Age-related Differences in Hippocampal Subfield Volumes Across the Human Lifespan: A Meta-analysis

Abstract

The human hippocampus (Hc) is critical for memory function across the lifespan. It is comprised of cytoarchitectonically-distinct subfields: dentate gyrus (DG), cornu ammonis sectors (CA) 1–4, and subiculum, each of which may be differentially susceptible to neurodevelopmental and neurodegenerative mechanisms. Identifying age-related differences in Hc subfield volumes can provide insights into neural mechanisms of memory function across the lifespan. Limited evidence suggests that DG and CA3 volumes differ across development while other regions remain relatively stable, and studies of adulthood implicate a downward trend in all subfield volumes with prominent age effects on CA1. Due to differences in methods and limited sampling for any single study, the magnitude of age effects on subfield volumes and their probable lifespan trajectories remain unclear. Here, we conducted a meta-analysis on cross-sectional studies (n=48,278 participants, ages=4–95 years) to examine the association between age and Hc subfield volumes in development (n=11 studies), adulthood (n=30 studies), and a combined lifespan sample (n=41 studies) while adjusting estimates for sample sizes. In development, age was positively associated with DG and CA3–4 volumes, whereas in adulthood a negative association was observed with all subfields. Notably, the observed age effects were not different across subfield volumes within each age group. All subfield volumes showed a non-linear age pattern across the lifespan with DG and CA3–4 volume showing more distinct age trajectory as compared to the other subfields. Lastly, among all the study-level variables, only female percentage of study sample moderated the age effect on CA1 volumes: a higher female-to-male ratio in the study sample was linked to the negative association between age and CA1 volumes. These results document that Hc subfield volumes differ as a function of age offering broader implications for constructing theoretical models of lifespan memory development.

INTRODUCTION

The hippocampus (Hc) is a fundamental neural substrate of episodic memory, with bilateral Hc lesions resulting in global amnesia (Squire, 1992; Scolville & Milner, 1957). The Hc is a complex structure comprised of cytoarchitectonically distinct subfields: dentate gyrus (DG), cornu ammonis sectors (CA) 1 to 4 ¹, and the subicular complex (Jones & Mchugh, 2011). Hc subfields have different connectivity patterns (Insausti & Amaral, 2012), phylogenetic and vasculature properties (Spallazzi et al., 2019; Nieuwenhuys et al., 2008), and may underlie complementary aspects of memory function (Azab, Stark, & Stark, 2014; Jones & Mchugh, 2011; Schapiro, Turk-Browne, Botvinick, & Norman, 2017).

Over the last two decades, advances in high-resolution neuroimaging techniques have opened the possibility of in vivo examinations of Hc subfield volumes to investigate its structural development across the human lifespan and its memory correlates. Findings from these investigations suggest that Hc subfield volumes are differentially associated with age (Daugherty, Bender, Raz, & Ofen, 2016; Bussy et al., 2021; Raz et al., 2015). This differential association may contribute to the mixed patterns of change and stability in memory function commonly observed across development and adulthood. For example, age-related differences in DG and CA3 volumes correlate with associative memory ability in children (Daugherty, Flinn, & Ofen, 2017) and older adults (Shing et al., 2011). Such observations continue to motivate studies of age-related differences in Hc subfield volumes, with the hope of gaining insights into the neural basis of memory function across the lifespan.

Investigations of Hc subfield volumes across the lifespan are typically cross-sectional and study the samples representing separate age groups (e.g., children, older adults). Overall, the results of these studies have limited consistency in the magnitude and direction of age-related differences in Hc subfield volumes. Piecing together evidence from all available studies that used various methodologies, we can infer a plausible developmental trajectory for the Hc subfield volumes. The general trend suggests that the volumes of DG and CA subfields (Canada, Ngo, Newcombe, Geng, & Riggins, 2019; Keresztes, Raffington, Bender, & Bögl, 2021; Krogsrud et al., 2014; Lee, Ekstrom, & Ghetti, 2014; Mu, Yuan, & Tan, 2020; Riggins et al., 2018) and subiculum (Canada et al., 2019; Krogsrud et al., 2014; Mu et al., 2020) enlarge from mid-childhood to adolescence and then remain stable until young adulthood (Krogsrud et al., 2014). Yet, some evidence suggests that the growth and development of DG and CA3 volumes continue through young adulthood, with the direction of the age-volume associations remaining debatable (Daugherty et al., 2017; Keresztes et al., 2017).

Few longitudinal studies have investigated changes of Hc subfield volumes in children and have yielded discrepant results. A sample of 4 to 8 year-old children revealed an increase over one year in CA1 volumes in the Hc head from age 4 to 5, followed by an increase in other subfield volumes in the Hc body from age 5 to 6 (Canada, Hancock, & Riggins, 2021). Another sample of 6 to 10 year-old children showed an increase over two years that was limited to the subiculum volume in Hc body (Keresztes et al., 2021). Over a wider age range, one study identified a decline in subfield volumes over a two-year period in participants ages 8 to 21 years old (Tamnes et al., 2014). Yet, another sample of 8 – 26 years old participants, tested twice over a two-year period, provided additional evidence of differential development across subfields: CA1 and subiculum volumes increased in younger participants while DG and CA2–4 volumes decreased across the entire age range (Tamnes, Bos, van de Kamp, Peters, & Crone, 2018). Altogether, findings of cross-sectional and longitudinal studies of Hc subfield development are difficult to integrate for clear identification of developmental trends.

The cross-sectional studies of healthy young and older adults also document age-related differences in Hc subfield volumes that are somewhat variable across studies. Negative correlations of advanced age and CA1 volume are fairly consistently reported (Bender, Daugherty, & Raz, 2013; Daugherty, Bender, Raz, & Ofen, 2016; de Flores, La Joie, & Chételat, 2015; Frisoni et al., 2008; Mueller et al., 2007; Mueller, Schuff, Raptentsetsang, Elman, & Weiner, 2008; Susanne G Mueller & Weiner, 2009; Shing et al., 2011; Wisse, Biessels, et al., 2014; Wisse, Jan, Heringa, & Kuijf, 2014), with some exceptions that report null effect (Amaral et al., 2018; Bussy et al., 2021). The age-related differences in the DG and subiculum volumes are less consistent across studies. Some studies report a significant negative association between the volume of these subfields and age (Mueller et al., 2007; Mueller & Weiner, 2009; Pereira et al., 2014; Wisse et al., 2014; Thomann et al., 2013; de Flores et al., 2015; Frisoni et al., 2008; Joie et al., 2010; Wang et al., 2003), but others report a null association (Bender et al., 2013; de Flores et al., 2015; Joie et al., 2010; Shing et al., 2011; Mueller et al., 2007; Mueller and Weiner, 2009). To the best of our knowledge, only one longitudinal study to date has investigated volumetric changes of Hc subfield on older adults (ages ≥ 65) over a period of 4 years. This study found a shrinkage in all Hc subfields, which was especially pronounced in DG and CA4 volumes (Nadal et al., 2020). Thus, with the inconsistent cross-sectional reports and limited longitudinal evidence, the patterns of Hc subfield aging remains unclear.

One of the main reasons for this inconsistency is a significant variance in the age ranges used in various studies. Importantly, lifespan developmental studies of Hc subfield volumes are rare. In a rare report on a lifespan sample (ages 8–82 years), we observed a negative non-linear association between age and combined DG-CA3 volumes, negative linear differences in combined CA1–2 volumes, and negligible age differences in subiculum volume (Daugherty et al., 2016). Aside from this exception, other accounts based on qualitative summaries of the literature (de Flores, La Joie, & Chételat, 2015) suggest a non-linear trajectory of DG and CA3 volumes: increase in childhood and young adulthood, and shrinkage in late adulthood. In contrast, volume of other subfields appears to enter a period of stability through adolescence followed by shrinkage across the adult lifespan.

