The trajectory of cortical GABA across the lifespan, an individual participant data meta-analysis of edited MRS studies.

γ-Aminobutyric acid (GABA) is the principal inhibitory neurotransmitter in the human brain and can be measured with magnetic resonance spectroscopy (MRS). Conflicting accounts report decreases and increases in cortical GABA levels across the lifespan. This incompatibility may be an artifact of the size and age range of the samples utilized in these studies. No single study to date has included the entire lifespan. In this study, eight suitable datasets were integrated to generate a model of the trajectory of frontal GABA estimates (as reported through edited MRS; both expressed as ratios and in institutional units) across the lifespan. Data were fit using both a log-normal curve and a nonparametric spline as regression models using a multi-level Bayesian model utilizing the Stan language. Integrated data show that an asymmetric lifespan trajectory of frontal GABA measures involves an early period of increase, followed by a period of stability during early adulthood, with a gradual decrease during adulthood and aging that is described well by both spline and log-normal models. The information gained will provide a general framework to inform expectations of future studies based on the age of the population being studied.

Introduction

Magnetic resonance spectroscopy (MRS) is a non-invasive imaging technique that allows for the measurement <span class="Gene">of levels <span class="Gene">of metabolites. Of particular interest to the neurosciences is the measurement of specific neurotransmitters such γ-aminobutyric acid (GABA) in vivo (Edden and Barker, 2007; Mescher et al., 1998; Mullins et al., 2014; Puts et al., 2011; Rothman et al., 1993). GABA is the main inhibitory neurotransmitter in the human nervous system and plays a fundamental role in central nervous system function (Buzsáki et al., 2007). A number of studies have explored the relationship between cortical GABA (as measured with MRS) and age in various contexts. These studies have found that aging-related changes in GABA are consistently associated with cognitive and neurophysiological outcomes that change across the lifespan. These findings have important implications for both healthy and pathological development and aging.

Most prior studies <span class="Gene">of aging-related differences in <span class="Chemical">GABA have utilized a restricted age range. Some have reported increases in GABA as age increases in left frontal cortex between 13 and 53 years (Ghisleni et al., 2015). Others have reported decreases in medial frontal GABA (Marenco et al., 2018, 18–55 years; Porges et al., 2017a, 43–92 years; Rowland et al., 2016, 16–62 years) and GABA/creatine plus phosphocreatine (Cr+PCr) (Gao et al., 2013, 20–76 years). Gao and Porges show the same aging-related decrease in midline parietal GABA/Cr+PCr and GABA, respectively, as does (Simmonite et al., 2019) for GABA/Cr+PCr in occipital regions (18–87 years). Still others reported no significant aging-related changes in GABA (Aufhaus et al., 2013, 21–53 years) or GABA/Cr+PCr (Mikkelsen et al., 2017, 18–48 years). This inconsistency makes the results difficult to interpret, as partially overlapping age ranges produce conflicting trajectories. For example, the age range of participants reported by Gao et al., 2013 has substantial overlap with those reported by Ghisleni et al., 2015, yet has apparently conflicting trajectories. Of interest, all studies were cross-sectional rather than looking at within-subject change in GABA with age.

However, conflicting results only exist when looking at different age ranges and consistent findings have been shown in studies <span class="Gene">of similar age range. For example, both Porges et al., 2017a and Gao et al., 2013 focus on adults through advanced age and both r<span class="Gene">eport aging-related decrease in GABA. Here, we predict that this apparent conflict in the relationship between GABA and age in the literature is the result of restricted age ranges within each study and that the various linear relationships are part of larger, non-linear asymmetric lifespan trajectory of GABA with age. This would be consistent with prior work showing non-linear relationships with age across the human lifespan including inhibition-dependent behavior (Williams et al., 1999), GAD65 expression (Pinto et al., 2010), cortical thinning (Fjell et al., 2019; Walhovd et al., 2016; Raznahan et al., 2011), and white matter development (Lebel et al., 2012). With respect to MRS, this prediction of a non-linear asymmetric lifespan trajectory, with an increase during childhood and aging-associated gradual decrease, is further suggested by an exploratory LOESS (locally estimated scatterplot smoothing) regression for macromolecule-suppressed medial-frontal GABA × age in healthy controls reported by Rowland et al., 2016. Based on this prior work, and that of individual study of the relationship between GABA and age, when collating data across studies, we predict to find a non-linear aging-related trajectory involving an increase of GABA in childhood, relative stability in adolescence, followed by a gradual decrease during adulthood. To date, no single study has explored the lifespan trajectory of cortical GABA spanning development, adulthood, and aging. In the absence of a lifespan study, we implemented an individual participant data meta-analytic (IPD-MA) approach following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Moher et al., 2009; see Figure 5) supplemented with data collected by the authors and previously published in summary form (Puts et al., 2017).

The majority <span class="Gene">of MRS studies <span class="Gene">of GABA at 3 T have utilized J-difference editing to selectively ‘edit’ the GABA signal (e.g. Mullins et al., 2014). Editing is necessary at 3 T due to the low concentration of GABA in the human brain (1–2 mM) (Harris et al., 2017). In unedited MRS, signal from higher concentration molecules like N-acetylaspartate (NAA) and Cr+PCr masks the GABA signal. The most widely used J-difference editing MRS technique is MEGA-PRESS (Mescher et al., 1998), in which a GABA-selective editing pulse at 1.9 ppm is applied in half of the experiment (edit-ON and is coupled to a GABA signal at 3 ppm) but not in the other half (edit-OFF, where ON and OFF acquisitions are typically interleaved). The difference spectrum (ON-OFF acquisitions) shows only those signals affected by the frequency-selective editing pulse, revealing a GABA signal at 3 ppm. The difference spectrum is further complicated by a macromolecule signal at 3 ppm which is coupled to another macromolecule signal at 1.7 ppm and thus falls within the envelope of the editing pulse, resulting in a co-edited macromolecule signal as part of the 3 ppm GABA signal (Edden et al., 2014; Henry et al., 2001). Consequently, most studies refer to the GABA signal as ‘GABA+’. Both macromolecule-suppressed and GABA+ measures are included here (Table 2). Due to its low concentration, measurement of the GABA-edited signal in humans requires a large voxel (most commonly 27 cm3; Mullins et al., 2014; Peek et al., 2020) to keep acquisition times reasonable (~10 min) and to provide an adequate signal-to-noise ratio (SNR) (Mikkelsen et al., 2018). This limitation constrains the spatial specificity of the measurement to coarse regions that often lack discrete functional specificity.

The functional relevance <span class="Gene">of MRS measures <span class="Gene">of GABA is important to note. Measured GABA levels include both intracellular (both somatic and synaptic) and extracellular contributions to the overall GABA concentration. However, these relative contributions are not well known; GABA levels measured via MRS at rest describe a physiological characteristic of the tissue measured and, while associated with functional metrics (e.g. neurophysiological response or behavior), are better interpreted in a manner similar to structural neuroimaging. However, studies have suggested that the majority of the GABA measured using MRS reflects intracellular somatic levels rather than synaptic or extracellular levels since MRS measures of GABA were associated with expression of the 67 isoform of glutamic acid decarboxylase (GAD) which is predominantly present in the soma of the neuron (Marenco et al., 2011; Rae, 2014; Stagg et al., 2011). Other than that, no study has shown clear discrimination between contributions of different pools to the GABA signal. As such, the GABA signal is considered to be reflective of inhibitory tone (Rae, 2014). GABA levels measured via MRS in young adults at rest have been reported to be stable for up to 7 months (Near et al., 2014) and do not exhibit a diurnal rhythm (Evans et al., 2010).

In this manus<span class="Chemical">cript, we statistically combine <span class="Chemical">datasets of published research that used MEGA-PRESS to measure GABA in discrete age ranges where individual data points were presented relative to age to present a non-linear model for GABA estimates over the human lifespan using an individual participant data approach. This work is motivated by several studies showing that MEGA-PRESS measures of cortical GABA are relevant to both development and aging with a specific emphasis on cognition and perception. We further discuss the lack of available data across the lifespan.

The importance of GABA in the context of aging in health and disease

<span class="Chemical">GABA as measured with MRS has been linked to clinical and cognitive outcomes. Alte<span class="Species">rations of GABA levels are seen in neurodevelopmental disorders such as ADHD (Bollmann et al., 2015; Edden et al., 2012), autism spectrum disorder (Cochran et al., 2015; Drenthen et al., 2016; Gaetz et al., 2014; Puts et al., 2017), and Tourette syndrome (Puts et al., 2015), as well as in other neurological and psychiatric disorders. These clinical differences are reviewed by Puts and Edden, 2012 and by Schür et al., 2016. Associations with GABA measures have been reported across a variety of sensory and cognitive domains (for a review, see Duncan et al., 2014) associations with other measures of brain function (for a review, see Duncan et al., 2014). Given the functional relevance of GABA in the context of both pathological and healthy cohorts (especially in the context of development and aging), understanding how GABA changes with age in a healthy cohort is important to understand typical and atypical progression of behavior. However, few studies have included the impact of age to their investigations. Our approach allows for a systematic review of existing work studying GABA across development and aging with immediate impact on future studies in the context of development and aging.

GABA across the lifespan

To <span class="Chemical">date, <span class="Chemical">cross-sectional and longitudinal investigations of cortical GABA across the entire human lifespan have yet to be published. However, there have been recent reports investigating the relationship between cortical GABA levels and discrete age ranges in humans that test a linear association focused on a specific population (e.g. ‘aging’). As discussed above, using these studies to make blanket statements about correlations between age and GABA is inappropriate as these studies report a variety of effects. Some suggest a positive correlation between GABA and age, others hint at a negative correlation, and still others conclude that there is no relationship at all. Here, we discuss these reports in the context of human development and divide them into three categories: developmental, adult, and aging.

Developmental

The developmental component of the lifespan <span class="Gene">of cortical GABA as measured by MRS of GABA is explored in less depth than in adult or aging cohorts. Port et al., 2017 and Silveri et al., 2013 report a maturational increase in GABA+/Cr+PCr in typically developing children and adolescents. The majority of evidence comes from non-MRS work, showing dramatic change in GABAergic function during early life. Human autopsy data describes large changes in both GABA synthesis and receptor expression (Pinto et al., 2010) and animal models describe a shift in GABA from excitatory to inhibitory (Leonzino et al., 2016). Yet, in vivo reports in healthy younger populations (particularly infants and young children) are sparse or missing. Several reasons exist for the absence of high-quality MRS data during development with technical challenges being one main challenge. These challenges are not unique to MRS but exist for most – if not all – magnetic resonance modalities. For example, it is well established that imaging of MRS of GABA is highly sensitive to motion (Edden et al., 2016; Mullins et al., 2014), thus compounding the challenges involved when imaging pediatric cohorts. Many studies in pediatric cohorts, including our own (Puts et al., 2017), also suffer from the limitation that age relationships are not reported due to individual studies often studying a restricted age range to minimize developmental effects within the cohort.

Adult

The vast majority <span class="Gene">of studies in healthy populations focus on age ranges between development and aging to minimize the effect <span class="Gene">of age on the measures of interest (and for ease of recruitment). However, this limits the reporting of GABA-age relationships within this range. Mikkelsen et al., 2017 conducted a multisite study collecting GABA+ levels in 272 participants between 18 and 35 years of age, providing a substantial dataset to assess this relationship. The sample size and restriction to healthy adults in this study provide a reasonable representation of normal GABA estimates in the target demographic. Their objective with this study was to report stability of the GABA+ measure across multiple 3 T MRI platforms with systems by GE, Phillips, and Siemens well represented. Voxel placement was selected for the medial parietal lobe. While their original manuscript contains neither a report nor a visualization of the GABA/age relationship, we are able to provide this information for our review (data are freely available from the Big GABA repository, Mikkelsen et al., 2017, https://www.nitrc.org/projects/biggaba/). There was no aging-related increase or decrease between age and GABA+/Cr+PCr [χ2(7)=3.52, pboot = 0.31] in this large cohort of adults between 18 and 35 years of age.

