Procedure and participants

This study belongs to the Physical Activity and Health Equity: Primary Prevention for Healthy Ageing (AEQUIPA) prevention research network [24] and uses data obtained in the second study phase of the PROMOTE-study [25]. The primary aim of the second study phase was to compare the effectiveness of two different physical activity intervention modalities (web- vs. print-based intervention) on changes in physical activity among older adults. Ethical approval for the study was obtained on July 3rd, 2018, from the Medical Association in Bremen. All study participants were fully informed about the study and provided informed consent. The data analyses reported in this paper are of exploratory nature.

A random sample of n = 3492 adults aged 60 years and above residing in Bremen, Northwestern Germany, was drawn from the residents’ registration office and contacted via mail. Additionally, press releases and public talks were used to recruit study participants who could contact the study team and choose to participate after receiving further information on the study. Older adults were included if they provided informed consent and were either inactive or recently active, meaning that they had not been sufficiently physically active for more than one year. Individuals with time and health constraints, as well as those not owning a mobile phone or not being able to use it regularly, were excluded. Further details on eligibility criteria were published in the study protocol [25]. The final baseline study sample consisted of n = 242 individuals (see Additional file 1 for the flow chart). Eligible older adults were randomly assigned to a print-based intervention or a web-based intervention during a telephone interview with a study nurse. The intervention groups were assigned to alternating weeks. During the telephone interview, participants were randomized by having them choose a weekly appointment while being blinded to the intervention condition assigned to the chosen week.

The print-based intervention group (n = 113) received physical activity recommendations based on the World Health Organization guidelines [1], a printed physical activity diary and a brochure with age-appropriate exercises. The web-based intervention group (n = 129) received the same program in the form of a website and an android smartphone-application. The interventions were designed to promote self-monitoring of physical activity, were based on health behavior change theory [14, 26] and incorporated behavior change techniques [27]. A subgroup (30% of the web-based intervention group, n = 38) additionally received an activity tracker (Fitbit Zip, Fitbit, San Francisco, USA), substituting the subjective self-monitoring intervention with an objective self-monitoring component. The interventions were mainly home-based but included face-to-face components. In the first intervention phase, each individual was offered to participate in ten weekly group sessions, covering 60 min of exercise training and 30 min of health education. During the second intervention phase lasting six months, four health education group sessions were offered. Older adults were not blinded to group affiliation once they were assigned to it, and neither were investigators.

Participants completed a self-administered questionnaire and wore an accelerometer for seven days during waking hours on their right hip at baseline (T0, January to April 2019), at the first follow-up (T1, April to July 2019) and at the second follow-up (T2, September 2019 to January 2020). A cognitive screening test was conducted during the first weekly group session. The dropout rate after T2 completion was 33.9% (see Additional file 1 for numbers per intervention group regarding loss to follow-up).

MeasuresPhysical activity and sedentary behavior

Physical activity and sitting time were assessed using accelerometers (GT3X+, ActiGraph, Pensacola, USA). Valid days were identified using the Actilife 6.8.0 software and R 3.6.1. Valid days were defined as having ≥8 h of valid wear-time, with no definition of maximum wear-time. Participants needed to have at least three valid days, including one weekend day. Total minutes of light (0–2690 counts per minute), moderate (2691–6166 counts per minute), and vigorous physical activity (6167–9642 counts per minute), as well as sitting time (0–99 counts per minute) were derived by using one-second epochs for the categorization of counts per minute according to cut-off values considering the vector magnitude [28]. Minutes per week were derived by dividing the total minutes spent in light, moderate or vigorous physical activity, respectively, by the days the accelerometer was worn and then multiplying the value by seven. Additionally, MVPA was derived using 2691–9642 counts per minute and counting only the time spent in bouts of at least ten minutes according to the physical activity recommendations given in the study. The average minutes spent with SB per day were calculated by dividing the total minutes spent with SB in bouts of at least 30 min by the number of the days the accelerometer was worn.

Baseline measures

Demographic information, including sex and date of birth, was assessed as reported in the study protocol [25]. The International Standard of Education (ISCED) [29] was used to code an educational status score, which was dichotomized into “low/moderate” and “high” educational status. Need-weighted income per capita was derived according to the German Microcensus [30] and tertiled into “low”, “moderate” and “high”. Employment was dichotomized into “fully retired” and “other than fully retired”. Body mass index was calculated from self-reported weight and height and dichotomized into “underweight/normal weight” and “overweight/obese”. Cognitive screening was administered using the Mini Mental State Examination 2 - brief version (MMSE-2-BV) [31, 32].

Social-cognitive predictors for engaging in the recommended levels of physical activity were assessed using validated measures as reported in the study protocol [25] and published results of the first study phase [33]. Older adults were asked to rate respective statements on Likert-scales from 1 (= totally disagree) to 7 (= totally agree). For example, intention was assessed with one item which consisted of the statement “I intend to engage in moderate endurance training for at least 150 minutes per week (not tiring, slightly sweating) and strength and balance training twice a week.” Furthermore, the following social-cognitive predictors were assessed: positive and negative outcome expectations (two items, respectively), self-efficacy (one item measuring action self-efficacy, two items measuring maintenance self-efficacy, and two items measuring recovery self-efficacy), action and coping planning (three items, respectively), and habit strength (two items). A detailed description of the assessed social-cognitive predictors has been provided in previous publications [33]. Mean scores were aggregated per social-cognitive predictor (Cronbach’s alpha ranged from .72 to .96) except for negative outcome expectations (Cronbach’s alpha = .65).