Differences in the implied lifespan developmental trajectories plausibly reflect selective sensitivity of the subfields to different neurodevelopmental processes that unfold across the lifespan. For example, variable rates of neurogenesis in the DG, from rapid in early development to slow and steady in adulthood (Aimone, Deng, & Gage, 2011), and declining at the later part of the lifespan (Seki & Aria, 1995; Seib & Martin-Villalba, 2015) may account for the non-linear growth and subsequent shrinkage in DG volume as suggested by the extant literature. Given the direct connection between DG and CA3 (i.e., trisynaptic pathway), similar pattern of change may also be observed in CA3. In contrast, strong negative link between age and CA1 volume across the adulthood may reflect the heightened vulnerability of this region to hypoxia and ischemia (Bartsch et al., 2015) compounded by age-related reduction in angiogenesis (Ingraham, Forbes, Riddle, & Sonntag, 2008) and age-related increase in cardiovascular risk and disease. Last, shrinkage of subiculum, a phylogenetically younger and structurally more complex Hc subfield (Nieuwenhuys et al., 2008), likely reflects neuronal loss (Šimić, Kostović, Winblad, & Bogdanović, 1997) due to lack of sufficient resources in combating the cumulative negative effects of age (i.e., a last in, first out hypothesis; Raz, 2000). Therefore, studying Hc subfield volumes across the lifespan may ultimately pave the way for understanding neurodevelopmental and neurodegenerative processes that modify the Hc-dependent functions, such as memory and spatial navigation, across lifespan.

A quantitative review is required to address the outlined limitations of the extant narrative reviews. In such a review, a meta-analysis, the reported effect sizes across developmental and adulthood studies can be pooled while accounting for the study sample characteristics. By compiling the effects across existing reports, we can expand the age range to include the entire lifespan and assess the extent to which estimated effect sizes vary across the subfields as a function of age. Further, a meta-analytic approach casts the estimated effects in the context of quantifiable study differences in methodological features (e.g., segmentation protocols, manual vs. automated demarcation, magnetic field strength, etc.) and sample characteristics (e.g., the degree of age, and sex representation). A quantitative review of the extant literature is an important first step to establish normative developmental trajectories and gauging the sources of inter-study variability for the benefit of the future studies.

Here, in a meta-analysis of cross-sectional studies, we assessed age-volume associations across Hc subfields to estimate age effects in development, adulthood, and a combined lifespan sample (ages 4–95 years old). We specified our hypotheses while considering the extant literature on Hc subfields development in diverse species, observed age-related differences in specific memory characteristics (i.e., memory complexity being linked to DG and CA3), and other age-related factors (i.e., cardiovascular disease association with smaller CA1 in adults). We hypothesized that DG and CA3 volumes would show a positive association with age across development, and the magnitude of this effect would be greater than those observed in CA1 and subiculum volumes. In adulthood, volumes of CA regions and DG would demonstrate a significant negative association with age and this effect would be greater than that observed in subiculum volume. Further, we hypothesized that across the lifespan, the sample mean age would moderate the direction of age-volume associations for CA regions and DG consistent with a non-linear trajectory of developmental increase followed by late life decline. The effect of methodological factors as well as sample characterizations on estimates of age differences in subfield volumes were evaluated in an exploratory analysis.

METHODS

Study Sample

A literature search of the computerized PubMed database was conducted on September 12th, 2022. For a valid and reliable estimation of effect sizes, the years of publication were limited to 2000 till the present day - roughly when high-resolution neuroimaging methods became common. Search terms were included (“hippocampal subfield” [tw] OR “HC subfield” [tw] OR subfield* [tw] OR “dentate gyrus” [tw] OR subiculum [tw] OR subicular [tw] OR “cornu ammonis” [tw] OR “CA1” [tw] OR “DG” [tw] OR “CA2” [tw] OR “CA3” [tw] OR “CA4” [tw]) AND (volume [tiab] OR volumetric [tiab]) AND (development* [tiab] OR children [tiab] OR aging[tiab] OR “age difference” [tiab] OR lifespan [tiab] OR adult* [tiab] OR adolescents [tiab]) NOT (rat [ti] OR mice [ti]). The initial search of PubMed database yielded 999 studies. An additional search of study references was performed for detecting the studies that were not included in PubMed (n=38). To address the issue of publication bias, a community-based inquiry by listserv (e.g., hippocampal subfield group) and social media (e.g., Twitter) was made to collect unpublished null findings. This search yielded two unpublished studies: Yu et al., 2021, and Geerlings, 2022 (for details on the survey see Supplementary Materials).

Inclusion and Exclusion Criteria

Titles and abstracts of the whole sample of studies (n=1039) were reviewed to eliminate the studies that did not meet the primary inclusion criteria (i.e., studies of human population), therefore all the animal studies, case studies, and duplicate studies were excluded at this stage. It is notable that, at this phase of screening, studies were not eliminated even if age was not a primary variable of interest (e.g., studies of genetic and environmental contributors of Hc subfield volumes, healthy and patient group comparisons of Hc subfield volumes, etc.) or if the study was primarily longitudinal in design. This ensured incorporating the potential age effect if it was included as a covariate in the statistical models. Through this screening procedure, 774 studies were excluded, and 265 studies were considered for full-text review.

Through the process of full-text review, 197 studies were excluded for the following reasons: age was not included in the statistical model or, if it was, the age effect was not reported (neither quantitively nor qualitatively), the age criteria were not met (ages <4 years old), age-related differences in total Hc or subregion (i.e., head, body, tail) volumes were reported without subfield measurements, different neuroimaging measures other than volumetry were used, the study sample included patients or was conducted ex vivo, or if the paper was a review or commentary without original data. At this stage of data screening, we further examined the studies for the availability of their data in open repository that would have allowed us to calculate the age effect on Hc subfield volumes on entire or sub-sample. Two studies met this criterion and were included in analyses. (Christidi, Karavasilis, & Rentzos, 2019; Eijk et al., 2020).

Of the 68 eligible articles, dependent-sample studies were identified (i.e., the whole same sample or subsample was overlapping between publications), and among these redundant samples, the study with a larger sample size was selected for the meta-analysis. We excluded findings from studies that included both children and adults but did not report separate age effects for children and adults. Statistically, this guarded against the possibility of averaging the likely opposite effect sizes in development and adulthood. Conceptually, effects from a single sample including children and older adults would be difficult to interpret because of the likely different neurobiological mechanisms underlying the age differences in childhood and adulthood. These selection steps further excluded 25 articles. It is notable that some of the adulthood studies reported only significant statistics for the subfield volumes and the null findings were verbally described as non-significant (CA1 n = 5, DG and CA3–4 n = 5, and subiculum n = 6). For all these cases with qualitive reports that could not be converted to exact effect size estimates, we contacted the authors for the missing information, and in the event of not responding, the effect sizes for these subfields were treated as zero and included in data analysis to avoid the bias of overestimating effect sizes. This decision was aimed at maximizing the validity of the analyses through representing the available literature and avoiding the inflation of the aggregated effect size estimate. We note that the inclusion of the null findings did not affect the standard error of the parameter estimates, suggesting negligible bias. For two studies we could not estimate age effects for the ROIs defined in the current study and thus excluded them at this stage. One of these studies (Carey et al., 2019) included age and Hc subfield volumes as predictors of two memory indices. It reported discrepant estimates of the age effect on Hc subfield volumes depending on the behavioral outcome of interest, thus rendering estimates of age differences non-independent from behavioral variables. The other study (Riphagen et al 2020) investigated a limited subset of Hc subfield volumes (i.e., CA1, CA3, DG) and failed to include sufficient information (i.e., significant vs. nonsignificant effects) or findings for other subfield volumes that were within the scope of our study (i.e., CA4, subiculum). After the last selection step, a total of 41 studies were selected for quantitative analysis (Figure 1).

Figure 1.

Illustrating the stepwise process for the selection and screening of the included studies.