Aging

Most, if not all, MRS <span class="Gene">of <span class="Chemical">GABA studies that investigate aging populations report a decrease in cortical GABA as a function of age in both frontal (Gao et al., 2013; Porges et al., 2017a) and parietal (Gao et al., 2013) voxels. Marenco et al., 2018 also show a decrease in GABA with aging. It is important to note that other reports (Hermans et al., 2018; Maes et al., 2018) have compared MRS of GABA between defined groups of older and younger adults rather than with continuity across the lifespan. These findings are consistent with continuous approaches, with older adults having reduced GABA. However, a categorical approach comparing two groups does little to elucidate the aging-related trajectory. Manuscripts that employ MEGA-PRESS methodology in a manner that is inconsistent with methods outlined in consensus papers (Mullins et al., 2014; Puts and Edden, 2012) have insufficient SNR or other technical limitations that have not been considered for this assessment.

In conclusion, an understanding <span class="Gene">of the link between <span class="Chemical">GABA and age is incredibly important for the study of inhibition across the lifespan, the study of development- and aging-related behavioral and cognitive processes, and the study of health and disease. No study has attempted to study GABA across the entire lifespan. Here, we utilize an IPD-MA approach for combining all existing and eligible cortical edited MRS of GABA data across the lifespan (from childhood development to aging) to build a model that informs us of the best-fit model of GABA across the lifespan. We hypothesize that the model of best fit would be consistent with that of other cortical measures of development and maturation, including indices of white matter (Lebel et al., 2012) and gray matter (Gilmore et al., 2012; Gogtay et al., 2004), as well as EEG power (Whitford et al., 2007), revealing non-linear trajectories, increasing during development and slowly decreasing during aging.

Results

A naïve approach to describing <span class="Chemical">GABA as a function of age is to assume a linear trend. The resulting regression models, fit separately for each of the eight datasets, are depicted in Figure 1 (scaled with respect to each dataset’s geometric mean), with corresponding regression statistics reported in Table 1. A cursory examination of this approach reveals that such an assumption is inappropriate, suggesting that changes in this relationship over time render extrapolation from any of the datasets ineffective. Furthermore, assuming a linear trend is unsuccessful even at describing the data within each study. In six of the eight datasets, the linear trend explained less than 20% of the variance. More dramatically, linear fits were even less successful at predicting extrapolated trend beyond each study’s age range, and although the slope in the dataset with the youngest participants was positive, slopes tended to become more negative as the age of the participants increased. When combined into a naïve aggregate of all studies, the extrapolated trendlines show no coherent pattern (Figure 1, right panel). Although a linear trend may provide an approximate summary over short periods of time, a linear trend over the entire lifespan is not appropriate. Since unqualified statements about the correlation between age and any neurophysiological measure are, effectively, linear models, Figure 1 also shows how poorly such statements extrapolate to larger datasets. With this in mind, our meta-analysis did not rely on a linear trend. Instead, we sought to balance our estimation of the overall non-linear function that described the common pattern across datasets with the need to control for systematic differences between studies.

Figure 1.

Linear relationships between age and γ-aminobutyric acid (GABA) signal, showing that linear extrapolation over the lifespan is not appropriate.

In each dataset, GABA was scaled relative to the geometric mean. Linear models were fit for each dataset separately. Dark shaded regions represent the 95% credible interval for the interpolated regression line, given the data from each study and the assumption of a linear effect, whereas the light shaded regions represent the 95% credible interval for the extrapolated regression line.

Table 1.

Regression statistics for simple linear fits.

Intercepts are omitted because rescaling causes them to be entirely determined by the slopes and the mean age. Further study details can be found in Tables 2 and 3.

GABA study	Mean slope	Lower and upper bounds (0.025–0.975 quantiles)	Mean residual σ	Lower and upper bounds (0.025–0.975 quantiles)	R2
Aufhaus	−0.0008	−0.0055 to 0.0041	0.155	0.125 to 0.194	0.003
Gao	−0.0075	−0.0092 to −0.0058	0.121	0.106 to 0.140	0.459
Ghisleni	0.0034	0.0007 to 0.0061	0.111	0.092 to 0.134	0.106
Mikkelsen	−0.0019	−0.0054 to 0.0016	0.129	0.117 to 0.142	0.005
Porges	−0.0099	−0.0133 to −0.0064	0.172	0.148 to 0.200	0.279
Puts	0.0436	0.0085 to 0.0795	0.215	0.188 to 0.249	0.058
Rowland	−0.0032	−0.0060 to −0.0004	0.176	0.151 to 0.207	0.061
Simmonite	−0.0023	−0.0040 to −0.0005	0.160	0.127 to 0.203	0.156

In each <span class="Chemical">dataset, <span class="Chemical">GABA was scaled relative to the geometric mean. Linear models were fit for each dataset separately. Dark shaded regions represent the 95% credible interval for the interpolated regression line, given the data from each study and the assumption of a linear effect, whereas the light shaded regions represent the 95% credible interval for the extrapolated regression line.

Interc<span class="Gene">epts are omitted because rescaling causes them to be entirely determined by the slopes and the mean age. Further study details can be found in Tables 2 and 3.

Table 2.

Neuroimaging acquisition and analysis details for eight studies included in the analysis.

Reference method refers to either reference to water (in estimated concentration/H2O) or as a ratio to creatine plus phosphocreatine (Cr+PCr) and describes whether data was acquired macromolecule-suppressed (γ-aminobutyric acid [GABA]) or as GABA+ macromolecules (GABA+). MRS averages refer to the number of ON + OFF transients. *The manuscript refers to 96 averages. It was clarified with the authors that this referred to 96 ON and 96 OFF averages.

GABA study	Type of scanner	Analysis software	Reference method	Voxel volume (ml3)	MRS means	TE (ms)	TR (ms)	Voxel location
Aufhaus et al., 2013*	3 T Siemens	jMRUI/LCModel	GABA/H2O	24	192*	68	3000	Medial frontal lobe
Gao et al., 2013	3 T Philips	jMRUI	GABA+/Cr+PCr	27	320	68	2000	Medial frontal lobe
Ghisleni et al., 2015	3 T GE	LCModel	GABA+/H2O	30	320	68	2000	Left dorsolateral prefrontal lobe
Mikkelsen et al., 2017	3 T GE/Philips/Siemens	Gannet	GABA+/Cr+PCr	27	320	68	2000	Medial parietal lobe
Porges et al., 2017a	3 T Philips	Gannet	GABA+/H2O	27	320	68	2000	Medial frontal lobe
Puts et al., 2017	3 T Philips	Gannet	GABA+/H2O	27	320	68	2000	Right precentral sulcus
Rowland et al., 2016	3 T Philips	Gannet	GABA/H2O	24	256	68	2000	Medial frontal lobe
Simmonite et al., 2019	3 T Philips	Gannet	GABA+/Cr+PCr	22.5	256	68	1800	Medial occipital lobe

Table 3.

Descriptive statistics for eight studies included in the analysis.

This gives a basic description of sample size and age range for the eight datasets. Additionally, Figure 6 depicts the distribution of ages using a raincloud plot (Allen et al., 2018).

GABA study	# of subjects	Mean age	Age (SD)	Age range	Reference
Aufhaus	44	35.5	10	21–53	Aufhaus et al., 2013
Gao	96	45.7	14.5	20–76	Gao et al., 2013
Ghisleni	55	27.2	11	13–53	Ghisleni et al., 2015
Mikkelsen	220	26.5	4.9	18–48	Mikkelsen et al., 2017
Porges	86	71.8	10.6	43–92	Porges et al., 2017a
Puts	101	10.3	1.2	8–13	Puts et al., 2017
Rowland	82	38.0	13.7	18–62	Rowland et al., 2016
Simmonite	38	50.1	29.2	18–87	Simmonite et al., 2019

To provide a meta-analytic synthesis <span class="Gene">of these <span class="Chemical">datasets without biasing our result through an arbitrary choice of our function  we fit two models: both non-linear and able to accommodate the pattern visible from the individual trendlines.

The estimate for our basis spline model is depicted in Figure 2 (left). Overall, this time course is characterized by an in<span class="Chemical">crease in GABA during childhood, followed by a plateau from adolescence through midlife, and then a gradual decline from approximately 40 years onward. Note that although Figure 2 (left) depicts the mean spline, any of the wide variety of smooth curves that fit within the shaded credible interval are in principle nominated by this analysis. A future expansion of this model to a larger assemblage of datasets would further refine which of those curves is a suitable candidate for the population’s canonical function. The estimate for our log-normal model is depicted in Figure 2 (right).

Figure 2.

Non-linear regression models of γ-aminobutyric acid (GABA) signal integrating all data simultaneously.

The shaded region depicts the 95% credible interval for the mean. (Left) Penalized basis spline model. (Right) Log-normal model.

The shaded region d<span class="Gene">epicts the 95% <span class="Chemical">credible interval for the mean. (Left) Penalized basis spline model. (Right) Log-normal model.

To confirm that the models in Figure 2 did a reasonable job <span class="Gene">of combining the <span class="Chemical">data despite their differing origins and methodologies, we examined the posterior estimates of the scaling factors. Figure 3 (left) shows the posterior estimates for each dataset’s scaling factor for our basis spline model, whereas Figure 3 (right) shows the estimates for the log-normal model. These scaling factors were given broad prior distributions. Rather than trying to characterize possible contributions to each scaling factor (e.g. by reference method or scanner manufacturer), we allowed the scaling factors to be estimated in a manner that was agnostic to any label beyond which dataset each observation came from. Thus, scaling factors were governed only by (1) the observed values of the data and (2) the constraints of the model.

Figure 3.

Posterior estimates of the relative scaling factor for each study in the penalized basis spline model (left) and the log-normal model (right), sorted by reference method.

Boxes represent the 80% credible interval for the posterior estimate, whereas whiskers represent the 95% credible interval.

Boxes represent the 80% <span class="Chemical">credible interval for the posterior estimate, whereas whiskers represent the 95% credible interval.

Despite our decision to ke<span class="Gene">ep our priors broad and agnostic, the model nevertheless recapitulates expected differences in measured outcomes. The Big <span class="Chemical">GABA series of meta-analyses (Mikkelsen et al., 2017; Mikkelsen et al., 2019) are the largest published multi-site studies to date characterizing GABA measurement, and our estimated scaling factors line up with the score ranges consistent with those reported in that series of studies. For example, our three Cr+PCr-referenced datasets (Gao, Mikkelsen, and Simmonite) all yielded scaling factors well below 1.0, consistent with the measurement scale typically observed in GABA+/Cr+PCr studies (Mikkelsen et al., 2017), whereas water-referenced data yielded scaling factors one or two orders of magnitude larger than Cr+PCr-referenced data, also consistent with typical measurement (Mikkelsen et al., 2019). In this respect, our estimated scaling factor not only put our observations on a comparable scale but did so in a way that provides a sanity check against published findings.

Importantly, however, the scaling factor estimates are themselves distributions, and those distributions d<span class="Gene">epend not only on the variation in the original <span class="Chemical">data but on their covariance with other variables and on the model assumptions. The scaling factor estimates display less uncertainty in the log-normal case because that model imposes much stronger assumptions on the time course of the data than would be expected from a spline model. The estimated scaling factor for the Mikkelsen data varies less than that of the Simmonite data in part because the former dataset is nearly six times as large. Since every parameter estimate within each model covaries with every other parameter estimate, and furthermore is relative to the other reported factors, it is important to treat these scaling factors as strictly relative and contingent on the current data and model assumptions.

