We found that both stronger baseline social isolation (values for models 1/2: β = −5.6/–5.7 mm³/point on the LSNS (pt), FDR-corrected q-value (q) = 0.0034/0.0078) and increases in social isolation (β = −4.7/–4.5 mm³/pt, q = 0.0035/0.0066) significantly predict smaller hippocampal volumes independent of confounders (Table 2, Figures 1—3). Significance of these findings is further underlined by Bayes factors of 15–19 for baseline social isolation and of 2–3 for change in social isolation. The effect size of one point on the LSNS is equivalent to a 2.5-month difference in baseline age.

Figure 2

Download asset Open asset

Figure 3

Download asset Open asset

In analogous linear mixed effects models, we tested the effects of social isolation on cognition, measured using domain-specific composite scores based on z-scored results of the trail-making test (TMT A and B) and the CERAD-plus test battery (CERAD – Consortium to Establish a Registry for Alzheimer’s Disease, RRID:SCR_003016) assessed under standardized conditions (Beyer et al., 2017). Overall, stronger baseline social isolation and to a lesser extent increases in social isolation were linked to worse cognitive performance (Table 3, Figure 1). Specifically, stronger social isolation at baseline significantly predicted lower executive functions (β = −0.028/–0.017 SD/pt, q = 9.6e-09/0.0014) and lower processing speed (β = −0.018/–0.017 SD/pt, q = 1.2e-05/3e-04). The link to lower memory (β = −0.014/–0.008 SD/pt, q = 0.0016/0.0914) was strong in model 1 but did not survive FDR-correction when controlling for additional covariates. Increases in social isolation over time significantly predicted lower memory in models 1 and 2 (β = −0.019/–0.0018 SD/pt, q = 0.0034/0.0142) but not processing speed (β = −0.007/–0.008 SD/pt, q = 0.238/0.198) and executive functions (β = –0.003/0.001 SD/pt, q = 0.41/0.69). Very high Bayes factors corroborate and substantiate the evidence for the negative effect of baseline social isolation on cognitive functions. Figures 2 and 3 allow comparisons of these effects with other predictors for the different dependent variables.

We did not observe interaction effects of social isolation on hippocampal volume or cognitive performance with age. Appendix 1—tables 4–6 provide a comprehensive summary of all LMEs and predictors including covariates.

To explore whether social isolation affects regional cortical thickness, we conducted whole-brain vertex-wise linear mixed effects analyses on FreeSurfer-derived 3D cortical maps (Reuter et al., 2012). In model 1, we found a total of eight clusters of significantly decreased cortical thickness associated with stronger baseline social isolation after FDR-correction with an α-level of 5% (Figure 4). The clusters were located in the left precuneus, cuneus, precentral gyrus and posterior cingulate gyrus, and right supramarginal gyrus and cuneus. Increases in social isolation over time were linked to decreased cortical thickness in one cluster in the right superior frontal gyrus (Figure 5). When additionally controlling for cardiovascular covariates (model 2), no significant clusters were detected. Table 4 lists these clusters, their locations, and sizes.

Figure 4

Download asset Open asset

Figure 5

Download asset Open asset

Turning to the stress-buffering hypothesis, we investigated whether perceived stress, measured using the Trierer Inventar zum chronischen Stress (TICS) (Schulz and Schlotz, 1999), mediated the relationship of social isolation and hippocampal volume. Moreover, we investigated whether hippocampal volume mediated the association between social isolation and cognitive functions. Specifically, we investigated the indirect path resulting from the regressions of follow-up mediator on baseline LSNS and follow-up dependent variable on baseline mediator.

Neither the mediation analyses with chronic stress as a mediator (n = 51 complete observations) nor the mediation analyses with hippocampal volume as a mediator (n = 341–360) yielded significant results. Due to the requirements of the model design and over 50% missingness in the stress questionnaire, the sample sizes of the mediation analyses were gravely diminished. Details on the mediation analyses are provided in Appendix 1—table 7.

In addition to these pre-registered analyses, we conducted sensitivity analyses to test the robustness of our results on hippocampal volume and cognitive functions. These included possible effects of the Covid-19 pandemic, effects related to the definition of exclusion criteria or confounder specificities. Analyses accounting for (a) potential effects of measurements before compared to during the Covid-19 pandemic, (b) reducing the exclusion criteria (i.e. not excluding cognitively impaired participants, participants taking centrally active medication, and participants with recent cancer treatment), (c) only including participants with two timepoints and using mean and within scores, (d) using a hypertension cut-off of 140 mmHg, (e) using an MMSE cut-off of <27, (f) additionally controlling for physical activity, (g) additionally controlling for sleep quality, and (h) standardizing cognitive functions using the baseline mean rather than the grand mean confirmed the regression coefficients of our models in terms of direction and size (Appendix 1—tables 8–15).

