Earnings among people with multiple sclerosis compared to references, in total and by educational level and type of occupation: a population-based cohort study at different points in time

Objectives To investigate earnings among people with multiple sclerosis (PwMS) before and after MS diagnosis compared with people without MS, and if identified differences were associated with educational levels and types of occupations. Furthermore, to assess the proportions on sickness absence (SA) and disability pension (DP) in both groups. Design Population-based longitudinal cohort study, 10 years before until 5 years after MS diagnosis. Setting Working-age population using microdata linked from nationwide Swedish registers. Participants Residents in Sweden in 2004 aged 30–54 years with MS diagnosed in 2003–2006 (n=2553), and references without MS (n=7584) randomly selected by stratified matching. Outcome measures Quartiles of earnings were calculated for each study year prior to and following the MS diagnosis. Mean earnings, by educational level and type of occupation, before and after diagnosis were compared using t-tests. Tobit regressions investigated the associations of earnings with individual characteristics. The proportions on SA and/or DP, by educational level and type of occupation, for the diagnosis year and 5 years later were compared. Results Differences in earnings between PwMS and references were observed beginning 1 year before diagnosis, and increased thereafter. PwMS had lower mean earnings for the diagnosis year (difference=SEK 28 000, p<0.05), and 5 years after diagnosis, this difference had more than doubled (p<0.05). These differences remained after including educational level and type of occupation. Overall, the earnings of PwMS with university education and/or more qualified occupations were most like their reference peers. The proportions on SA and DP were higher among PwMS than the references. Conclusions The results suggest that the PwMS’ earnings are lower than the references’ beginning shortly before MS diagnosis, with this gap increasing thereafter. Besides SA and DP, the results indicate that educational level and type of occupation are influential determinants of the large heterogeneity of PwMS’ earnings.

l 50: the estimate is SEK28,000 but the CI limits are negative, this is a bit awkward, I would just make them all positive and note that you did say "lower earnings" to give direction. l 54 alike should be like l 86 "may" I think you can cite literature to make a stronger claim that it does.

GENERAL COMMENTS
Dear authors Thank you for this nicely written and systematically constructed paper on "earnings among people with multiple sclerosis compared to references, in total and by educational level and type of occupation: a population-based cohort study at different points in time" Your findings are much in line with previous research and adds important knowledge to this interesting topic.