To ensure that the directionality <span class="Gene">of the ‘late-age’ component was not driven by differences in Cr+PCr, we performed an additional ‘leave-one-out’ (LOO) cross-validation approach (Arlot and Celisse, 2010) for each of the three datasets that included late-age participants. Figure 4 shows that when datasets are left out, the general shape of both models remains the same, with wider 95% credible intervals for the oldest of age due to limited available data. This suggests that the overall pattern of a negative slope in late age is unlikely to be driven by Cr+PCr differences, as the data contributed by Porges et al., 2017a made the largest contribution to this slope (having the most observations to contribute) and relied on a water-based reference method. That said, since the Porges dataset contains most of the observations over 60 in this analysis, its removal leaves only a handful of cases from the Simmonite dataset as evidence for late-life model estimates, resulting in substantial uncertainty in the spline model. By contrast, the age ranges represented by Gao and Simmonite have good coverage by other studies, so their removal does not have as large an impact on the estimate.

Figure 4.

Non-linear regression models of γ-aminobutyric acid (GABA) signal performed using a leave-one-out (LOO) cross-validation approach.

The gray shaded region and black line depict the 95% credible interval for the mean for the model and the colored shaded region and dotted line show the model with the respective data left out. In all cases, a similar overall trajectory to the full data is implied by each of the subsets, albeit with greater variation in the posterior estimates.

These data show no substantial impact on the direction of the slope over time.

The gray shaded region and black line d<span class="Gene">epict the 95% <span class="Chemical">credible interval for the mean for the model and the colored shaded region and dotted line show the model with the respective data left out. In all cases, a similar overall trajectory to the full data is implied by each of the subsets, albeit with greater variation in the posterior estimates.

These <span class="Chemical">data show no substantial impact on the direction <span class="Gene">of the slope over time.

Figure 4 also helps to demonst<span class="Species">rate the importance <span class="Gene">of model assumptions: The removal of the Porges dataset affords the spline model the possibility that the model of best fit might level off or even become positive in late life, whereas the log-normal model’s rigid shape is only able to predict a downward trajectory. On the one hand, the log-normal model can probably be criticized for its inflexibility in this respect; however, on the other, this demonstrates that without much data to guide it, the spline model is perhaps too agnostic about the direction of the effect. The spline model’s advantage is that it can be persuaded to turn, provided the evidence is consistent with that conclusion. The log-normal, by contrast, is too inflexible to accommodate a change in direction in late life regardless of the volume of data it is provided.

We additionally also explored the potential effect <span class="Gene">of region by performing the analysis on only frontal <span class="Chemical">data (removing Mikkelsen et al. and Simmonite et al., respectively) with no substantial impact on the direction of the slope over time. This information can be found in Figure 4—figure supplement 1.

Figure 4—figure supplement 1.

To accompany Figure 4, we performed additional analysis on frontal data only (removing Mikkelsen et al. and Simmonite et al., respectively) to investigate whether the inclusion of non-frontal regions biased our data.

These data show no substantial impact on the direction of the slope over time.

Discussion

Here, we show that the aging-related trajectory <span class="Gene">of <span class="Chemical">GABA across the lifespan is characterized by a fairly quick increase in GABA during development, followed by a flattening during adolescence and by a subsequent slow decrease with aging. It should be noted that we included studies reporting GABA levels in institutional units and as measured ratios. No single study with a limited age range reveals such a relationship (as evidenced by Figure 1) – it is only by meta-analysis of multiple datasets that we are able to identify this relationship. Here, we will discuss the case for a biological mechanism that drives a non-linear trajectory of GABA measures with age and then discuss methodological and biological factors that may influence these results. Finally, we review the implications of our evaluation and suggest potential directions of future research characterizing the trajectories of GABA over the lifetime.

Biological mechanisms

Although previous work, including the authors’ own, has generally r<span class="Gene">eported a linear relationship between <span class="Chemical">GABA and age within a given stage of life, it is far more plausible that differences across the lifespan are non-linear, in line with other biological effects. For example, while Gao et al., 2013 might capture a decrease in GABA with age, the broad age range studied in Ghisleni et al., 2015 might prevent observation of a linear relationship. Indeed, a decline in GABAergic interneurons with age has been widely reported in animal models (Hua et al., 2008; Stanley et al., 2012). Post-mortem data from human samples showed a reduction in the 65 isoform of GAD (the enzyme responsible for the production of GABA, GAD) in visual cortex, suggesting reduced GABA production with aging (Pinto et al., 2010). MRS of GABA by itself cannot provide sufficient resolution to determine what these reductions in GABA reflect; we can only make vague assumptions and conclusions on the relationship between brain structure, cognitive function, and potential molecular mechanisms.

In considering a non-linear model for <span class="Chemical">GABA a<span class="Chemical">cross the human lifespan, we find three major stages: (1) a developmental stage where GABA measures increase, (2) a stabilization phase during adulthood where GABA concentrations remain mostly stable, and (3) a gradual descending period of GABA with advanced aging. This is consistent with previous studies of brain structural and cognitive function across the lifespan. Non-linear trajectories have been reported in aging effects of total gray matter (Lenroot et al., 2007; Sussman et al., 2016) and cortical thickness (Shaw et al., 2008). Diffusion tensor imaging studies show non-linear aging-related differences through childhood and adolescence (Lebel and Beaulieu, 2011), with an ‘inverted U-shaped’ trajectory that peaks at approximately 40 years of age (Bendlin et al., 2010; Good et al., 2001; Lebel et al., 2012; Westlye et al., 2010). A similar trajectory pattern for cognitive abilities is reported in memory, verbal ability, and inductive reasoning (Kobayashi et al., 2015), as well as word recall, verbal fluency, math skills (Whitley et al., 2016), and behavioral inhibition (Williams et al., 1999).

GABA differences throughout development

An increase in cortical <span class="Chemical">GABA could be extrapolated from the proliferation of new GABAergic neurons during development. GABA is thought to be linked to the myelination of frontal white matter trajectories by controlling oligodendrocyte precursor cell activity through the developmental phase (Ghisleni et al., 2015; Vélez-Fort et al., 2012). An increase in cortical GABA concentrations could also be the result of increased synaptic activity of GABAergic neurons during development. This potentiation-like mechanism of cortical GABA is supported by studies where astrocytes have been shown to mediate plasticity of rodent hippocampus (Kang et al., 1998) and visual cortex (Chen et al., 2012), suggesting a possible potentiation-like mechanism of cortical GABA concentrations. In addition, upregulation of GAD could lead to increased production of GABA and indeed, GABA levels as measured with MRS have been shown to relate to expression of the GAD1 gene (Marenco et al., 2010). GABAergic neuronal function is reported to become more efficient through synaptic pruning and long-term depression during development (Paolicelli et al., 2011; Wagner and Alger, 1995; Wu et al., 2012). It should be pointed out that no eligible MRS studies of GABA were available for infancy and early development, which is a significant gap in the literature that should be addressed in future work.

GABA differences throughout aging

De<span class="Chemical">creased <span class="Chemical">GABA concentrations during aging are most likely linked to gray matter atrophy and demyelination associated with pathology and normal aging. Animal models have explored the physiological underpinnings and functional implications of these changes. Rodent studies have shown a reduction in the number of interneurons expressing GAD in the medial prefrontal cortex. These reductions were accompanied by altered spatial working memory, linking altered GABA function to altered behavior in aging (Spiegel et al., 2013). Additionally, rodent models of cognitive decline have shown altered function at both GABA-A and GABA-B receptors (McQuail et al., 2015), as well as a reduction in GABAergic neurons in cats and monkeys (Hua et al., 2008; Leventhal et al., 2003).

GABA differences across the lifespan

Very few non-MRS studies have assessed changes or differences in <span class="Chemical">GABA function a<span class="Chemical">cross the lifespan. In those that report differences in the GABAergic system across the lifespan, a pattern of change similar to our findings has been presented. Using GAD labeling methodology in the human visual cortex, GAD65 has been reported to increase early in life and gradually decrease during aging (Pinto et al., 2010). Provocatively, they show GAD67 to be stable across the lifespan. Similarly, a recent cross-sectional exploration of sex differences in young and old adults found reduction in GAD65, but not GAD67 in the superior temporal gyrus of females (but not males). No change in GAD65 or GAD67 was found in other regions, though prefrontal cortex was not reported (Pandya et al., 2019). We are unaware of other reports of normal human aging-related differences in GAD; however medial temporal lobe reductions (prefrontal cortex was not reported) in GAD65, but not GAD67, have also been reported in Alzheimer’s disease (Schwab et al., 2013). Given the more specific relationship between GAD65 and neurotransmission, this may underlie reports of GABA-associated alterations in cognitive function that occur during periods of GABAergic change (Porges et al., 2017a) throughout the lifespan. Future studies should aim to address the relationship between these two different functional isoforms, brain GABA levels, and function in both health and disease.

Regional differences

It can be assumed that regional differences in <span class="Chemical">GABA are not homogeneous. We do not presume that the model we have presented is characteristic <span class="Gene">of aging-related differences in GABA in all neural tissues due to well-known ontological and aging-related regional variation in tissue that influences GABA (Lebel et al., 2012). One could even argue that regional differences are likely to exist within smaller regions of the frontal lobe. However, limiting this regional selectivity even further would have not allowed for our analysis, as (1) we would not be able to correct for covariation between frontal and parietal regions, and (2) inclusion of multiple observations per participant would violate the independence of individual data points in our analysis. Because all data we included from a particular study were restricted to the same cortical region, differences between regions would be absorbed by the studies’ scaling factors  and would be indistinguishable from methodological systematics (further discussed below). We therefore decided to include frontal data when available, and other regions when not available, but given the potential regional differences, we also show our aging-related trajectory without the inclusion of Mikkelsen et al., 2017 and show that the model shows a similar trajectory for frontal-only data (Figure 4—figure supplement 1). The majority of published studies on MRS of GABA focus on cortical, rather than subcortical, regions of interest. The few publications that describe subcortical GABA often show trends that are difficult to associate with cortical GABA levels. For instance, GABA was negatively associated with age in subcortical voxels but positively associated with age in anterior and posterior cortical voxels (Ghisleni et al., 2015). Subcortical GABA undoubtedly plays an important role in brain function, as evidenced by increased GABA concentrations in subcortical basal ganglia that have been associated with schizophrenia and depression (Puts and Edden, 2012). However, the role of subcortical GABA—measured via MRS—in cognitive function is not thoroughly investigated in the literature, and thus there are too few studies to perform a suitable meta-analysis. In the current review, we limit our discussion to GABA concentrations in cortical voxels in order to provide a meaningful initial evaluation of GABA over the lifespan. Both the publication of future participant-level data and the release of participant-level data from past studies will increase the sample available to the field, which in turn will make possible the rigorous examination of the contributions of these covariates.

Methodological differences

Inconsistencies in aging-related differences in <span class="Chemical">GABA estimates may stem from differences between study methodologies or inherent structural differences a<span class="Chemical">cross the lifespan. We discuss each of these in turn below.