Moreover, we found that treating social isolation as a dichotomous variable, using the standard LSNS cut-off of 12 points, led to results very similar to those of our analyses with continuous LSNS scores (Appendix 1—tables 16–18). However, we found no evidence for an interaction effect of continuous and categorical LSNS variables (Appendix 1—tables 19 and 20).

Of note, neuroscience has historically neglected sex and gender differences, predominantly resulting in increased misdiagnoses of and relatively worse treatments for women (Shansky and Murphy, 2021). Therefore, we recalculated analyses in gender-stratified samples (n women = 1110 observations, n male = 1137 observations) to test for differences in the effects of social isolation (Appendix 1—table 21). No clear pattern of difference emerged between women and men. A minor observable difference was that the effect of social isolation on hippocampal volume was mostly driven by baseline social isolation amongst women and by change in social isolation amongst men. This pattern was reversed for other outcomes, though.

In order to further investigate the nature of the correlations, we calculated bivariate latent change score (BLCS) models. In these models we simultaneously tested for an effect of baseline social isolation on change in cognitive functions or hippocampal volume and vice versa (see Appendix 1—figure 2 for a visualization). The BLCS models did not produce solid evidence regarding directionality (Appendix 1—table 22). As in the mediation analyses, the design requirements of the BLCS resulted in smaller sample sizes (n = 362–585 complete observations).

Our findings indicate that social isolation contributes to grey matter loss in the hippocampus, a focal point of atrophy in mild cognitive impairment (Devanand et al., 2007), and Alzheimer’s dementia (Fox et al., 1996).

Notably, not only baseline social isolation (a between-subject effect) but also change in social isolation (a within-subject effect) significantly predicted hippocampal volume. Through the employment of statistical LMEs, we were able to distinguish and study effects at these different levels (van de Pol and Wright, 2009) and the design helped us to avoid fallacious inferences from single-level data (Robinson, 1950) to which simple linear regressions would have been susceptible. Specifically for the study of social isolation as a risk factor for dementia, it is crucial to disentangle between- and within-subject effects. Social isolation has both been described as a trait (Noonan et al., 2021), implying it to be an invariant between-subject characteristic and as a potential target for interventions (Hussenoeder and Riedel-Heller, 2018), implying it to be a modifiable within-subject effect. The finding of a significant within-subject effect of change in social isolation therefore offers hope for modifiability as it implies that the observed associations are not (exclusively) the effect of an invariant trait. Thus, our data point towards that reducing social isolation could help to maintain hippocampus integrity in ageing.

However, this assumes a causal effect of social isolation. As associations with social isolation could also have resulted from reverse causation through health selection, that is, that participants with accelerated brain ageing are more likely to become socially isolated, this assumption needs careful consideration. Bayes factors imply the absence of an interaction effect of baseline social isolation with change in age and the bivariate latent change score models did not provide evidence in favour of causality in the hypothesized direction either. However, neither did they provide evidence for reverse causality. This inconclusiveness might result from our reduced follow-up sample size and thus related lower power, especially in the latent score models. For example, data from the English Longitudinal Study of Aging from >6000 older adults measured at up to 6 two-year intervals supports the assumed causality of social isolation with regards to memory performance (Read et al., 2020). In our study, presence of considerable effect sizes and the high statistical confidence in these estimates on multiple outcomes in our healthy sample without cognitive impairment speaks against the competing hypothesis of reverse causality through health selection and in favour of a causal role of social isolation. Furthermore, the lack of any strong increase in effect size when including health-impaired participants or decrease when applying more stringent exclusion criteria for cognitive health corroborate this interpretation. Still, overall these results only add a modicum of corroboration to the case for a causal role of social isolation.

Baseline social isolation, and to a lesser extent, change in social isolation, were significantly associated with cognitive performance, that is, executive functions, processing speed, and memory, all of which undergo decline in (pathological) ageing (Blazer et al., 2015). Again, our results thus imply a detrimental role of social isolation on cognitive functions. We could however not observe that social isolation lowered memory performance through reductions in hippocampal volume, a hypothesis raised by considerations of the central role of the hippocampus in memory (Buzsáki and Moser, 2013). Similarly, we could not find evidence that social isolation affected hippocampal volume through higher chronic stress measured with questionnaires, a hypothesis put forward by the stress buffering theory (Kawachi and Berkman, 2001). These latter analyses suffered from small sample sizes and a limited number of timepoints. Nonetheless, the lack of any significant link between chronic stress and social isolation (see Appendix 1—table 2) is hard to align with the stress-buffering hypothesis in spite of the missingness in the TICS.

