A retrospective analysis was conducted using an existing database of responses to the National Health and Wellness Survey (NHWS), a self-reported cross-sectional survey designed to reflect the general US population including individuals who report diagnosis of MS. The NHWS is an internet-based general health questionnaire distributed to a sample of the adult population. Respondents were qualified if they were ≥18 years of age, able to read and write English, and electronically provided informed consent. Respondents were recruited through opt-in email, co-registration with MySurvey.com partners, eNewsletter campaigns, banner placements, and both internal and external affiliate networks, using a stratified random sampling framework to ensure representativeness of the US population in terms of age and gender. Additional details about the NHWS have previously been published.23-25
Data from 2015 to 2016 (2015 NHWS, N=97,700 and 2016 NHWS, N=97,503; total N=195,203) were analyzed. If an individual completed the survey in both years, the response in the most recent year was used. Respondents who reported being diagnosed with MS by a doctor and indicated RRMS as the type of MS were included in the RRMS group. Potential control respondents were selected from those who self-reported as not having a diagnosis of MS.
The following patient and disease characteristics were evaluated: age, gender, employment status, annual household income, marital status, education, possession of health insurance, body mass index (BMI), smoking status, alcohol use, exercise, Charlson comorbidity index score (CCI),26 and emotional issues such as anxiety and depression, and sleep problems. The CCI allows for adjustment of baseline comorbidity between groups, and is a widely used comorbidity index in studies that use administrative health data.27 The higher the score, the more likely the predicted outcome will result in mortality or higher resource use.28, 29 MS characteristics included severity of MS, symptoms, fatigue, and perceived cognitive impairment.
The HRQOL was measured using Short Form (SF)-36v2 and EQ-5D. In SF-36v2, the HRQoL was captured by the physical (PCS) and mental (MCS) component summary scores 30, 31 Both the MCS and PCS have a theoretical range of 0–100.25 Higher scores on these measures indicate better HRQoL. The EQ-5D was used as a utility measure of health and was expressed as a health utility index score.25 Previously, minimally important differences (MIDs) have been defined by differences of 5.0 points for MCS and PCS scores and 0.074 for the EQ-5D.25, 31-33 The impact on labor force participation was measured by defining employment status as currently in the labor force (full-time employed, part-time employed, self-employed, or not unemployed but looking for work) or not currently in the labor force (retired, disabled, not employed and not looking for work). The Work Productivity and Activity Impairment-General Health scale (WPAI-GH) assessed work productivity loss and activity impairment.34 The WPAI-GH contains six questions.35, 36 The WPAI-GH captures absenteeism (% of work time missed because of one’s health), presenteeism (% impairment while at work because of one’s health), overall work impairment (% of overall work impairment due to health; a combination of absenteeism and presenteeism), and activity impairment (% of impairment in daily activities because of one’s health).35 WPAI outcomes are expressed as impairment percentages, with higher numbers indicating greater impairment and less productivity. Absenteeism, presenteeism, and overall work impairment were calculated for employed respondents only, whereas activity impairment was calculated for all respondents.
HCRU included visits to healthcare providers (HCPs), general physician (GP) or primary care physician (PCP), specialists (e.g., neurologists), emergency rooms (ER) and hospitalizations in the prior six months.
Two-sample comparisons using Chi-square tests for categorical variables and one-way ANOVAs for continuous and count variables were conducted between employed respondents diagnosed with RRMS and those not diagnosed with MS to characterize the two populations and determine baseline variables for propensity score matching.
Propensity Score Matching
Propensity score matching is used to obtain similar groups of treatment and control subjects by matching individual observations on their propensity scores.37 Employed individuals who reported a diagnosis of RRMS were propensity matched to employed individuals without a diagnosis of MS at a 1:4 ratio based on the survey year, age, gender, education, type of health insurance, BMI, and comorbidity burden as assessed by the CCI. These demographic and patient characteristics were included as criteria for the propensity score match in order to control for differences between the two groups. Balance post propensity match was examined using ANOVA, chi-square tests and p-values for variables which were significant at >0.05 were considered to not be balanced.
Variables included in the match were entered into a logistic regression to predict presence of RRMS (vs. no MS) and propensity scores were saved from this model to match each individual with RRMS to four individuals without MS using a greedy-matching algorithm. This identified controls to match to a single case at up to eight decimal places of the propensity score (and as little as one decimal place, if no other suitable control was identified).38, 39
Bivariate analyses using Chi-square tests for categorical variables and one-way ANOVAs for continuous variables were conducted for the employed RRMS vs. no MS groups on patient characteristics to determine whether balance was achieved post-match. Then, outcomes (e.g., SF-36, EQ5D, WPAI, HCRU) were compared between groups (employed RRMS vs. no MS) using one-way ANOVAs. This was repeated in an analysis comparing RRMS and non-MS controls.
An additional analysis was conducted among employed RRMS individuals. MS characteristics, symptoms, and outcomes (e.g., SF-36, EQ5D, WPAI, HCRU) were described by level of work impairment. Levels of work impairment were defined by tertiles based on the observed distribution of the response variable. Chi-square tests (for categorical variables) and one-way ANOVAs (for continuous variables) were used to compare demographics, health characteristics, and health and economic outcomes by levels of work impairment. All multiple pairwise comparisons were conducted with t-tests (continuous variables) or z-tests of column proportions (categorical variables) and adjusted using the Bonferroni correction. P<0.05 between groups is considered as the level of significance. All analyses were performed using SPSS 23.0 and SAS 9.4.