Quantification <span class="Gene">of MRS is relative and expresses the <span class="Species">ratio between the signal of interest and an internal reference signal. The most widely used references for GABA measures are the Cr+PCr signal in editing-off spectra and the unsuppressed water signal from the same volume (Alger, 2010; Mullins et al., 2014). Each quantification approach has its advantages and disadvantages. For instance, the Cr+PCr at 3.05 ppm has minimal chemical shift from the GABA signal at 3 ppm (Mullins et al., 2014) and is acquired during the MEGA-PRESS sequence. Therefore, its signal comes from the same location as the GABA signal, and it does not require a separate acquisition. In contrast, the water signal represents a more concentrated chemical yielding a higher SNR, but it may also introduce error in estimates of location due to chemical shift effects (Mullins et al., 2014) when not acquired from the same voxel as GABA and on some scanners requires a separate acquisition (see Choi et al. for a discussion). Furthermore, the Cr+PCr signal arises only from tissue, whereas water signal arises from tissue and cerebrospinal fluid (CSF) with substantially different relaxation behavior. A small number of studies have looked at the relationship between Cr+PCr and age (Ding et al., 2016; Lind et al., 2020) showing an increase or no change with age (for a review, see Cleeland et al., 2019). The impact of increasing Cr+PCr during aging would be to make the any aging-related decrease in GABA/Cr+PCr more pronounced, and the potential of this to contribution should not be ruled out. However, Gao et al. investigated both GABA/Cr+PCr and GABA/NAA and found similar aging-related reductions in GABA referenced to both molecules. Given that NAA has, in contrast to the increase of Cr+PCr, been shown to consistently decrease in aging (Cleeland et al., 2019), it makes it unlikely that Gao’s aging-related reduction was solely driven by a Cr+PCr increase during aging. Still the potential for a Cr-bias on the final model that needs to be considered. Our LOO approach to the ‘late-age’ data (Figure 4) found no substantial differences in the final models, providing some support to the notion that differences in Cr+PCr are not driving the GABA findings presented here, but the impact of reference compound needs to be considered.

One <span class="Gene">of the cornerstone assumptions <span class="Gene">of our analysis is that, although the use of different references results in different fundamental units, participants within any given study using a given method can be compared to one another in relative terms. If, for example, participants in the oldest quartile of a study show only 80% of the GABA signal on average of participants in the youngest quartile, then there is a basis for reporting a reduction in GABA regardless of the reference method used. In other words, our analysis presumes that the measures’ scores in any given study share a common scaling factor (which we denote as ) that can be estimated in order to factor out its influence, thus allowing studies to be compared in terms of their relative expressions of GABA. This approach thus relies on different reference methods displaying consistent sensitivity within each study; that is, participants with comparatively low scores in a study should be measured with a precision that is similar to that of participants with comparatively high scores. If this assumption is violated, a more complex analysis would be needed to address it, such that each participant’s contribution is given a weight as a function of their measurement error relative to other participants in the same study. Provided this assumption is met, however, simultaneous estimation of between-list scaling factors and within-list participant differences provides a robust means of integrating multiple studies into a single overall trajectory.

Another limitation is that voxel location can be inconsistent between studies (e.g. the medial prefrontal cortex in one study may not be localized the same way as in another study; also see a recent review: Peek et al., 2020). This becomes more problematic in younger cohorts where, due to smaller intra<span class="Chemical">cranial volumes, the methodologically limited size <span class="Gene">of the voxel (Mullins et al., 2014) necessarily incorporates a proportionally larger fraction of the brain.

It is well known that voxel tissue composition has a significant impact on the quantification <span class="Gene">of <span class="Chemical">GABA levels (Harris et al., 2015a). In many cases, tissue correction is appropriate due to existing partial volume effects (Barker et al., 1993; Christiansen et al., 1993; Danielsen and Henriksen, 1994; Ernst et al., 1993; Hennig et al., 1992; Kreis et al., 1993; Thulborn and Ackerman, 1983). Researchers frequently tissue-correct GABA values by segmenting the T1 weighted structural images (Ghisleni et al., 2015). Interestingly, none of the studies included in this review showed significant differences in segmented tissue content with age. This is surprising because other studies (Maes et al., 2018; Porges et al., 2017b) showed that aging-related differences in GABA depend on tissue correction due to atrophy being common in older participants. Unfortunately, there is no consensus on the appropriate tissue correction approach except that accounting for different compartments when using a water reference is strongly advised. The potential effect of tissue composition on GABA levels is limited if a study focused on a young cohort where aging-related atrophy would be negligible. All of the studies included in this review assume a linear model to represent change in GABA. Thus, while tissue correction could potentially contribute to the variance and differences between MRS-GABA studies, they are unlikely to explain the discrepancies between study populations of different age ranges.

In <span class="Chemical">water-referenced studies <span class="Gene">of aging-related differences in older cohorts (where bulk tissue changes are most likely to impact this relationship), accounting for CSF in the tissue correction approach does not remove a significant relationship between age and cortical GABA (Porges et al., 2017b). Given our concern was with relative differences in GABA, our analysis approach used a feature scaling approach to examine within-study variation as an estimation problem that was simultaneous with the estimation of the time course. Given this approach, the scaling factor  for each study s is conceptually very similar to a method of correcting ‘house effects’ in the analysis of political polling data (Jackman, 2005), allowing pollsters that show systematic bias (e.g. to a particular party) to be included in aggregate measures of public opinion. Furthermore, a limited effect of tissue is also suggested by consistency between findings of GABA+/Cr+PCr and estimated concentrations. There is no significant difference in the correlation for GABA × age in the overlapping age range of 43–76 years between Porges et al. (GABA+/H2O) and Gao et al. (GABA+/Cr+PCr) who looked at comparable samples (Fisher R-to-Z [z = 1.48, p = 0.19]) showing consistency between studies focusing on the same population (see also the discussion above).

As for other methodological differences (Table 2 and Table 3), studies used a variety <span class="Gene">of scanner vendors: a variation that is known to contribute to between-site variability (Mikkelsen et al., 2017) but is unlikely to lead to substantial difference in the age relationship. Furthermore, a variety <span class="Gene">of analysis techniques have been used, making direct comparison between studies problematic. A given analysis pipeline is unlikely to be biased to substantial differences in the age relationship. By rescaling the data in a way that emphasized scale ratios within each dataset, we minimized the impact of differences in site and vendor-driven GABA magnitude estimations. Furthermore, inconsistencies between reference compound (e.g. water or Cr+PCr) additionally complicate interpretation of the findings. Future studies should provide concentration values across the lifespan; we recommend reporting of values both as metabolite and water ratios to look at within-study consistency and robustness of the effect of GABA. Recent developments in MRS of GABA methodology will allow for this, even when data is collected on multiple scanner platforms (Saleh et al., 2019).

Reference method refers to either reference to <span class="Chemical">water (in estimated concent<span class="Species">ration/H2O) or as a ratio to creatine plus phosphocreatine (Cr+PCr) and describes whether data was acquired macromolecule-suppressed (γ-aminobutyric acid [GABA]) or as GABA+ macromolecules (GABA+). MRS averages refer to the number of ON + OFF transients. *The manuscript refers to 96 averages. It was clarified with the authors that this referred to 96 ON and 96 OFF averages.

This gives a basic des<span class="Chemical">cription <span class="Gene">of sample size and age range for the eight datasets. Additionally, Figure 6 depicts the distribution of ages using a raincloud plot (Allen et al., 2018).

Figure 6.

Raincloud plot depicting the age distribution in each of the eight included datasets.

Plotted densities are scaled within study. Boxes represent the first and third quartiles, while whiskers represent the range of the data. The central notch corresponds to the bootstrapped confidence interval for median. Kernel density estimates for each dataset were computed using a Gaussian kernel according to Silverman, 1986. The relevant analysis script for these densities is included as supplementary material.

It should be noted that all <span class="Chemical">datasets used for this meta-analysis were <span class="Chemical">cross-sectional. Although it is tempting to draw longitudinal inferences from our analysis, there are several limitations in doing so. For example, the present data provide no insight into long-term survival trends, so the population that is represented at age 20 likely differs in various ways from the population that is represented at age 70 and we cannot rule out a relationship between GABA and survival-related confounds. Additionally, because our method simultaneously estimated the scaling factor and the time course for each dataset, areas of minimal overlap between datasets (e.g. at around 13 years) are particularly uncertain.

While here we discuss potential methodological origins <span class="Gene">of bias, it should be noted that many <span class="Gene">of these cannot be extracted from the literature. We opted to use the MRS-Q as a guideline to assess quality, and recent consensus (Lin et al., 2021) also addresses issues in reporting important data acquisition, analysis, and quality information moving forward, including measures that may impact quantification, including proxy measures of motion, fit errors/CRLB, and shim.

The present analysis provides a model <span class="Gene">of the lifespan based on the <span class="Chemical">data that is currently available. It would be valuable to test these in a longitudinal study of MRS of GABA across the lifespan. Even if it is not feasible to measure GABA in the same individual over a 70-year span, obtaining multiple estimates per participant over a moderate length of time (e.g. 5 years) would greatly facilitate estimating rate of change over time.

Conclusion and future research

This review considered research <span class="Gene">of cortical <span class="Chemical">GABA concentrations over the lifespan. After combining datasets, we conclude that a linear model of GABA over the lifetime is not supported. Instead, consistent with other developmental aging studies of neurophysiology and cognitive function, we propose non-linear models to describe lifespan GABA measurements. A log-normal trajectory provides a satisfactory parametric description of the life course for the time being, but large and longitudinal datasets may necessitate the use of nonparametric regression strategies to best characterize the age-GABA relationship.

In the future, it will be important to investigate the neurophysiological and anatomical processes that drive apparent differences in bulk metabolite and neurotransmitter levels. While it is clear that GABA changes with age and seems to follow trends r<span class="Gene">eported in other lifespan datasets, non-linear relationships in GABA and other neurotransmitter and neurometabolite (e.g. NAA, myoinositol, and choline) concentrations merit further exploration. Glutamate in particular, given its close functional and metabolic relationship with GABA, would warrant a similar lifespan approach. In fact, Glx (combined glutamate and glutamine) can typically be quantified from the edited difference spectrum as a result of MEGA-PRESS but most studies do not report Glx. This may be due to the absence of clear hypotheses regarding Glx and a mechanistic focus on GABA, or simple omission of these data. Furthermore, it would be important to link these findings to anatomical measures underlying the different neurochemical measures, such as evidence of non-linear cortical thinning and histological studies of GABAergic neurons (Pandya et al., 2019; Vidal-Pineiro et al., 2020).

Future inquiries would benefit from re<span class="Chemical">cruiting cohorts that encompass the entire lifespan, and sufficient r<span class="Gene">epresentation and characterization of males and females, as sex differences have been reported in both GABA (O'Gorman et al., 2011) and GAD65 (Pandya et al., 2019). This would address that a limitation of the present study was the inability to distinguish males from females in the figures data was extracted from. Studies of development including those of infants and young children are virtually non-existent but are crucial given the importance of GABA in early development. Furthermore, alterations in GABA have been seen in neurodevelopmental and neurodegenerative disorders—a better understanding of abnormal GABAergic function early in development may elucidate this relationship, point to potential early-intervention targets, or explain variability in response to pharmacological treatments for these conditions.

Here, we r<span class="Gene">eport on the relationship between <span class="Chemical">GABA and age across the lifespan. It is well known that GABA contributes to cognition and perception, which in and of themselves change with age. It would be extremely interesting to apply an IPD-MA approach to the investigation of how aging-related differences in GABA might correlate with cognition. However, this is greatly complicated by the wide variety of cognitive measures used across studies, which would be much more difficult to standardize across studies than feature-scaled GABA as measured using an MRI machine. While pursuit of this work is notably challenging, conducting such research is undeniably crucial going forward.