Overall, comparing our brain morphometric results with those of existing cross-sectional studies on social isolation, detected brain regions coincide. A rather small-sampled study did not find a link between social isolation and grey matter volumes (Lin et al., 2020) but James et al., 2012 (occipital lobe), Blumen and Verghese, 2019 (hippocampus, precuneus, medial frontal gyrus) and Shen et al., 2022 (hippocampus, right supramarginal gyrus) found decreased volumes in regions we detected, too.

Several of the cortical regions identified in our study (precuneus) belong to the pattern of exacerbated regional atrophy found in Alzheimer’s disease. Furthermore, we detected regions known for increased cortical thinning in the healthy process of ageing (cuneus) and both in healthy and pathological ageing (supramarginal gyrus) (Bakkour et al., 2013; Pini et al., 2016). This indicates an aggravating role of social isolation in cortical thinning that may contribute to normal and accelerated brain ageing processes. However, the findings of lower cortical thickness must be interpreted cautiously due to the limited consistency between cross-sectional and longitudinal effects and the exploratory approach of whole-brain analyses.

A limitation of this study is its uncertain generalizability to the general population because the sample was probably affected by selection and attrition bias common to longitudinal studies (Chatfield et al., 2005). Attrition bias might have mostly affected the mediation and BLCS models that thus offered reduced interpretability, despite the comparatively large neuroimaging cohort. However, the LMEs were mostly unscathed by this problem due to their ability to make use of datapoints of participants with only one full observation. In addition, our population represents a WEIRD sample (i.e. western, educated, industrialized, rich, democratic) which might skew our understanding of how social isolation affects brain health (Laird, 2021). As we found higher SES to be associated with lower LSNS scores, this relatively high SES sample might have led to underestimation of the detrimental effects of social isolation and increases in social isolation in the ageing process. Considering hippocampus segmentations, it has been argued that FreeSurfer systematically overestimates volumes compared to manual volumetry; however, this difference did barely emerge in participants over the age of 50 (Wenger et al., 2014). A further limitation are ceiling effects in the CERAD word list memory task in healthy adults, potentially limiting the sensitivity to detect subtle differences. In addition, time of day during testing might have affected cognitive performance (Schmidt et al., 2007), yet we did not control for this. Almost all cognitive tests were performed between 9 a.m. and 1 p.m., though. Covariance of social isolation with other variables such as hypertension or diabetes could have influenced the results. However, note that all variance inflation factors (VIFs) were acceptable, indicating low reason for concern regarding multicollinearity. Lastly, inferences from our results regarding dementia aetiology must be made with caution as we did not investigate clinically diagnosed dementia patients.

In quantitative studies, despite its importance in shaping the research process and conclusions, for example, in functional MRI analysis (Botvinik-Nezer et al., 2020), researchers’ influence is often disregarded. In the supplementary text, we offer a brief reflexivity section to make relevant influences on this study transparent and to shortly discuss the value of reflexivity for quantitative science.

This pre-registered large-scale population neuroimaging analysis adds robust support to the view that social isolation is associated with accelerated brain ageing and cognitive decline in non-demented adults in mid- to late-life. Our findings further imply that social contact protects from detrimental processes and thereby preserves brain structure and function. Henceforth, targeting social isolation through tailored strategies might contribute to maintaining brain health into old age.

We showed that the established LSNS cut-off can be employed by clinicians to identify subjects likely to suffer adverse effects due to social isolation. However, the absence of evidence for more pronounced negative effects of less social contact amongst those that are deemed socially isolated by the cut-off renders a public health strategy focused on high-risk individuals questionable.

While we could not observe significant contributions of physical activity or sleep quality measured using questionnaires in a smaller subsample on brain and cognitive outcomes, previous studies suggested that physical activity (Musich et al., 2022) and sleepiness (Holding et al., 2020) interact with social isolation and could protect against negative health effects of social isolation, and should therefore be explored in future studies that incorporated these outcomes more systematically.

While we see evidence converging on social isolation as a causal risk factor for dementia and cognitive decline, future neuroimaging studies should pay particular attention to questions of temporality in their design to clear up remaining uncertainties. Studies with more numerous timepoints will be of importance to this end and will furthermore allow us to model important aspects like slopes for individual participants (van Doorn et al., 2021). Intervention studies will be the gold standard to provide evidence with regards to the causal role and effect size of social isolation. Multidomain interventions for dementia prevention justifiably become the norm (Stephen et al., 2019), so that effects of reduced social isolation must be investigated as a likely contribution to an aggregate effect.

