Early-life education may help bolster declarative memory in old age, especially for women

Abstract

Declarative memory is critical for everyday learning. Although declarative memory abilities decline with age, evidence suggests that sex and early-life education might moderate these weaknesses. However, limitations in previous studies, such as small education ranges and unexamined potential interactions, have stymied our understanding of these possibly ‘protective’ variables. We investigated effects of sex and education on nonverbal declarative memory in 704 older adults (aged 58–98, 0–17 years of education) in a non-Western (Taiwanese) population, by probing recognition memory following incidental encoding. Items were drawings of real (e.g., hairbrush) and made-up objects. Mixed-effects regression revealed that age negatively impacted declarative memory, though this age effect was moderated by sex and object type. It was steeper for males than females (yielding a female advantage from age 70), though only for real objects. Remembering made-up objects showed shallower declines with no sex differences. Education was positively associated with memory, but also interacted with sex and object type: education benefited women more than men, and remembering real more than made-up objects. In men, the overall memory gains associated with each year of education were two times larger than the losses experienced during each year of aging; in women, they were five times larger. The findings suggest that nonverbal memory in older adults is associated negatively with age but positively with education, though both effects are modulated by sex, and by whether learning relates to pre-existing knowledge or new information. The study suggests downstream benefits from education, especially for girls.

Declarative memory is a key learning system rooted in the medial temporal lobe, including the hippocampus (Cabeza & Moscovitch, 2013; Davachi, 2006; Eichenbaum, 2012; Henke, 2010; Ullman, 2016; Ullman & Pullman, 2015; Wixted & Squire, 2011). It underlies a host of diverse tasks involving the learning of various types of information about events (episodic knowledge) and/or facts (semantic knowledge), such as remembering a shopping list, where you put your keys, when your doctor’s appointment was, the name of your new neighbor, or that fact you learned last week while watching Jeopardy. (Here we use the term ‘declarative memory’ to refer to the learning stages of this system—not to long-established semantic or autobiographical knowledge or its processing—and specifically to the learning that relies on the medial temporal lobe and its associated circuitry; Ullman, Earle, Walenski, and Janacsek, 2020.) While declines in declarative memory can be a symptom of various neurological disorders, such as Alzheimer’s disease, epilepsy, and hypoxia (Dickerson & Eichenbaum, 2010), they are also a hallmark of healthy aging. Indeed, a common complaint as people age is that they experience difficulties in remembering recently encountered information of the sort listed above.

Multiple studies have found that older adults are worse than younger adults at tasks probing learning in declarative memory, such as word list learning, paired associates, and story recall (Berenbaum, Baxter, Seidenberg, & Hermann, 1997; Craik & McDowd, 1987; Danckert & Craik, 2013; De Chastelaine, Mattson, Wang, Donley, & Rugg, 2015, 2016; Mattson, Wang, De Chastelaine, & Rugg, 2014; C. Murphy, Nordin, & Acosta, 1997; Naveh-Benjamin, 2000; Old & Naveh-Benjamin, 2008; Ratcliff & McKoon, 2015; Verhaeghen, Marcoen, & Goossens, 1993; Wang, Johnson, De Chastelaine, Donley, & Rugg, 2016; Zelinski & Burnight, 1997) (In contrast, retrieving long-established knowledge, in particular semantic knowledge, shows weaker declines and may in fact improve with age; Levine, Svoboda, Hay, Winocur, & Moscovitch, 2002; Mitchell, 1989; Nyberg, Lövdén, Riklund, Lindenberger, & Bäckman, 2012; Piolino, Desgranges, Benali, & Eustache, 2002.) The learning declines may start as early as during one’s 20s (S.-C. Li et al., 2004; Park et al., 1996; Verhaeghen & Salthouse, 1997), though recent claims based on longitudinal studies suggest a somewhat later onset, around middle adulthood (Nyberg et al., 2012). Some evidence suggests that the declines are non-linear, with the greatest declines during older adulthood (Nyberg et al., 2012). (A note on terminology: we use ‘decline’ and similar terms to refer to all age-related differences, including in cross-sectional studies, which constitute the vast majority of the extant literature on this topic; Pliatsikas et al., 2018.)

Given the importance of declarative memory in everyday life and the evidence suggesting aging declines in this domain, a thorough understanding of aging and declarative memory seems warranted, especially considering the rapidly aging population globally (Phillips, 2002; Rechel et al., 2013). Examining the roles of potentially ‘protective’ factors such sex and education that may moderate such age-related declines seems particularly important. Moreover, because healthy aging typically constitutes the baseline comparison for disorders that are associated with aging and declarative memory deficits (e.g., Alzheimer’s disease), elucidating declarative memory in healthy aging may have important translational impacts.

However, a number of gaps and weaknesses in previous research have stymied a fuller understanding of declarative memory in aging, including regarding factors that may mitigate (or aggravate) declines. In the remainder of the Introduction we discuss these gaps and weaknesses.

First, it remains unclear whether and how declarative memory aging declines hold for different types of information. Almost all previous studies (e.g., as cited above) have probed the learning of verbal information (or have collapsed verbal and nonverbal tasks; Herlitz, Nilsson, & Bäckman, 1997), despite the likely importance of nonverbal learning in everyday life, for example when encountering new objects (e.g., tools, animals), smells, tastes (e.g., “Have I smelled/tasted this before?”), faces, or places (e.g., “Do I recognize this face or place as one I have seen or been to before?”). A few declarative memory aging studies have tested the encoding of nonverbal stimuli but then probed recognition with verbal stimuli or recall with verbal responses (e.g., participants encoded pictures of objects and then recalled them verbally) (Maitland, Herlitz, Nyberg, Bäckman, & Nilsson, 2004; Rönnlund & Nilsson, 2008; Springer, McIntosh, Winocur, & Grady, 2005). We are aware of three studies that have examined aging effects in purely nonverbal declarative memory tasks, in which participants were presented with geometric designs and were then asked to draw or recognize them (Gale, Baxter, Connor, Herring, & Comer, 2007; Park et al., 2002; Pauls, Petermann, & Lepach, 2013). Though declarative memory abilities in these partially or purely nonverbal declarative memory studies also showed aging declines, the status of nonverbal declarative memory clearly remains understudied, including regarding the factors and confounds described below.

In addition to the distinction between verbal and nonverbal information, different types of nonverbal information might yield different patterns for declarative memory in aging. Of interest here, remembering having encountered a familiar object, person, or context may be expected to involve representations of established knowledge whose semantic richness and/or verbalizability (even if no specific verbal label is attached to it) could lead to better performance than remembering having encountered a brand new item or information not linked to prior knowledge. This distinction between remembering recently encountered information linked to established knowledge versus recently encountered information not linked to such knowledge (or linked to less such knowledge) may be an important factor in declarative memory performance and other aspects of cognition, with the former type of knowledge leading to better performance in a variety of tasks and paradigms (Badham, Hay, Foxon, Kaur, & Elizabeth, 2016; Duñabeitia, Avilés, & Carreiras, 2008; Duñabeitia, Marín, & Carreiras, 2009; Grondin, Lupker, & McRae, 2009; Kounios et al., 2009; Pexman, Hargreaves, Siakaluk, Bodner, & Pope, 2008; Pexman, Holyk, & Monfils, 2003; Pexman, Lupker, & Hino, 2002; Rabovsky, Schad, & Abdel Rahman, 2016; Springer et al., 2005; Umanath & Marsh, 2014). Indeed, previous evidence suggests that real objects are remembered better than novel (made-up) objects, for example in recognition memory tasks (Hedenius, Ullman, Alm, Jennische, & Persson, 2013; Lukács, Kemény, Lum, & Ullman, 2017), though we are not aware of any studies examining whether or how this difference may change during aging. Given that semantic knowledge, which may bolster the recognition of real objects, does not decrease and may in fact increase with increasing age (see above), it would be interesting to examine whether recognition memory for real versus for novel objects shows differential aging trajectories.