Data Extraction and Coding

Five raters extracted and coded the variables from the selected articles, with high inter-rater agreement of 91% average across all items coded. The extracted variables included sample size, age range, age mean, age standard deviation (SD), female percentage, ethnical and racial composition, education, percentage of right-handedness, IQ, measures of depression, image sequence (i.e., T₁, T₂, or T₁ & T₂), image resolution, magnetic field strength, segmentation methods (i.e., manual vs. FreeSurfer, ASHS, etc.), intracranial volume (ICV) correction method, and applied segmentation protocol. For all coded variables, the inconsistency among ratings were sporadic and did not follow any discernable pattern. To further guard against potential errors in coding process, the lead author spot-checked all information extracted by the raters and included in the quantitative analysis. In the instances that mean age was not reported, the median was used as a measure of central tendency (Parker et al., 2019). Age groups were defined as developmental (age range including 4 to 18 years, and upper age limit of 33 years; n = 11) or adult (age range including 18 to 95 years, and lower age limit of 18; n = 30, Figure 2). Only few studies reported both linear and non-linear age estimates (n = 3), and therefore we extracted and used only the linear age estimates. To assess the effect of segmenting methods, manual delineation was contrasted with automated segmentation methods (i.e., FreeSurfer, ASHS, MAGet). We further assessed if the age estimates obtained from FreeSurfer segmentation differ when compared to other segmentation methods. When multiple segmentation methods were used in a single study, manual segmentation estimates were selected for analysis (de Flores, La Joie, Landeau, et al., 2015). In-plane image resolution was coded as a 3-level ordinal variable: level 3 was defined as typical high-resolution (≤ 0.5 mm), followed by level 2 (0.51 – 0.99 mm) and the lowest resolution was defined as level 1 (≥ 1 mm). The ICV correction method reported in the original studies was used when possible. However, in some cases and despite the original study design (n = 3), the age effect were available, or feasible to calculate, only for unadjusted volumes (i.e., reported correlation matrix, Botdorf, Dunstan, Sorcher, Dougherty, & Riggins, 2022; Picci et al., 2022; or calculated from descriptive statistics, Malhi et al., 2019). In one study that reported multiple ICV correction methods, we selected a method that enabled us to calculate exact effect sizes (Samara, Raji, Li, & Hershey, 2021). Two studies (Dalton, McCormick, De Luca, Clark, & Maguire, 2019; Dounavi et al., 2020) that used ICV-normalized volumes but did not specify the applied method were excluded from subsequent moderation analyses. For two studies that made their dataset available in open repository (Christidi et al., 2019; Eijk et al., 2020), ICV correction was performed using residual method.

Figure 2.

Age distribution for developmental (n=11) and adulthood (n=30) subsamples. Circles indicate the mean age of each study sample, lines indicate the age range of study samples.

Meta-analytic Procedures and Computation of Effect Sizes

All reported statistics were converted to Fisher z-scores (Z_r) as observed effect sizes for analysis. The computations were based on reported Pearson correlations (r), t-tests (i.e., two age-group comparisons), F-ratio (i.e., three or more age-group comparisons), means and SD of age groups, p-values, and sample sizes, and linear regression statistics (i.e., standardized coefficient). When calculating effect sizes based on the means and SD of the age groups, if three age groups were defined by a study, we calculated the effect sizes using the two extreme age groups for the purpose of independency of the samples and analysis of a wider age range (i.e., youngest (12–13 years old) vs. oldest (16–17 years old); (Malhi et al., 2019). In Malhi et al. (2019), descriptive statistics of age groups were separately reported for the participants of minimal and higher emotional trauma. For this study, the effect sizes were calculated only for minimal emotional exposure group to avoid the possibility of inclusion of clinical participants.

Three regions of interests (ROIs) were defined as follows: CA1, combined DG and CA3–4, and subiculum. Published segmentation protocols vary considerably in the selection of labels with regards to the number of CA regions combined with CA1 thus affecting CA2-specific estimates. In the meta-analysis, the CA2 estimate was combined with DG and CA3–4 when possible, yet there were 7 reports of CA1–2 combined volumes and 3 reports of CA1–2-3 combined volumes (see Table 1). Because CA2 was inconsistently represented across protocols and was inconsistently lumped with CA1 or DG and CA3–4 labels, no specific analysis for the region was feasible and therefore the variance it contributed would be treated as noise distributed across the other ROIs. For data analysis, we defined a covariate that labeled the type of segmentation protocol based on the number of anatomical labels combined with CA1 (e.g., CA1 = 1, CA1–2 = 2, CA1–2-3 = 3). This allowed us to include all studies regardless of segmentation protocols and assess if the number of anatomical labels combined contributed to the heterogeneity of estimated effect sizes in defined ROIs (i.e., CA1 and DG and CA3–4). In calculating the effect sizes, when the values were separately reported for the subfields (e.g., CA3, CA4, DG), or for the left and right hemispheres and Hc subregions (i.e., head, body, tail), the correlation values were averaged for a whole ROI estimate prior to calculating Fisher Z scores. When the subfield information was available for whole Hc and Hc subregions, only the former was used in analysis (Malykhin, Huang, Hrybouski, & Olsen, 2017).

Table 1.