Illuminating the mechanistic underpinnings of the association should be another focus for future research. Studies might prioritize obtaining reliable proxies for the hypothesized mediators. As elevated cortisol levels, in line with the stress-buffering hypothesis, may exert detrimental effects on cognition and contribute to AD pathology (Ouanes and Popp, 2019), using hair cortisol, a reliable measure of chronic stress (Staufenbiel et al., 2013), could be a promising choice to further investigate this proposed mechanism. In light of the lack of evidence for the stress-buffering hypothesis in our data, alternative mechanistic theories should be pursued, too. The main-effect theory postulates that social relationships foster beneficial health behaviours, affective states and neuroendocrine responses, ultimately protecting neuronal tissue (Kawachi and Berkman, 2001). Others point out that socializing is cognitively demanding and requires engagement with complex environments. In the ‘use-it-or-lose-it’ theory, this is crucial for the maintenance of cognitive function (Hultsch et al., 1999). Promising approaches to answer this research question could be interventions specifically targeting one of the hypothesized detrimental processes in isolated individuals and mediation analyses of multi-wave studies with larger sample sizes. Lastly, reverse causality or simultaneity cannot be completely ruled out yet. However, the observed solid correlations in our healthy sample and the lack of an increase in effect sizes when including participants with dementia or low MMSE scores renders this alternative hypothesis to a causal role of social isolation unlikely.

Moreover, studies investigating social isolation due to lockdown measures and its impact on cognitive and brain health will be of great significance.

In light of the relevance of social isolation for cognitive and general health and well-being (National Academies of Sciences, 2020), its pervasiveness in the elderly population of the global north (Livingston et al., 2020) is alarming. Physical distancing measures have caused an unprecedented rise in the attention to the impact of social isolation but social isolation has been a grave problem before Covid-19 and it will remain a central public health concern thereafter. Existing and future research on reasons for and the role of social isolation in health and disease should provide guidance for the urgently needed development and evaluation of tailored strategies against social isolation and its detrimental effects. These should address social isolation both through intervention strategies on the individual but also societal level, leveraging values like solidarity and communality.

We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) and Committee on Best Practices in Data Analysis and Sharing (COBIDAS) on MRI guidelines in our reporting wherever appropriately applicable.

The study’s preregistration can be found at https://osf.io/8h5v3/. Please refer to it for information on the authors’ previous knowledge of the data and a comprehensive overview of our pre-specified hypotheses and models.

We used longitudinal data from the ‘Health Study of the Leipzig Research Centre for Civilization Diseases’ (LIFE). The study was approved by the institutional ethics board of the Medical Faculty of the University of Leipzig and conducted according to the Declaration of Helsinki. The LIFE-Adult-Study is a population-based panel study of around 10,000 randomly selected participants from Leipzig, a major city with 550,000 inhabitants in Germany. A subgroup of around 2600 participants underwent MRI testing at baseline. The baseline examination was conducted from August 2011 to November 2014. Follow-up assessments were performed around 6–7 y after the respective first examinations (Engel et al., 2023). Around 1000 participants of the MRI-subsample returned for follow-up testing.

We included all participants over 50 with MRI data that did not fulfil any of the following exclusion criteria:

• Anamnestic history of stroke

• Any medical condition (i.e. epilepsy, multiple sclerosis, Parkinson’s disease)/chronic medication use that would compromise cognitive testing (i.e. cancer treatment in the past 12 mo or drugs affecting the central nervous system)

• Diagnosed dementia or Mini-Mental State Examination (MMSE)-score < 24

• A trained radiologist considered the MRI scans unusable due to brain tumours, or acute ischaemic, haemorrhagic or traumatic lesions

If no MMSE data were available, the participants were excluded if their overall performance in cognitive tests negatively deviated from the wave’s mean by 2 standard deviations (SDs) which is a stricter criterion excluding ~2.6% of the sample compared to ~0.8% excluded based on the MMSE. The exclusion criteria were chosen to reduce the potential of reverse causality, that is, dementia symptoms leading to a loss of social connections, as correlations observed in this cognitively intact sample should not stem from dementia symptoms.

We obtained T1-weighted images on a 3 Tesla Siemens Verio MRI scanner (Siemens Healthcare, Erlangen, Germany) with a 3D MPRAGE protocol and the following parameters: inversion time, 900 ms; repetition time, 2300 ms; echo time, 2.98 ms; flip angle, 9°; field of view, 256 × 240 × 176 mm³; voxel size, 1 × 1 × 1 mm³, shimming: tune-up shim, no fat suppression, whole-brain coverage. We processed the scans with FreeSurfer (FreeSurfer, V5.3.0, RRID:SCR_001847) and the standard cross-sectional pipeline recon-all. FreeSurfer automatically measures hippocampal volume, vertex-wise cortical thickness, and intracranial volume. To ensure high within-subject reliability, we employed FreeSurfer’s longitudinal pipeline on all scans, including those of participants without a follow-up scan. Please see Reuter et al., 2012 for details. Moreover, we smoothed the cortical thickness surfaces with a 10 mm kernel to improve reliability and power (Liem et al., 2015). Different Linux kernels and Ubuntu versions constituted the computational infrastructure during the data acquisition and processing.

