Differential effect of contributory factors in patients and controls

To assess the differential impact of mastery on employment in patients or controls, a regression model was developed with employment (yes/no) as dependent variable. As incidence rate ratios (IRRs, generated by modified Poisson regression) more accurately reflect risks than ORs (generated by logistic regression) when the prevalence of the outcome is >10%, modified Poisson regression was preferred over logistic regression.²¹ Exploratory analysis revealed that patients on TNFi were in a worse health state compared with patients not on TNFi. Likely, these patients had even worse disease at the start of TNFi, and the role of TNFI at the individual level cannot further be explored as this covariate is subject to confounding by indication. Consequently, TNFi use was not included in the multivariable models to prevent biased results. Also, as the aim of this analysis was to compare the differential effect of factors in patients and controls, generic variables (such as SF-36PCS) were preferred over disease-specific factors (such as BASDAI). Variables of interest were thus education (dichotomised, higher education/university versus other), smoking, alcohol use, body mass index, comorbidity, health status (SF36-PCS, SF36-MCS) and mastery. These were first explored in univariable analysis, correcting for gender and age. Next, a basic multivariable model was computed in the total sample including age, gender and group-membership (patient versus control). Subsequently, demographic variables and health variables (SF-36PCS, comorbidity) that were associated with being employed in univariable analyses (p<0.20) were added using a manual forward method, after ruling out collinearity between variables. Variables were retained if significantly associated with the outcome (p<0.05) and/or a confounder (changed the IRR of included variables>10%). As a final step, mastery was added.

Interactions between group-membership and all variables in the model were tested, and p<0.10 was considered sufficient reason to explore stratified analyses. As significant interactions were found between group-membership and mastery (p=0.02) as well as SF-36PCS (p=0.08), further analyses were carried out in patients and controls separately. As an additional interaction was observed between education and mastery in the patient group (p=0.08), the final analyses were carried out in three separate subsamples: controls, patients with lower education and patients with higher education. All statistical analyses were performed with SPSS V.23.0 (IBM, Armonk, New York, USA) and Stata Release 14 (Stata, College Station, Texas, USA).