VERSION 1 -AUTHOR RESPONSE
Reviewer: 1 Reviewer Name: John Pearson Institution and Country: University Otago Christchurch -New Zealand Please state any competing interests or state 'None declared': None declared 1. Very nice dataset that directly quantifies the loss of earnings between people with MS and their peers. Longitudinal data like this is a particularly strong addition to the literature. The data has been appropriately analysed and conclusions are well supported by the data. The authors have struck a good balance between sufficiently describing their statistical methods and maintaining readability, although I have asked below for a little more detail. It is also usual to include what software was used for the analysis. As it stands this is an interesting addition to the literature, however I'll take the opportunity to ask the authors to dig a little deeper.
Authors' response: Thank you very much.
We have now revised the Methods section (See page 12-13, lines 280-283) to include the software used: SAS v9.2 in the data management and regression models and Excel for the differences in mean earnings calculations presented in Table 2. 2. Can you include an anova type analysis for each model to demonstrate statistical support for each factor in the models. Were there any interaction effects, particularly between MS and Age, MS and Education, MS and Occupation? That is, does your data support a difference in the earnings~age curve for MS and non-MS? It is possible and likely given figure 2, that the relationship is non linear so perhaps GEEs or non linear effects might be worth trying. It would also be worth exploring these differential effects for the other covariates. There is some evidence in the literature for sex differences too, and I think your data could be used to address this, for example do females with higher qualifications have a greater loss of income than their male counterparts in this dataset?
Authors' Response: Thank you for the suggestion to delve deeper into the data with your engaged response.
We have now included models which consider the interactions between MS*Education (Model 1b) and MS*Occupation (Model 2b) as requested (See page 12, lines 274-278 in Methods, and page 20, lines 389-397 in Results). We have included the table presenting these interactions below for your interest, however, in the interest of brevity we have summarised the key findings in text within the revised manuscript.   Year of inclusion; for people with MS: year of first MS diagnosis and for references: year of inclusion to one of four panels. Estimates from Tobit-models with lower bound set at zero (i.e., no earnings).
The covariates included in the regression models were (MS (yes/no), panel membership, sex, age at inclusion, country of birth, and type of living area), in addition to level of education and type of occupation. We did not use a statistical method to motivate the selection of these 'base' covariates, rather they were selected in light of the literature, for example, age at inclusion as a proxy for age at onset given that progression of MS is to some degree age-dependent (Confavreux & Vukusic, 2006 Brain). Age, alongside sex, was also used in selecting our matched reference group.
Regarding your comments about the possibility of GEE analyses, the aim of this study was to conduct an initial explorative test to investigate if there were any differences. The exact specifications of such potential differences are something that can be investigated further in other studies. The data allowed for a wide choice of statistical methods, and we agreed on Tobit regression models for this study. The main reason was that Tobit regression models are often considered in cases with clustering of values above and/or below certain thresholds. In cases without such clustering, results from a Tobit model are near identical to the corresponding results from an ordinary least squares regression (OLS). In this study, the proportion of clustered values (i.e., earnings=0) was about 12%. Further, given that all analysed covariates were categorical the results should be interpreted as a shift in level of mean earnings vs. the reference category. Since this is in line with the aim of the study we argue that this was the most appropriate method to explore the data.
However, we have crudely tried GEE models from year of diagnosis until end of study (T0-T+5). We included time as a categorical variable and an interaction time*MS, otherwise all other variables were the same as in the Tobit models, i.e. educational level and type of occupation in separate models. The GEE models were estimated using a one-step autoregressive correlation within the individual (AR (1)). This assumed that an individual's observations are less correlated the longer time has passed between the two observations. For example: For individual X, the earnings observed at T+2 are assumed to be more highly correlated with T+3 than T+4 or T+5. Overall, the estimates from the GEE models (presented here) would thus have resulted in the same interpretations of the differences between the two groups in T0 and T+5. Both regression techniques resulted in negative estimates which were in similar ranges at T0 and T+5. However, unlike the presented Tobit regression results in the manuscript, these results do not take into account the relatively high proportion of zeros (i.e., no earnings).
The GEE models we have presented here used time as a categorical variable, which is not ideal for describing the linearity. Our results give basis for suggesting that the specific nature of potential linearity could be further investigated in future studies which focus on longitudinal statistical methods, potentially by using piece-wise regressions with knots at 1 year before diagnosis and 1 year after.
We also agree that age is a very important variable in MS studies, given the age-dependency of MS prognosis (Confavreux C &Vukusic S. 2006 Brain). However, age has duly received much attention in the literature, and given our focus on education and occupation, we decided to match on age (and sex) when selecting the reference group, and age was accordingly included as a covariate in the regression models.
With regards to the comment concerning interactions between sex and education/occupation: We agree that such analysis would be of interest, especially given that the prevalence of MS is higher among women and sex differences in labour market attachment/earnings. However, to achieve this, either a complex three-way interaction is required in the Tobit regressions (MS*Sex*Education/occupation) or a stratified analysis including only the MS group. Such focus within the MS group is beyond the scope of this study. Future studies, stratifying by sex are needed, based on other aims than of this study. Hopefully, our results will inspire to conduct such studies, preferably based in gender theories. minor points: 3. l 50: the estimate is SEK28,000 but the CI limits are negative, this is a bit awkward, I would just make them all positive and note that you did say "lower earnings" to give direction.
Authors' response: We have now substituted the confidence intervals in the abstract with p-values, to avoid confusion (See page 3, lines 52-53). In addition, we noticed a discrepancy in our Results section where we presented p-values in text and 95% confidence intervals in Table 4 . Of course this can vary greatly from the situation in countries with other social security systems. However, to maintain our current focus of this paper on earnings from paid work, we agreed that adding such information so early in the paper could be misleading. Brundin et al, (2017 Mult Scler) present the findings for Sweden in a multicountry questionnaire, notably finding high numbers of part-time workers, and most (78%) of those in paid work self-reported that MS affects their productivity at work.
Secondly, there is large heterogeneity within the population with MS in any given setting, with regards to disease characteristics, for example, severity. Severity can be chronic, progressive but also with periods of small/undetected disability. Disability in terms of EDSS score has been found to be associated directly with earnings (Kavaliunas et al, 2015 PLoS One).
Thus we claim that may is the appropriate wording. Please leave your comments for the authors below Dear authors Thank you for this nicely written and systematically constructed paper on "earnings among people with multiple sclerosis compared to references, in total and by educational level and type of occupation: a population-based cohort study at different points in time" Your findings are much in line with previous research and adds important knowledge to this interesting topic.