Visual quality control was based on the recommendations of Klapwijk et al., 2019. After the baseline data were acquired, our team visually controlled all results of the cross-sectional recon-all pipeline. Additionally, we controlled the outputs of the longitudinal stream of all participants with follow-up data and those whose cross-sectional runs required editing. If we detected errors in the processed scans, we manually edited them (N = 262). We excluded participants from analyses using MRI measures if we deemed the processed scans to be unusable (n = 98).

We excluded outliers for our core variables based on a cut-off of 3 SDs (LSNS-score, adjusted hippocampal volume, cognitive functions). Please see Figure 6 (flowchart) for the limited effect of outlier exclusion on sample sizes of the different models. For further details on outlier detection and handling regarding covariates, please see the supplementary text.

Figure 6

Download asset Open asset

To avoid an excessive reduction in sample size due to missing data, we performed imputations for missing predictor variables using the sample mean, distributions based on existing data, or the participant’s mean. Please see the supplementary text for information on our procedures of the respective measures.

Furthermore, we used FIML for analyses using structural equation modelling.

All code can be found at https://github.com/LaurenzLammer/socialisolation, (copy archived at Lammer, 2023). Please see the supplementary text for information on the software used for the analyses.

Our threshold for significance for all tests was p<0.05. To control for multiple hypothesis testing we FDR-corrected families of tests and each individual whole-brain analysis (see the supplementary text for definition of families).

BFs of 3–10 and BFs of 10–30 are commonly considered to be moderate or strong evidence in favour of a hypothesis. To evaluate these thresholds in light of multiplicity, we conducted two simulation studies described in the supplementary text that revealed that using a BF threshold of 10.75 rather than 3 would keep α below 5% and that this would not substantially decrease power.

To ensure that our continuous predictors are normally distributed, we plotted their histograms. We had to log-transform the CES-D, IPAQ, and PSQI scores to obtain a normal distribution.

To rule out major collinearity, we calculated VIFs. The VIFs did not surpass the threshold of 10 (Myers, 1990) in any model.

Furthermore, we tested the stability of our LMEs in R by comparing the estimates obtained from the model based on all data with those obtained from models with the levels of the random effects excluded one at a time. This revealed the models to be fairly stable. Moreover, we visually controlled them for heteroskedasticity with both a histogram and a qq-plot. The qq-plots show a heavy-tailed distribution of the residuals in some models. This is only a minor deficit as the models are not intended to make accurate predictions at specific points (Gelman and Hill, 2006).

Fit indices providing further information on the quality of a model fit using structural equation modelling can be found in Appendix 1—tables 23 and 24 (Schermelleh-Engel et al., 2003). Fit index thresholds were surpassed by multiple mediation models. As the BLCS models are saturated, fit indices are uninformative.

We excluded the datapoints (all measures of the timepoint) of all participants with measures deviating from the mean by 3 SD for our core variables (LSNS-score, adjusted hippocampal volume, cognitive functions). In case of TICS-score deviations by 3 SD we replaced the values with ‘NA’ and hence did not include them in mediation analyses.

Considering confounders, highly implausible values (±4 SD) for CES-D-score or BMI were treated as missing datapoints and we replaced them with values imputed according to our imputation plans listed below in order not to overly reduce the sample size.

All outlier analyses were conducted separately for baseline and follow-up measurements.

The data on the control variables education, BMI, diabetes, hypertension, age, and gender were complete or mostly complete. Henceforth, we could impute missing datapoints without inducing severe bias by using the sample mean for continuous variables or values drawn from a distribution determined by the existing data for categorical variables.

However, CES-D-scores were an exception amongst our control variables because the questionnaires often missed a single or a few items. As suggested by Bono et al., 2007, we imputed up to four missing items per participant using the person mean. Similarly, we imputed up to one item in the LSNS and up to six items in the TICS using the person mean.

If results from one of the cognitive tests required to calculate a composite score for a cognitive function was missing, we calculated the score based on the average performance in the remainder of available tests contributing to the composite score, if at least two tests were available.

Appendix 1—figure 4 provides an overview of missingness in relevant variables at different LSNS scores.

The LMEs with hippocampal volume and the cognitive functions as dependent variables form one large family except for models regressing on the interaction of baseline LSNS and change in LSNS. In each family, we separately corrected model 1 and model 2 analyses resulting in two families of 12 tests. Additionally, we FDR-corrected each individual whole-brain analysis using the sided two-stage adaptive FDR-correction in the FreeSurfer-toolbox (Bernal-Rusiel et al., 2013b). All other analyses and the whole-brain analyses were considered to be exploratory and must be evaluated as such.