Sample characterization, methods, effect size

	N	[Age Range] (M±SD)	%Female	Age Group	Education (M±SD)	MFS	Seq.	Resolution (mm)3	Segmenting Method	Labels Segmented	ROIs	ICV Correction	CA1	DG & CA3-4	Sub
Lee et al., 2014	38	[8–14] (11.3 ± 2.8)	48.7	Dev.	NA	3T	T2	.4 × .4 × 1.9	Manual	CA1, CA3/DG, Sub	CA1, CA3/DG, Sub	Residuals	0.26	0.25	0.18
Picci et al., 2015	49	[9–16] (12.53 ± 1.51)	53	Dev.	NA	3T	T1 & T2	1 × 1 × 1 & .4 × .4 × 2	FreeSurfer 7.1.1	-	CA1, [CA3, CA4/DG], Sub	Raw	−0.11	0.01	−0.27
Malhi et al., 2019	92*	[12–17] (14.41 ± 1.79)	100	Dev.	NA	3T	T1	NA	FreeSurfer 6.0	-	CA1, [CA3, CA4, GC/ML/DG], Sub	Raw	−0.78	−0.14	−0.13
Yu. & Ofen. 2021	97	[7–20] (12.52 ± 3.62)	53	Dev.	NA	3T	T2	.4 × .4 × 2	Manual	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	0.08	0.14	0.10
Keresztes et al., 2017	103	[6–27] (14.09 ± 2.44)	51	Dev.	NA	3T	T2	.4 × .4 × 2	Manual	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	0.13	0.22	−0.12
Keresztes et al., 2020	147	[6–8] (7.19 ± 0.45)	45	Dev.	NA	3T	T2	.4 × .4 × 2	ASHS	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	0.18	0.15	−0.03
Coughlin et al., 2022	149	[7–33] (15.26 ± 2.93)	52	Dev.	NA	3T	T2	.4 × .4 × 1.5	ASHS	CA1, CA2/3, DG, Sub	CA1, [CA2/3/ DG], Sub	Residual	0.00	0.24	0.11
Mu et al., 2020	198	[6–26] (11.65 ± 4.58)	43	Dev.	NA	3T	T1	1 × 1 × 1	FreeSurfer 6.0	-	CA1, [CA2/3/4/GC/ML/DG], Sub	Covariate	0.30	0.28	0.33
Riggins & Spencer., 2020	137	[4–8] (6.19 ± 1.5)	50	Dev.	NA	3T	T2	.4 × .4 × 2	ASHS	CA1, CA2/3/4/DG, Sub	CA1, CA2/3/4/DG, Sub	Residual	0.14	0.12	0.07
Krogsgud et al., 2014	244	[4–22] (12.3 ± 4.8)	52	Dev.	NA	1.5T	T1	1.25 × 1.25 × 1.20	FreeSurfer 5.1	-	CA1, [CA2/3/4/DG], Sub	Covariate	0.16	0.14	0.22
Botdorf et al., 2022 ^	5284	[9–13] (9.92 ± .62)	51.7	Dev.	NA	3T	T1 & T2	1 × 1 × 1	FreeSurfer 7.1	-	CA1, [CA2/3, CA4, GC/DG], Sub	Raw	0.10	0.08	0.09
Hanseeuw et al., 2011	15	[NA] (69.4 ± 4.8)	33.3	Adults	14.9 ± 2.4	3T	T1	.81 × .95 × 1	FreeSurfer	-	CA1, [CA2/3, CA4/DG], Sub	Raw	0.21	−0.47	−0.49
Shing et al., 2011	29	[20–78] (57.33 ± 2.56)	44.8	Adults	NA	3T	T2	.4 × .4 × 2	Manual	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	−0.38	NS	NS
Lim et al., 2013	33	[NA] (75.6 ± 4.2)	57	Adults	9.1 ± 4.7	3T	T1	NA × NA × .8	FreeSurfer 5.0.1	-	CA1, [CA2/3/4/DG], Sub	Residual	NS	−0.24	NS
Buch et al., 2022	39	[18–81] (35.4 ± 14.2)	51.2	Adults	NA	3T	T1	.44 × .44 ×.1	FreeSurfer 6.0.0	-	CA1, [CA2/3, CA4, GC/DG], Sub	NA	−0.10	−0.08	0.02
Periera et al. 2013	50	[50–75] (63.7 ± 7)	60	Adults	10.9 ± 3.9	3T	T1	1 × 1 × 1	FreeSurfer 5.1	-	CA1, [CA2/3/4/DG], Sub	Residual	−0.25	−0.47	−0.37
Doxy & Kirwan, 2015	54*	[18–63] (52.8 ± 5.72)	54	Adults	NA	3T	T1	1.15 × 1.15 × 1.20	Manual	CA1, CA3/DG, Sub	CA1, CA3/DG, Sub	Ratio	−0.68	−0.27	−0.59
Colenutt et al., 2017	60	[49–87] (67.8 ± NA)	61.7	Adults	NA	3T	T2	.34 × .34 × 1.72	Manual	CA1, CA2, CA3, DG, Sub, SRLM	CA1, [CA2/3/DG], Sub	Ratio	−0.44	−0.21	−0.16
Chi, Yang, Chang., 2022	60	[20–85] (45.70 ± 4.52)	52	Adults	15.44, 1.63	3T	T1	1 × 1 × 1	FreeSurfer 6.0	-	CA1, CA2/3, CA4, DG, Sub	Residual	−0.36	−0.21	−0.40
Mueller & Weiner., 2009	119	[22–85] (53.4 ± 17.2)	54	Adults	NA	4T	T3	.4 × .4 × 2	Manual	CA1, CA1/2 transition, DG/CA3, Sub	CA1, [CA1/2 transition, CA3/DG], Sub	Ratio	−0.18	−0.09	NS
Raz et al., 2015	80	[22–82] (57.84 ± 14.27)	75	Adults	16.28 ± 2.45	3T	T2	.4 × .4 × 2	Manual	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	−0.31	0.07	0.14
Malykhin et al., 2017	129	[18–85] (47.6 ± 18.9)	54.2	Adults	15.86 ± 2.46	4.7T	T2	.52 × .68 × 1	Manual	CA1/3, CA4/DG, Sub	CA1/3, CA4/DG, Sub	Ratio	NS	−0.17	−0.17
de Flores et al., 2015	98	[19–84] (45.7 ± 19.2)	59	Adults	13.5 ± 3.5	3T	T2	.375 × .375 × 2	Manual	CA1, CA2/3/4/DG, Sub	CA1, CA2/3/4/DG, Sub	Residual	−0.41	−0.17	−0.59
Wearn et al., 2021	99	[NA] (69.2 ± 8.55)	52.5	Adults	15.8 ± 3.13	3T	T1	.34 × .34 × 1.5 & 1.7	ASHS	CA1/3, DG, Sub	CA1/3, DG, Sub	Residual	−0.34	−0.27	−0.21
Christidi et al., 2019	108	[30–83] (63.13 ± 10.25)	55	Adults	NA	3T	T1	1 × 1 × 1	FreeSurfer	-	CA1, [CA2/3/4/GC/DG], Sub	Residual	−0.52	−0.58	−0.40
Voineskos et al., 2015	137	[18–86] (45.39 ± 19.02)	47	Adults	15.42 ± 1.95	1.5T	T1	.78 × .78 × 1.5	MAGeT	CA1, CA2/CA3, CA4/DG, SRLM, Sub.	CA1, [CA2/3/4/GC/DG], Sub	Covariate	−0.13	−0.29	−0.18
Radhakrishnan et al., 2022	154	[NA] (46.29 ± 4.57)	59.1	Adults	14.26 ± 2.28	3T	T1	.8 × .8 × .8	Manual	CA1, CA3/DG, Sub	CA1, CA3/DG	Raw	−0.08	−0.37	−0.11
Foster et al., 2019	177	[20–94] (53.66 ± 19.09)	59.3	Adults	15.56 ± 2.49	3T	T2	.42 × .42 × 2	ASHS	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	−0.25	−0.39	−0.33
Zheng et al., 2018	275	[20–89] (54.85 ± 20.61)	63.3	Adults	NA	3T	T1	NA × NA × 1	FreeSurfer 6.0	-	CA1, [CA2/3/4/GC/DG], Sub	Covariate	−0.31	−0.31	−0.25
Amaral et al., 2018 ^	316	[18–94] (45.17 ± 23.88)	62.3	Adults	NA	3T	T1 & T2	.3 × .3 × .3	MAGeT	CA1, CA2/3, CA4/DG, SRLM, Sub	CA1, [CA2/3/4/DG], Sub	Covariate	0.43	0.027	0.01
Geerlings et al., 2022	333	[27–90] (68 ± 9)	30	Adults	NA	7T	T2	.35 ×.35 × .35	ASHS	CA1, CA2, CA3, CA4/DG, Sub, tail	CA1, [CA2/3/4/DG], Sub	Ratio	−0.05	−0.07	−0.07
Bender et al., 2020	300*	[61–82] (69.99 ± 3.92)	38.3	Adults	14.07 ± 2.90	3T	T2	.4 × .4 × 2	ASHS	CA1/2, CA3/DG, Sub	CA1/2, CA3/DG, Sub	Residual	−0.18	−0.16	−0.19
Parker et al., 2019	408	[69–71.9] (70 ± NA)	49.5	Adults	NA	3T	T1	1 × 1 × 1	FreeSurfer 6.0	-	CA1, [CA2/3/4/GC/DG], Sub	Covariate	−0.10	−0.07	−0.10
Bussy et al., 2021 ^	665	[18–93] (55.4 ± 18.5)	58.4	Adults	NA	3T	T1	1 × 1 × 1	MAGeT	CA1, CA2/3, CA4/DG, SRLM, Sub	CA1, [CA2/3/4/DG], Sub	Covariate	−0.08	−0.21	−0.11
Eijk et al., 2020	727	[21–30] (23.94 ± 2.44)	63.5	Adults	NA	4T	T1	.94 × .94 × .9	FreeSurfer 6.0	-	CA1, [CA2/3/4/GC/DG], Sub	Residual	−0.11	−0.05	−0.036
Veldsman et al., 2021^	36653	[44–82] (64.04 ±7.54)	NA	Adults	NA	3T	T1 & T2	1 × 1 × 1 & 1.05 × 1 × 1	FreeSurfer 6.0	-	CA1, [CA2/3/4], GC/ML/DG], Sub	Residual	−0.57	−0.43	−0.50
Sahayakan et al., 2021	195	[18–40] (23.7 ± 3.9)	67.7	Adults	NA	3T	T1	1 × 1 × 1	FreeSurfer 6.0	-	[CA1/3, ML, Sub], [GC/DG, CA4], [PreSub]	Covariate	0.11	0.05	0.08
Samara et al., 2021	172	[20–79] (47.59 ± 16.52)	51.7	Adults	NA	3T	T1 & T2	.9 × .9 × 1.2 & .9 × .9 × 1.2	FreeSurfer 6.0	-	CA1, [CA2/3/4/GC/DG], Sub	Covariate	−0.19	−0.22	−0.04
Aslaksen et al., 2018	47	[20–71] (38.36 ± 20.16)	66	Adults	13.78 ± 2.02	1.5T	T1	.94 × .94 × 1.25	FreeSurfer 6.0	-	-	Residual	NS	NS	NS
Dounavi et al., 2020 ^	178	[40–59] (51.77 ± 5.4)	70.2	Adults	15.95 ± 3.36	3T	T1 & T2	1 × 1 × 1 & .4 × .4 × 2	FreeSurfer 6.0	-	-	Normalized	NS	NS	NS
Dalton et al., 2019	30	[NA] (46.7± 3.75)	40	Adults	NA	3T	T2	.52 × .52 × .5	Manual	CA1, CA3/2, DG/CA4, Sub, PreSub, ParaSub	-	Normalized	NS	NS	NS

Note:

N= sample size used in this analysis

(M±SD) = mean and standard deviation

NA=Not available

Dev. = developmental

MFS=magnetic field strength

Seq.= Imaging sequence

ROIs = regions of interests defined in this study, CA = Cornu ammonis, DG = Dentate gyrus, PreSub= Presubiculum, ParaSub = Parasubiculum, Sub = Subiculum

Hc= Hippocampus, ML= Molecular layer, HATA = Hc-Amygdala transition area, SRLM = Stratum-Radiatum-Lacunosum-Moleculare

^{^}Open access data

^*Indicates adjusted sample sizes based on the authors comments or data used for effect size calculations.