Second, the role of sex in age-related declines of declarative memory remains unclear, including regarding the distinction between learning verbal and nonverbal material. Verbal declarative memory tasks have shown female advantages, as compared to males, in numerous studies across the adult lifespan (Bleecker et al., 1988; Jack et al., 2015; Kimura & Seal, 2003; Maitland et al., 2004; Portin, Saarijärvi, Joukamaa, & Salokangas, 1995; Weiss, Kemmler, Deisenhammer, Fleischhacker, & Delazer, 2003; Youngjohn et al., 1991; Zelinski, Gilewski, & Schaie, 1993; but see Bonsang, Skirbekk, & Staudinger, 2017). The extent of this advantage seems to change during aging, though the relative performance and rate of decline for males and females vary across studies (Bleecker et al., 1988; De Frias, Nilsson, & Herlitz, 2006; Jack et al., 2015; Maitland et al., 2004; Pauls et al., 2013).

Of particular importance for the present study, the pattern for nonverbal declarative memory tasks is understood even less well. Across adulthood, nonverbal declarative memory studies have reported male advantages (visuospatial memory: Kail & Siegel, 1978; Lewin, Wolgers, & Herlitz, 2001; Pauls et al., 2013), no sex differences (visuospatial: Lewin et al., 2001; McGivern et al., 1997; Portin et al., 1995; unfamiliar odors: Öberg, Larsson, & Bäckman, 2002) and, apparently most frequently, female advantages (visuospatial: Alexander, Packard, & Peterson, 2002; Mcburney, Gaulin, Devineni, & Adams, 1997; McGivern et al., 1997, 1998; faces: Eals & Silverman, 1994; Herlitz & Yonker, 2002; Lewin & Herlitz, 2002; Lewin et al., 2001). The effect of aging on sex differences in nonverbal declarative memory has received little attention. We are aware of two aging studies examining sex differences in purely nonverbal declarative memory. One found steeper age-related declines for females than for males (with no sex difference among young adults but a male advantage in older adults; Pauls et al., 2013), whereas the other found similar declines between males and females (with a female advantage among middle-aged adults that persisted into old age; Gale et al., 2007). Thus, further work on how sex may modulate nonverbal declarative memory in aging seems warranted.

Third, education (i.e., years of education early in life) appears to have a positive effect on declarative memory in adulthood (Fuh, Wang, Lee, Lu, & Juang, 2006; Giogkaraki, Michaelides, & Constantinidou, 2013; Kremen et al., 2019; Lachman, Agrigoroaei, Murphy, & Tun, 2010; Lövdén et al., 2000; Magalhães & Hamdan, 2010; Malloy-Diniz, Lasmar, Gazinelli, Fuentes, & Salgado, 2007; Portin et al., 1995; Rönnlund & Nilsson, 2008; Van der Elst, Van Boxtel, Van Breukelen, & Jolles, 2005; Weber, Rubin, & Maki, 2013). However, this factor has not been well examined in studies of aging. It is currently unclear whether the observed education advantages for declarative memory change with increasing age, since the majority of studies comparing different age groups do not report whether they tested for age-education interactions. One study (Magalhães & Hamdan, 2010) specifically reported the absence of such an interaction, suggesting that the positive effects of education do not change with age. Another study examined only older adults, and did not find an effect of education on declarative memory; the authors suggested that the influence of demographic factors, including education, decreases in old age owing to the selective survival of participants in this age range, and rigorous health screening that is usually undertaken when studying these participants (Hassing, Wahlin, & Bäckman, 1998). Moreover, we are aware of only one study of declarative memory in older adults that tested for an interaction between education and sex (and did not find one) (Portin et al., 1995). Finally, to the best of our knowledge all studies investigating the role of education on declarative memory in middle-aged and older adults have employed verbal tasks, suggesting the possibility that positive effects of education could be restricted to verbal declarative memory. Thus, the examination of the role of education on nonverbal declarative memory in aging seems desirable.

Fourth, task-related gaps and confounding factors may also have impeded progress in our understanding of declarative memory in aging. Most studies of declarative memory, including in the investigation of aging effects, have employed explicit or intentional encoding tasks, in which participants are explicitly told to memorize items or other information (moreover, often in list learning paradigms, which can be problematic; see below). Incidental encoding, another important declarative memory paradigm, has been much less widely used in studies of aging. In this approach participants are given a task in which they encode information without knowing that they will later be tested on their memory of that information (Chipman & Kimura, 1998; De Chastelaine et al., 2015; Mattson et al., 2014; Park et al., 2002; Springer et al., 2005; Wang et al., 2016; see also Maitland et al., 2004). It has been suggested that in the real world most learning in declarative memory is incidental (Buckner, Kelley, & Petersen, 1999, p. 311; Dos Santos et al., 2010, p. 451; Plancher, Gyselinck, Nicolas, & Piolino, 2010, p. 381; Vingerhoets, Vermeule, & Santens, 2005, p. 675). Interestingly, sex differences in declarative memory might be influenced by the type of encoding: the one incidental encoding study we are aware of examining sex differences found similar performance in the two sexes (in young adults; Cherney & Ryalls, 1999). Given these patterns, and the fact that intentional encoding seems to rely in part on working memory (see next paragraph), further examining declarative memory with an incidental encoding paradigm appears to be an important step to further elucidate aging effects in this domain.

The current picture of declarative memory in aging may also be partly confounded with the status of working memory. By far the most commonly used task paradigm to study declarative memory in aging is some form of list learning (usually in an intentional learning paradigm), in which participants are presented with a series of items (often written or spoken words, but sometimes images of objects or shapes), which they are later asked to recall, or, less often, to recognize. However, list learning is heavily dependent on working memory abilities (e.g., Lum, Ullman, & Conti-Ramsden, 2015), likely due to the involvement of working memory in rehearsal as well as in strategy formation, which (among other factors) are also at play in other intentional encoding paradigms (Blumenfeld, Parks, Yonelinas, & Ranganath, 2010; Craik & Rose, 2012; Logan, Sanders, Snyder, Morris, & Buckner, 2002; Ranganath & Knight, 2003; Takashima et al., 2006). Working memory abilities are highly age-sensitive and are affected by both sex and education in aging (Bopp & Verhaeghen, 2018; Pliatsikas et al., 2018), potentially confounding the effects of these factors on declarative memory when it is assessed in a list learning task. Moreover, some non-list learning studies (including one of the purely nonverbal declarative memory aging studies; Pauls et al., 2013) have only reported dependent measures that collapse immediate and delayed recall, again confounding declarative memory with working memory, since the latter is particularly important for immediate recall.

Lastly, as with most experimental work on cognition, the overwhelming majority of studies of declarative memory in aging has examined participants in Western societies. We are aware of two studies assessing the effects of adult aging on declarative memory outside of North America or Europe, both in Brazil (Magalhães & Hamdan, 2010; Malloy-Diniz et al., 2007), as well as one recent study examining verbal declarative memory in multiple countries around the world (Bonsang et al., 2017). Such research is of importance given evidence that multiple cognitive abilities, including memory, can show substantial inter-population variability (Allaire & Whitfield, 2004; Bonsang et al., 2017; Henrich, Heine, & Norenzayan, 2010). Thus, findings applicable to one population may not necessarily hold for another group, leading to the call for research on different populations (Allaire & Whitfield, 2004; Bonsang et al., 2017; Henrich et al., 2010).

The present study

Here we present data from a study of nonverbal declarative memory (episodic memory) in a representative sample of 704 Taiwanese participants between 58 and 98 years of age, with a relatively balanced sex ratio and a wide range of years of education (0–17 years). We assessed declarative memory with a recognition memory task: five minutes after participants were given an incidental encoding paradigm (in which they were presented with drawings of real and made-up objects and were asked to decide which ones were real), they were given a recognition memory test (in which they were presented with the same real/made-up items as in encoding, as well as new real/made-up items as foils, and were asked to decide which ones they had seen before). Based on the previous literature, we predicted that recognition memory abilities should decline with age. Object type was also expected to show an effect, with better recognition memory performance at real than made-up objects. We expected that higher levels of education might mitigate the age-related memory weaknesses; the nature of potential moderating effects of sex was less clear. We did not have specific predictions about more complex interactions among (various combinations of) age, sex, education, and object type.