The participants’ education was assessed using an extensive questionnaire and given a score ranging from 1 (no degree at all) to 7 (A-levels + master’s degree [or equivalent] or promotion) according to prior research (Lampert et al., 2013a). The effects of education and the significance of different degrees are likely to be culture specific. Fortunately, a recent study examined the effects of education in a population of elderly residents of the city of Leipzig. In this study education operationalized as having a tertiary degree or not was found to be a significant predictor of dementia incidence (Then et al., 2016). This is approximated with a cut-off at a score <3.6.

Although it is sometimes claimed that Bayesian statistics do not require any multiplicity control (Gelman et al., 2012), we do not believe that this is the case in our study. A truly Bayesian approach would require researchers to adjust the priors to all other tests with non-independent hypotheses or datasets (Sjölander and Vansteelandt, 2019). This is hardly feasible and hence, in practice, Bayesian statistics are usually employed without taking all dependencies into account and their results are measured against thresholds similar to those of frequentist statistics. Appendix 1—figure 5 shows how this results in an increasing familywise error rate (FWER) with an increasing number of tests in both Bayesian and frequentist statistics using an example from Keysers et al., 2020. De Jong has provided a solution for this problem for ANOVAs that has been implemented in the JASP software (de Jong, 2019) but there is still a great lack of available tools for researchers using other statistical methods. Henceforth, we decided to conduct a simulation study to find a Bayes factor threshold adjustment that should control our FWER similar to α-adjustments in frequentist statistics.

To find the expected number of false positives for a given number of tests and threshold, we replaced the variables for baseline social isolation and change in social isolation with random normally distributed values with the same SD and kept the original dataset otherwise untouched. Then we calculated our 24 LMEs belonging to the families of tests with the modified dataset and repeated this process 42 times. At a BF threshold of 3, 14 of the 1008 tests were false positives and 881 were detected as true negatives. Appendix 1—figure 6 shows a histogram of the resulting Bayes factors. The study suggests that for the family size of 12 tests in our study a threshold of about 10.75 would ensure a FWER below 5%. Appendix 1—table 25 gives an overview of the false positives and FWERs.

Furthermore, we wanted to see how this threshold adjustment would affect the power of our study. For this simulation study we generated a dataset that closely resembles the actual dataset but has different regression coefficients for baseline social isolation and change in social isolation. Instead of the actual coefficients we set the effect size per point on the LSNS to 0.1, 0.2, or 0.5 y of baseline age. We simulated a dataset and calculated a Bayes factor for each model and each effect size. As we only calculated the LMEs without interaction terms for reasons of simplicity, this resulted in a number of 48 Bayes factors from simulated data for each of our 13 runs totalling 624 tests. While our power for the smallest effect sizes was generally small (<10%), it was 85.6% for baseline social isolation with an effect size of half a year of baseline age. Increasing the threshold to 10.75 would not substantially decrease it (81.7%). Appendix 1—tables 26 and 27 provide an overview of the percentages of false negatives and true positives using the thresholds 3 and 10.75.

For the most part, we stuck closely to our preregistered plan in this study but departed from it at some points for different reasons.

We used the function q-value instead of p.adjust for the FDR correction for the simple reason that it provides us with a more comprehensive output. As we set the argument pi to 1, q-value is equivalent to the classic procedure (Storey, 2002).

We originally intended to first perform a full-null model comparison using an ANOVA and only follow this up with the function drop1 in case of a significant value for the respective predictor of interest. Our intention was to avoid any multiplicity problems due to testing all predictors. Using the scope argument of drop1 solved the problem more parsimoniously.

Our plan to exclude participants with two or more lesions in their MRI was the result of an internal equivocation regarding the meaning of an abbreviation. We excluded participants based on the type of lesions but not based on lesion count.

Furthermore, we used FIML for analyses using structural equation modelling. The similar results obtained using our preregistered approach can be found in the pre-print (Lammer et al., 2021). Similarly, results based on R version 3.6.1 can be found there. The final analyses were conducted in R version 4.2.2 because 3.6.1 was discontinued at our institute.

Lastly, we changed from the term sex to gender as it seems more appropriate.

We performed most analyses using R (R Project for Statistical Computing, V4.2.2, RRID:SCR_001905). For the whole-brain analyses we used MATLAB (MATLAB, V9.13 (2022b) RRID:SCR_001622).

We used the package lme4 (R package: lme4, RRID:SCR_015654) to calculate LMEs in R. To obtain reliable p-values, we used the Satterthwaite option from the lmerTest package (R package: lmerTest, RRID:SCR_015656; Kuznetsova et al., 2017). In the whole-brain analyses we employed the MATLAB-toolbox provided by FreeSurfer to calculate vertex-wise LMEs (Bernal-Rusiel et al., 2013a). For mediation analyses and BLCS models we used the sem function from the lavaan package (Rosseel, 2012).