Outcome and analysis sample definition

The primary outcome was minutes of MVPA in bouts of at least ten minutes per week, in line with the physical activity recommendations given to study participants. The secondary outcome was the average minutes spent sitting in at least 30-min bouts per day. Subgroups were not defined a-priori as this study’s objective was the identification of unobserved subgroups in terms of latent trajectories (not directly observed). However, based on a systematic review on physical activity trajectories [21], the maximum possible number of trajectory classes was set to six, possibly including the following categories: increasing, stable high, stable sufficient, decreasing moderate, stable insufficient, and decreasing low physical activity.

Older adults were included in the analysis sample if they were cognitively healthy (MMSE-2-BV ≥ 13) and had existing values for the primary outcome on at least one timepoint. The inclusion value for the MMSE-2-BV was ≥15 originally, but it was changed to ≥13 based on previous studies [34, 35]. The analyzed sample (n = 215, see Additional file 1) did not differ from the recruited sample, which was tested considering a set of socio-demographic, psychological and health-related characteristics (effect sizes were all < .20).

Statistical analysesLatent trajectory analysis strategy

Finite mixture models were calculated using an expectation-maximization algorithm for maximum likelihood estimation of model parameters in Mplus version 8.4 [36]. The best-fitting latent trajectory model was determined following the steps proposed by van der Nest et al. [17], and the recommendations provided by the Guidelines for Reporting on Latent Trajectory Studies (GRoLTS) Checklist [37]. The slopes for the three timepoints were set to be 0, 2.66 and 8.35 – according to the median months the measurements lay apart. Latent class growth analysis (LCGA) was conducted to identify latent MVPA and SB trajectories, respectively. LCGA for SB was adjusted for wear-time, as the amount of time the accelerometer was worn correlated with SB.

Investigated fit indices to determine LCGA model fit were the Bayesian Information Criterion (BIC), Akaike’s Information Criterion (AIC), and sample-size adjusted BIC (SABIC). An elbow plot of fit indices was created to visualize the point at which the decrease in fit indices became less in extent. The p-values of the Lo-Mendell-Rubin likelihood ratio test (LMR-LRT) and the bootstrapped likelihood ratio test (BLRT) were considered to determine whether the respective model provided a better fit than the model with one class less. To validate the selected model, the minimum class size was evaluated with the cut-off at 5% and an entropy approaching 1.000 indicating higher certainty. The selected model was critically reviewed for clinical and theoretical plausibility and meaningfulness. The models were rerun using different starting values to ensure that the estimation did not result in local maxima. The dataset including the categorical variable indicating the latent trajectory class was exported to SPSS 26 (IBM Corp. Released 2019. IBM SPSS Statistics for Windows, Version 26.0. Armonk, NY: IBM Corp) to investigate changes within latent trajectory classes and associations with baseline social-cognitive predictors.

Changes by Timepoint and activity-type and associations with social-cognitive predictors

To analyze whether a linear function could describe the data well, changes within the latent trajectory classes between the three timepoints were investigated using paired samples t-tests. Changes in MVPA, SB, light, moderate and vigorous physical activity were analyzed. Associations of latent trajectory class membership with social-cognitive predictors were investigated with independent samples t-tests. An investigation of social-cognitive indicators as predictors of latent trajectory class membership in logistic regression was deliberately omitted. Calculating odds ratios for social-cognitive indicators would provide information on the likelihood of belonging to a latent change trajectory given a one-unit increase in a social-cognitive indicator. Comparing the mean values between groups and testing for statistical significance between them, on the other hand, was deemed more relevant and more suitable with the aims of this manuscript.

Missing data handling and interpretation of effects

Finite mixture models were calculated using full-maximum likelihood estimation, as missing value analysis indicated that the precondition of data missing at random was met. For analyses of changes within latent trajectory classes and associations with baseline variables, missing values were imputed using multiple imputation with predictive mean matching. For imputed data, the mean and standard error (SE) were calculated to report continuous indicators by latent trajectory classes for each assessment timepoint. Cohen’s d was calculated as a measure of effect size based on the pooled mean differences and standard deviations and Cramer’s V was averaged across all datasets. All analyses were carried out under the intention-to-treat assumption. We would like to stress that the analyses of this exploratory study did not serve to evaluate intervention effectiveness by comparing web- and print-based components of the physical activity intervention. These primary outcomes are reported elsewhere [38]. As primary outcome analyses showed that there was no substantial difference between the intervention groups in terms of effects on MVPA or SB [38], the analyses reported here considered all intervention groups as a joint group under the assumption of no differential effect present between the intervention groups.