Methods

Participants

This study presents data collected as part of the Social Environment and Biomarkers of Aging Study (SEBAS). Together with its parent study (the Taiwan Longitudinal Study of Aging), SEBAS has collected a wide range of social, demographic, and health-related data, as well as performance and biomarker measures, on older adults in Taiwan, across several collection waves (Cornman et al., 2016; Goldman et al., 2004; Weinstein et al., 2014). During the 2011 SEBAS data collection wave, three computer-based cognitive tasks were also included: the attentional network task (ANT; Fan, Mccandliss, Fossella, Flombaum, & Posner, 2005; results reported in Veríssimo et al., in prep), the n-back task of working memory (Owen, McMillan, Laird, & Bullmore, 2005; results reported in Pliatsikas et al., 2018), and a recognition memory task to examine learning in declarative memory (Hedenius et al., 2013; Lukács et al., 2017), which is reported in the present paper. A variety of demographic and related information was also acquired. This included sex, date of birth, total years of education (0–17, with 17 also representing higher levels of education), handedness as measured by four questions modified from the Edinburgh Handedness Inventory (Oldfield, 1971) for the population being tested (specifically targeting writing, the use of chopsticks, scissors, and brushing teeth), and information on any history of neurological, psychiatric, learning, cognitive, or other brain-related conditions. The research was approved by the Institutional Review Boards at Georgetown University, Princeton University, and the Ministry of Health in Taiwan. Data requests for this study should be sent to: Health Promotion Administration; Ministry of Health and Welfare; 6th Floor, No 95 Mincyuan Road; West District, Taichung City; Taiwan, 40341 ROC.

A cohort of 1031 individuals participated in the 2011 wave of SEBAS. All were native speakers of Chinese, in particular Hakka, Mandarin, or Taiwanese (Hokkien). Of these individuals, 989 agreed to perform the recognition memory task. Fifty-six were given the encoding phase of the task (see below for details), but then did not perform the recognition phase; these participants were excluded from analysis. Of the remaining 933 participants who performed both the encoding and recognition phases, 5 were excluded because they did not perform the recognition phase until the end without interruptions, and 1 owing to a coding error, which made it impossible to match their data with their demographic measures. Of the remaining 927 participants, 73 were excluded due to the presence of neurological, psychiatric, learning, cognitive, or other brain-related conditions (e.g., Parkinson’s disease, stroke, depression). A further 50 were excluded due to missing date of birth, and 12 owing to missing handedness information (analyses were performed both without and with handedness included as a covariate; see Results). Of the remaining participants, 82 were then excluded due to recognition phase d’ scores of zero or below (zero d’ prime scores reflect chance performance; negative d’ scores indicate that the participants may have been performing the task incorrectly), and a further 6 were removed because 50% or fewer of their trials in the recognition phase were valid: valid trials were defined as those for which a yes or no response was given at least 400ms post-stimulus (faster responses are outliers in our dataset; such fast responses may distort the magnitude of effects; Ratcliff, 1993), and within 5000ms, which was the time limit for each item (see below). Statistical analyses, as reported below, were performed on the data of the resulting 704 participants. Years of education for these participants, categorized by 5-year age brackets and by sex, are presented in Table 1.

Table 1:

Number of participants and mean years of education (and SDs) presented in 5-year age brackets, for each sex and across both sexes.

	N			Mean years of education (SD)
	Female	Male	Total	Female	Male	Total
55–59 (min. age 58)	38	44	82	9.08 (4.06)	10.48 (3.61)	9.83 (4.71)
60–64	120	123	243	8.06 (4.17)	10.11 (3.93)	9.10 (4.79)
65–69	58	58	116	6.29 (3.99)	8.47 (4.01)	7.38 (4.87)
70–74	40	56	96	2.80 (4.32)	8.00 (4.23)	5.83 (4.88)
75–79	30	33	63	3.50 (3.80)	6.03 (4.36)	4.83 (4.64)
80–84	24	42	66	4.21 (3.08)	7.69 (4.13)	6.42 (4.47)
85+	17	21	38	2.29 (3.68)	7.95 (5.56)	5.42 (4.48)
Whole sample	327	377	704	6.22 (4.61)	8.84 (4.31)	7.62 (4.64)

Note. This table displays sample sizes and educational information in 5-year age brackets, for informational purposes only; we remind readers that all analyses were performed with age as a continuous variable. Mean age of the full sample: 67.99 years (SD 8.72); males: 68.82 years (SD 8.99); females: 66.99 years (SD 8.38). The male and female participants differed significantly in education (t(702) = 7.79, p < .0001) but not in age (t(702) = 1.38, p = .169). The correlation between age and education was weak to moderate (r = −0.33, p < .001; male: r = −0.27, p < .001; female: r = −0.46, p < .001). Also see Data Analysis.

Materials

The recognition memory task consisted of two phases: incidental encoding and recognition. During the encoding phase, participants were presented with 64 black and white line drawings: 32 images that depict real objects (referred to as ‘real’ in this paper; e.g., a hippopotamus) and 32 images that depict made-up objects (referred to below as ‘novel’). See Figure 1 for examples. These experimental items were preceded by three practice trials. Participants were asked to decide if the objects are real or made-up. No indication was given that they would be subsequently tested on their memory of having seen these objects. The recognition phase took place about 5 minutes after the encoding phase finished. During recognition the participants were presented with 128 drawings, half of which they had seen before during encoding, while the other half were foils that they had not seen during encoding, which again were composed of 32 real and 32 novel objects. The target images were preceded by six practice trials (half seen during encoding practice, half not). Participants were asked to decide if they saw the objects in the previous phase or not. Images for both the real and novel items used across the phases were taken from various sources, and modified as necessary, including to minimize nameability of the novel objects; this was confirmed through piloting. For more detailed information on this task (e.g., item creation), see Hedenius et al. (2013) and Lukács et al. (2017).

Figure 1:

Examples of the real and novel objects used as stimulus materials for the encoding phase (top two images, in A) and the recognition phase (bottom two images, in B). Note the reminder symbols at the bottom of the screen, which differ between the encoding and recognition phases; see Procedure.

Procedure

Both phases of the task were presented in black on a white background on a laptop with Windows XP, using E-Prime Version 2.0 (Schneider, Eschman, & Zuccolotto, 2002a, 2002b).

Encoding phase.

In this phase, participants were asked to decide if each item was real or not. Preceding each item, a fixation cross appeared in the center of the screen for 1000 milliseconds, followed by a real or novel item for 500 ms. When the item disappeared, the cross reappeared on the screen until the participant responded (button press on a Psychology Software Tools Serial Response Box; SRBox), or after the timeout of 5000 ms. After the subject responded, a 200 ms advance tone was played, followed by 800 ms of fixation. If instead the participant gave no response within the 5000 ms period, participants heard a 400 ms time-out tone sound followed by 600 ms of fixation, after which the 200 ms advance tone and 800 ms fixation occurred. The next item then began with 1000 ms of fixation. Note that if the participant responded before the 500 ms item presentation ended, the response time was recorded, but the program did not advance until the end of the 500 ms. This fixed stimulus duration ensured that the time the object was displayed for encoding was held constant across participants. The specific presentation time of 500 ms was selected on the basis of piloting with the target population, and ensured that the total task duration was reasonable, in order to minimize fatigue (which may tend to disadvantage the older participants).

Participants were instructed to indicate, via a button press, whether or not an item was real. A reminder appeared at the bottom of the screen throughout the task indicating which button corresponded to which decision; “real” was indicated by a green circle, and “novel” by a red X (Figure 1A); symbols rather than verbal reminders were used to avoid problems with individuals with low education/reading abilities. Left/right mapping of real/novel was counterbalanced. The items were presented in a (single) pseudo-randomized order, with no more than three consecutive real or novel items.