We calculated BFs for all LMEs in R using the BayesFactor package and the functions posterior and generalTestBF with default priors (Rouder and Morey, 2012).

FDR-correction was performed using the q-value function (R package: Qvalue, RRID:SCR_001073) in R and the sided two-stage adaptive FDR-correction in the FreeSurfer-toolbox (Bernal-Rusiel et al., 2013b).

VIFs were calculated using the package car (Fox and Weisberg, 2019).

Reflexivity, a sensitivity to and acknowledgement of the ways in which scientists shape the collected data and research findings, is an established hallmark of scientific rigour in qualitative research (Mays and Pope, 2000; Sandelowski and Barroso, 2002). The challenges addressed by reflexivity are perhaps more pronounced in but by no means exclusive to qualitative studies. Nevertheless (at least in an openly conducted form), it is largely absent from quantitative studies (Ryan and Golden, 2006). Methodological reforms in quantitative research like preregistrations and registered reports (Nosek et al., 2018; Nosek and Lakens, 2014) are valuable tools to limit the researchers’ potential to make data fit their prior assumptions but their scope is limited. They do not address some of the most fundamental issues in epidemiology: Which analogies are used to make sense of the data, which questions are being raised and answered, and which theories are chosen to explain phenomena (Krieger, 2011)? Disclosing personal characteristics, researchers’ values, and positionality relative to the object of research (Berger, 2015) thus helps readers assess a study and its findings more thoroughly. Additionally, an external evaluation of the presence and prevalence of non-empirical decision vectors (Solomon, 2001) in a field of research can be greatly facilitated. Furthermore, as Stephen J. Gould has put it: “It is dangerous for a scholar even to imagine that he might attain complete neutrality, for then one stops being vigilant about personal preferences and their influences – and then one truly falls victim to the dictates of prejudice” (Gould, 1996).

Henceforth, I, as the first author, want to expand this study by a brief reflection on influences that might have played a role in the formation of this study. I am a medical doctoral student with no prior experience in research and conducted this study as the centrepiece of my planned dissertation. Thus, I entered this project with little prior knowledge. I believe that this both made me more flexible and restricted in my choices. On the one hand I was not dedicated to any specific research programme or topic, but on the other hand my reliance on the advice and support from more senior researchers made me emulate their work and methods in many aspects. Further, my worldview has probably made me tend to epidemiological theories (social epidemiology, eco-social theory) (Berkman et al., 2015; Krieger, 2014) broader than the study of lifestyle factors and hence made me choose social isolation as my research topic. A further characteristic that might be of interest to readers is that during the course of the research, two of my relatives struggled with dementia. Ultimately, this reflexivity is inherently limited, as the use of secondary data precludes me from reflecting on the pivotal processes of data acquisition and participant recruitment.

Variables in bold are dropped in the null model.

H 1.1 Social isolation is negatively associated with hippocampal volume across individuals.

Model111: HCV ~LSNS_bl +LSNS_change +age_bl +age_change +sex + (1|subject)

Model112: HCV ~LSNS_bl +LSNS_change +age_bl +age_change +sex + hypertension

+diabetes + BMI+CESD + education + (1|subject)

H 1.3 Social isolation is negatively associated with hippocampal volume within individuals.

Model131: HCV ~LSNS_bl +LSNS_change +age_bl +age_change +sex + (1|subject)

Model132: HCV ~LSNS_bl +LSNS_change +age_bl +age_change +sex + hypertension

+diabetes + BMI+CESD + education + (1|subject)

H 1.5 Participants that are socially more isolated at baseline will experience aggravated age- related changes in hippocampal volume over the follow-up period.

Model151: HCV ~LSNS_bl +LSNS_change +age_bl +age_change +

LSNS_bl*age_change +sex + (1|subject)

Model152: HCV ~LSNS_bl +LSNS_change +age_bl +age_change +

LSNS_bl*age_change +sex + hypertension +diabetes + BMI+CES.D +

education + (1|subject)

H 2.1 Social isolation is negatively associated with cognitive functions across individuals.

Model211a: executive function ~LSNS_bl +LSNS_change +age_bl +age_change +sex +

(1|subject)

Model212a: executive function ~LSNS_bl +LSNS_change +age_bl +age_change +sex +

hypertension +diabetes + BMI+CES.D+education + (1|subject)

Model211b: memory performance ~LSNS_bl +LSNS_change +age_bl +age_change +

sex + (1|subject)

Model212b: memory performance ~LSNS_bl +LSNS_change +age_bl +age_change +sex

+ hypertension + diabetes +BMI + CES.D+education + (1|subject)

Model211c: processing speed ~LSNS_bl +LSNS_change +age_bl +age_change +

sex + (1|subject)

Model212c: processing speed ~LSNS_bl +LSNS_change +age_bl +age_change +sex +

hypertension +diabetes + BMI+CES.D+education + (1|subject)

H 2.2 Social isolation is negatively associated with cognitive functions within individuals.