When studies reported a “significant” or “non-significant” relationship (i.e., for a certain ROI, hemisphere, or subregion) without reporting the actual statistics, the effect sizes were conservatively estimated based on p = 0.049 or r = 0, respectively (Aslaksen, Bystad, Ørbo, & Vangberg, 2018; Dalton, Mccormick, Luca, Clark, & Maguire, 2019; Dounavi et al., 2020; Lim et al., 2013; Malykhin et al., 2017; Mueller & Weiner, 2009; Parker et al., 2019; Shing et al., 2011). For one study (Lee et al., 2014) that reported the upper and lower limits for statistics (i.e., F > or p <), the lowest limit of effect size was estimated using p-value. These considerations were applied to guard against type I error or prevent overestimating effect sizes.

Statistical Methods

The association between age and subfield volumes was analyzed using metafor package in R with random-effects model estimation and based on the DerSimonian-Laird method (DerSimonian & Laird, 1986). The assumption of the random-effects model is that different true effects underlie the observed effect sizes across studies, and the true effect size may or may not vary from one study to another. The statistical analyses were conducted separately for three ROIs (i.e., CA1, DG and CA3–4, subiculum) across developmental and adulthood studies (for details see Table 2). In addition to estimating the average effect sizes for each ROI within each age group, we assessed the source of potential variability in estimated effect sizes by conducting a moderation analysis. Mixed effect models were implemented to evaluate heterogeneity (Q and I²) on various study-level variables including sample characteristics (age mean, age SD, age range, female percentage) and imaging methods (image resolution, image sequence (i.e., T₁, T₂, T₁ & T₂), segmentation method (i.e., manual vs. automated), segmentation protocol (i.e., CA1 = 1, CA1–2 = 2, CA1–2-3 = 3), correction for ICV volume method (i.e., covariate, residuals, ratio)). Other factors – ethnical and racial composition, education, percentage of right-handedness, IQ, measures of depression – were excluded due to insufficient availability. The analysis of moderation was conducted in the combined lifespan sample to maximize the number of studies (n=41) and variability in the sample. To assess the robustness of our findings and possibility of publication bias, we constructed funnel plots and performed the trim-and-fill analysis (Duval & Tweedie, 2000). The funnel plots compare the effect sizes to the standard error of individual studies (as an index of study size) to detect any publication bias across small and large studies based on their results. To this end, we visually inspected the funnel plots for asymmetry and conducted regression analysis to examine relationship between effect sizes and standard errors of the studies. We further used the nonparametric trim-and-fill method, to better understand what hypothetical studies may be missing due to publication bias and the extent to which the addition of these missing studies might have affected the age estimates.

Table 2.

Descriptors of the studies in the developmental, adulthood, and lifespan sample

	Age Group
	Development	Adulthood	Lifespan
# of studies Total # of participants	11 6,538	30 41,740	41 48,278
Mean age of study sample (mean, SD)	(11.58, 2.8) years	(53.65, 13.18) years	(42.36, 22) years
Age range of study sample [min-max] (mean)	[4–33] (11.45) years	[18–94] (49.96) years	[4–94] (38.19) years
Mean % Female of study sample (SD)	54.49% (15.44)	54.83%) (10.50)	54.74% (11.84)
MRI Sequence Weighting, study frequency	T1=3 T2=8	T1=16 T2=14	T1=19 T2=22
Segmentation Methods, study frequency	FreeSurfer=5 MAGeT=0 ASHS=3 Manual=3	FreeSurfer=14 MAGeT=3 ASHS=4 Manual=9	FreeSurfer=19 MAGeT=3 ASHS=7 Manual=12

RESULTS

Age Linked to the Larger DG and CA3–4 Volumes Across Development

The result of a random effect model across 11 developmental studies showed a significant positive association between age and volumes of DG and CA3–4 (b = 0.14; se = 0.03, Z_test = 4.2, p <. 0001, 95% CI: 0.07/0.20, Cohen’s d = 0.28) but not with CA1 (b = 0.02; se = 0.11, Z_test = .19, p = .84, 95% CI: −0.20/0.24, Cohen’s d = 0.04) or subiculum (b = 0.06; se = 0.05, Z_test = 1.3, p = 0.18, 95% CI: −0.03/0.16, Cohen’s d = 0.12, Figure 3). Comparing 95% confidence intervals that were largely overlapping indicated no statistical difference in the magnitude of age-volume associations between the subfields, consistent with a general age effect (Figure 3). The fail-safe analysis (Orwin, 1983) revealed that equal number of studies (n = 11) with null findings would be needed to offset the observed effect sizes. The test for funnel plot asymmetry and trim-fill analysis indicated an unbiased report of publications [CA1: z = −.44, p = 0.65; DG and CA3–4: z = 0.17, p = 0.82, subiculum: z = −1.23, p = 0.22, Figure S1).

Figure 3.

Forest plot of effect sizes for age differences in the volumes of CA1 (top), DG and CA3–4 (middle), and subiculum (bottom) across child development. Effect sizes are arranged based on the age mean of the study sample.

Age Linked to the Smaller Hc Subfield Volumes Across Adulthood

Across adulthood, age was negatively associated with volume with the magnitude of association similar across all regions: CA1 (b = −0.19; se = 0.05, Z_test = −4.00, p < .0001, 95% CI: −0.28/−0.09, Cohen’s d = 0.38, Figure 4), DG and CA3–4 (b = −.19; se = 0.03, Z_test = −5.66, p < .0001, 95% CI: −0.26/−0.13, Cohen’s d = 0.38, Figure 5) and subiculum (b = −0.18; se = 0.04, Z_test = −4.45, p <. 0001, 95% CI: −0.25/−0.09, Cohen’s d = 0.37, Figure 6). Comparing 95% confidence intervals that were largely overlapping indicated no statistical difference in the magnitude of age-volume associations between the subfields, which is consistent with a global age effect. The fail-safe analysis (Orwin, 1983) revealed that 30 additional studies with null findings is needed to offset the effect size observed for each ROI. The test for Funnel plot asymmetry and trim-fill analysis also indicated an unbiased report of publications [CA1: z =.01, p = 0.99; DG and CA3–4: z = 0.007, p = 0.99, subiculum: z = −0.11, p = 0.90, Figure S2).

Figure 4.

Forest plot of effect sizes for age differences in the volumes of CA1 across adulthood. Effect sizes are arranged based on the age mean of the study sample.

Figure 5.

Forest plot of effect sizes for age differences in the volumes of DG and CA3–4 across adulthood. Effect sizes are arranged based on the age mean of the study sample.

Figure 6.

Forest plot of effect sizes for age differences in the volumes of subiculum across adulthood. Effect sizes are arranged based on the age mean of the study sample.

Hc Subfield Volumes Showed a Non-linear Age Pattern Across the Lifespan

Before assessing the hypothesis of non-linear age trends of DG and CA volumes across the lifespan, we first confirmed significant heterogeneity within the observed effect sizes in our combined lifespan sample for each ROI: CA1 (Q χ² (40) = 4182.20, I² = 97.56%, p < 0.001), DG and CA3–4 (Q χ² (40) = 2057.67, I² = 95.33%, p < 0.001), and subiculum (Q χ² (40) = 2907.99, I² = 95.45%, p < 0.001). Because the sample’s mean age correlated both with the sample’s age range and with the sample’s age SD (rs = 0.54 & 0.41, respectively, p < 0.05 for both), two separate moderation analyses were conducted for each ROI. In each analysis, the effect of mean age was assessed together with either age range or age SD, as well as their respective interaction terms (age mean × age range, or age mean × age SD) as covariates. We centered the three variables at the respective grand mean to reduce potential collinearity. In both analyses, the sample’s mean age significantly moderated the age effect for all subfield volumes (i.e., mean age × age), consistent with a non-linear lifespan trajectory (ps < 0.01, Table 3). This effect was not observed for age range, age SD, or their interaction terms (ps > 0.18, Table 3). These results suggest that the effect of the sample’s mean age on observed effect sizes is independent of the age variability in the sample.

Table 3.