Recognition phase.

Participants were told they would see pictures of real and novel (made-up) objects again, some of which they had seen earlier and some of which they had not. They were asked to indicate, via a button press, whether or not they had seen the item earlier. The timing and presentation of the items were identical to the encoding phase, but the reminder at the bottom of the screen was Yes/No (or No/Yes, depending on the counterbalanced version), where “Yes” was indicated by an image of an eye, and “No” by the same image of an eye, but circled in red and with a line through it (Figure 1B). These symbols were selected after piloting in Taiwan. Item order was pseudo-randomized, with no more than 3 consecutive real or novel items, no more than 3 consecutive seen-before or not-seen-before items, and no more than 2 consecutive items with the same real/novel and seen-before/not-seen-before values (e.g., seen-before real items).

Data analysis

In line with previous work (Lukács et al., 2017), the dependent variable of interest here is d-prime (d’) in the recognition phase. According to signal detection theory (Stanislaw & Todorov, 1999), d’ scores measure discrimination independent of response bias, that is, independent of any tendencies for participants to give one or the other type of response (in this case, seen-before or not). D-prime was computed, separately for real and novel items for each participant, as the z-score of the hit rate minus the z-score of the false alarm rate, with both the hit rate and false alarm rate adjusted by the loglinear method to avoid infinite or indeterminate d’ scores (Hautus, 1995; Stanislaw & Todorov, 1999). We also computed d’ scores for the encoding phase (e.g., the correct identification of a real item was a hit). Performance on the encoding phase is of secondary interest, in particular because encoding was incidental, and thus performance in this phase is not clearly indicative of whether or not an item was actually encoded (Lukács et al., 2017); also see Results.

The recognition phase d’ scores were analyzed in a linear mixed-effects regression model (quadratic effects were also examined; see Results), with participant as a random effect, using the lme4 package in R (Bates, Mächler, Bolker, & Walker, 2015). Fixed effects were age (continuous), sex (2 levels: male, female), object type (2 levels: real, novel), and education (continuous), as well as all their interactions. (Random slopes could not be included because there was only one observation per condition per subject, given the use of d’.) In order to obtain estimates of ‘main effects’ for all predictors (analogous to those obtained for main effects in ANOVAs), in the main analyses the continuous predictors (age and education) were mean-centered, and the categorical predictors (sex and object type) were assigned sum-coded contrasts (e.g., −0.5 and 0.5) (Barr, Levy, Scheepers, & Tily, 2013; Levy, 2014). Thus, all effects (e.g., for age) hold at mean values of other continuous variables (e.g., at the mean value of education) and constitute the average of the effect of interest across different levels of other categorical variables (e.g., for males and for females). Follow-ups to interactions with categorical predictors were obtained by relevelling such predictors and refitting the model. For example, in the follow-ups for the sex by education interaction (see Results), sex was assigned 0/1 contrasts (with either male or female as the reference level of 0, depending on which sex was examined in the follow-up), such that all other model estimates, including the education effect, refer to the reference level. Note that another type of ‘main effect’ coding for categorical predictors is to convert them to numeric variables and then to mean-center them (e.g., Fraundorf & Jaeger, 2016; Montero-Melis, Jaeger, & Bylund, 2016). We also ran the regression model using this coding approach, in alternate analyses; see Results.

All predictors were simultaneously included in the regression analyses, allowing us to control for any correlations among them. Specifically, estimates in multiple regression, including with mixed-effects models, reflect the unique variance of each predictor (i.e., the part of each variable that cannot be predicted by all others in the regression model). Effects should therefore be interpreted as the contribution of each variable beyond any correlations with the others (e.g. Wurm & Fisicaro, 2014).¹

Results

Table 2 presents mean d’ scores and standard deviations (computed by-participants) across all participants and both object types, as well as separately for males and females and for real and novel objects.

Table 2:

Mean performance in d’ scores (and SDs) in the recognition phase.

	Females	Males	All participants
Real objects	1.03 (0.70)	1.13 (0.72)	1.08 (0.71)
Novel objects	0.50 (0.41)	0.49 (0.38)	0.50 (0.39)
All objects	0.76 (0.63)	0.81 (0.66)	0.79 (0.65)

The results of the linear mixed-effects regression model are shown in Table 3, which presents regression estimates (b), standard errors (SE), t-values, and p-values for all main effects and interactions.

Table 3:

Results from the mixed-effects linear regression model on recognition memory d’ scores.

Random effects:	Name	Variance	SD
Subject	Intercept	0.0677	0.2601
Residual		0.1697	0.4119

Fixed effects:	b	SE	t	p
Intercept (estimated grand mean)	0.8050	0.0172	46.75	<.00001	***
Age	−0.0144	0.0021	−6.87	<.00001	***
Sexa	−0.0570	0.0344	−1.65	.099	#
Education	0.0418	0.0036	11.46	<.00001	***
Object typeb	−0.5843	0.0257	−22.75	<.00001	***
Age × Sex	−0.0095	0.0042	−2.28	.023	*
Age × Education	<0.0001	0.0004	−0.01	.992
Age × Object type	0.0134	0.0031	4.30	<.00001	***
Sex × Education	−0.0151	0.0073	−2.07	.039	*
Sex × Object type	0.0366	0.0514	0.71	.478
Education × Object type	−0.0530	0.0054	−9.74	<.00001	***
Age × Sex × Education	−0.0006	0.0008	−0.69	.490
Age × Sex × Object type	0.0135	0.0062	2.16	.031	*
Age × Education × Object type	0.0004	0.0006	0.59	.555
Sex × Education × Object type	−0.0086	0.0109	−0.79	.430
Age × Sex × Education × Object type	0.0012	0.0012	0.99	.322

Notes.

^***p < .00001;

^*p < .05;

^#p < .10.

P-values were obtained from t-tests with 1392 degrees of freedom, calculated as the number of data points (i.e., 1408) minus the number of fixed effect estimates (i.e., 16) (Baayen, Davidson, & Bates, 2008). All continuous predictors were mean-centered; all categorical predictors were assigned sum-coded contrasts (see Methods).

^aSex is coded as −0.5 for females and 0.5 for males.

^bObject type is coded as −0.5 for real objects and 0.5 for novel objects.

These mixed-effects models yielded main effects of age (worse performance with increasing age), education (better performance with greater numbers of years of education), and object type (better performance with real objects than with novel objects), as well as a marginal main effect of sex (marginally better performance for males than females).

These main effects were qualified by several significant interactions. First, a significant interaction between sex and education reflected the fact that (independent of age or object type) an increase in education had a greater effect on females’ than on males’ recognition performance (see Figure 2), though both effects were significant (females: b = 0.0494, SE = 0.0054, t = 9.21, p < .0001; males: b = 0.0343, SE = 0.0049, t = 6.92, p < .0001). Moreover, at the minimum level of education in our sample (0 years), the predicted d’ scores (i.e., from the regression model) did not differ between males and females (b = 0.0581, SE = 0.0631, t = 0.92, p = .358). In contrast, at the maximum number of years of education (17 years), the predicted performance of females was significantly higher than that of males (b = −0.1985, SE = 0.0789, t = −2.52, p = .012). (This ‘endpoint’ comparison was computed by subtracting the maximum education value, i.e., 17, from each participant’s number of years of education; the endpoint analysis just above and those below were computed in an analogous manner.) The lowest education level at which males and females showed a significant difference on their predicted d’ scores was 9 years (b = −0.0778, SE = 0.0366, t = −2.12, p = .034).

Figure 2:

Performance at recognition memory (d’) as a function of education and sex, controlling for other predictors (i.e., controlling both for main effects and for interactions). In all figures, regression lines represent partial effects, that is, the effects of interest while holding all other continuous predictors constant at their means, and as the average of the effects of interest at different levels of other categorical variables. Shaded bands represent pointwise standard errors (95% confidence intervals are approximately twice the width of standard error bands).