While many papers exist that r<span class="Gene">eport on aging-related differences that are consistent with our asymmetrical lifespan trajectory, these are <span class="Gene">often reported as group or cohort differences without the presentation of individual data points necessary for such meta-analyses as applied here (Hermans et al., 2018; Port et al., 2017). Collaborative and group science approaches are becoming increasingly important in generating large datasets that allow for a broader and larger-scale application of this work and we hope that sharing data, or at least reporting individual data points, becomes more common in the future even in studies where age may not be the main focus.

Materials and methods

We conducted and r<span class="Gene">eported this systematic review in accor<span class="Chemical">dance with the PRISMA statement (Moher et al., 2009) and we used an IPD-MA approach (Debray et al., 2015).

A systematic lite<span class="Species">rature search was performed in two ite<span class="Species">rations by BF and EP to retrieve studies in which MRS of GABA using the MEGA-PRESS method was collected in the human brain from voxels that included the frontal lobe. In the first iteration, a search was performed using Google Scholar and Medline with the following combination of terms: (GABA OR GABA+ OR γ-aminobutyric acid OR gamma-Aminobutyric acid) AND (MRS OR Magnetic Resonance Spectroscopy) AND (MEGAPRESS OR MEGA-PRESS OR MEshcher-GArwood Point RESolved Spectroscopy OR edited). Both GABA+ and macromolecule-suppressed measures were deemed inclusive. We purposely did not include unedited measures of GABA as it is not clear whether these allow for specific and reliable estimation of GABA as per consensus (Choi et al., 2021; Mullins et al., 2014). The following constraints were applied to limit results: the result should be (1) a full-text article or a conference abstract, (2) peer-reviewed, (3) written in English, (4) included in the publication must be a scatter plot with GABA by age suitable format for extraction of individual human subject data via WebPlotDigitizer. The search was conducted on April 2, 2019, resulting in a total of 271 studies. Out of these, 55 were relevant.

Additional criteria

Study design

A second st<span class="Gene">ep was performed by BF, <span class="Gene">EP, and NP to exclude based on the following criteria: studies must report GABA, acquired using MEGA-PRESS, in at least one cortical voxel (subcortical voxels were excluded) in populations reported as being ‘healthy’, ‘normal’, or free of reported relevant disorders. To exclude a potential region effect, when a frontal voxel (any inclusion of frontal cortex) was available, we included that voxel. We chose to focus on frontal regions given this region is most-widely studied with respect to aging-related differences (see Introduction), making it suitable for an individual data meta-analysis; we are first and foremost interested in its relation to behavioral and cognitive associations with GABA. Furthermore, we can link neurochemical aging-related differences to other established brain indices in frontal regions during aging (e.g. white matter). Additionally, we were interested in the apparent contradictory reports in the change trajectory in frontal regions which we attempt to reconcile and finally, frontal regions allowed us to use data-informed predictions on aging-related differences in GABA to support the design and interpretation of future work.

However, to maximize the available <span class="Chemical">data, if no frontal regions were r<span class="Gene">eported, we used the other cortical voxel only as was the case for Mikkelsen et al. If a study sampled multiple voxels, we only extracted data from a frontal voxel to prevent multiple sampling of a single subject as this would include in inclusion of non-independent datasets. If multiple frontal voxels were available, we chose the dataset with the largest sample. Then, data were only included when they originated from published figures of sufficient quality for data extraction of individual data points, and psychiatric or neurologic populations, if presented in the same figure could be distinguished. Finally, the data extracted required contiguous age ranges of 5 or more years for inclusion in the analysis. Studies failing to satisfy all criteria were deemed incompatible with the IPD-MA approach (as described below). Duplicate datasets were excluded as well.

Data quality

A third st<span class="Gene">ep was performed to assess whether studies adhered to consensus quality assurance <span class="Chemical">criteria for data collection, analysis, and reporting. For this purpose, two coauthors (NP and EP) evaluated all remaining studies using the MRS-Q, which was specifically designed for MRS and based upon consensus documentation (Peek et al., 2020). MRS-Q identifies a limited number of edited MRS approaches to report and is specific for edited MRS and consistent with the recently published MRS standards in reporting (Lin et al., 2021). These criteria are necessary methodological minimum requirements (e.g. voxel size) as reported in existing consensus work for GABA quantification (Mullins et al., 2014; Wilson et al., 2019; Peek et al., 2020). Spectral quality was not a criterion for exclusion, as individual data spectra or quality metrics (e.g. linewidth, CRLB, etc.) are rarely if ever published in manuscripts used for extraction in our IPD-MA approach.

<span class="Gene">Of the 55 relevant studies retrieved in the systematic search, only seven included figures suitable for <span class="Chemical">data extraction or had data freely available in available online repositories. One dataset was only partly published (Puts et al., 2017) but was supplied in full by authors of this manuscript NP and RAEE, for a total of eight datasets. While full-text article and conference abstracts were included in the initial search, no conference abstracts meet our inclusion criteria, thus all included datasets were from peer-reviewed manuscript. At final review, these datasets were reviewed for consistency in research methods (see Figure 5 and Table 2), evaluated for age distributions (see Figure 6 and Table 3), and combined in aggregate (see Figure 2 and Table 1).

Figure 5.

PRISMA 2009 flow diagram of study identification and inclusion.

Plotted densities are scaled within study. Boxes r<span class="Gene">epresent the first and third quartiles, while whiskers r<span class="Gene">epresent the range of the data. The central notch corresponds to the bootstrapped confidence interval for median. Kernel density estimates for each dataset were computed using a Gaussian kernel according to Silverman, 1986. The relevant analysis script for these densities is included as supplementary material.

Risk of bias

All eight studies included in the systematic review included both male and female <span class="Species">participants, although the pediatric <span class="Chemical">data were somewhat skewed to include more males. No data were excluded based on acquisition parameters as assessed with the MRS-Q. Most studies were well matched for acquisition and followed published recommendations (Mullins et al., 2014). No exclusions were made based on the direction of the correlations. With the exception of two publications (Gao et al., 2013; Porges et al., 2017a), the age-by-GABA relationship was not a primary outcome and was therefore unlikely to have been a driver of publication bias in the majority of studies included. However, a risk of publication bias cannot be ruled out.

Individual <span class="Chemical">data points were extracted from figures using WebPlotDigitizer (Rohatgi, 2019). None <span class="Gene">of the data included were corrected for voxel tissue fractions (see Discussion). MEGA-PRESS sequences can vary between and within MRI vendors; these can impact editing efficiency and in turn absolute quantification (Harris et al., 2015b; Saleh et al., 2019). However, the consequence of this will not impact the within-site relationship to age (the metric used in this review) as the consequences of such variation are stable within-site and function as a scaling factor (Mikkelsen et al., 2018).

Statistical methodology

Our meta-analysis made use <span class="Gene">of an <span class="Chemical">IPD-MA approach (Debray et al., 2015). The chief advantage of this approach is that it allows the analyst to account for the evidence provided by the individual observations recorded in each study while also accounting for any systematic differences between studies using the framework of a multi-level model that evaluates both the data overall and within the context of each separate study as a simultaneous estimation problem. As a result, methodological steps like estimating an overall ‘weight’ associated with each study in order to determine how to combine reported statistics is rendered unnecessary, as the relative effect size and uncertainty of each study is communicated to the model by the data themselves (Riley et al., 2010). Such an approach is especially important in non-linear regression paradigms because the goodness of fit provided by each value of a parameter is interdependent with every other parameter. Put another way, optimal parameter values cannot be estimated for each parameter in isolation of the others, as the posterior distributions for the parameters covary with one another. As such, the meta-analytic models’ uncertainty for any value of a continuous predictor depends on the complex covariance of parameters. In the present analysis, we are concerned with scaling every study relative to every other study, so it would be inappropriate to convert the data to some kind of standard score prior to combining them; instead, this standardization should happen simultaneously with the estimation of other model parameters.

Our general statistical framework in this analysis was to assume that some unknown ‘canonical function’ des<span class="Chemical">cribes the average change in the feature-scaled <span class="Chemical">GABA signal over the lifespan as a function of age. In other words, given participant i in study s, their age is denoted by and the relative change in GABA over the lifespan is given by the function  This change cannot be directly observed though imaging, but is instead inferred indirectly from a measurable reference (in our case, either water or Cr+PCr). As such, the mean observed effect in each study has an unknown feature scaling factor , which is expected to vary by orders of magnitude as a function of reference method, variations in equipment, and other systematics that may not be documented in all (or even any) of the included studies. By estimating both  and , simultaneously, the model was able to find the optimal balance between the shape of the function and the alignment of observations along zones of overlapping age. Finally, individual observations are assumed to vary with respect to the lifespan function given normally distributed noise  with an unknown error term , chosen as a minimally informative maximum entropy distribution for the error (Jaynes, 2003). In total, this gives the following general form for each observation in our data :

The scaling factor is intended to act as a feature-scaled stan<span class="Chemical">dardization <span class="Gene">of the unitless function . Because of this, the scaling factors were constrained in two ways. The first was a hard constraint that truncated their distributions to exclude negative values. The second was a soft constraint imposed by using a weakly regularizing prior to favor values close to unity. Since each of the datasets had at least a few dozen subjects, the lion’s share of the scaling factor’s effect size was determined by the data themselves, with regularization serving mainly to keep the tails of the posterior distribution from considering unreasonably large values, effectively penalizing the likelihood of such extreme values. The net effect of the scaling factor was that the geometric mean of all values  was equal to 1.0, despite the unscaled values from different studies varying dramatically in their original units. In addition to allowing different reference methods to be combined, the relative values of act as a study-by-study correction for any systematics that might otherwise shift one study’s observations relative to any other study, including both known and unknown factors that might introduce a bias. As with all other parameters in our models, the scaling factor estimates are not point values, but rather are posterior probability distributions with a lower bound of zero but no proscribed upper bound.

This model is highly general, accommo<span class="Chemical">dating any function  the analyst deems appropriate. The parameters that must be estimated are one scaling factor for each study included in the meta-analysis, a global error term , and whichever parameters the function  requires to specify its shape. Because parameter estimates in any non-linear regression model necessarily covary, it is essential that all parameters be estimated simultaneously (McElreath, 2020). If, for example, each <span class="Chemical">dataset was ‘feature scaled’ prior to the meta-analysis and then subsequently stitched together, any vertical shift needed to maximize the overlap of outcomes recorded at overlapping age ranges would necessarily be ad hoc and would not be able to balance the relative weight of the evidence from each study in that area of overlapping age. By this same token, it was important that all studies included in our analysis included a range of ages that overlapped with at least one other study.

We chose two models as candi<span class="Chemical">dates for . Our first function  was a penalized cubic basis spline model, a<span class="Chemical">dapting the procedure described by Kharratzadeh, 2017. This provided a non-linear and semiparametric estimate of GABA differences over the lifespan as described by these eight datasets. Since it does not make strong assumptions of its own, such a model requires considerable data to make good inferences. Our second candidate was a parametric function that followed the shape of a log-normal distribution, which has been chosen in the past as a way to characterize the pattern of increases in GABA in childhood, followed by a gradual decline that describes other neurophysiological changes across the lifespan (Lebel et al., 2012). We include this second function primarily to contrast how powerfully the choice of functions, and of their underlying assumptions, can influence the estimate.

In order to ensure that all estimates were permitted to covary appropriately, we obtained posterior probability distributions for each parameter numerically using a Bayesian paradigm (Gelman et al., 2014) and implemented with the Stan programming language (Carpenter et al., 2017). Note that a ‘posterior probability distribution’ is a statistical term in this context and is not intended for use in anatomical orientation. Details <span class="Gene">of these analyses can be found in the supplementary information.

