A longitudinal cohort study was performed, using data from four time points from the Longitudinal Aging Study Amsterdam (LASA)(39), collected from 2001 through 2012. In this period, participants were measured four times; in 2001-2002 (cycle E), in 2005-2006 (cycle F), in 2008-2009 (cycle G) and in 2011-2012 (cycle H). While LASA has been collecting data since 1992, we only included these cycles because visual acuity was only measured at these time points.
LASA’s first cohort was formed in 1992 from a random sample of people aged 55 to 85 years, drawn from population registers in eleven municipalities in the Netherlands. The acquired sample was stratified for age, gender and level of urbanisation. This sample was first used in the NESTOR study on Living Arrangements and Social Networks (LSN)(40). In 2002-2003, a second cohort was formed from an identical sampling frame. This process has shaped representative samples of the older Dutch population and has further been described in detail in previous publications(41-43).In total, data were collected on 2599 unique participants. Selection bias was kept to a minimum by including a very large population, by recruiting participants from three culturally distinct areas with different levels of urbanisation and by contacting members of a general population rather than clinical recruitment(44).
Visual acuity was reported in terms of visual acuity rating (VAR), measured using a Colenbrander 1-meter chart with a +1.00 dioptre magnifying glass (45). For analysis, all obtained VAR scores were converted to log units (logarithm of the Minimal Angle of Resolution, logMAR)(46). Visual acuity of the better eye was used for analysis, regardless of lateralisation.
Validated Dutch translations of widely-used questionnaires were deployed to assess different aspects of mental health. For depressive symptoms, the Center for Epidemiologic Studies – Depression scale (CES-D)(47) was used. The CES-D questionnaire contains 20 items, measuring depressive symptoms on a 4-point Likert scale (scored 0-3). For symptoms of anxiety, the Hospital Anxiety and Depression Scale – Anxiety Subscale (HADS-A)(48) was used. The HADS-A contains 7 items, measuring symptoms of anxiety on a 4-point Likert scale (scored 0-3). For mastery, the Pearlin Mastery Scale (PMS) (30) was used. The PMS questionnaire contains 7 items, measuring mastery on a 5-point Likert scale (scored 0-4). For self-esteem, the Rosenberg Self-Esteem Scale (RSES)(49) was used. The RSES questionnaire contains 10 items, of which the first 4 were included, measuring self-esteem on a 5-point Likert scale (scored 0-4). Items using adverse wording were coded reversely. Thus, higher scores corresponded with greater levels of depression, anxiety, mastery and self-esteem, respectively.
Additionally, other variables were obtained, including age, education, nationality, living arrangement (independent or not), marital status (currently married or not), partner status (partner being present or not), personal network size, functional limitations (i.e. dressing and undressing, chair use, clipping one’s toenails, walking, transportation and stair use), special housing adjustments (none or one and more) and chronic somatic comorbid diseases (i.e. chronic non-specific lung disease, cardiac disease, peripheral artery disease, stroke, diabetes mellitus, arthritis and malignancies). These variables were chosen either to describe essential characteristics of the research population, or because previous literature showed them to be factors of importance in the studied associations. During statistical analysis, age was found to grossly violate the linearity assumption, rendering the variable unfit to be included as a continuous measure. Therefore, age was divided into three groups to facilitate separate analysis for three clinically relevant groups: 1) a working-age population (up to 65 years old in the Netherlands), 2) a middle old population (65 to 90 years old) and 3) the oldest old(90 years and older). For the oldest old we chose a cut-off of 90 years and older since this is the fastest growing segment of the Dutch population and this age group constitutes a growing and distinct group of patients with visual impairment (50). To address the possible issue of information bias, data collection on outcomes occurred in a highly structured fashion and similarly for all participants.
IRT-analyses – also called latent trait analyses – were performed on the questionnaires at all measurement time-points to estimate individual latent trait (θ-)scores per item. These statistical models incorporate the characteristics of questionnaire items and all obtained responses, rendering a more accurate representation of the respondent’s score on the latent construct, which was originally set out to be measured. Item-response models provide certain compelling advantages, describing the relationship between a latent trait, the characteristics of the items in the scale and the answers of respondents to the individual items(51). This results in a more accurate estimation of one’s true latent trait (e.g. level of depression), increasing the validity of the used questionnaires and the accuracy of the obtained results. A Graded Response Model (GRM) was chosen as the preferred IRT-model for its flexibility regarding to item goodness-of-fit(52). In order for IRT-analyses to be accurate, questionnaires should meet the criteria for three crucial assumptions; unidimensionality, local independence of items and monotonicity(53). Unidimensionality was tested using standard indices. Local independence of items and monotonicity were checked by analysing residual covariance and plotting results of Mokken analyses, respectively(54). Confirmatory Factor Analysis (CFA) was conducted to assess goodness-of-fit and the estimated number of fundamental factors in the model. Based upon the retrieved parameters, the acquired θ (ranging from -4.0 to +4.0) was used in further analysis. Data preparation was conducted in RStudio, Version 0.99.896.
Linear mixed modelling (LMM) with maximum likelihood estimation was used to estimate the associations between logMAR visual acuity and mastery, self-esteem, depression and anxiety. LMM is a preferred statistical method for analysing longitudinal data, using random intercepts and fixed variables. This particular model was chosen for its superior properties in dealing with missing values – which were inherent to the design of the study – and its integration of both interpersonal and intrapersonal variance (55). Possible confounding(44) was analysed and adjusted for. Analyses were carried out using logMAR visual acuity as a continuous independent measure for visual impairment. IBM SPSS Statistics, Version 22.0 was used to conduct the analyses.
We hypothesise that mastery and self-esteem mediate the effect between visual acuity (the independent variable) and depression and anxiety (dependent variables) over time. First, an LMM with maximum likelihood estimation was performed to describe the total effect of the independent variable on the dependent variables. Second, LMM analysis was performed to calculate the ‘a-path’; the association between the independent variable and the potential mediators. Third, a final LMM with maximum likelihood estimation was performed to estimate the direct effect (c’) of the independent variable on the dependent variables, whilst controlling for the potential mediators (b), by including it in the model. Subsequently, these three pathways were compared. Again, potential confounding was analysed and adjusted for.