Second, a significant interaction between education and object type (see Figure 3) indicated that education had a more positive effect on participants’ performance with real objects (b = 0.0683, SE = 0.0046, t = 15.01, p < .0001) than with novel objects (b = 0.0153, SE = 0.0046, t = 3.37, p < .0001), though for both object types the positive effect of education was significant. Moreover, participants’ advantage at recognizing real objects as compared to novel objects was larger at the maximum number of years of education (17) than at the minimum number of years (0), though both differences were significant (maximum: b = −1.0810, SE = 0.0589, t = −18.37, p < .0001; minimum: b = −0.1803, SE = 0.0471, t = −3.83, p < .0005).

Figure 3:

Performance at recognition memory (d’) as a function of education and object type, controlling for other predictors.

The model also yielded significant interactions between age and sex (with a larger age decline for males than females), and between age and object type (with a larger age decline for real than novel objects), though both of these were qualified by a three-way interaction between age, sex, and object type (Table 3). Following up on this three-way interaction, we found an interaction between age and sex for real objects (b = −0.0163, SE = 0.0052, t = −3.12, p = .002), but not for novel objects (b = −0.0028, SE = 0.0052, t = −0.54, p = .589). Rather, for novel objects (Figure 4, Panel B) there was simply a main effect of age on performance, across both sexes (b = −0.0077, SE = 0.0026, t = −2.94, p = .003), with no main effect of sex (b = −0.0387, SE = 0.0430, t = −0.90, p = .368).

Figure 4:

Performance at recognition memory (d’) as a function of sex and age, controlling for other predictors separately for (A) real objects and (B) novel objects.

Following up on the significant two-way interaction for real objects (Figure 4, Panel A), we found that increasing age had a stronger negative effect on males’ than females’ recognition performance on real objects, though both age effects were significant (males: b = −0.0292, SE = 0.0030, t = −9.80, p < .00001; females: b = −0.0129, SE = 0.0043, t = −3.02, p = .003). Moreover, at the minimum end of the age range (58 years), the males and females did not differ on their predicted d’ scores for real objects (b = 0.0941, SE = 0.0622, t = 1.51, p = .131), whereas at the maximum end of the age range (98), females had significantly higher predicted d’ scores for real objects than males (b = −0.5572, SE = 0.1683, t = −3.31, p = .001). The lowest age at which males and females showed a significant difference on their predicted d’ scores for real objects was 70 years (b = −0.1013, SE = 0.0453, t = −2.23, p = .026).

These patterns were robust, with the exact same pattern of significance (i.e., p < .05) for main effects and interactions (as shown in Table 3) obtained in a range of different alternative (sensitivity) analyses. First, the same pattern was obtained with two types of ‘main effect’ coding, that is, regardless of whether categorical predictors were assigned sum-coded contrasts or were converted to numerical variables and then mean-centered; see Methods. Second, it might be argued that the non-significant higher-order interactions should not be retained in the regression model, since they might reduce the statistical power of lower-order effects. However, when the four-way interaction (Age × Sex × Education × Object type) and the three non-significant three-way interactions (Age × Sex × Education, Age × Education × Object type, Sex × Education × Object type) were removed from the model, again the same pattern of significance was obtained for the remaining effects. Third, it is possible that handedness, which was not included as a covariate, might bias the results. However, when handedness (coded as a continuous variable from − 100 to 100; see Methods and Oldfield, 1971 was included as a covariate, again the same pattern was obtained.

Finally, we tested for potential non-linearities in the relation between age and d’ scores, by including an additional quadratic term for age in the mixed-effects regression model. In order to eliminate the correlation between the quadratic and linear terms of age, the quadratic term was included in the model as an orthogonal polynomial. A likelihood ratio test revealed that this model did not have a significantly higher goodness-of-fit than the linear model presented above (χ²(1) = 0.6329, p = .426); that is, the quadratic term for age failed to reach significance (b = −0.4619, SE = 0.5842, t = −0.79, p = .430). Because the inclusion of a quadratic term of age failed to improve model fit, cubic and other higher-order polynomials were not tested for inclusion.

Incidental encoding

This paper investigates nonverbal declarative memory in aging and the factors that moderate it. Hence our analyses have focused on performance on the recognition phase of the recognition memory task. However, as described above, this task also includes an incidental encoding phase, in which participants had to judge whether the stimulus they saw was real or novel (made-up). Performance at the encoding phase (i.e., discriminating between real and novel objects) for males and females is shown in Table 4. Note that data for the encoding phase were missing for 8 participants.

Table 4:

Performance in the incidental encoding phase

	d’
Males	1.53 (0.69)
Females	1.43 (0.65)
Total	1.48 (0.68)

Note. d’ refers to the discriminability between objects as real versus novel.

We computed linear regression models to assess whether any of the variables affecting recognition accuracy might have influenced encoding accuracy. These analyses were analogous to the ones that were performed for recognition accuracy with the exception that object type, which was a within-subjects factor for the recognition accuracy analysis, was not a factor in the encoding analysis (since d’ is computed from the decision of real vs. novel). Due to the absence of a within-subjects factor (i.e., object type), the encoding analyses were run as linear regressions rather than mixed-effects models. The results of the encoding analyses are shown in Table 5.

Table 5:

Results from the linear regression model on encoding d’ scores

Coefficients:	b	SE	t	p
Intercept (estimated grand mean)	1.4700	0.0312	47.09	<.00001	***
Age	−0.0191	0.0038	−5.03	<.00001	***
Sexa	0.0601	0.0624	0.96	.336
Education	0.0301	0.0067	4.52	<.00001	***
Age × Sex	−0.0054	0.0076	−0.71	.479
Age × Education	0.0005	0.0007	0.67	.505
Sex × Education	<0.0001	0.0133	<0.01	.998
Age × Sex × Education	−0.0003	0.0015	−0.18	.858

Notes.

^***p < .00001.

P-values were obtained from t-tests with 688 degrees of freedom. All continuous predictors were mean-centered; all categorical predictors were assigned sum-coded contrasts (see Methods).

^aSex is coded as −0.5 for females and 0.5 for males.

The analysis revealed main effects of age and of education, and no significant interactions. In particular, the results indicate that discriminating between the real and novel objects was negatively impacted by age but positively impacted by education.

These differences in incidental encoding could potentially affect performance at the recognition phase – though note that because the encoding was incidental, the relation between success at this phase (distinguishing real and novel objects) and learning the material that is tested later in recognition is not entirely clear. Nevertheless, we reran the recognition memory analyses with each participant’s encoding-phase d’ as a covariate (omitting the 8 participants for whom encoding data were missing). Although a significant main effect of encoding d’ indicated that participants with better discrimination abilities during the encoding task had better recognition accuracy (b = 0.1749, SE = 0.0201, t = 8.72, p < .0001), crucially the inclusion of this covariate did not affect the significance of the other effects: these showed the exact same pattern of significance (i.e., ps < .05) as reported above, with the exception of sex, which showed a significant main effect (b = −0.0664, SE = 0.0329, t = −2.02, p = .045) whereas in the original analysis it was marginally significant (Table 3). In other words, while better encoding was associated with better recognition memory performance, controlling for encoding differences did not affect the pattern of our results.

Discussion

We investigated the influence of age, education, sex, and object type (real vs. novel objects) on nonverbal declarative memory in a representative sample of 704 middle-aged to older Taiwanese adults, aged 58–98 years, with 0–17 years of education. Declarative memory was assessed with a recognition memory paradigm, following incidental encoding. Using mixed-effects regression models, we assessed the unique effects (that is, beyond the influence of all other factors) of the four predictors and their interactions on the outcome variable (recognition memory).