Overall categorical and moderator analysis within CA1, DG and CA3-4, and subiculum

			CA1
			Parameter Estimate			Test of Moderation
Moderator	k		b	se	95% CI for M		Qm	p
Overall Heterogeneity	41		−0.13	0.05	−0.22	−0.04	4182.2	<.0001*
Age Mean	41		−0.005	0.002	−0.009	−0.002	7.44	0.006*
Age SD	39		−0.002	0.008	−0.016	0.012	0.09	0.75
Age Range	36		−0.002	0.002	−0.006	0.001	1.34	0.24
Female%	40		−0.01	0.003	−0.017	−0.003	8.39	0.003*
Image Resolution	38		0.04	0.051	−0.058	0.139	0.64	0.42
Manual vs. Automated Segmentations	41		−0.034	0.110	−0.251	0.182	0..10	0.75
FreeSurfer vs. Other Segmentations	41		−0.056	.096	−0.243	0.133	0.33	0.56
		Post-hoc
Sequence	41						1.26	0.53
		T1 vs. T2	0.100	0.105	−0.106	0.306
		T1 vs. T1&T2	0.125	0.141	−0.151	0.402
		T2 vs. T1&T2	0.025	0.144	−0.258	0.308
ICV Correction	33						4.99	0.08
		Covariate vs Residual	−0.193	0.109	−0.407	0.021
		Covariate vs Ratio	−0.309	0.152	−0.608	−0.010
		Residual vs Ratio	−0.116	0.140	−0.390	0.158
Protocol	41						0.24	0.88
		CA1 vs. CA1-2	0.047	0.131	−0.209	0.303
		CA1 vs. CA1-3	0.070	0.185	−0.294	0.433
		CA1-2 vs. CA1-3	0.022	0.212	−0.394	0.438
			DG and CA3-4
			Parameter Estimate				Test of Moderation
Moderator	k		b	se	95% CI for M		Qm	p
Overall Heterogeneity	41		−0.103	0.035	−0.173	−0.033	2057.67	<.0001*
Age Mean	41		−0.007	0.001	−0.009	−0.005	36.04	0.001*
Age SD	39		−0.011	0.005	−0.021	−0.002	5.72	0.016*
Age Range	36		−0.004	0.013	−0.006	−0.001	11.08	0.0009*
Female%	40		−0.001	0.003	−0.007	0.003	0.42	0.51
Image Resolution	38		0.047	0.042	−0.035	0.131	1.26	0.26
Manual vs. Automated Segmentations	41		0.0738	0.083	−0.089	0.18	0.78	0.37
FreeSurfer vs. Other Segmentations	41		−0.057	.071	−0.196	0.081	0.65	0.42
		Post-hoc
Sequence	41						4.82	0.09
		T1 vs. T2	0.169	0.077	0.018	0.319
		T1 vs. T1&T2	0.078	0.1	−0.118	0.276
		T2 vs. T1&T2	0.12	0.14	−0.154	0.394
ICV Correction	33						4.99	0.08
		Covariate vs. Residual	−0.039	0.096	−0.227	0.148
		Covariate vs. Ratio	−0.089	0.134	−0.351	0.173
		Residual vs. Ratio	−0.050	0.123	−0.291	0.192
Protocol	41						1.47	0.47
		CA1 vs. CA1-2	0.116	0.097	−0.074	0.306
		CA1 vs. CA1-3	−0.005	0.136	−0.271	0.262
		CA1-2 vs. CA1-3	−0.121	0.156	−0.427	0.186
			Subiculum
			Parameter Estimate			Test of Moderation
Moderator	k		b	se	95% CI for M		Qm	p
Overall Heterogeneity	41		−0.111	0.036	−0.182	−0.04	2907.98	<.0001*
Age Mean	41		−0.005	0.001	−0.008	−0.002	13.76	0.0002*
Age SD	39		−0.007	0.005	−0.018	0.002	2.10	0.1467
Age Range	36		−0.003	0.001	−0.006	−0.004	4.87	0.02*
%Female	40		0.000	0.003	−0.006	0.005	0.005	0.94
Image Resolution	38		0.02	0.044	−0.065	0.106	0.21	0.64
Manual vs. Automated Segmentations	41		−0.024	0.084	−0.19	0.141	0.08	0.77
FreeSurfer vs. Other Segmentations	41		−0.004	.073	−0.139	0.148	0.003	0.95
		Post-hoc
Sequence	41						0.43	0.8
		T1 vs. T2	0.052	0.081	−0.108	0.211
		T1 vs. T1&T2	0.008	0.107	−0.202	0.218
		T2 vs. T1&T2	−0.044	0.110	−0.259	0.171
ICV Correction	33						3.34	0.18
		Covariate vs Residual	−0.154	0.094	−0.338	0.030
		Covariate vs Ratio	−0.194	0.131	−0.451	0.063
		Residual vs Ratio	−0.040	0.121	−0.278	0.197

Note: k= the number of effect sizes included in the analysis; se= standard error, CI = confidence Intervals; M=moderator

Significant results are bolded

The 95% confidence intervals of moderation analysis for age mean for each ROI were largely overlapping indicating no statistical difference in the observed age effect on Hc subfield volumes across the lifespan. Nevertheless, using the estimated parameters from the fit linear regression (Figure 7), we estimated the approximate age, at which the slope of age-related differences becomes negative. From the summary estimates, CA1 and subiculum volumes likely to display a similar age trajectory across the lifespan with negative correlations with age beginning approximately at the age of 20 (Z = 0) and a moderate age-related effect (Z = −0.3) by about the 8th decade. Unlike these two subfields, DG and CA3–4 volumes were negatively correlated with age beginning around the age of 28 years and moderate age-related effects by the early 7th decade. This reflects the summary of age-related differences in the literature but cannot be interpreted as volume change. Such conclusion can be drawn only in longitudinal studies.

Figure 7.

Scatterplots of effect size vs. sample mean age for CA1 (top), DG and CA3–4 (middle), and subiculum (bottom). The red dashed line indicates the reference line (Z=0). The areas of circles are proportional to sample sizes.

Before interpreting a lifespan trajectory as stitched together across studies in the analysis, we must consider if different methodological features affect the estimated effect size and direction. We found that, among all the study-level variables, only the percentage of female participants in the study sample moderated estimated effect sizes for CA1 volume (Q_M < 8.3, p = 0.003) such that a higher female-to-male ratio in the study sample was linked to the greater negative association between age and CA1 volumes. (Figure 8). No other methodological difference in imaging data collection or segmentation method significantly accounted for differences in observed effect sizes (for details see Table 3). Taken together, the moderator analysis strengthens interpretation of a non-linear lifespan trajectory of Hc subfield volumes that cannot be ascribed to method differences across studies.

Figure 8.

Scatter plot of effect size against female percentage of study sample for CA1. The areas of circles are proportional to sample sizes.

DISCUSSION

In this meta-analysis, we aimed to assess the associations between age and Hc subfield volumes across the human lifespan. We conclude that across the examined studies, age was positively associated with DG and CA3–4 volumes across development and negatively associated with all subfield volumes across adulthood. No difference in age-volume associations across the subfields was found within the age groups, suggesting a general effect of age on the Hc. A non-linear age-volume association was supported for all Hc subfield volumes across the lifespan: the relationship was positive or null in development and negative in adulthood. Last, a higher female-to-male ratio in the study sample was linked to a greater negative effect sizes for CA1 volume, suggesting that females’ CA1 volumes may be selectively affected by age.

The observed pattern of age-related differences generally supports the notion of Hc structure as dynamic and susceptible to change throughout the life-course. MRI-derived volumes are treated as a proxy for the underlying neural mechanisms thought relevant to brain health and function. The proxy measure is crude and with a poor understanding of the neurobiological basis of volume measures, any interpretation of the effects based on these measures remains speculative. Importantly, the volumetric measures may reflect different neurobiological mechanisms across development and adulthood that may also depend on brain regions that are known to develop and age on discrepant schedules. For example, shrinkage of the cerebral cortex is proposed to reflect pruning and increase in myelin fraction in children, yet the same effect is thought to reflect the loss of neuropil in older adults (Fjell et al., 2009; Giedd et al., 1999; Gogtay et al., 2004; Raz et al., 2005). Likewise, volume stability may reflect the cumulative, dynamic effects of various neural processes such as pruning, myelination, neurogenesis, and fluctuation in dendritic arborization in children compared to other processes that result in neuropil maintenance in older adults. Despite this general limitation, a few possibilities can be put forward for interpreting the pattern of effects we observed as reflective of neurodevelopmental and neurodegeneration mechanisms unfolding across the lifespan.

