The association between control of interference and white-matter integrity: A cross-sectional and longitudinal investigation

Proactive interference (PI) occurs when old information interferes with newly acquired information and has been suggested as a major cause of forgetting in working memory. In this study, we investigate cross-sectional (N = 267) and longitudinal (N = 148) associations between PI and white-matter integrity (WMI) using diffusion-weighted imaging in an adult life-span sample (25-80 years; Mage = 60.15; 138 female). Older age was related to higher PI and lower WMI. Cross-sectional analyses showed associations between PI and WMI spanning several white-matter tracts as well as globally, suggesting that the age-related decline in PI may be driven primarily by global changes in WMI. Furthermore, longitudinal changes in PI were shown to be negatively correlated with concurrent changes in WMI in the fornix. Mediation analyses showed that WMI mediated the relationship between age and PI only in older adults, indicating that WMI becomes increasingly connected to cognitive functioning with increasing age. This is the first demonstration of WMI decline contributing to the age-related decline in PI.

Increased PI in working memory has been found to correlate with worse performance in several other cognitive domains including episodic memory, block design, fluency, and processing speed. Furthermore, PI has been found to explain variance in age -cognition associations in these cognitive domains, over, and above the effects of processing speed ( Samrani & Persson, 2021 ). Defective interference control, indicated by increased PI, may therefore contribute to age-related impairments across several cognitive domains.
While increased PI with increasing age is a well-established finding ( Loosli et al., 2016 ;Samrani & Persson, 2021 ), the underlying contributing factors of this association are not yet fully understood. Age-related differences in PI have been linked to both altered functional connectivity ( Oren et al., 2017 ) and task-related BOLD activation ( Loosli et al., 2016 ), as well as smaller gray-matter volume in the prefrontal cortex ( Samrani et al., 2019 ). Despite some previous observations linking age-related impairments in the ability to control PI in working memory to brain structure and function, the underlying neural mechanisms for such deficits remain largely unknown.
Concurrent with an age-related cognitive decline, aging is also associated with a deterioration of the brain's white matter (WM). WM consists of fiber bundles, primarily made up of myelinated axons, connecting different gray matter areas of the brain, and is essential for inter-neuronal communication (e.g Liu et al., 2017 .). WM is believed to be important for normal cognitive functioning as such functions rely on communication between regions across the brain ( Bennett & Madden, 2014 ;Rabin et al., 2019 ). Thus, studies of WM microstructural properties might shed light on the nature of the cognitive decline associated with aging. Diffusion-weighted imaging (DTI) enables in vivo investigation of WM microstructural properties by generating indices of the rate and the directionality of water diffusion ( Basser & Jones, 2002 ;Beaulieu, 2002 ;Mori & Zhang, 2006 ).
Both PI and other executive control processes as well as working memory rely on a network of brain regions, including the prefrontal cortex, parietal cortex, and the striatum ( Assem et al., 2020 ;Fedorenko et al., 2013 ;Nee et al., 2007 ;Niendam et al., 2012 ;Majid et al., 2013 ). Optimal behavioral performance is dependent on an efficient within-network communication and connectivity ( Filley, 2005 ;Schmahmann et al., 2008 ). Thus, age-related cognitive decline can, at least partly, be explained by corticocortical disconnection caused by WM deterioration ( Charlton et al., 2010 ;Madden et al., 2009 ). It has been postulated that working memory and executive control processes may be particularly sensitive to such age-related WM disruption ( Charlton et al., 20 06 , 20 08 ;Madden et al., 2009 ), compared to other cognitive functions.
To our knowledge, no study to date has investigated the relationship between PI and WMI. Moreover, there is a lack of longitudinal studies examining the relationship between PI, and WMI across the whole age-span. The current study is designed to fill this gap.
The purpose of this study was, thus, to examine the association between PI and WMI in aging using both cross-sectional (N = 267) and longitudinal (N = 148) data. The analyses were based on a population-based sample (N = 267; 25-80 years; M age : 60.15; SD:13.8) across 2 time points (5 years interval). Tract-based spatial statistics (TBSS) was used to estimate mean fractional anisotropy (FA), which is a commonly used index of WMI, across eleven WM tracts as well as global WM. We hypothesized that (1) there would be negative associations between PI and FA, such that greater PI is associated with less WMI; (2) changes in WMI would be associated with changes in PI over 5 years; and (3) that these relationships would be more pronounced in older age (more and/or stronger associations). In addition, we hypothesized that the previously established negative association between age, and PI ( Samrani et al., 2017 ;Samrani & Persson, 2021 ) would be at least partly mediated by WMI.
Given previous demonstrations of age-differential relationships between brain function and structure in younger and older adults ( Burzynska et al., 2012 ;Koen & Rugg, 2019 ;Rieckmann et al., 2018 ;Van Petten, 2004 ), we performed age-stratified analyses with 2 groups of younger ( < 65 years) and older (65 years and above), in addition to whole sample analyses. In addition, moderation analyses were conducted, investigating age as a moderator of the WMI -PI relationships.
An investigation of the relationship between PI and WMI may provide important clues about the neural mechanisms of impaired interference control in aging. Such an investigation can also provide insights into the decline in other cognitive functions characterized by diminished interference control. Furthermore, looking at how this relationship changes over time can provide details about the causal nature of the relationship.

