences; each sequence represents a timeline of implementation for clusters in that sequence, with the transition to implementation for clusters in each sequence occurring in a different step. The 15 cancer care teams represented 11 teams at Mayo Clinic Rochester (RST) and 4 Mayo health system community clinics in Southwest Wisconsin [SWWI], Northwest Wisconsin [NWWI], Southwest Minnesota [SWMN] and Southeast Minnesota [SEMN]. These 15 cancer care teams were randomized to 5 sequences, stratified on 3 factors:
- annual volume, categorized as <4000, 4001-8000, and 8000+ using the calendar year prior to E2C2 initiation.
- RST team cancer type: head/neck, genitourinary (GU), lung, sarcoma, gastrointestinal (GI), breast, gynecological, brain, and general/endocrine, and mixed (the 4 community sites)
- RST GI team: Mayo Rochester has 2 GI cancer care teams; internally (and in Figure 1) they are denoted the “red” and “blue” GI teams.
The interactions of these three factors included many empty cells, resulting in only 4 distinct strata. Randomization was performed using a computer program which allocated clusters by strata to sequences. All clusters were randomized by the study statistician (JH) prior to study start, and the allocations revealed to all clusters at the beginning of the study. The assignment to strata and sequences is illustrated in Figure 1.
ADDITIONAL FEATURES
Details on sample size, data collection and management, secondary outcome determination, and implementation assessment were detailed in our prior protocol [1]
Statistical Analyses
What follows represents an updated analytic approach. The original analytic plan described weekly cluster rates as the unit of analysis, while also stating that the effect of the intervention on the weekly rates would be adjusted for patient clinical and demographic factors. Recognizing that these assertions were inconsistent, we have revised the statistical model to one which uses each patient symptom score as the unit of observation. The updated analytic approach also addresses the additional primary hypotheses described above. All methods continue to be appropriate for SW-CRTs and will be performed on an intention to treat basis. [5] All assessments are for equivalence. All analyses will be performed at the conclusion of the study period.
Aim 1 Analyses
The unit of analysis is the patient encounter. We will summarize patient characteristics (e.g., sex, age, cancer type and stage, insurance status) by intervention status (pre-E2C2, E2C2) of the encounter. All patient encounters will be analyzed on an intention to treat status; this principle will be extended to the cluster status, so that delays in implementation of an intervention will not affect the intervention status of patients.
We will include a common set of variables in all models testing the effectiveness of the intervention. Cancer type and stage will be included, as well as 7 variables selected as likely to impact the use of the portal and completion of the surveys: from the EHR we will capture age (categorized into 10-year increments), sex (male or female), and severity of illness (Elixhauser score as a categorical variable). We will also link the patient 9-digit zip code to Census data to get estimates of educational level, income, and broadband access, which will be categorized into quartiles.
Hypotheses H1a
To test the primary hypothesis H1a for our primary outcomes, we will use generalized linear mixed models to assess the effects of the intervention on each of the 6 SPPADE scores [5]. All models will account for correlation of outcomes within clusters and across patients, and for changes in this correlation over time. We will adjust for time, measured by study period, to account for temporal effects. Each score will be adjusted for the prior score, with a coefficient that depends on the time t since the prior score; to account for potential changes in correlation of outcomes by over time we will include autoregressive terms at the patient level. [6] To account for changing correlation within clusters over time we will incorporate a discrete-time correlation structure. [7]
Our primary analysis will treat each SPADE score as a continuous variable and employ a mixed effects generalized multivariate model; we will model all 6 scores jointly, which will allow us to account for the correlation between them and avoid correcting for multiple hypothesis tests.
Specifically, let Yijkm be the ith score (i = 1,…,6) of the jth encounter for the kth patient in the mth cluster (m = 1,…,15), for j = 2,… ; the first score for each patient (j = 1) is a baseline score collected during the first visit, not an outcome. Let w = 1,…,6 indicate the trial step during which the survey is collected.
Then to test the hypothesis that the intervention impacts the 6 symptom scores we will estimate the joint models:
where
\({\text{X}}_{jkm}\) is an intervention indicator which equals 1 if encounter j for patient k in cluster m was post intervention and 0 otherwise;
\({\varvec{B}}_{\varvec{i}}\) is a vector of fixed effects for the intervention on symptom score i, i = 1,…,6;
\({\varvec{L}}_{\varvec{i}}\) is a vector of time dependent lags, representing the effect of the most recent measurement on the current one, as a function of time t in days between measures;
Z k is a vector of patient factors for patient k, F is a corresponding vector of fixed effects;
S m is a cluster level set of indicators for randomization strata, D is a vector of fixed effects;
Q is a vector representing the secular time effect, where qjmk is the calendar quarter; and,
the time elapsed since encounter j-1; for j = 2,….
In addition
A ikm is a vector of random effects, [A1km,…,A5km] specified as varying across patients nested within clusters
![](data:image/png;base64,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)
where Ss are exchangeable correlation matrices.