In the interests <span class="Gene">of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acc<span class="Gene">eptance summary:

This meta-analysis combines <span class="Chemical">GABA magnetic resonance spectroscopy (MRS) <span class="Chemical">data from multiple studies to estimate the lifespan trajectory of GABA levels. The manuscript is well-motivated and clear in its aim to untangle apparent contradictions which appear in the literature of age-dependent GABA concentration trajectories. The process of combining MRS datasets from different studies acquired using different MRS protocols is extremely challenging and difficult to achieve, but the approach presented and the GABA trajectories revealed should be of broad interest.

Decision letter after peer review:

Thank you for submitting your article "The trajectory <span class="Gene">of cortical <span class="Chemical">GABA across the lifespan: An individual participant data meta-analysis of edited MRS studies" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by Chris Baker as the Reviewing and Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: William Clarke (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing/Senior Editor has drafted this decision to help you pr<span class="Gene">epare a revised submission. There was vigorous discussion and while all reviewers appreciated the attempt to estimate the trajectory <span class="Gene">of GABA across the lifespan, there were many concerns raised about the difficulties of combining MRS studies.

As the editors have judged that your manuscript is <span class="Gene">of interest, but as described below that additional data and analyses are required before it is published, we would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). First, because many researchers have temporarily lost access to the labs, we will give authors as much time as they need to submit revised manuscripts. We are also offering, if you choose, to post the manuscript to bioRxiv (if it is not already there) along with this decision letter and a formal designation that the manuscript is "in revision at eLife". Please let us know if you would like to pursue this option. (If your work is more suitable for medRxiv, you will need to post the preprint yourself, as the mechanisms for us to do so are still in development.)

Summary:

The aim <span class="Gene">of this meta-analysis is to combine <span class="Chemical">GABA MRS data from multiple studies to estimate the lifespan trajectory of GABA levels. The manuscript is well-motivated and clear in its aim to untangle apparent contradictions which appear in the literature of age-dependent GABA concentration trajectories. The methods are well written and do a superb job to explain the potentially tricky to understand IPD-MA approach. However, the process of combining MRS datasets from different studies acquired using different MRS protocols is extremely challenging and difficult to achieve.

Essential revisions:

There are three major issues the reviewers would like to see the authors address. The comments below reflect the discussion between the reviewers. In general, the reviewers would like to see much more detail and justification <span class="Gene">of the approach throughout the manus<span class="Chemical">cript.

1. Exclusion <span class="Gene">of a large number <span class="Gene">of publications

The reviewers appreciate that the method used requires that for a study to be included it contains a range <span class="Gene">of ages, and that this individual subject <span class="Chemical">data is available for the meta-analysis. Thus the exclusion of a large number of studies is not unexpected. However, out of 273 MRS GABA publications the authors only identify 57 as being relevant and ultimately only use data from 8.

The meta-analysis would be much stronger if more <span class="Chemical">data were included. In the text it is stated that, "<span class="Gene">Of the 55 relevant studies retrieved in the systematic search, only 7 included figures suitable for data extraction or had data freely available in available online repositories." Did the authors make an attempt to contact the corresponding authors of the identified studies to try and obtain the data that was not readily available from the publications? This would also potentially allow the calculation of comparable GABA values across studies.

We would also like to see more detail regarding the reasons for exclusion <span class="Gene">of so many studies including the precise spectral quality exclusion metrics.

2. Comparability between studies

There are many differences between the included studies: voxel location (many different areas across the PFC, parietal and occipital areas), voxel size (impacting the varying portion <span class="Gene">of GM), TR (implying that the degree of T1-weighted is different across studies) and quantification including wether the water signal was fully relaxed.

The logic <span class="Gene">of the <span class="Chemical">IPD-MA approach is that data collection methods only need to be consistent across all subjects in a study (i.e. internally consistent). All other differences are lumped into the single scaling factor. Even those factors that have non-linear functions (e.g. partial saturation) should only result in independent fixed scalings between datasets.

But how can we be confident that the approach is able to discount all those difference between studies? Is the method sufficient for the authors to be able to draw the conclusions they do? The authors don't give bounds on <span class="Chemical">dataset scaling factors. While the method is in principal sound it may be that the single point-estimate currently shown disguises a lack <span class="Gene">of sensitivity in determining correct scaling factors.

3. Clarification on the validity <span class="Gene">of Figures 4/5

The authors should justify the suitability <span class="Gene">of the log-normal, which is not clearly explained.

How do the authors account for the clear discontinuity in Figure 4 around 20-30 years <span class="Gene">of age? It suggests that the individual scaling factors applied were ineffective and may reflect the difficulty <span class="Gene">of accounting for all the differences between studies as noted above and hence, questions the legitimacy of combining different MRS studies in this way. Is it because datasets with narrower age ranges have many reasonable/possible scaling values?

[Editors' note: further revisions were suggested prior to acc<span class="Gene">eptance, as des<span class="Chemical">cribed below.]

Thank you for submitting your article "The trajectory <span class="Gene">of cortical <span class="Chemical">GABA across the lifespan, an individual participant data meta-analysis of edited MRS studies" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by Chris Baker as Reviewing and Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Jeffrey Stanley (Reviewer #1); William Clarke (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you pr<span class="Gene">epare a revised submission.

Essential revisions:

While the reviewers think the manus<span class="Chemical">cript has improved, there are still some remaining concerns that need to be addressed as outlined in the Reviewer comments below. Please carefully consider and respond to these concerns.

Reviewer #1:

Overall, the revised manus<span class="Chemical">cript is an improvement and more focused but still has many issues to addressed, which includes the following.

The trajectory <span class="Gene">of cortical <span class="Chemical">GABA across the lifespan, an individual participant data meta-analysis of edited MRS studies

In the abstract, it is important to state how <span class="Chemical">GABA is being expressed and from where in the brain with respect to the main outcome <span class="Gene">of the modelling.

In the 2nd paragraph <span class="Gene">of the introduction, it would be helpful to state whether cited studies are based on <span class="Chemical">cross-sectional data or longitudinal data.

In the 3rd paragraph <span class="Gene">of the introduction, the text is confusing, verbose with r<span class="Gene">epeating statements, and hence, difficult to follow.

The statement des<span class="Chemical">cribing MEGAPRESS, "The majority <span class="Gene">of MRS studies of GABA at 3 Tesla have utilized the unique structure of the GABA molecule to selectively edit the GABA signal using a typical MRS acquisition with frequency-selective pulses", is confusing and inaccurate – the isolation of GABA is derived by taking the difference between two spectra with and without the selective saturation – technically, MEGAPRESS is not spectral editing approach but a "difference editing" approach by proper definition.

The statement, "studies have shown that the majority <span class="Gene">of the <span class="Chemical">GABA measured using MRS reflects intracellular levels rather than synaptic levels", is confusing – GABA is initially defined as being localized either in the intracellular or extracellular space – where does "synaptic level" fit in? It should be viewed as part of the intracellular space.

The section, "The importance <span class="Gene">of <span class="Chemical">GABA in cognition motivates an understanding of the age-relationship", in the introduction does not help to focus the narrative or help to motivate the aims of the manuscript – it should be removed.

The statement, "the <span class="Chemical">GABA signal is considered to be reflective <span class="Gene">of inhibitory tone rather than dynamic synaptic inhibition", is confusing and needs clarification – "dynamic synaptic inhibition" is too convoluted.

The information in Table 1 can be extended to also include <span class="Species">ratio or absolute, the voxel location(s) and sample size for each study.

In Figure 1, the 95% CI bands are misleading and should only cover the range <span class="Gene">of <span class="Chemical">data points in each graph and not the full range of the age scale – it's inappropriate to imply that the modelling in each is generalizable across the full age range – the text also conflicts this as noted below.

In the result section, first paragraph, the conclusion from the "cursory examination" is confusing with conflicting messages. The authors are arguing that studies r<span class="Gene">eporting linear differences with age are inaccu<span class="Species">rate and non-linear is better. Then note that linear trends over short intervals are OK. Also, the fact that the "linear trends explained less than 20% of the variance" does not help the author's argument whether linear or no-linear is better – the assumption that GABA must show age effects across all ages may not be correct. This needs to b better articulated.

In Figure 3, the <span class="Species">rationale to only focus on <span class="Chemical">data from the posterior is unclear. What is the significance of the posterior vs frontal data or using both? This needs to be explained.

The results to support similar age effects when expressed as <span class="Species">ratio vs absolute is weak and not compelling. The bulk <span class="Gene">of the data is dominated by studies reporting ratios, which does not help. If the GABA age effects are not driven by age effects with PCr+Cr, then it would imply that PCr+Cr levels are not different across age, which is not the case as supported

The Figure 4 caption is confusing – it refers to models shown in Figure 4.

Regarding the main finding <span class="Gene">of a "sharp" or "rapid" early in<span class="Chemical">crease in GABA is driven by data from one study (weak) and results are expressed as a ratio. For example, a miss adjustment in the scaling factor on that one study could artificially enhance this effect. Also, the bulk of the data from those 8 studies expressed GABA as a ratio and GABA/PCr+Cr ratio values do not equate to GABA levels especially when PCr+Cr levels differ across age as noted in the literature (far deeper than what is noted in the text and including ratios relative to NAA does not help). Therefore, greater caution is needed in reporting this observation – a softer tone is needed including in the abstract.

In the discussion, it must be emphasized at the beginning that <span class="Chemical">GABA/P<span class="Chemical">Cr+Cr ratio values do not equate to GABA levels or concentration – this needs to be acknowledged and corrected in the discussion – one must be careful not to over interpret the data.

In the discussion, the statement, "Cr+P<span class="Chemical">Cr is acquired during the same MEGAPRESS scan as GABA, limiting effects of chemical shift: Cr+PCr has a minimal shift from the GABA signal", is confusing and needs to be clarified.

The statement, "the <span class="Chemical">water signal r<span class="Gene">epresents a more concentrated chemical yielding a higher SNR, but it may also introduce error in estimates of location due to chemical shift effects", is also confusing – unclear how chemical shift is a factor.

The discussion on the CSF correction applied in the Porges et al. 2017b completely ignores the fact that the correct was not applied correctly – it should not be treated as a co-variate in the regression but as a volume correction factor (as noted earlier in the text). It would help to address this.

Reviewer #2:

The authors have thoroughly addressed my one major concern, regarding the uncertainty in the per-study scaling factors at the heart <span class="Gene">of their method. The authors have introduced figure 3 to present clearly the range <span class="Gene">of potential values that each scaling factor could take. It is apparent that the overall non-linear trend identified by this work is consistent within the bounds presented in this new figure.

Furthermore, it appears the authors have engaged constructively with all reviewer comments (though I defer to my co-reviewers to determine if their comments have been addressed sufficiently). Therefore, I have no further concerns with this manus<span class="Chemical">cript. I believe that it makes a substantial contribution to the field. And with the revisions, the authors are forthright in addressing the potential limitations <span class="Gene">of their work.

Essential revisions:

There are three major issues the reviewers would like to see the authors address. The comments below reflect the discussion between the reviewers. In general, the reviewers would like to see much more detail and justification <span class="Gene">of the approach throughout the manus<span class="Chemical">cript.

1. Exclusion <span class="Gene">of a large number <span class="Gene">of publications

The reviewers appreciate that the method used requires that for a study to be included it contains a range <span class="Gene">of ages, and that this individual subject <span class="Chemical">data is available for the meta-analysis. Thus the exclusion of a large number of studies is not unexpected. However, out of 273 MRS GABA publications the authors only identify 57 as being relevant and ultimately only use data from 8.