Participants
The study sample was drawn from the Betula prospective cohort study: a longitudinal, population-based investigation of memory, health, and aging ( Nilsson et al., 1997 ;Nyberg et al., 2020 ). Participants in the Betula study were recruited from the city of Umeå on the northeast coast of Sweden and its vicinity, and were randomly sampled from the population registry, stratified by age, and gender. The first wave of data collection started in 1988 and in total 7 main waves of data collections have been completed. Participants were included from samples for which MRI measures were collected in 2008-2010 (baseline) and 2013-2014 (follow-up), and consisted of 12 age cohorts (25,30,35,40,45,50,55,60,65,70,75,80 years of age at baseline. Dementia status was assessed at baseline and reassessed every 5 years using a 3-step procedure. First, an overall evaluation was performed by an examining physician according to the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSM-IV; American Psychiatric Association, 1994 ). Second, using a composite measure based on scores from several cognitive tests (episodic memory, working memory, speed, semantic memory, and fluid intelligence), each participant is compared to the mean cognitive score for his/her age cohort. If an individual scores more than 2 standard deviations below the mean of the age cohort, he/she will be flagged for further assessment of dementia by a clinical psychiatrist. Third, all participants scored at or above the cut-off score for dementia of 25 using MMSE (Mini Mental State Examination, Folstein et al., 1975 ). Moreover, T2-weighted Fluid-Attenuated Inversion Recovery (FLAIR) images were acquired with a 2D T2 FLAIR sequence (48 slices with 3 mm thickness; TR: 80 0 0 ms, TE: 120 ms, field of view: 24 × 24 cm). This sequence was included as part of a clinical protocol, and all FLAIR images were inspected by a neuroradiologist for deviations.
The sample consisted of 371 participants at baseline (T5). Participants were excluded if they had an MMSE score of below 25 (32 participants), had a history of a neurologic condition (12 participants; 1 epilepsy, 3 dementia, 1 stroke, 1 Parkinson, 1 Alzheimer's, 5 neurologic remark), had incomplete MRI data (5 participants), or abnormal brain structure (2 participants). Participants were also excluded if they did not complete the 2-back task at T5 (41 participants) or had a hits-fa rate of below 10 (12 participants). Positive responses to previously presented items were scored as hits and positive responses to lure trials were scored as false alarms. Hit rate minus false alarm rate was used as a measure of working memory performance. The final sample consisted of 267 participants. From this sample, 148 had complete data for follow-up (T6), and thus formed the sample for the longitudinal analyses.
Analyses were performed across the whole sample, as well as separately for younger ( < 65 years), and older adults ( > 65 years). It should be noted that while these age groups are arbitrary constructions, many prior studies, including those that included participants from the Betula study ( Gorbach et al., 2017 ), have used a similar age-stratification. Demographics and cognitive scores for T5 and T6 are presented in Table 1 . The Betula project was approved by the Regional Ethical Review Board (2013/92-31) and written consent was obtained from all participants.