µ imw , i = 1,…,6 is a discrete-time vector of random effects for each cluster m and step w; and,
n it , i = 1,…,6 is a multivariate autoregressive process of order 1, AR(1) with Gaussian noise, to account for time dependent correlation of responses within each patient over time; and,
Prior to estimating this model, the distributions of the SPPADE scores will be examined and, if appropriate, transformed to have approximately normal distributions. Covariates will be similarly reviewed and either transformed or categorized before inclusion in the mode.
In general, fitting the above models (1) which has several random effects (in both time and across clusters) can be challenging. However, Markov Chain Monte Carlo estimation is very effective at fitting such models and are appropriate for assessing stepped-wedge cluster randomized designs. [8, 9, 10] Moreover, by incorporating historical time trends into our priors for the effect of calendar time, a Bayesian model can increase statistical power relative to a frequentist approach.[11] All other priors will be flat. In the event that model (1) can’t be estimated, we will simplify the error structure and other features using review of variable distributions and correlations, including intra-cluster correlations and time dependent correlation, to inform the simplifications. Then, using our final model we will calculate posterior probabilities and credible intervals to assess whether coefficients differ from zero; specifically, we will achieve the primary aim of the study by calculating the posterior probabilities that each of \({B}_{1},\dots ,{B}_{6}\) is greater than 0; we will accept that the intervention is effective if PP[abs(Bi ) > 0, i = 1,..6] ≥ 0.80.
We will estimate models using the STAN software package.[12]
Hypotheses H1b
To test whether the intervention has a differential effect on those patients with at least one elevated symptom score, we will construct an indicator, Eijkm, varying by measurement time, which indicates whether the ith prior symptom score was elevated (> 3). By interacting these indicators with the main intervention effects Bi we can assess the impact of the intervention on the corresponding scores of patients with at least one elevated score:
We will then test the hypothesis as above by calculating the PP[Abs(Di) > 0, i = 1,…,6]
Secondary And Sensitivity Analyses
In a secondary analysis, we will classify each symptom score into low (0–6) or high (7–10) and estimate a model similar to (1) and (2) using a logit link. In sensitivity analyses we will replicate the main analyses where models (1) and (2) have been modified to include a) time as calendar week, implemented with splines b) an interaction between step (1–5) and intervention.
Secondary Outcomes
For secondary outcomes of unplanned hospitalization or ED visit, therapy adherence, quality of life, and survival, we will estimate generalized linear models analogous to (1) but with single dependent variable (i.e., non-multivariate) and appropriate link functions.
To address a secondary objective of this aim, to better understand the effectiveness of components of the E2C2 intervention in reducing symptoms, we will perform a separate set of analyses including only the intervention patients to assess the impact of specific process measures on SPADE scores and other clinical outcomes. For each process measure we will first examine the distribution of values (frequency or duration) and, if highly skewed, categorize into 2 or more categories. We will then estimate for each outcome a mixed effects linear model similar to (1) above, though with a single dependent variable. By testing for an overall process measure effect, we can assess which processes contribute most to reducing outcomes; we will also report measures of variance explained to enable relative comparison of process measure effects.
For all models we will report measures of fit such as predictive R2 values and variance estimates, as well as convergence measures for MCMC estimation.
Specific Aim 2 Analyses
The analysis for Aim 2, which is restricted to only intervention encounters, is updated only to include the same patient factors included in Aim 1. For evaluation of the implementation bundle, we will summarize patient characteristics by intervention status (baseline and intervention). We will also summarize all survey responses by cluster implementation status, including clinician characteristics (age, sex, and years of practice). As with Aim 1, analyses will be based on intention to treat at the patient level and at the cluster level.
We will assess the effect of process measures on SPADE scores and other clinical outcomes. To assess the effect of clinician experience on implementation processes, we will assess the relationship between clinician scores and implementation outcomes by modeling each outcome as dependent on clinician scores. For count data, we will use appropriate models and select standard or zero-inflated Poisson or negative binomial models according to the Akaike information criterion. For binary measures, we will use logit models. For each outcome we will estimate bivariate models, including only one survey score, a model with all scores, and a final model that also includes clinician characteristics. These models will include the patient factors included in model (1): cancer type and stage, as well as 7 variables selected as likely to impact the use of the portal and completion of the surveys: age, sex, and severity of illness (Elixhauser), and, according to 9-digit zip code, area estimates of educational level, income, and broadband access.
We will use the measures of adoption, fidelity, and penetration (described above) to assess the differential impact of E2C2 components on patient and clinician outcomes. Formally, we will undertake a mediation analysis with structural equation modeling to assess the extent to which each component of E2C2 mediates the intervention. We will use models appropriate for count data, selecting standard or zero-inflated Poisson or negative binomial models according to Aikake’s Information Criteria, or logit models for binary measures. For each outcome we will estimate a) bivariate models including only one survey score; b) a model with all scores; and c) a final model which also includes clinician characteristics.
Specific Aim 3 Analyses
Analyses for this aim will be analogous to those for Aim 2, though restricted to elderly and rural dwelling patients.