The meta-analysis would be much stronger if more <span class="Chemical">data were included. In the text it is stated that, "<span class="Gene">Of the 55 relevant studies retrieved in the systematic search, only 7 included figures suitable for data extraction or had data freely available in available online repositories." Did the authors make an attempt to contact the corresponding authors of the identified studies to try and obtain the data that was not readily available from the publications? This would also potentially allow the calculation of comparable GABA values across studies.

We would like to thank the reviewer(s) for drawing our attention to this. We were not sufficiently clear in the submitted version. In pr<span class="Gene">eparing this manus<span class="Chemical">cript we used available data (either via extraction from figures where individual subject data points were presented on figures or from data repositories), and we are sharing all code used in the analysis. This approach allows for complete transparency of our analysis and potential to replicate and extend our work. If other manuscripts had otherwise met our inclusion criteria but did not provide extractable or shared data, 1) these would be incompatible with our proposed approach, and 2) the transparency and replication objectives of our previously set PRISMA framework would be violated.

We consider the more involved process <span class="Gene">of managing raw <span class="Chemical">data requests from other authors to be an exciting future direction for this method and are considering pursuing it as part of a larger consortium-approach, such as http://enigma.ini.usc.edu which Dr. Porges is affiliated with. However, this is beyond the scope of the current manuscript and not in line with the PRISMA framework established at the outset.

We would also like to see more detail regarding the reasons for exclusion <span class="Gene">of so many studies including the precise spectral quality exclusion metrics.

<span class="Chemical">Data were not excluded based on spectral quality, as this would not be possible to apply to extracted <span class="Chemical">data given it is rarely included for all subjects. However, datasets were excluded based on acquisition methods if they did not meet field consensus for GABA quantification using MRS.

To address the above concerns <span class="Gene">of the reviewers, we have added the following text to the <span class="Chemical">Data Quality Section:

“A third st<span class="Gene">ep was performed to assess whether studies adhered to consensus quality assurance <span class="Chemical">criteria for data collection, analysis, and reporting. […] Spectral quality was not a criterion for exclusion, as individual data spectra or quality metrics (e.g., linewidth, CRLB, etc.) are rarely if ever published in manuscripts used for extraction in our IPD-MA.”

2. Comparability between studies

There are many differences between the included studies: voxel location (many different areas across the PFC, parietal and occipital areas), voxel size (impacting the varying portion <span class="Gene">of GM), TR (implying that the degree of T1-weighted is different across studies) and quantification including wether the water signal was fully relaxed.

The logic <span class="Gene">of the <span class="Chemical">IPD-MA approach is that data collection methods only need to be consistent across all subjects in a study (i.e. internally consistent). All other differences are lumped into the single scaling factor. Even those factors that have non-linear functions (e.g. partial saturation) should only result in independent fixed scalings between datasets.

But how can we be confident that the approach is able to discount all those difference between studies? Is the method sufficient for the authors to be able to draw the conclusions they do? The authors don't give bounds on <span class="Chemical">dataset scaling factors. While the method is in principal sound it may be that the single point-estimate currently shown disguises a lack <span class="Gene">of sensitivity in determining correct scaling factors.

This is a valid concern, and we appreciate the reviewers(s) for expressing it and giving us the chance to elabo<span class="Species">rate. Using a per-study scaling factor is not a way <span class="Gene">of uncovering, or even identifying the factors that result in systematic bias in a dataset. Rather, such a factor seeks to estimate, relative to the other studies that are included, the multipliers needed to bring the included studies into alignment with one another. This has been clarified in the revision. Additionally, we now articulate explicitly that, although these scaling factors have a lower bound of zero (since negative MRS signals are a priori impossible), they have no strict upper bound. Further, we express explicitly that our scaling factor estimates, like all parameters in the model, are posterior distributions and not point-estimates. If there is a lot of uncertainty in which scaling factor is ideal, this results in a more uncertain posterior estimate of the scaling factor, as an automatic consequence of the principles of Bayesian inference.

3. Clarification on the validity <span class="Gene">of Figures 4/5

The authors should justify the suitability <span class="Gene">of the log-normal, which is not clearly explained.

How do the authors account for the clear discontinuity in Figure 4 around 20-30 years <span class="Gene">of age? It suggests that the individual scaling factors applied were ineffective and may reflect the difficulty <span class="Gene">of accounting for all the differences between studies as noted above and hence, questions the legitimacy of combining different MRS studies in this way. Is it because datasets with narrower age ranges have many reasonable/possible scaling values?

As we would like to incorpo<span class="Species">rate this useful feedback, we now note in the manus<span class="Chemical">cript that, although the mean of the splines shows a slight upward trend in this age range, this does not mean that every spline that is nominated by this model must include this feature. Any doubly-differentiable curve that passes through the shaded region in Figure 4 (Left) is in principle nominated by the spline model, and it would not be at all surprising if certain kinks and wrinkles in the present estimate smoothed out as additional datasets were added. What is expected given the model’s assumptions is that the true curve, whatever its shape, probably mostly resides within the shaded region.

[Editors' note: further revisions were suggested prior to acc<span class="Gene">eptance, as des<span class="Chemical">cribed below.]

Reviewer #1:

[…] In the abstract, it is important to state how <span class="Chemical">GABA is being expressed and from where in the brain with respect to the main outcome <span class="Gene">of the modelling.

We now explicitly state we studied <span class="Chemical">GABA from frontal regions for the meta-analysis and explicitly state that any measure <span class="Gene">of GABA (but only from editing) was accepted in the search criteria within the confines of the limited abstract length.

Explicitly, we have amended the following:

“No single study to <span class="Chemical">date has included the entire lifespan. In this study, 8 suitable <span class="Chemical">datasets were integrated to generate a model of the trajectory of frontal GABA levels (as reported through edited MRS; both expressed as ratios and in institutional units).”

“Integ<span class="Species">rated <span class="Chemical">data show the lifespan trajectory of frontal GABA measures involves an early period of increase,[…]”

In the 2nd paragraph <span class="Gene">of the introduction, it would be helpful to state whether cited studies are based on <span class="Chemical">cross-sectional data or longitudinal data.

None <span class="Gene">of the studies were longitudinal. We now explicitly state all studie were <span class="Chemical">cross-sectional in the last sentence of paragraph 2:

“<span class="Gene">Of interest, all studies were <span class="Chemical">cross-sectional rather than looking at within-subject change in GABA with age.”

In the 3rd paragraph <span class="Gene">of the introduction, the text is confusing, verbose with r<span class="Gene">epeating statements, and hence, difficult to follow.

We thank the reviewer for this comment and have tidied this section up as below:

“However, conflicting results only exist when looking at different age-ranges and consistent findings have been shown in studies <span class="Gene">of similar age-range. […] In the absence <span class="Gene">of a lifespan study, we implemented an individual participant data meta-analytic (IPD-MA) approach following PRISMA guidelines (Moher, Liberati, Tetzlaff, and Altman, 2009; see Figure 5) supplemented with data collected by the authors and previously published in summary form (Puts et al., 2017).”

The statement des<span class="Chemical">cribing MEGAPRESS, "The majority <span class="Gene">of MRS studies of GABA at 3 Tesla have utilized the unique structure of the GABA molecule to selectively edit the GABA signal using a typical MRS acquisition with frequency-selective pulses", is confusing and inaccurate – the isolation of GABA is derived by taking the difference between two spectra with and without the selective saturation – technically, MEGAPRESS is not spectral editing approach but a "difference editing" approach by proper definition.

We have r<span class="Gene">ephrased this section and have expressed a clearer introduction to MEGA-PRESS as a J-difference editing technique. We have also re-ordered the paragraph to enhance flow. However, we based our des<span class="Chemical">cription of that of the MRS of GABA consensus (Choi et al. NMR Biomed. 2020) which states the definition as follows:

“<span class="Chemical">Spectral editing in in vivo 1H‐MRS provides an effective means to measure low‐concent<span class="Species">ration metabolite signals that cannot be reliably measured by conventional MRS techniques due to signal overlap, for example, γ‐aminobutyric acid, glutathione and D‐2‐hydroxyglutarate. Spectral editing strategies utilize known J‐coupling relationships within the metabolite of interest to discriminate their resonances from overlying signals." under which definition this would be considered spectral editing.

Our revised introduction now states the following:

“The majority <span class="Gene">of MRS studies <span class="Gene">of GABA at 3 Tesla have utilized J-difference editing to selectively ‘edit’ the GABA signal (e.g. Mullins et al., 2014). […] This limitation constrains the spatial specificity of the measurement to coarse regions that often lack discrete functional specificity.”

The statement, "studies have shown that the majority <span class="Gene">of the <span class="Chemical">GABA measured using MRS reflects intracellular levels rather than synaptic levels", is confusing – GABA is initially defined as being localized either in the intracellular or extracellular space – where does "synaptic level" fit in? It should be viewed as part of the intracellular space.

We thank the reviewer for this comment. We have rewritten this section for clarity which now reads:

“The functional relevance <span class="Gene">of MRS measures <span class="Gene">of GABA is important to note. Measured GABA levels include both intracellular (both somatic and synaptic) and extracellular contributions to the overall GABA concentration. […] GABA levels measured via MRS in young adults at rest have been reported to be stable for up to 7 months (Near et al., 2014) and do not exhibit a diurnal rhythm (Evans et al., 2010).”

The section, "The importance <span class="Gene">of <span class="Chemical">GABA in cognition motivates an understanding of the age-relationship", in the introduction does not help to focus the narrative or help to motivate the aims of the manuscript – it should be removed.

In the nature <span class="Gene">of transparency, this in fact did motivate our study initially, given we have a strong interest in lifespan-related and <span class="Chemical">GABA-dependent changes in cognition and sensory processing. We are therefore reluctant to remove this statement (since it does in fact motivate this study) but we agree it was overly lengthy. We have substantially reduced this paragraph and the next, to emphasize that age-related changes are important to account for in studies of the functional role of GABA. It now reads:

“The importance <span class="Gene">of <span class="Chemical">GABA in the context of ageing in health and disease

<span class="Chemical">GABA levels measured with MRS have been linked to clinical and cognitive outcomes. […] Our approach allows for a systematic review <span class="Gene">of existing work studying GABA across development and aging with immediate impact on future studies in the context of development and aging.”

The statement, "the <span class="Chemical">GABA signal is considered to be reflective <span class="Gene">of inhibitory tone rather than dynamic synaptic inhibition", is confusing and needs clarification – "dynamic synaptic inhibition" is too convoluted.

As per our fifth response, this section was revised including this statement. We now refer to the interpretation <span class="Gene">of Rae (2014).

The information in Table 1 can be extended to also include <span class="Species">ratio or absolute, the voxel location(s) and sample size for each study.

Table 1 refers to the regression coefficients (in the results first, as per eLife gui<span class="Chemical">dance). However, we have added the text:

“Further study details including sample sizes can be found in Tables 2 and 3.”

In Table 2, which references acquisition and analysis details, <span class="Species">ratio was previously r<span class="Gene">eported and is now detailed further for clarity. Voxel location has been added to Table 2.

Sample sizes were already included in the Table 3 and we believe they fit better there. The following text has been added to the Table 2 captions:

“Reference method refers to either reference to <span class="Chemical">water (in estimated concent<span class="Species">ration / H2O) or as a ratio to PCr + Cr (Cr+PCr) and describes whether data was acquired macromolecule-suppressed (GABA) or as GABA GABA+ macromolecules (GABA+).”

In Figure 1, the 95% CI bands are misleading and should only cover the range <span class="Gene">of <span class="Chemical">data points in each graph and not the full range of the age scale – it's inappropriate to imply that the modelling in each is generalizable across the full age range – the text also conflicts this as noted below.