Interpretation of results

Here we discuss various potential explanatory accounts for the findings. First, a main effect of age indicated a decline in nonverbal declarative learning abilities between middle and old adulthood, that is, between 58 and 98 years of age. This effect was robust, holding in all follow-up analyses to the interaction between age, sex, and real/novel, that is, for both males and females for both real and novel objects. Moreover, the effect of age held not only in the primary analyses, but also across the alternative analyses. This result is consistent with prior research that has found reliable age-related declines in verbal declarative memory (see Introduction), and adds to the few previous studies of nonverbal declarative memory (Gale et al., 2007; Park et al., 2002; Pauls et al., 2013) in demonstrating aging effects in this aspect of learning as well. The findings underscore the presence of aging declines in (nonverbal) declarative memory even in recognition memory tasks (which appear to yield smaller and/or less consistent declines than recall tasks probing declarative memory; Craik & McDowd, 1987; Danckert & Craik, 2013; Schonfield & Robertson, 1966; Zelinski & Burnight, 1997) and with incidental encoding (which has been less examined in aging studies than tasks with intentional encoding; see Introduction). We found evidence for only linear, not non-linear, aging declines in our study. This result is consistent with the findings of Park et al. (2002), who reported only linear nonverbal declarative memory declines from young to old adulthood. (Non-linearity was not examined by Pauls et al., 2013, or Gale et al., 2007.) In contrast, some previous work on verbal declarative memory (especially but not only from longitudinal studies) has suggested non-linear aging effects across the adult lifespan (Nyberg et al., 2012; Rönnlund, Nyberg, Bäckman, & Nilsson, 2005; Salthouse, 1998). Note however that non-linear declines across adulthood do not necessarily imply non-linear effects within older ages; indeed, previous work has suggested steeper declines in older adulthood, which could be linear. In other words, the participants in our study may well have shown shallower declines between young and middle adulthood, which would have resulted in a non-linear decline across the lifespan.

The observed age-related declines in our study as well as in previous verbal and nonverbal declarative memory tasks are likely to be at least partially explained by neuroanatomical declines of the hippocampal and other medial temporal lobe structures (e.g., perirhinal cortex) that underlie various aspects of (verbal and nonverbal) declarative memory (also see below). Not only have previous studies reported (linear and non-linear) declines in metrics of the gray and white matter of these structures (Allen, Bruss, Brown, & Damasio, 2005; Giorgio et al., 2010; Jack et al., 2015; Pelletier et al., 2013; Raz, 2000; Raz, Ghisletta, Rodrigue, Kennedy, & Lindenberger, 2010; Šimic, Kostovic, Winblad, & Bogdanovi, 1997; Yang, Goh, Annabel Chen, & Qiu, 2013), including hippocampal volumes, but these neuroanatomical declines have been tied to performance decreases in declarative memory tasks, including in recognition memory (Diana, Yonelinas, & Ranganath, 2007; Douet & Chang, 2015; Fletcher et al., 2013; Hackert et al., 2002; Koen & Yonelinas, 2014; Metzler-Baddeley, Jones, Belaroussi, Aggleton, & O’Sullivan, 2011; Raz, 2000; Reiman et al., 1998; Stadlbauer, Salomonowitz, Strunk, Hammen, & Ganslandt, 2008; Troyer et al., 2012; Wolk, Dunfee, Dickerson, Aizenstein, & Dekosky, 2011; Yasmin et al., 2009; Yonelinas et al., 2007).

Whereas list learning and explicit encoding paradigms appear to involve working memory (see Introduction), the incidental encoding paradigm used here is designed to minimize this confounding factor, decreasing the likelihood that the observed aging declines were explained by declines in working memory (Bopp & Verhaeghen, 2018; Pliatsikas et al., 2018). Thus, the present study suggests that nonverbal declarative memory itself declines with aging, independent of working memory influences.

Second, a main effect of education indicated that declarative memory performance in this sample of older adults improved with an increasing number of years of education during childhood and young adulthood. Despite interactions, this effect was found for both sexes and for both object types, and held robustly across analyses. While several studies have reported positive effects of education on declarative learning in older adults for verbal tasks (Fuh et al., 2006; Giogkaraki et al., 2013; Lachman et al., 2010; Lövdén et al., 2000; Magalhães & Hamdan, 2010; Malloy-Diniz et al., 2007; Portin et al., 1995; Rönnlund & Nilsson, 2008), we believe that this is the first study to extend this finding to (incidental or intentional) nonverbal learning.

The positive association between the number of early years of education and declarative memory performance in older adults may be explained by various factors (for findings related to effects of later-life education, see Kremen et al., 2019). First of all, education during one’s younger years may directly or indirectly confer downstream benefits for declarative memory. It has been suggested that, in general, cognitive stimulation offered by schooling may – in particular during an early critical period – promote the development of neural substrates underlying cognitive abilities (Garlick, 2002; Rueda, Rothbart, Mccandliss, Saccomanno, & Posner, 2005). Moreover, some evidence suggests that academic studying in young adults leads to larger hippocampi (Draganski et al., 2006), which are linked to better declarative memory (Protopopescu et al., 2008; Schofield et al., 2009). In Draganski et al. (2006) these hippocampal increases were also found several months after studying, suggesting that such changes can persist at least to some extent past the point of educational stimulation. Additionally, greater education often leads in turn to occupational outcomes that have more cognitively stimulating environments. Importantly, such occupations have been shown to improve cognitive functioning (including declarative memory) later in life (Baldivia, Andrade, & Bueno, 2008; Jorm et al., 1998; Karp et al., 2009; Stern, 2002; see C.-Y. Li, Wu, & Sung, 2002, for data from Taiwan).

Note that it has been suggested that education can lead to better performance in explicit memory tests by improving one’s ability to organize the stimuli to be recalled (Cole, Gay, Glick, & Sharp, 1971; Sharp et al., 1979) and/or by teaching strategies such as recoding, maintenance rehearsal, and chunking (Fahrmeier, 1975; Rogoff, 1981; Sharp et al., 1979; Wagner, 1978). However, such accounts would be unlikely to explain the education effects in our incidental encoding paradigm.

Alternative explanations for the main effect of education on declarative memory are nevertheless plausible. On one view, the positive relationship between education and declarative memory performance may be due not (only) to education improving downstream learning, but, conversely, to better declarative learning abilities in early years leading to greater educational attainment, that is, to more years of schooling. Note that this account would require declarative memory abilities in childhood and adolescence (when education occurred) to correlate with declarative memory abilities in older adulthood (when it was tested). Although there appears to be some lifespan stability in some cognitive functions (e.g., as measured by IQ; Deary, Pattie, & Starr, 2013), to our knowledge this remains to be ascertained for (nonverbal) declarative memory. Moreover, this account may be more likely to hold in some sociopolitical contexts than in others. For example, the account seems plausible in many Western countries these days, in which wide access to education can result in more gifted children attending school for a longer period (e.g., in certain European countries, where one may graduate from secondary school at different grades, with better students leaving school later). It is less clear whether this might have held in Taiwan during much of the time when the participants in the present study went to school (about 1920–1960), especially since many Taiwanese people were effectively banned from receiving secondary (let alone tertiary) education during the Japanese colonial period (indeed, even primary education became mandatory in Taiwan only after 1945) (Chou & Ching, 2012; Zimmer, Liu, Hermalln, & Chuang, 1998). Alternatively, it is also possible that one or more additional factors, such as motivation or socio-economic status, could result in both better declarative memory performance and higher levels of education. In line with this view, socio-economic status correlates with both the level of educational attainment (White, 1982) and declarative memory abilities (Hackman & Farah, 2009; Noble, Houston, Kan, & Sowell, 2012). Thus, the positive association between education and declarative memory must be interpreted with caution.

