Study design

This study used data from the Portuguese Epidemiology of Chronic Diseases (EpiDoC) cohort, a nationwide prospective cohort that enrolled a nationally representative random sample of non-institutionalized Portuguese adults (≥ 18 years old) between 2011 and 2013 [11]. There were four total waves of evaluation. The baseline evaluation (EpiDoC 1; 2011–2013) included 10,661 participants who were representative of the Portuguese adult population and was performed in two phases. First, a structured face-to-face interview was undertaken by a trained research assistant to screen for rheumatic and musculoskeletal diseases and collect sociodemographic and health-related data. In the second phase (n = 3877), a structured evaluation was conducted by a rheumatologist during a clinical appointment to validate the diagnosis of rheumatic and musculoskeletal diseases, as described elsewhere [3]. The three subsequent follow-up waves—EpiDoC 2, March 2013 to July 2015, n = 7591; EpiDoC 3, September 2015 to July 2016, n = 5653; EpiDoC 4, March to August 2021, n = 3757—were conducted using semi-structured phone call interviews in which a computer-assisted personal interview (CAPI) system delivered a core questionnaire, similar to that used in the first wave.

Study population

This study included participants with a diagnosis of HKOA that was validated according to the American College of Rheumatology HKOA classification criteria, as described elsewhere [12]. The exclusion criteria were non-responses to the question, “Do you practice regular exercise/sports?,” or answering, “Doesn’t know/doesn’t answer,” and reporting a low physical activity frequency of less than once per week (rarely, sporadically, or occasionally).

Outcomes assessment and definitionPhysical function

Physical function was assessed in the four waves through the Health Assessment Questionnaire (HAQ), which evaluates physical limitations in daily activities through 20 questions with four levels each (without difficulty, some difficulty, with much difficulty, unable to do). A total score was computed for HAQ, ranging from 0 (no disabilities) to 3 (complete disability) [13].

Health-related quality of life

In the four waves, HRQoL was measured using the Portuguese validated version of the EQ-5D-3L questionnaire, which is composed of five dimensions (mobility, self-care, usual activities, pain/discomfort, and anxiety/depression) with three levels each (without problems, some problems, extreme problems). The descriptive system was converted into a summary index score ranging from 0 (equivalent to death; negative values correspond to states worse than death) to 1 (full health) [14].

Exposure assessment and definitionPhysical activity

Self-reported regular physical activity was assessed in each wave through the question, “Do you practice regular exercise/sports?,” with possible responses of “yes,” “no,” and “doesn’t know/doesn’t answer.” The frequency of intentional physical activity per week was defined as the number of days of exercise per week. Regular physical activity was defined as intentional exercise occurring at least once per week. Physical activity frequency was categorized into three subgroups: non-frequent (0 times per week), frequent (1 or 2 times per week), or very frequent (at least 3 times per week).

Covariates assessment and definition

Sex, age, nomenclature of territorial units for statistics II (NUTS II) region (Lisbon, North, Centre, Algarve, Alentejo, Madeira, and Azores), marital status, and education level were considered as sociodemographic variables. The NUTS II regions Madeira and Azores were considered as one Islands region. Marital status was categorized as “with partner” (married or consensual union) or “no partner” (single, widowed, or divorced). Education level was categorized as “ < 4 years” (less than primary education), “4–9 years” (primary or secondary education), or “ ≥ 10 years” (secondary or superior education). Body mass index (BMI) was categorized as “underweight/healthy weight” (≤ 24.99 kg/m2; combined due to the lack of representation in the underweight category), “overweight” (≥ 25 and ≤ 29.99 kg/m2), or “obese” (≥ 30 kg/m2), according to self-reported height and weight. Smoking habits (“never,” “in the past,” or “occasionally or daily”; the latter were combined due to few observations in the “occasionally” category) were also noted.

We defined multimorbidity as having ≥ 2 self-reported chronic non-communicable diseases of the following list of diseases considered at baseline: high blood pressure, high cholesterol, cardiac disease, diabetes mellitus, chronic lung disease, problems in the digestive tract, neurological disease, mental health disorders, allergies, cancer, and hyperuricemia. Self-reported hospitalizations in the previous year were coded as “yes/no.”

Clinical severity was evaluated at baseline with the Portuguese versions of the Knee Injury and Osteoarthritis Outcome Score (KOOS) [15] and the Hip Disability and Osteoarthritis Outcome Score (HOOS) [16]. A composite score encompassing the mean scores of each dimension of the assessments (pain, other symptoms, activities of daily living, sports and leisure, and quality of life) was computed and transformed into a 0–100 scale [17]. For easier interpretation, the inverted normalized mean score (0–100) was used, with higher values corresponding to higher clinical severity, as previously reported [18]. Pain intensity was measured as the mean pain intensity in the previous week with the 11-point Numeric Pain Rating Scale (NPRS) at baseline, and the population was divided into two subgroups: manageable pain levels (< 5 points) and unmanageable pain levels (≥ 5 points), according to the cutoff point for a manageable pain day in OA found by Zelman et al. [19]. In the case of both knee and hip being affected, the worse score of the two was considered.

Time-dependent variables were collected in all waves and included HRQoL, physical function, BMI, regular exercise, smoking habits, multimorbidity, and hospitalizations. The variables considered time-independent (only collected at baseline) were sex, NUTS II region, marital status, education level, disease severity, and unmanageable pain levels. The time in years since the baseline assessment was computed; therefore, only the age at baseline was considered to avoid multicollinearity.

Statistical analysis

A descriptive analysis of the participants by their baseline frequency of physical activity was conducted using absolute (n) and relative frequencies (%) for categorical variables and mean ± standard deviation for continuous variables. Physical activity groups were compared using the chi-squared test (categorical variables) and Kruskal–Wallis test (continuous variables). Linear mixed models were used to assess the association of physical activity frequency (non-frequent, frequent, and very frequent) with physical function and HRQoL over time, considering varying intercepts for each participant and an independent covariance structure. Random slopes and other variance structures were tested but did not improve the models.

Univariate models were computed first to test the significance of potential predictors, with p < 0.25 as the selection criterion (Additional file 1: Table S1). The model-building process included comparing the models through likelihood ratio tests. Potential confounding variables were kept in the multivariate model if the literature supported their effects on physical function and HRQoL or if they achieved statistical significance of p < 0.05. Four models were built and adjusted for years from baseline. The interaction between physical activity and years from baseline was non-significant and thus not considered. NUTS II region, marital status, and smoking habits were also not included. Model 1 shows the crude effect of physical activity frequency. Model 2 was adjusted for sex, age at baseline, and education level. Model 3 was additionally adjusted for BMI. Model 4 was further adjusted for multimorbidity, hospitalizations, clinical severity, and unmanageable pain levels. The equation for the fully adjusted linear mixed model is:

$$_=_+_}_+_\mathrm_+_}_+_}_+_}_+_}_+_}_+_}_+_\mathrm}_+_\mathrm}_+ _+_$$

where physical function and HRQoL scores (\(y_\)) for observation \(i\) in cluster \(j\) depend on the overall mean intercept \(_\), the coefficient \(_\) for the fixed effect \(_\), and the cluster-specific varying intercept \(_\). The overall error term is \(_\). Fewer than 10% of the data were missing, so no imputation techniques were used. A sensitivity analysis was performed considering unmanageable and manageable pain levels subgroups, as well as clinical severity tercile subgroups (low, medium, and high). Analyses were carried out in STATA 17, and statistical significance was assumed at p < 0.05. Plots were derived using R version 4.1.1.