Potential Neural Mechanisms of Volumetric Differences in Hc Subfields Across Lifespan

The observed positive age-volume associations for DG and CA3–4 subfields across development align with the findings of animal studies reporting protracted development of these regions as compared to CA1 and subiculum (Lavenex & Banta Lavenex, 2013). The positive associations between age and DG and CA3–4 volumes during development likely reflect the inter-dependent, progressive processes of neurogenesis, dendritic growth, synaptogenesis, and myelination that are more influential than the refining processes such as pruning. Particularly relevant here is evidence pointing to the role of neurogenesis and myelination in the continued structural reorganization of DG and CA3 over development (Ábrahám et al., 2010; Seib & Martin-Villalba, 2015). The ongoing process of neurogenesis adds new granule cells to DG with their axons extending to CA3 (Seib & Martin-Villalba, 2015). The hilus of DG also shows a delayed process of myelination, known to be a potent inhibitor of neurite outgrowth and sprouting (Ábrahám et al., 2010; Arnold & Trojanowski, 1996; Benes, 1994).

The smaller subfield volumes among adults likely reflect the diminished efficiency or eventual failures of the same neural processes implicated in development (Kirkwood & Holliday, 1979) such as neuronal loss (Bobinski et al., 1999; Terry, DeTeresa, & Hansen, 1987), reduced neurogenesis (Seib & Martin-Villalba, 2015), shrinkage of neuronal elements or neuropil (Scheibel, Lindsay, Tomiyasu, & Scheibel, 1975; Qiu et al., 2013), or reduced myelin content (Courchesne et al., 2000; Arshad et al, 2016). The relative contributions of these components to age-related differences in volume may differ among the subfields based on cytoarchitecture and physiological vulnerability.

For example, smaller CA1 volume in older age may represent mostly neuronal loss (Šimić et al., 1997). Pyramidal cells of CA1 are selectively vulnerable to neuroinflammation (Courchesne et al., 2000) and have a relatively high oxidative stress threshold that may ultimately put this subfield at risk of oxidative damage and neuronal death (Wang & Michaelis, 2010; Wang et al., 2005, 2009). Risk factors that develop with age and worsen inflammation and oxidative damage differentially impact this region. For example, cardiovascular risk factors, hypoperfusion, and ischemia are shown to exacerbate age-related differences in CA1 volume and its cognitive correlates (Bender et al., 2013; Shing et al., 2011). The source of this vulnerability plausibly rests in the high concentration of glutamate receptors and increased calcium channels in CA1, which causes a temporary mismatch between energy demand and oxygen supply that leads to excitotoxicity and subsequent neurodegeneration (Bartsch et al., 2015; Shaw et al., 2021). Similarly, smaller subiculum volumes in clinical populations suggest glia-associated oxidative damage and elevated inflammatory cytokines as primary mechanisms of neurodegeneration (Rodriguez, Tai, LaDu, & Rebeck, 2014). However, a mendelian randomization study of healthy adults reports that homozygosity for a proinflammatory IL1βC-511T variant is associated with smaller volumes across all Hc subfields (Raz et al., 2015). The reduced subiculum volume in late life may also reflect different cell maintenance demands due to their phylogenetical and cytoarchitectonic characteristics (Nieuwenhuys et al., 2008; Hill et al., 2010).

DG and CA3, on the other hand, are suggested to have protection against these same mechanisms (Jackson, Rani, Kumar, Fooster, 2009; Lalonde & Mielke, 2014; Scharfman & Schwartzkroin, 1989), and the volumetric shrinkage of these regions beginning in mid-life instead may reflect an attenuated rate of neurogenesis and white matter degradation (Aimone, Deng, & Gage, 2011; Lee, Clemenson, & Gage, 2012; Seib & Martin-Villalba, 2015; Yassa, Muftuler, & Stark, 2010). The unique nature of neurogenesis in the DG that when preserved in late life may be protective against neurodegenerative disease (Babcock, Page, Fallon, & Webb, 2021), but in unregulated excess can be detrimental to memory function (Johnston, Shtrahman, Parylak, Gonçalves, & Gage, 2016), will need to be considered in future applied studies that include cognitive or functional measures in addition to volumetry.

Behavioral Implications of Volumetric Differences in Hc Subfields Across Lifespan

Interpreting the functional relevance of Hc subfield volumes across the lifespan is hindered by the lack of memory indices in the current meta-analysis. The quantitative review of cross-sectional reports points to a positive relationship between age and DG and CA3–4 volumes across development. Given that memory ability improves in parallel in young individuals, we may speculate that larger volumes and thus, greater neurogenesis, dendritic arborization, or myelination are beneficial to memory ability. As these neurodevelopmental processes unfold in time, new synapses are established within and across these subfields, and information transmission becomes efficient. The slow maturation of DG and CA3 subfields likely supports the gradual addition of spatial and temporal precision to memory events (Canada et al., 2021; Daugherty et al., 2017; Keresztes et al., 2017) to become more detailed and specific later in development (Keresztes et al., 2020; Ofen, Tang, Yu, & Johnson, 2019). In fact, the pattern of age-related differences in memory ability in children is consistent with hypothesized memory processes attributed to DG and CA3: DG is postulated to support memory specificity to minimize interference (i.e., pattern separation), whereas CA3 to be involved in binding the memory events to contribute to memory complexity and enrichment (i.e., pattern completion).

Longitudinal studies yield mixed findings with respect to the directionality of volume-memory associations across development. Two studies of children (ages 4–8 years, Canada et al., 2021; Keresztes et al., 2021, ages 6–10 years) report that the volumetric increase in CA1 and subiculum is advantageous for memory functioning in early- and mid-development. In contrast, in a sample of a wider age range (8–21 years), volume reduction of DG and CA2–4 has been linked to better verbal learning (Tamnes et al., 2014). Negative volume-memory associations plausibly highlight the role of pruning as a dominant, beneficial process across development that eliminates extra synapses to optimize the efficiency of information processing. The discrepancy in the reports might be an artifact of the differences in methodology (i.e., segmentation from T₁-weighted vs. T₂-weighted images, manual vs. automatic segmentations, sample size, etc.) or the reliability of the assessments, but it also may reflect the differential timing of underlying neurobiological mechanisms that unfold at different ages that may not be detected by cross-sectional studies.

The evidence linking smaller volume to better memory scores in children is intriguing as smaller subfield volume has been consistently implicated in lower memory scores in adults. It remains unknown, at what point of the lifespan the transition in volume-memory associations may occur. Older adults show difficulty in associative memory and pattern separation compared to younger adults unless more dissimilarities in the input are provided (Bender et al., 2013; Mueller et al., 2011; Shing et al., 2011; Yassa et al., 2011). This is likely due to impairment in DG and CA3 function and the respective memory process attributed to these subfields (i.e., pattern separation and completion, respectively). Notably, the perforant path extending from the entorhinal cortex to DG is vulnerable to aging and related diseases, which may disrupt the input to DG and downstream encoding process, and in turn pattern separation. When considering the finding of decreased neurogenesis rate in the DG, the cumulative effect on memory function may lead to reduction in the selectivity of representation and lesser differentiation among mnemonic representations (Aimone et al., 2011; Deng, Aimone, & Gage, 2010; Seib & Martin-Villalba, 2015), causing memory saturation, interference and familiarity features that are commonly reported in older adults. Importantly, the structural and functional changes in DG, as the main gateway of Hc, may potentially affect the Hc function at the network level (Burke & Barnes, 2006; Smith, Adams, Gallagher, Morrison, & Rapp, 2000). In fact, the evidence points to diminished inhibitory input to CA3 and hyperactivity in the auto-associative fibers of collaterals, which may explain older individuals’ reliance on pattern completion and familiarity (Geinisman, de Toledo-Morrell, Morrell, Persina, & Rossi, 1992). The memory correlates of CA1 and subiculum decline in aging have been less often studied, but limited evidence implicates CA1 in memory consolidation, delayed retrieval (Mueller, Chao, Berman, & Weiner, 2011) and subiculum in loss of efficiency and precision in spatio-temporal information (Daugherty et al., 2017).