Cognitive assessment
Full details about recruitment, cognitive and health assessment, and testing procedures have been described elsewhere ( Nilsson et al., 1997 ). PI in working memory was measured using the n-back task. In the n-back task, participants are instructed to indicate whether the word presented in the current trial matches the one presented 2 trials earlier by pressing a button for "yes" and another button for "no." In this study, an interference version of the n-back task was used ( Gray et al., 2003 ;Jonides & Nee, 2006 ;Marklund & Persson, 2012 ). In this version of the task, target trials are intermixed with 2 kinds of non-target trials, lure trials, and unfamiliar trials . In lure trials, the currently presented word matches one presented 3 or 4 trials earlier, thus inducing familiarity but requiring a negative response. Thus, lure trials require participants resolve PI in order to give an accurate response. In unfamiliar trials, the currently presented word has not been presented previously, thus requiring a negative response. In target trials the currently presented word matches the one presented 2 trials earlier (2-back), thus requiring a positive response.
Participants were instructed to answer as quickly and accurately as possible. Each word was presented for 2500 ms, with a 2000 ms inter-stimulus interval, and a total of 40 words were presented (9 target trials, 8 3-back trials, 2 4-back trials, and 21 unfamiliar trials). The first 2 trials in the task were excluded from the analysis due to high predictability.
Difference scores were calculated by dividing performance on lure trials with performance on new trials for both accuracy and reaction time (RT). These scores, hereinafter referred to as PI scores, represent the proportional difference between lure trials (familiar, negative) and unfamiliar trials (unfamiliar, negative). Calculating the scores in this way overcomes the problem of overestimation of age-related deficits caused by general between-group baseline differences where older adults generally perform worse ( Faust et al., 1999 ). Independent scores for accuracy and RT were initially calculated, where positive PI scores reflect more proactive interference (i.e., reduced ability to control interference). Previous observations from the Betula sample ( Samrani & Persson, 2021 ) show that RT-based scores are modestly affected by chronological age, and are only weakly associated with other cognitive functions. This may indicate low reliability for RT-based estimates of PI. Accuracy-based PI measures demonstrate both robust age-effects and are also strongly associated with other cognitive functions and may therefore be a more reliable estimate of PI. Consequently, all statistical analyses are based on accuracy-derived estimates of PI.

Image preprocessing
Full details of DTI preprocessing have been described elsewhere ( Nyberg & Salami, 2014 ;Salami et al., 2012 ). Briefly, diffusionweighted data were analyzed using the University of Oxford's Center for Functional Magnetic Resonance Imaging of the Brain (FM-RIB) Software Library (FSL) package ( http://www.fmrib.ox.ac.uk/fsl ). Before imaging preprocessing, 3 sessions of the subject-specific diffusion acquisitions were concatenated into a 4D file for each subject. Then the raw images were corrected for eddy-current induced distortions and head movement by full affine aligning to the first no-diffusion weighted image (b = 0). The transformation matrix was then used to rotate bval and bves files. A binary brain mask was then generated using the first B0 image with the Brain Extraction Tool (BET). Finally, the preprocessed diffusion-weighted images were fitted using the tensor model. The eigenvalues in 3 directions were obtained by matrix diagonalization for each voxel within the brain mask. Voxel-wise maps of FA were generated using the 3 eigenvalues.

Tract-based spatial statistics (TBSS)
First, all FA images were registered to MNI space using the high-resolution standardized image (FMRIB58_FA) as a target image. Next, a mean FA image was generated by averaging all the transformed FA images. The transformed FA images were concatenated into a single 4D file. The mean FA image was fed into the skeleton-generation algorithm to produce a WM skeleton. Finally, each subject's FA image was skeletonized by projecting onto the group WM skeleton mask. Voxel-wise statistics were run on skeletonized FA images.

