Sample and procedure
This study recruited participants via Rare Patient Voice, LLC; patient-advocacy groups; and word of mouth (i.e., snowball technique). Eligible participants were age 18 or older, able to complete an online questionnaire, and were providing caregiving support to a family-member with DMD at least two years old, usually their son. Ineligibility criteria included not being of age 18, not being able to complete an online questionnaire, not providing caregiving support to a family-member with DMD who was at least two years old, and not being able to provide informed consent. Caregiver-participants with motor, visual, and/or other problems that made it difficult for them to complete the web-based survey instrument, enlisted the assistance of a household member to enter their survey answers. This survey was administered through the HIPAA-compliant, secure Alchemer engine (www.alchemer.com) from June to November 2020. Dillman’s Tailored Design Method [26] was followed to maximize response and data quality. This method specifies a process regarding all communication to study participants (e.g., structure and content of all emails) and survey implementation to reflect professionalism and attention to detail (e.g., end-user experience of survey, pretesting, tracking responses and persevering in email invitations, etc.).
Recruitment was stratified by age of the caregiver’s child with DMD: 2–7, 8–12, 13–17, and > = 18. These strata broadly correspond to the disease-related phases of progression: ambulatory (age 2–7), transitional (up to age 12), and non-ambulatory (age > = 13), with increasing dependence and involvement of other systems as the person ages into adulthood (age > = 18). If caregivers had more than one person with DMD for whom they were providing caregiving support, they were asked to report on the eldest or most disabled person (the index patient). Caregivers were paid $75 honoraria for their time completing the survey. The protocol was reviewed and approved by the New England Independent Review Board (NEIRB #20201623), and all participants provided informed consent prior to beginning the survey.
Measures
Person-reported outcomes (PROs) were used to characterize and describe caregiver-impact groups. Characterizing the groups focused on caregiver impact, DMD care recipient’s disability, and out-of-pocket expenditures.
Caregiver Impact was assessed using a DMD adaptation of the Hemophilia Caregiver Impact measure [27, 28]. Items were modified to refer to DMD and were pretested for content via telephone interviews. Evidence for the content validity, construct validity, and internal consistency reliability of this DMD adaptation of the measure were obtained through initial analyses of this data set [25]. The resulting 39-item DMD Caregiver Impact measure (DCI) includes seven negative-impact subscales (Practical, Physical, Financial, Symptom, Lifestyle, Social, and Emotional) and one positive-impact subscale (Positive Emotions). Scores use a standardized T-score metric, with a mean of 50 and a standard deviation of 10.
Care recipient’s disability was measured using the Patient-Reported Outcome Measurement Information System (PROMIS) Parent Proxy (PPP) item banks [29]. These were adapted for use with and validated in this same sample of DMD caregivers in an earlier study [30]. The PPP measure included four domain scores where higher scores reflect better outcomes (mobility, upper extremity, cognitive function, and positive emotion) and four where higher scores reflect worse outcomes (fatigue impact, strength impact, negative affect, and sleep device symptoms). In the present work, all scores used a standardized T-score metric, with a mean of 50 and a standard deviation of 10.
Out-of-pocket expenditures assessed included binary variables reflecting the following 11 home and vehicle modifications that might have been implemented to accommodate their child’s DMD: entrance, bathroom, doorways, van, bedroom, kitchen, elevator, new home, ceiling lift, scooter, and other (open text option to specify). We refer to these modifications as “expenditures” because they involve financial costs, in contrast to other accommodations that might not be directly related to financial costs, such as choosing clothing items that do not involve buttons.
Statistical analysis
Initial analyses sought to use Rasch modeling to summarize the out-of-pocket expenditures on a single unidimensional scale, but these analyses did not result in good item- or person-fit or strong correlations with the DCI subscales. Accordingly, we used a simpler approach, computing independent-sample t-tests and related effect sizes for group differences on each out-of-pocket category and DCI subscale. Using Cohen’s criteria, a d of 0.2 to 0.49 is considered a small effect size, 0.5 to 0.79 is medium, and 0.8 or greater is large[31].
To investigate the interplay of DMD caregiver impact and patient disability domains, two multivariate analyses were implemented and compared. Hierarchical cluster analysis [32] was used because it accommodates binary, categorical, and continuous variables and skewed distributions. It is flexible and effective even in small samples. Three- and four-cluster solutions were examined, and a four-cluster solution was selected because it better discriminated groups, had a reasonable sample size within each cluster, had large between-cluster variance relative to within-cluster variance (i.e., eta-squared in analysis of variance (ANOVA)), and because multinomial logistic regression showed good ability of the 16 DCI and PPP scores to predict cluster membership. A radar chart was used to display the clusters by DCI and PPP scores.
The second multivariate approach utilized Latent Profile Analysis (LPA) [33]. LPA is a latent variable model based, like factor analysis and modeling, on item response theory (IRT) [33]. The latent variable is conceptualized as continuous in factor analysis and IRT, but as categorical in Latent Class Analysis (LCA) and LPA. While LCA uses categorical indicators to measure the latent trait, LPA uses continuous indicators. LPA models were tested evaluating one-, two-, three-, and four-profile solutions. Final model selection was based on entropy values (criterion of > 0.80), an index based on the uncertainty of classification [34] that indicates the level of separation between classes [35, 36]; p-values of the Lo-Mendell-Rubin (LMR) [37] and bootstrap likelihood ratio (BLRT) [38, 39] tests; and the Akaike and Bayesian Information Criteria (AIC and BIC) [40,41,42], both of which are based on the maximum likelihood estimates of the model parameters for selecting the most parsimonious and informative model [35]. Cross-tabulations were obtained to examine profiles by child age.Footnote 1 The two multivariate approaches were compared on the basis of average explained variance of the 16 included scores, and on the strength of the association between membership categories (Cramer’s V).
Statistical considerations and power related to sample recruitment
We aimed to recruit a minimum of 130 caregivers per age stratum, allowing for 20% attrition while still yielding 80% power, α = 0.05, to detect a medium effect size in a multiple regression model with seven covariates [32]. A medium effect size is a common standard for clinical significance [43].
IBM SPSS version 27 [44] and Mplus version 8.5 [36] were used for all analyses.