Model221a: executive function ~LSNS_bl +LSNS_change +age_bl +age_change +sex +

(1|subject)

Model222a: executive function ~LSNS_bl +LSNS_change +age_bl +age_change +sex +

hypertension +diabetes + BMI+CES.D+education + (1|subject)

Model221b: memory performance ~LSNS_bl +LSNS_change +age_bl +age_change +sex + (1|subject)

Model222b: memory performance ~LSNS_bl +LSNS_change +age_bl +age_change +sex

+hypertension + diabetes +BMI + CES.D+education + (1|subject)

Model221c: processing speed ~LSNS_bl +LSNS_change +age_bl +age_change +

sex + (1|subject)

Model222c: processing speed ~LSNS_bl +LSNS_change +age_bl +age_change +sex +

hypertension +diabetes + BMI+CES.D+education + (1|subject)

H 2.3 Participants that are socially more baseline will experience aggravated age-related changes in cognitive function over the follow-up period.

Model231a: executive function ~LSNS_bl +age_bl +age_change +LSNS_bl*age_change

+sex + (1|subject)

Model231a: executive function ~LSNS_bl +age_bl +age_change +LSNS_bl*age_change

+sex + hypertension +diabetes + BMI+CES.D+education + (1|subject)

Model231b: memory performance ~LSNS_bl +age_bl +age_change +

LSNS_bl*age_change +sex + (1|subject)

Model231b: memory performance ~LSNS_bl +age_bl +age_change +

LSNS_bl*age_change +sex + hypertension +diabetes + BMI+CES.D+education + (1|subject)

Model231c: processing speed ~LSNS_bl +age_bl +age_change +LSNS_bl*age_change

+sex + (1|subject)

Model231c: processing speed ~LSNS_bl +age_bl +age_change +LSNS_bl*age_change

+sex + hypertension +diabetes + BMI+CES.D+education + (1|subject)

H 5.1 In people who are socially more isolated at baseline, an increase in social isolation

from baseline to follow-up will have a stronger negative association with HCV than in people who are less socially isolated at baseline.

Model511: HCV ~LSNS_bl +LSNS_change +LSNS_bl*LSNS_change +age_bl +

age_change +sex + (1|subject)

Model512: HCV ~LSNS_bl +LSNS_change +LSNS_bl*LSNS_change +

age_bl +age_change +sex + hypertenison +diabetes + BMI+CES.D +

education + (1|subject)

H 1.2 Social isolation is negatively associated with vertex-wise cortical thickness across individuals.

For model 1 we built a matrix consisting of six columns: intercept (all ones), age_bl,

age_change, sex, LSNS_bl and LSNS_change.

The corresponding contrast matrix was [0 0 0 0 1 0].

For model 2 we built a matrix consisting of eleven columns: intercept (all ones), age_bl,

age_change, sex, hypertension, diabetes, education, BMI, CES_D, LSNS_bl and

LSNS_change.

The corresponding contrast matrix was [0 0 0 0 0 0 0 0 0 1 0].

H 1.4 Social isolation is negatively associated with vertex-wise cortical thickness within individuals.

For model 1 we built a matrix consisting of six columns: intercept (all ones), age_bl,

age_change, sex, LSNS_bl and LSNS_change.

The corresponding contrast matrix was [0 0 0 0 0 1].

For model 2 we built a matrix consisting of eleven columns: intercept (all ones), age_bl,

age_change, sex, hypertension, diabetes, education, BMI, CES_D, LSNS_bl and

LSNS_change.

The corresponding contrast matrix was [0 0 0 0 0 0 0 0 0 0 1].

H 1.6 Participants that are socially more isolated at baseline, will experience aggravated age-related changes in cortical thickness over the follow-up period.

For model 1 we built a matrix consisting of seven columns: intercept (all ones), age_bl,

age_change, sex, LSNS_bl, LSNS_change and LSNS_bl*age_change. The last term is an

interaction between baseline LSNS and age_change.

The corresponding contrast matrix was [0 0 0 0 0 0 1].

For model 2 we built a matrix consisting of twelve columns: intercept (all ones),

age_bl, age_change, sex, hypertension, diabetes, education, BMI, CES_D, LSNS_bl,

LSNS_change and LSNS_bl*age_change. The last term is an interaction between baseline

LSNS and age_change.

The corresponding contrast matrix was [0 0 0 0 0 0 0 0 0 0 0 1].