We apologize for the confusion and fully agree with the Reviewer. Figure 1 explicitly shows what not to do, which is extrapolate beyond the <span class="Chemical">data, as a visualization <span class="Gene">of what many studies do end up doing. Here we show that these data are not generalizable. In response, we 1) clarify the goal of this figure and address that generalizability is inappropriate. 2) We also rephrase the last sub figure as “extrapolated aggregate” and 3) We have the figure to reflect the correlation that covers the sample and which that does not by making the 95%CI transparent beyond the sample of the data.

Figure caption 1 now reads:

“Figure 1: Linear relationships between age and <span class="Chemical">GABA signal, showing that linear extrapolation over the lifespan is not appropriate. In each <span class="Chemical">dataset, GABA was scaled relative to the geometric mean. Linear models were fit for each dataset separately. Dark shaded regions represent the 95% credible interval for the interpolated regression line, given the data from each study and the assumption of a linear effect, whereas the light shaded regions represent the 95% credible interval for the extrapolated regression line.”

The text in the first paragraph <span class="Gene">of the Results section now reads:

“Furthermore, assuming a linear trend is unsuccessful even at des<span class="Chemical">cribing the <span class="Chemical">data within each study. […] [PN1] Although a linear trend may provide an an approximate summary over short periods of time, a linear trend over the entire lifespan is not appropriate.”

In the result section, first paragraph, the conclusion from the "cursory examination" is confusing with conflicting messages. The authors are arguing that studies r<span class="Gene">eporting linear differences with age are inaccu<span class="Species">rate and non-linear is better. Then note that linear trends over short intervals are OK. Also, the fact that the "linear trends explained less than 20% of the variance" does not help the author's argument whether linear or no-linear is better – the assumption that GABA must show age effects across all ages may not be correct. This needs to be better articulated.

We thank the reviewer for this comment and have clarified this section, also with respect to the figure discussed in our ninth response. It is now phrased as:

“Furthermore, assuming a linear trend is unsuccessful even at des<span class="Chemical">cribing the <span class="Chemical">data within each study. […] [PN1] Although a linear trend may provide an approximate summary over short periods of time, a linear trend over the entire lifespan is not appropriate.”

In Figure 3, the <span class="Species">rationale to only focus on <span class="Chemical">data from the posterior is unclear. What is the significance of the posterior vs frontal data or using both? This needs to be explained.

Figure 3 refers to the Bayesian posterior probability distributions for parameter estimates; that is, parameter estimates that have taken both the prior and the likelihood into account (since posterior = (likelihood * prior)/constant). We have clarified our use <span class="Gene">of this statistical term, which coincidentally shares a name with the anatomical region, in both the figure legend and in the body <span class="Gene">of the manuscript.

With respect to Figure 4 and Supplementary Figure 1 (which does refer to Frontal and Posterior cortex measures); this was in response to a prior reviewer comment who wanted us to also show <span class="Chemical">data for frontal-only <span class="Chemical">data only given that we focus on frontal regions but also include posterior regions where possible. This reviewer was concerned about the potential impact of posterior regions due to regional differences. We now justify this figure more clearly in the text.

“We additionally also explored the potential effect <span class="Gene">of region by performing the analysis on only frontal <span class="Chemical">data (removing Mikkelsen et al. and Simmonite et al. respectively) with no substantial impact on the direction of the slope over time. This information can be found in Supplemental figure 1.”

The results to support similar age effects when expressed as <span class="Species">ratio vs absolute is weak and not compelling. The bulk <span class="Gene">of the data is dominated by studies reporting ratios, which does not help. If the GABA age effects are not driven by age effects with PCr+Cr, then it would imply that PCr+Cr levels are not different across age, which is not the case as supported

We agree that variability in the methods aggregated in this <span class="Chemical">IPD-MA are a limitation but we can only rely on the available <span class="Chemical">data with respect to comparison between ratio versus “absolute” measures of GABA. Further; we are aware of, and have enriched the discussion of, the work attempting to characterize age related changes in PCr+Cr during aging. The most common, but certainly not universal results, have shown an age related increase in PCr+Cr with a great deal of variability.

While we can’t actually comment on whether /<span class="Chemical">Cr+P<span class="Chemical">Cr or water referenced data are consistent with each other, we do have overlapping data available for Gao et al. and Porges et al. and these do support similar age effects across methods. Further we report that when looking at overlapping age ranges from the two studies, there is no significant difference in an Age x Gaba correlation between the two. Based on the reviewer comment we have limited and softened our interpretation of a non-significant result (P17), but it does not support the notion that age-related increases in PCr+Cr are driving the age-related decrease in GABA which is discussed in the following text (P19).

“Furthermore, a limited effect <span class="Gene">of tissue is also suggested by consistency between findings <span class="Gene">of GABA+/Cr+PCr and estimated concentrations. There is no significant difference in the correlation for GABA x age in the overlapping age range of 43 to 76 years between Porges et al. (GABA+/H2O) and Gao et al. (GABA+/Cr+PCr) who looked at comparable samples (Fisher R-to-Z; (z = 1.48, p = 0.19)) showing consistency between studies focusing on the same population (see also the discussion above).”

A 2019 systematic review (Cleeland et al. 2019) provides mixed evidence that generally supports an age related increase in /Cr+PCr. Since the publication of that review, a 2020 study at 7T has produced further data suggesting a modest increase of /Cr+PCr in some but not all brain regions during aging (Lind et al. J. Neurosci 2020).

Importantly, an age related in<span class="Chemical">crease in /<span class="Chemical">Cr+PCr would only enhance the GABA-effect seen in aging as we report in the discussion. No comparable data are available in early development.

Thus there is indeed some, but limited, support that this is driven by <span class="Chemical">GABA and not the reference compound and we are now more careful in interpreting this as strong evidence and discuss this further in the discussion (see also our fourteenth response). We have now modified the section on page 18 to read:

“The impact <span class="Gene">of in<span class="Chemical">creasing Cr+PCr during aging would be to make the any age-related decrease in GABA/Cr+PCr more pronounced, and the potential of this to contribution should not be ruled out. […] Our leave-one-out approach to the ‘late-age’ data (Figure 4) found no substantial differences in the final models, providing some support to the notion that differences in Cr+PCr are not driving the GABA findings presented here but the impact of reference compound needs to be considered.”

As is noted in our fourteenth response, we also recommend r<span class="Gene">eporting <span class="Gene">of both ratios and estimated concentrations when possible.

The Figure 4 caption is confusing – it refers to models shown in Figure 4.

We now ensure figure captions match the right figures. We have renamed the Supplemental Figure to Supplemental Figure 1 and explain this accompanies Figure 4 in the main manus<span class="Chemical">cript.

Regarding the main finding <span class="Gene">of a "sharp" or "rapid" early in<span class="Chemical">crease in GABA is driven by data from one study (weak) and results are expressed as a ratio. For example, a miss adjustment in the scaling factor on that one study could artificially enhance this effect. Also, the bulk of the data from those 8 studies expressed GABA as a ratio and GABA/PCr+Cr ratio values do not equate to GABA levels especially when PCr+Cr levels differ across age as noted in the literature (far deeper than what is noted in the text and including ratios relative to NAA does not help). Therefore, greater caution is needed in reporting this observation – a softer tone is needed including in the abstract.

We have s<span class="Gene">oftened the language on a sharp in<span class="Chemical">crease in development in both the main body of the manuscript and in the text and removed any reference to ‘rapid’. With respect to the ratios, our approach is designed to handle differences in ratio and we have added a substantial paragraph (see below). It should be noted that at least in one of our recent papers, correlations between GABA/Glx and correlations hold between GABA IU and GABA/Cr+PCr. We now are more explicit (see also our twelfth response) in detailing this as a limitation of our meta-analysis and of the literature and recommend either reporting both GABA IU and GABA/Cr+PCr unless authors have clear expectations/justification regarding reporting of their data.

Our clarifying language in the discussion now reads:

“One <span class="Gene">of the cornerstone assumptions <span class="Gene">of our analysis is that, although the use of different references results in different fundamental units, participants within any given study using a given method can be compared to one another in relative terms. […] Provided this assumption is met, however, simultaneous estimation of between-list scaling factors and within-list participant differences provides a robust means of integrating multiple studies into a single overall trajectory.”

And we also added the following:

“Future studies should provide concent<span class="Species">ration values a<span class="Chemical">cross the lifespan; we recommend reporting of values both as metabolite as well as water ratios to look at within-study consistency and robustness of the effect of GABA. Recent developments in MRS of GABA methodology will allow for this, even when data is collected on multiple scanner platforms (Saleh et al., 2019).”

In the discussion, it must be emphasized at the beginning that <span class="Chemical">GABA/P<span class="Chemical">Cr+Cr ratio values do not equate to GABA levels or concentration – this needs to be acknowledged and corrected in the discussion – one must be careful not to over interpret the data.

We agree with the reviewer and incorrectly generalized these findings to “<span class="Chemical">GABA levels” in the discussion and are more careful throughout the manus<span class="Chemical">cript. We now refer to these as GABA “measures” and explicitly state that we included studies in UI and as ratios and as ‘estimates’ elsewhere. See also our fourteenth response.

In the discussion, the statement, "Cr+P<span class="Chemical">Cr is acquired during the same MEGAPRESS scan as GABA, limiting effects of chemical shift: Cr+PCr has a minimal shift from the GABA signal", is confusing and needs to be clarified.

We have clarified this statement to state:

“For instance, the <span class="Chemical">Cr+P<span class="Chemical">Cr at 3.05 ppm has minimal chemical shift from the GABA signal at 3 ppm (Mullins et al. 2014) and is acquired during the MEGA-PRESS sequence. […] In contrast, the water signal represents a more concentrated chemical yielding a higher SNR, but it may also introduce error in estimates of location due to chemical shift effects (Mullins et al., 2014) when not acquired from the same voxel as GABA and on some scanners requires a separate acquisition (see Choi et al. for a discussion).”

We now also refer to prior discussions on reference compound that are beyond our paper.

The statement, "the <span class="Chemical">water signal r<span class="Gene">epresents a more concentrated chemical yielding a higher SNR, but it may also introduce error in estimates of location due to chemical shift effects", is also confusing – unclear how chemical shift is a factor.

We now clarify this statement (see our sixteenth response). As discussed in the various consensus papers that came out over the past 2 years, acquisition <span class="Gene">of a <span class="Chemical">water unsuppressed scan brings in several difficulties. E.g. sometimes water is acquired from the same voxel (where the voxel is centred at 3 ppm) thus water is acquired from outside of the voxel that GABA was collected in. As per our sixteenth response, this is beyond the scope of this manuscript and we now refer to Mullins et al. 2014 and Choi et al. 2020.

The discussion on the CSF correction applied in the Porges et al. 2017b completely ignores the fact that the correct was not applied correctly – it should not be treated as a co-variate in the regression but as a volume correction factor (as noted earlier in the text). It would help to address this.

In Porges et al. 2017b, which is a study <span class="Gene">of the impact <span class="Gene">of different tissue-correction approaches, CSF correction was applied as volume correction, not as part of the regression model. Additionally α-corrected tissue correction proposed by Harris et al. was applied. We now clarify the impact of CSF volume correction in the manuscript:

“In <span class="Chemical">water-referenced studies <span class="Gene">of age-related differences in older cohorts (where bulk tissue changes are most likely to impact this relationship), accounting for cerebrospinal fluid (CSF) in the tissue-correction approach does not remove a significant relationship between age and cortical GABA (Porges et al., 2017b).”

In Porges et al. 2017a CSF was part <span class="Gene">of the regression model, but this does not impact the analysis performed here, as individual <span class="Chemical">data points were extracted from that study, no data from the regression model were used.