Statistical analyses
In the current study, relationships between cross-sectional estimates of FA and PI, as well as relationships between changes of brain markers and change in PI were assessed using partial correlation coefficients. Moderation analyses were conducted using the SPSS plugin PROCESS version 3.5 (model 1) with FA as the independent variable, PI as the outcome variable, and age as the moderator. The models were conducted with 10,0 0 0 bootstrap samples, a confidence interval of 95%, and employed the Johnson-Neyman technique to check for specific regions of significance. Analyses were conducted separately for each of the investigated tracts (N = 12).
Linear mixed models (LMMs) were conducted to examine change in PI and global WMI, respectively, between baseline and follow-up. These models included a random slope, timepoint and age as fixed effects, and an age x timepoint interaction term. Partial eta squared ( ƞ p 2 ) was used to measure effect size. False discovery rate (FDR; Benjamini & Hochberg, 1995 ) was used to adjust for multiple comparisons, and FDR-corrected p -values are reported in addition to uncorrected p -values. FDR-correction were based on number of WM tracts (n = 12) included in the analysis. Since age is associated with both reduced PI and lower WM integrity, and sex is associated with WM integrity, we controlled for age and sex in all analyses of relationships between PI and WM integrity. Statistical analyses were performed using SPSS software ver. 26.0 (IBM, Armonk, NY, USA).
Complementary analyses were conducted between PI and mean diffusivity (MD), radial diffusivity (RD), and axial diffusivity (AD). Results from these analyses were in line with the results reported from the PI -FA analyses, except for AD which showed weaker associations with PI. This pattern of results across indices of WMI are consistent with previous DTI aging studies showing that FA correlates significantly with RD, but not AD (e.g., Bhagat & Beaulieu, 2004 ;Davis et al., 2009 ;Madden et al., 2009;Zhang et al., 2010 ). Full details of the PI -MD, PI -RD, and PI -AD analyses are provided as supplementary tables 1, 2, and 3, respectively.
A mediation analysis was conducted using the SPSS plugin PRO-CESS version 3.5 (model 4) with age as independent variable, PI as the outcome variable, and global WMI as the mediator variable at baseline. The mediation model was conducted with 10,0 0 0 bootstrap samples and a confidence interval of 95%. Mediation analyses were also conducted separately for the 2 age groups with the same set up.

White matter tracts
Global WMI as well as 11 specific WM tracts were selected for the analyses. These tracts were defined based on the JHU-ICBM-DTI-81 atlas and included: the corpus callosum (1) genu, (2) body, (3) splenium), (4) the internal capsule, (5) the corona radiata, (6) the fornix, (7) the external capsule, (8) the posterior thalamic radiation, (9) the cingulum cingulate gyrus, (10) the cingulum hippocampus, and (11) the sagittal stratum. Since we did not have any specific hypothesis regarding regional specificity or laterality, we averaged the FA-values from both left and right hemisphere, as well as also combing estimated from different parts of some tracts (e.g., internal capsule and corona radiata) in order to reduce the number of statistical comparisons and the likelihood of type I errors. Global FA was estimated from the whole WM skeleton generated in TBSS.

Voxel-wise analysis
We performed voxel-wise analysis for FA images using the Randomize tool (part of FSL). General linear models (GLM) were fitted to the FA images to investigate the association between PI scores and WMI, with age and sex as nuisance variables. For each model, the Randomize toolbox ( https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/ Randomise ) was used to assess the regression coefficients for each voxel and to generate t-statistic maps. Threshold-free cluster enhancement (TFCE) was then applied on the t-maps. 50 0 0 permutations were performed for each contrast, resulting in a minimum possible p -value of 0.0 0 02 (1/P). The p -value less than 0.05 was considered statistically significant. Results from these analyses are provided in Supplementary Fig. 1.

Older age is associated with lower WMI and higher PI
In line with previous reports ( Samrani & Persson, 2021 ), crosssectional partial correlation analyses, controlling for sex, demonstrated a positive correlation between proactive interference and age ( r [265] = .447; p < 0.001), indicating that older age is associated with increased PI. Full details can be found in Table 2 . Similarly, in line with previous reports (e.g., Gunning-Dixon & Raz, 2003 ;Head et al., 2002 ;Kennedy & Raz, 2009 ;Nyberg et al., 2012 ;Peters et al., 2014 ), WMI was found to decrease as a function of age both globally ( r [260] = -0.612, p < 0.001) and in all investigated WM tracts ( p uncorrected < 0.0 01; p FDR < 0.0 01). Thus, indicating that older age is associated with less WMI. Full details are presented in Table 3 .
The LMM with PI as outcome variable, age and time (baseline, follow-up) as fixed effects and an age × time interaction showed a significant effect of age (F = 100.36; df = 279.58; p < 0 . 001) but no significant main effect of time or age × timepoint interaction. The LMM with WMI as the outcome variable, age and time as fixed effects and an age × time interaction showed a significant main effect of time (F = 14.19; df = 259,18; p < 0 . 001) and age (F = 96.15; df = 282.64; p < 0 . 001), as well as a significant age × time interaction (F = 11.88; df = 261.97; p = 0.001).