Third, we found a main effect of object type, with participants recognizing real objects more accurately than novel objects. Although interactions with real/novel were observed, the real object advantage held quite broadly: across all levels of education (see Figure 3), for females of all ages, and for males except at the upper end of the age continuum (see Figure 4), at which point the recognition of real and novel objects did not differ (see subsection ‘Implications and conclusion’ below). This pattern of better recognition memory on real than novel objects is consistent with previous research using the same experimental paradigm in children (Hedenius et al., 2013; Lukács et al., 2017) as well as with recognition memory findings in young adults for nameable versus non-nameable odors (Öberg et al., 2002), for nameable objects versus abstract shapes (McGivern et al., 1998), and for famous versus non-famous faces (La Corte, Dalla Barba, Lemaréchal, Garnero, & George, 2012; Liu, Grady, & Moscovitch, 2017; Zion-Golumbic, Kutas, & Bentin, 2009). Similar patterns have been reported for older adults, who indeed may show a particular benefit, as compared to younger adults, for items supported by existing knowledge in such tasks (Badham et al., 2016; Castel, 2005; Umanath & Marsh, 2014).

One likely reason for better declarative memory on real than novel objects is that the former are strongly supported by existing (semantic) associations, while the latter are not. (Indeed, this distinction was a factor influencing the design of our task.) This is consistent with the view that, more generally, pre-existing semantic knowledge benefits the creation of new episodic memories (Tulving, 1983, 1995), perhaps especially in older adults, whose semantic knowledge is usually largely preserved (Badham et al. 2016; Umanath and Marsh 2014). Additionally, real objects are associated with existing verbal labels, while novel objects are not (and those in the study were selected for low nameability), a difference that could also lead to improved performance on the former, for example, by remembering the verbal label.

Fourth, an interaction between sex and education reflected the finding that increasing years of education benefited declarative memory more for women than men (across all ages in our sample). Moreover, while at zero years of education there were no sex differences in declarative memory, a female advantage emerged at 9 years of schooling, and remained at higher levels of education. This pattern is consistent with previous studies, which have predominantly suggested a female advantage for nonverbal declarative memory (Alexander et al., 2002; Eals & Silverman, 1994; Herlitz & Yonker, 2002; Lewin & Herlitz, 2002; Lewin et al., 2001; McBurney et al., 1997; McGivern et al., 1997, 1998), despite a somewhat mixed pattern in the literature (see Introduction). One possibility is that this commonly observed female advantage may have been due in part to a relatively high level of education in these previous studies’ participants, since most if not all of those studies tested adults in Western countries, most of whom are likely to have had at least 9 years of education (indeed, most of the studies targeted university students). Consistent with this view, a recent large-scale study of verbal declarative memory in older adults (aged 50 and older) found a robust female advantage in countries with high levels of female education and employment, such as Sweden and the U.S., but not in countries with lower levels of female education and employment, such as India or Ghana (Bonsang et al., 2017). This latter finding underscores the view that investigating cognitive effects in more rarely examined populations, including in non-Western countries, can be highly informative. To the best of our knowledge, the present study is the first to report an interaction between sex and education for (verbal or nonverbal) declarative memory in older adults. We are aware of only one study that reported this interaction as non-significant in an investigation of declarative memory in older adults (Portin et al., 1995), though it is unclear whether other aging studies did not test for such an effect or did not report it if it was not significant. Nevertheless, one study of declarative memory in children found a similar sex by education effect as that observed here (Stevenson, Chen, & Booth, 1990). Finally, a recent study of working memory in older adults also reported a similar sex-by-education interaction (Pliatsikas et al., 2018), suggesting that this effect may not be specific to declarative memory, but may hold for other aspects of cognition as well.

A plausible account for the observed sex-by-education interaction is that, all else being equal, females generally show a declarative memory advantage, but at lower levels of education this advantage is offset by other factors. Indeed, prior evidence suggests a broad female advantage across adulthood at verbal declarative memory; this pattern may also hold for nonverbal declarative memory, though the data are less clear (see previous paragraph and Introduction). In the present study, women showed a marginally significant overall advantage compared to men (i.e., across all levels of education and other factors), which reached significance when incidental encoding performance was covaried out. We suggest that such general female advantages at declarative memory may be countered at lower levels of education by factors that may elevate men’s performance relative to women’s. In particular, at lower levels of education, men in Taiwan have been more likely to be employed than women, who have instead tended to stay at home (Thornton, Chang, & Sun, 1984; Tsai, Gates, & Chiu, 1994). Homemaking has been associated with greater downstream cognitive and other health declines in aging, including in declarative memory, as compared to employment, perhaps due to lower levels of cognitive and social stimulation (Leist, Glymour, Mackenbach, Van Lenthe, & Avendano, 2013; Ross & Mirowsky, 1995). Thus, at lower but not higher levels of education, such factors could outweigh the apparent overall female advantages. Indeed, this view is consistent with the findings of Bonsang et al. (2017) suggesting a robust female verbal declarative memory advantage particularly in countries with higher levels of education and cognitive stimulation (and greater gender equality more generally).

Fifth, the analyses yielded a robust interaction between education and object type: increasing years of education conferred greater benefits for remembering real than novel objects, even though real items were better remembered than novel items across the entire spectrum of education. We are not aware of any other studies reporting such an effect (or the lack thereof).

This interaction may be due to various factors. One possibility is that since higher education might be expected to lead to more knowledge and richer semantic networks, the greater dependence of real than of novel objects on such networks (see above) results in education benefiting the former more than the latter. Similarly, more education may be important for enhancing knowledge of the verbal labels of real objects, particularly for some less common items (e.g., cymbals or aqueduct), increasing the likelihood that verbal labels could facilitate declarative memory for those items (see above). These explanations are consistent with evidence suggesting that higher education alleviates age-related declines in tests of semantic knowledge (Christensen et al., 1997). Alternatively (or in addition), as discussed above, at a neurobiological level education may lead to hippocampal volume increases, and thus improvements in aspects of declarative memory that depend especially on the hippocampus. Some evidence suggests that the hippocampus is more engaged in the recognition memory of known than novel stimuli (Barense, Henson, & Graham, 2011; Leveroni et al., 2000; Liu et al., 2017), suggesting that education-moderated hippocampal increases may benefit real more than novel objects. Indeed, this finding is consistent with a large literature indicating that, within the medial temporal lobe, the hippocampus is particularly important for learning associations (as opposed to individual items) and/or for retrieving learned information through ‘recollection’ (retrieval of qualitative information about a prior episode, as opposed to the experience of a contextual sense of familiarity) (Davachi, 2006; Eichenbaum, Yonelinas, & Ranganath, 2007): after all, real items contain more associations to prior knowledge than novel items; moreover, evidence suggests that items with many such associations are more likely to be recollected than those with few associations (Kan, Alexander, & Verfaellie, 2009; Poppenk, Köhler, & Moscovitch, 2010; Reder et al., 2013).

Sixth and finally, we found an interaction between age, sex, and object type (real/novel), which reflected different age-by-sex patterns for real and novel objects. Recognition memory for real objects showed an age-by-sex interaction, with steeper age declines for males than females. Additionally, for real objects there was no sex difference at the minimum age in the sample (58 years), while a significant female advantage emerged at age 70, and continued at later ages. In contrast, novel objects showed a shallow decline that did not differ between the sexes. We are not aware of any previous declarative memory studies examining interactions between age, sex, and real/novel objects.

The observed pattern could be explained as follows. The preponderance of evidence seems to suggest that hippocampal volumes decrease during aging more in males than females (Blatter et al., 1995; Christiansen, Larsson, Thomsen, & Wieslander, 1994; Golomb et al., 1993; Jack et al., 2015; Pruessner, Collins, & Evans, 2001; Raz, Gunning-Dixon, et al., 2004; but see D. G. M. Murphy et al., 1996). Since the hippocampus may be particularly important for real objects in declarative memory tasks (see above), these items should show a greater age-related decline for males than females, as was observed. The lack of a sex difference in our task prior to age 70, but a female advantage from this age onwards, is intriguingly similar to the hippocampal patterns observed in a recent large-scale study, which found that males and females have similar hippocampal volumes from young adulthood up to age 60, at which point the sexes diverge, with males showing a greater decline than females (Jack et al., 2015). Broadly consistent with this finding, most studies examining sex differences in hippocampal volumes across adulthood report either a female advantage or no sex differences (Persson et al., 2014). However, the relation between hippocampal volumes and declarative memory abilities is still somewhat uncertain. On the one hand, the two appear to be correlated (Protopopescu et al., 2008; Schofield et al., 2009). On the other hand, the possible absence of larger female hippocampi earlier in adulthood, together with the widely reported female advantages across adulthood at verbal declarative memory abilities, as well as the preponderance of evidence suggesting a female advantage at nonverbal declarative memory in adulthood (see Introduction), suggests that the relation between hippocampal volumes and declarative memory is not straightforward.