Methodological Considerations, Limitations, and Future Direction

In this study, we found that a higher female-to-male ratio was linked to greater negative effect sizes for CA1 volume. The finding of significant effect of female representation in a sample on CA1 volume is intriguing and suggests that females’ CA1 volumes may be selectively affected by age. Nevertheless, this finding should be interpreted with caution and in the context of secondary data analysis. In our sample, the female-to-male ratio were relatively equally distributed across studies with only one study implementing a sample of 100% female (Malhi et al., 2019), the exclusion of which rendered a null effect of female percentage on CA1 volume (Q_M = 1.17, p = 0.27). Nevertheless, even without this study, the association between estimates of age effect for CA1 and percentage of females points to the negative direction, which was not observed in other subfields.

Excluding Malhi et al., (2019) from the developmental sample of studies resulted in the age differences in CA1 to be significant (b = 0.13), while not affecting the significance of age effects for other ROIs as indicated by overlapping confidence intervals. The estimates of age effect for CA1 volume obtained from this study emerged as an extreme value in reference to the distributions of the rest of the literature. Nevertheless, this study met all inclusion criteria and thus was retained in the meta-analysis. The difference in the observed effect size could reflect the nature of its sample, which consisted entirely of females. Although the presence of studies with disparate sample characteristics may produce outliers, including all eligible studies improves the external validity of a meta-analysis and provides the opportunity to test the potential effect of modifiers on observed effect sizes (e.g., sex-related differences).

Other study-level variables did not significantly modify estimated effect sizes. However, we note that the studies that qualified for inclusion in our analysis had relatively little variability in the factors considered. Additional studies are warranted using the broader literature and employing these methods to be able to address other research questions that cannot be adequately addressed here given the limited variability of some study-level parameters. We also did not have information about important life-stage events in the included studies, and therefore could not directly evaluate these age-related factors in the current meta-analysis. For example, the developmental trajectories of Hc subfield volumes may be affected by hormonal changes across the lifespan (e.g., puberty or menopause), and this effect may vary across subfields (Satterthwaite et al., 2014). Future investigations can use the non-linear age differences across the lifespan mentioned here to plan studies for sensitive periods in puberty and late-life hormonal changes in relation to Hc subfield volumes.

The inferential scope of this study is constrained by several limitations. First, numerous and highly variable Hc subfield segmentation protocols are used in the literature which was a challenge in aggregating across different findings. Hc subfields are segmented and measured using different protocols and methods (Yushkevich et al., 2015). This causes heterogeneity in the subfield delineation in terms of the number of anatomical labels and boundary definition, therefore undermining direct comparisons. For example, some laboratories delineate CA1 as an isolated subfield, likely for the sake of their research questions (de Flores et al., 2020), and some combine it with CA2 to ensure high reliability of volume estimations (Daugherty et al., 2016). Importantly, cytoarchitectural definitions of the subfields are established (Insausti & Amaral, 2003), but there is no consensus on specific operational definitions in histology or in vivo imaging to denote subfield boundaries. The current study combined the findings obtained from different protocols in respect to research hypotheses (i.e., CA1, DG and CA3–4, and subiculum) and any variance CA2 contributed to the other CA regions and DG was treated as noise. The test of CA label non-specificity as a moderator suggested no significant bias as a result of combining labels when interpreting the aggregate effect sizes, however, this approach bars any conclusion about age difference in CA2.

Our report on age trajectories of Hc subfield volumes is based on a summary of published evidence from cross-sectional studies with wide and variable age distributions (see Figure 2). In our moderation analysis of non-linear age trends across the lifespan (i.e., mean age moderating age estimates), we controlled for the age range and age SD of the study sample to account for the variable and overlapping age distributions. Nevertheless, such analysis is limited to test a smoothed continuous effect of age on the subfield volumes and thus our analysis of non-linear age trajectory in the lifespan sample was an approximation of this non-linearity. Further, our analysis is based on the reported linear estimates of age-volume associations. Notably, one of the included studies (Mu et al., 2019) reported on linear age effects (r values) based on the description in the methods or results sections but the figure captions indicated implementing locally smoothed models. Our approach to use linear age effects was in accord with our hypothesis testing of a non-linear age trajectory across the lifespan (i.e., mean age moderating linear age estimates). Further, given that most studies only report linear age trends, this decision allowed to maintain consistent estimates of age effect across studies. Importantly, the inclusion of non-linear polynomial terms testing age differences in the few studies (n = 3) could not be interpreted for the direction of effect similar to the linear age terms from other studies. To avoid this confound to interpretation, we excluded non-linear age estimates from analysis. This is an unavoidable limitation given the paucity of existing studies that systematically test continuous, non-linear slope terms.

Further, this meta-analysis aggregated the findings of cross-sectional studies, and thus the observed volume differences are not indicating age-related changes (Molenaar, 2009; Lindenberger, von Oertzen, Ghisletta, & Hertzog, 2011). In fact, our findings are at odds with the longitudinal reports on the specificity, direction, or magnitude of age effect (Canada et al., 2021; Tamnes et al., 2014; Keresztes et al., 2022; Tamnes et al., 2018). The observed discrepancies between the cross-sectional and longitudinal reports may reflect the major limitation of the former in detecting the developmental changes in the subfield volumes while accounting for individual differences (Lindenberger, von Oertzen, Ghisletta, & Hertzog, 2011; Pfefferbaum and Sullivan, 2015). The estimates of the age effect obtained from a cross-sectional study do not reflect temporal changes in the construct of interest but the accumulation of changes in the past. Individuals arrive to the assessment point of a cross-sectional study via various trajectories, and these individual differences in change are completely inaccessible to cross-sectional analyses. Moreover, cross-sectional estimates of age-related change are confounded with individual differences (i.e., cohort effect), the effect of which may not be disentangled from the true developmental changes. In a recent study, Keresztes et al. (2022) assessed developmental changes of the subfield volumes over two years in children aged 6–10 years old and compared it to the age-related volume differences obtained from the cross-sectional comparison in the same sample. Longitudinal analyses yielded an increase in the subiculum volume (Keresztes et al., 2022), which contradicts cross-sectional findings indicating larger volumes of DG and CA regions in older participants. Considering that that study used a reliable and sensitive method for Hc subfield volumetry (Homayouni et al., 2021), the discrepancy between cross-sectional and longitudinal findings likely goes beyond measurement error and highlights the limitations of the former in gauging developmental changes. These observations suggest that the cross-sectional framework is insufficient for studying the developmental changes of Hc subfield volumes, but instead, it may be applied to informing the hypotheses to be tested using a longitudinal framework. As has been repeatedly demonstrated in empirical studies of aging (Raz et al., 2005; Pfefferbaum and Sullivan, 2015) and theoretical discussions (Hofer & Sliwinski, 2001; Lindenberger et al., 2011) longitudinal study is essential to detect temporal changes across subfield volumes while accounting for the potential sources of individual differences.

Another limitation of this study is publication bias, which is inherent to any meta-analysis. Publication bias refers to the fact that the studies with significant and larger effect sizes are more likely to be published (Lipsey & Wilson, 2017), thus leading to overestimating the overall effect size in a meta-analysis (i.e., type I error). To address this issue, we sought unpublished findings by contacting the research community, which led to two reports with null effect (Yu, 2021; Geerlings, 2022). Due to the comparative nature of Hc subfields examination, studies of age effects are likely to publish even if only one of the subfields has a statistically significant result, which we speculate improves representation of null findings in this literature. In the event that null effects were verbally described but not quantitatively reported, we represented null effects as a zero-effect size to be conservative. Additionally, we estimated effect sizes from available descriptive statistics even when age was not a variable of primary interest in the study. All these steps were taken to maximize external validity of the meta-analysis to include plausible small or null effect sizes; however, it remains likely that additional null results exist and were never published. Our evaluation of potential publication bias indicated that at least 11 and 30 additional studies with null findings would be required to negate the observed overall effect of age on the subfields across development and adulthood, respectively, which further supports the meta-analysis estimated a robust effect.

CONCLUSION

In conclusion, the results of this meta-analysis revealed a robust association between age and Hc subfield volumes across the lifespan. The analysis suggests a non-linear trajectory across the lifespan, with a positive or stable association between age and subfield volumes across development, followed by a negative association across adulthood. The age-volume associations across subfields were not significantly different across development or adulthood, suggesting a global age effect on Hc. These findings shed light on the neural correlates of memory development across the lifespan and put forward the intriguing possibility for constructing theoretical models of lifespan memory development.

Data Availability Statement

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