Across all participants, global, and tract-specific FA was associated with ability to control PI
Investigating the relationship between PI and WMI at baseline

Analyses of moderation
The moderation analyses conducted with FA as the independent variable, PI as the outcome variable, and age as the moderator did not find age to be a significant moderator of the FA -PI relationship with only one tract (cingulum cingulate) approaching significance ( p = 0. 07). Results from the Johnson-Neyman procedure showed a trend toward significant FA -PI associations after specific age-points. Specifically, global FA was significantly associated with PI in participants above 48.97 years. Similarly, FA -PI associations for the corona radiata, the posterior thalamic radiation, the external capsule, and the cingulum cingulate were significant in participants above 55.69, 57, 48.63, and 60.73 years, respectively. FA in the internal capsule was significantly associated with PI in participants between 40.9 and 71.53 years. Results from PI -WMI correlations for each age group are displayed in Fig. 2 .

Longitudinal changes in PI and association with FA changes
Change (PI) -Change (FA) correlations revealed a negative correlation between changes in PI and FA in the fornix ( r [143] = -0.188, p uncorrected = 0 .023, p FDR = 0.28), indicating that increased PI over time was associated with decreased FA over time in this tract. In the age-stratified analyses, this correlation persisted with increased strength ( r [61] = -0.392, p uncorrected = .001, p FDR = 0.012) in the older group but was not present in the younger group. However, comparisons of the strength of correlation coefficients, using Fisher r-to-z transformation, between age groups revealed no significant difference (z = , 1.7, p = 0. 09).

Analyses of mediation
The mediation analysis conducted with age as the independent variable, PI as the outcome variable, and global FA as the mediator, revealed FA to be a significant mediator in the age-PI relationship ( Fig. 3 ) Follow-up age-stratified analyses showed that WMI had a mediating effect on the age -PI relationship in the older ( Fig. 4 a; N = 128; estimated indirect effect: 0.0877, LL CI: 0.0284, HL CI: 0.1541), but not the younger age-group ( Fig. 4 b; N = 139; estimated indirect effect: 0.0427, LL CI: -0.0220, HL CI: 0.1155). In older adults, all direct effects were significant. The effect of age, when WMI was included in the regression model, was reduced below significance (estimated direct effect: 0.1614), indicating a mediat-  ing effect of WMI on the age -PI relationship in older adults. WMI accounted for 35% of the total variance in the model.
In younger adults, the direct effect of age -FA and age -PI were significant while the FA -PI effect was not significant (estimated direct effect: -0.1085). The effect of age on PI was highly significant both before and after inclusion of WMI in the regression model, indicating that WMI did not mediate the relationship between the 2 variables. Furthermore, the indirect effect of WMI was not significant, and could accounted for 12% of the variance in the age -PI relationship.

Control analyses
Two sets of control analyses were performed. First, partial correlational analyses controlling for age, sex, and education were performed to investigate if including education as a covariate had a significant impact on the main findings in the study. Secondly, partial correlational analyses with a sample excluding participants diagnosed with dementia between 2015 and 2019 (3 participants) were performed to control for possible impact of preclinical de-mentia. Both sets of analyses replicated the main findings of the study. Full details on the results from these analyses are available in Supplementary table 4 (controlled for education), and Supplementary table 5 (controlled for preclinical dementia).

Discussion
In the present study we set out to investigate the relationship between PI, WMI, and aging. Four main findings were obtained from the current study. First, we demonstrate that higher levels of PI in working memory can be linked to lower WMI across multiple WM tracts, as well as with global WMI. Second, a negative association was found between changes in the ability to control PI over 5-years and 5-year change in WMI in the fornix. Third, Johnson-Neyman results and age-stratified analyses revealed that associations were primarily found in the older group, in line with our prediction that the relationships would be more pronounced in this group. Fourth, we show that the relationship between age, and PI in working memory was mediated by WMI.
The present study is the first investigation of associations between PI and WMI in a large population-based sample across the adult life span and, thus, the first study demonstrating a link between PI and WMI in aging. While the relation between WMI and PI has not been the target of previous investigations, several studies have targeted WM -cognition relationships in related cognitive domains including working memory ( Bendlin et al., 2010 ;Charlton et al., 2010 ;Kennedy & Raz, 2009 ), executive functioning ( Bendlin et al., 2010 ;Davis et al., 2009 ;Halliday et al., 2019 ;Jacobs et al., 2013 ;Kennedy & Raz, 2009 ;Li et al., 2018 ;Madden et al., 2009 ), and processing speed ( Bendlin et al., 2010 ;Halliday et al., 2019 ;Jacobs et al., 2013 ;Kennedy & Raz, 2009 ). The current results extend these previous demonstrations to the domain of PI in working memory.
Comparing the findings from these studies with the present results reveal some overlapping WM tracts. For example, the corona radiata, internal capsule, and cingulum have been found to correlate significantly with both processing speed ( Bendlin et al., 2010 ;Jacobs et al., 2013 ), executive functioning ( Bendlin et al., 2010 ;Halliday et al., 2019 ;Jacobs et al., 2013 ) and working memory performance ( Bendlin et al., 2010 ), and significant associations have additionally been found between executive functioning and WMI in the external capsule ( Jacobs et al., 2013 ) and the posterior thalamic radiation ( Halliday et al., 2019 ). While this might indicate a specific role for these tracts in higher order cognition, current observations of WM -PI associations across multiple tracts rather suggests that the relationships are global in nature. This is further supported by the finding of a significant association between PI and average global WMI across all investigated tracts. Thus, the fact that the significant PI -WMI associations were not restricted to specific tract(s) or regions of the brain and that there was a strong correlation with global WMI, indicates that PI is related to global WMI rather than the integrity in specific tracts.
The change(PI) -change(FA) analyses revealed a negative association between PI and FA in the fornix which persisted in the older group with increased robustness but was absent in the younger group. Thus, individuals with a reduced ability to control PI also showed reduced WMI in the fornix over 5 years. The fornix is typically associated with memory performance due to its bidirectional connection between hippocampus and subcortical structures. While human studies on fornix damage are scarce due to the rarity of such focal damage, the available studies report that such damage results in anterograde amnesia while not affecting either working or procedural memory or executive, and language functions (e.g., Park et al., 20 0 0 ;reviewed in Benear et al., 2020 ). Animal studies of fornix damage show complementary results, with impairment in performance in tasks relying on the hippocampus ( Kwok and Buckley, 2006;Maren and Fanselow, 1997;Wilson et al., 2007 ). Indeed, it has recently been suggested that the effects of forniceal damage on cognitive functioning resemble the deficits obtained from hippocampal lesions ( Benear et al., 2020 ).
While control of PI has typically been related to functions of the prefrontal cortex, recent evidence also highlights a role of the medial temporal lobe, including the hippocampus. For example, taskrelated fMRI has implicated that the medial temporal lobe as critical for the control of PI ( Oztekin et al., 2009;Samrani & Persson, 2022 ). By this view, resolving PI that is associated with a lure item requires reinstatement of item -context information in order to successfully reject the item based on non-matching temporal context. Since hippocampal functioning relies on intact neural communication with other cortical regions, reduced WMI of the fornix, which serves as its major output tract, could have a negative impact in the ability to control PI. Thus, previous reports of agerelated reductions in fornix WMI (e.g., Bennett et al., 2015 ;Bennett & Stark, 2016 ;Gunbey et al., 2014 ;Henson et al., 2016 ;Jang et al., 2011 ;Metzler-Baddeley et al., 2011 ;Metzler-Baddeley et al., 2012 ;Pelletier et al., 2013;Stadlbauer et al., 2008;Zahr et al., 2009 ;Zhuang et al., 2012Zhuang et al., , 2013, as well as in amnesic mild cognitive impairment ( Bozoki et al., 2012 ;Rémy et al., 2015 ;Zhuang et al., 2012Zhuang et al., , 2013 and preclinical Alzheimer's disease could explain not only impairments in associative memory in these groups, but may also underlie the ability to control interference using task-relevant item -context information. Indeed, fornix WMI has even been suggested as a potential biomarker for developing a pathologic memory condition ( Fletcher et al., 2013( Fletcher et al., , 2014Yu et al., 2017 a;Zhuang et al., 2013 ).
Moderation analyses, applying the Johnson-Neyman technique, revealed stronger associations between PI in working memory, and WMI in older adults compared to younger adults in several of the investigated WM tracts. These results corroborate previous structural MRI findings that brain volume -cognition interactions may be stronger in older adults ( Burzynska et al., 2012 ;Kaup et al., 2011 ). Furthermore, these results indicate that brain -behavior associations may be non-linear and only detectable in certain subgroups when a certain threshold of brain deterioration has been reached, such as in older adults. However, these results should be interpreted with caution since age was not found to be a significant moderator of the WMI -PI associations. Nonetheless, this finding is in line with the prediction of a more pronounced PI -WMI relationship in older age as the WM starts to deteriorate.
Mediation analyses revealed that global WMI was a significant mediator of the negative relationship between age and PI. Agestratified analyses showed that this mediating effect of WMI was present in the older, but not in the younger age-group, indicating that global WMI contributes to the age-related decline seen in PI. In the full sample, WMI explained 28% of the variance in the age -PI relationship, while the corresponding explanatory power when restricting to the older age group was 35%, and 12% when analyzing the younger age-group. This demonstrates a greater explanatory power of WMI in older adults, accounting for more than a third of the total effect.
Only a handful of studies have examined the mediating effect of WMI on age -cognition associations, and no study to date has specifically investigated if WMI mediates the age -PI relationship. However, some studies have found that WMI mediate age -cognition relationships in other, related, cognitive domains ( Bennett et al., 2011 ;Borghesani et al., 2013 ;Brickman et al., 2012 ;Burgmans et al., 2011;Gold et al., 2010;Johnson et al., 2015;Madden et al., 2009 ;Li et al., 2018 ;Perry et al., 2009 ;Salami et al., 2012 ;Zahr et al., 2009 ). For example, Burgmans et al. (2011) were able to demonstrate that WM axonal integrity partially mediated the age -processing speed relationship. Similarly, Salami and colleagues (2012) found a partial mediation effect of WMI, as indexed by both FA and mean diffusivity, on processing speed. Such mediating effects have also been found for executive functioning ( Brickman et al., 2012 ;Borghesani et al., 2013 ;Li et al., 2018 ) and working memory ( Zahr et al., 2009 ). Considering these findings together with our results indicate that WM can mediate age-related differences in speed-dependent cognitive functions.
There are several explanatory models for classifying the nature of mediating relationships. The most popular being the modal model, according to which WMI mediates the age -cognition relationship. This model is consistent with the disconnection theory and receives support from observations that the age -cognition relationship is weakened when controlling for WMI (e.g., Borghesani et al., 2013 ;Madden et al., 2009 ;Salami et al., 2012 ;Samanez-Larkin et al., 2012 ). Despite the popularity of the modal model, there are alternative models that also have received support. One model proposes that cognitive performance and WMI are independent and only related to each other through their sep-arate correlation with age. This model receives support from observations of reduced WMI -cognition relationships when controlling for age (e.g., Johnson et al., 2015;Lockhart et al., 2012 ;Salami et al., 2012 ). Another possible explanation is that the relation between age, WMI, and the cognitive variable is mediated by some other, unknown, variable such as a lifestyle variable. In the present analysis, only the modal model was tested, and supported through the observed decrease of age -PI effect when including WMI as a mediator in the regression model.
The observation of a mediating effect in older adults, in parallel with an absent mediating effect in younger adults, is in line with predictions, and can be understood as reflecting the fact that WMI remains quite stable until late adulthood in which there is a substantial decline (e.g., Salat, Kaye, & Janowsky, 1999 ;Guttmann et al., 1998 ). WMI is likely influenced by 2 different factors -genetic variability and life-style factors, and variability related to age-related decline. Before the onset of age-related individual differences there is less variability. Thus, in the older population we see a greater variability as the WMI is affected by more factors.
While we believe that the current study has several strengths, including a relatively large sample size with a wide age-range, and the availability of both cross-sectional and longitudinal data, a few limitations should be highlighted. First, while the finding of a change -change association between WMI in the fornix and PI is an interesting one, it should be noted that estimation of FA in this particular tract can be sensitive to artefacts. The fornix is located in close proximity to cerebrospinal fluid (CSF), thereby subjecting the tract to the risk of partial volume effects ( Concha et al., 2005 ;Jones & Cercignane, 2010 ). This may affect the estimation of FA in the tract which may have influenced the observed changes in FA between baseline and follow-up. Second, the measure used to estimate PI could be affected by practice and/or regression to the mean effects. Such effects could be reduced by the inclusion of more test occasions or by including more measures of PI. Additional studies spanning a larger time period with increased testoccasions could help reduce retest effects and/or distortions on the behavioral measure. Thus, such studies could help clarify how the relationship changes over time thereby providing causal insights on how the PI -WM relationship changes across aging. Third, the mediation analyses display interesting dedifferention patterns between older and younger adults with WMI observed to mediate the age -PI relationship in the former, but not the latter age group. While these results are in line with expectations, there may be confounding variables not accounted for in the regression model. Other factors could be related to either WMI or PI or both and thereby influence the mediating effect in different directions and/or ways. It is thus possible that some portion of the variance attributed to the mediating effect of WMI in fact should be attributed to another variable not controlled for in the present model. For example, episodic memory performance has been found to be correlated with both aging (e.g., Nyberg & Tulving, 1996 ), and PI in working memory (e.g., Samrani & Persson, 2021 ). Episodic memory is also closely related with the fornix with forniceal damage resulting in anterograde amnesia (reviewed in Benear et al., 2020 . In conclusion, the present results contribute to the knowledge of the neurobiological basis of proactive interference by demonstrating that increased PI in older adults is associated with reduced WMI. The results demonstrate, for the first time, associations between PI and WMI both cross-sectionally and longitudinally. We observed negative 5-year change -change associations between PI and WMI in the fornix that could be related to decreased memory performance but finding needs to be further investigated. Additionally, results demonstrated stronger PI -WMI associations amongst older adults, in line with expectations that WMI become increasingly important for cognitive functioning in older age. Finally, mediation analyses revealed WMI to be a significant mediator of the age -PI relationship in older, but not younger adults. This provides evidence for a relationship between WMI and PI in line with the cortical disconnection explanation for cognitive decline during aging. Collectively, these findings extend upon the existent body of research on age-related changes in higher cognitive functions by providing new insights about the neurobiological underpinnings of how control of interference is affected by increasing age.

Supplementary material
Supplementary fig. 1. Results from the voxel-wise analyses showing negative associations between FA and PI in all participants (top), older adults (middle), and younger adults (bottom), controlling for age, and sex. Significant voxels ( p < 0.05 are overlaid on a T1-weighted image (coordinates according to the MNI152 template), with yellow color indicating lower p values.
Supplementary table 1. Correlations coefficients for PI -MD associations at T5.
Supplementary table 2. Correlations coefficients for PI -RD associations at T5.
Supplementary table 3. Correlations coefficients for PI -AD associations at T5.
Supplementary table 4. Correlations coefficients for PI -FA associations at T5.
Supplementary table 5. Correlations coefficients for PI -FA associations at T5, controlled for age and sex, excluding participants diagnosed with dementia between 2015-2019.

Disclosure statement
Pernilla Andersson: Formal Analysis, Writing -original draft, Writing -review & editing; Xin Li: Formal analysis, Writing -review & editing and Jonas Persson: Writing -original draft, Writing -review & editing, Supervision, Funding acquisition, Conceptualization. The authors have no actual or potential conflicts of interest.

Data availability statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.

Verification
The work described in the present manuscript submission has not been published previously, is not under consideration for publication elsewhere, its publication is approved by all authors and tacitly or explicitly by the responsible authorities where the work was carried out, and, if accepted, it will not be published elsewhere in the same form, in English or in any other language, including electronically without the written consent of the copyright-holder.