In contrast, since novel objects are less likely to depend on the hippocampus (see above), they may also be less likely to show the sex difference pattern found for real objects. Indeed, no sex differences were observed for novel objects in the present study. Moreover, since the novel objects are not linked to (m)any prior associations, and thus may be remembered as isolated items and retrieved via familiarity rather than recollection (see above), recognition memory for these items might have depended heavily on perirhinal cortex, since this key medial temporal lobe structure appears to play an important role in familiarity-based memory for individual items (Bowles et al., 2007; Brown & Aggleton, 2001; Davachi, 2006). Intriguingly, perirhinal cortex volumes show less reliable aging declines than hippocampal volumes (Koen & Yonelinas, 2014; Raz, Rodrigue, Head, Kennedy, & Acker, 2004; Rodrigue & Raz, 2004), consistent with the shallow declines observed here for novel objects. Additionally, declarative memory for items seems to decline less in aging than for associations (Old & Naveh-Benjamin, 2008). Moreover, perirhinal cortical volumes do not show steeper aging declines in males than females or smaller volumes in males (Insausti et al., 1998; Raz, Gunning-Dixon, et al., 2004), consistent with the absence of a female advantage for novel objects in the present study. The possibility that recognition memory for the novel items depended particularly on perirhinal cortex is underscored by evidence suggesting that recognition memory for novel information, or information with fewer semantic associations, appears to rely more on familiarity than recollection (Belleville, Ménard, & Lepage, 2011; Ozubko & Yonelinas, 2012), since familiarity relies especially on perirhinal cortex (Eichenbaum et al., 2007; Yonelinas, Aly, Wang, & Koen, 2010). However, we emphasize caution in accepting the account given here, and further studies directly examining these issues are warranted.

Limitations

Our study also had limitations. First, as in the majority of previous work examining effects of aging on declarative memory, the study design treated age as a between-subjects factor, comparing participants of different ages with one another. The underlying assumption of such cross-sectional studies is that participants (are selected to) differ as little as possible from one another, except in their ages (and age-related changes to cognition that might be of interest). Other factors that may differ across the participants should be controlled for, as was done in the present study for education. However, there may of course have been other variables that our participants also differed on, such as physiological health (apart from the brain-related conditions that constituted exclusionary factors), socio-economic status, motivation, or the quality of the education they received, any or all of which could have influenced our results. Future studies could control for such factors, or could follow a combined cross-sectional longitudinal design (see e.g. Nyberg et al., 2012)—though note that longitudinal designs might be difficult to implement when investigating incidental encoding, as participants may begin to use strategies after the first session.

Second, the study attempted to minimize the influence of working memory in encoding, especially since working memory can be affected by both sex and education (see Introduction). Most previous research on age effects on declarative memory has employed intentional encoding paradigms, moreover often in list learning tasks, both of which appear to rely heavily on working memory (see Introduction, and Lum et al., 2015). In contrast, we used an incidental encoding (non-list learning) paradigm that was designed to be less susceptible to working memory confounds and thus may constitute a more process-pure way of assessing declarative memory (Hedenius et al., 2013; Lukacs et al., 2017). (Note that, consistent with this notion of process-purity, this paper takes a systems-based view of declarative memory, as the learning and memory rooted in the medial temporal lobe; however, more process- or task-based views of memory have also been posited (e.g., Craik & Grady, 2002), and may also be considered in future work.) Nevertheless, it is possible that working memory might have played a role in the observed effects. Future studies could address this issue by covarying out this factor.

Implications and conclusion

The findings have a number of implications and suggest various lines of future research. First, the study suggests that nonverbal declarative memory indeed weakens with age, even when tested in a recognition memory paradigm with incidental encoding. Thus, aging declines in declarative memory do not appear to be restricted to verbal information, nor to recall paradigms, explicit encoding, or list learning paradigms. The evidence therefore suggests that declarative memory itself seems to decline during aging, apparently independent of declines in working memory, verbal abilities, or other factors.

Second, the findings emphasize the importance of early education – especially for girls –for later cognitive functioning, in particular for declarative memory. The sex-by-education interaction also elucidates the nature of sex differences in declarative memory. As discussed above, based on evidence from this and other studies it seems plausible that the female advantage at declarative memory is found mainly at higher levels of education, in particular above about 9 years of schooling. This in turn highlights the importance of testing diverse populations, including in non-Western countries, in which education levels can vary substantially. The present study builds on Bonsang et al. (2017), who found a robust female advantage at verbal declarative memory in countries with higher education but not in countries with lower education. Whereas in that study countries being compared likely differed in factors other than female education, in the present study – which examined a relatively homogeneous population within a single country – potential contributions of such other factors were minimized, underscoring the likelihood that female advantages at declarative memory may emerge only at higher levels of education for both sexes.

Third, the real/novel object differences, and their interaction with education, are of interest. The results indicate that older adults (like children and younger adults; see above) are more likely to remember recent encounters related to existing information than those related to new information. Together with our finding that education improves the former more than the latter, this highlights the notion that learning begets learning. That is, it appears as though the greater one’s existing knowledge (from more prior learning in declarative memory), the better one’s declarative memory, since it can rely more on that knowledge. This may be thought of as a form of the Matthew Principle (“For whosoever hath, to him shall be given, and he shall have more abundance: but whosoever hath not, from him shall be taken away even that he hath.”) (Merton, 1968). Note that the males’ steeper aging decline in recognition memory for real than novel objects (e.g., see Figure 4) does not obviate this principle, since even at very high ages the males were no worse at real than novel objects (b = 0.0304, SE = 0.1047, t = 0.29, p = .772). The real/novel findings thus indicate that the distinction between real and novel items is important in the study of declarative memory. We emphasize that it remains unclear to what extent the real/novel distinction is confounded with the verbal(izable)/nonverbal(izable) distinction that has been the focus of many previous studies, since real objects are generally verbalizable, and most verbal material in verbal declarative memory studies (i.e., real words) represents real world information. This seems like an important topic for further research.

The findings also have real-world and translational implications. The study suggests the possibility of partially countering or delaying declarative memory declines in aging with greater early education as well as further learning and cognitive stimulation. Indeed, based on the (unstandardized) regression estimates (b) in the current study, over all participants each additional year of education had about as much positive impact on nonverbal memory abilities in older adults as the negative impact of three years of aging (b for the main effect of education is about three times larger than b for the main effect of age; 0.0418 vs. −0.0144, see Table 3). Moreover, whereas for males each additional year of education was associated with downstream memory benefits (b = 0.0343, SE = 0.0049) equivalent to the negative impact of about two years of aging (b = −0.0191, SE = 0.0024), for females the memory increases linked to each year of education (b = 0.0494, SE = 0.0054) corresponded to the losses from five years of aging (b = −0.0096, SE = 0.0034) (regression estimates over both object types). For example (and strikingly), according to the regression model, in the present study women with 16 years of education (corresponding to a typical bachelor’s degree in Taiwan as well as the U.S.) would have similar overall nonverbal declarative memory abilities at age 80 (d’ = 1.16, 95% CI [0.59, 1.73]) as women with 12 years of education would have at age 60 (d’ = 1.12, [0.59, 1.65]). These women would also have similar nonverbal declarative memory abilities specifically for real objects (d’ of 1.64, [1.07, 2.21] and 1.55, [1.01, 2.09], respectively) and novel objects (d’ of 0.68, [0.11, 1.25] and 0.69, [0.16, 1.22], respectively). Testing the replication of this finding is clearly important (e.g., with larger sample sizes, especially with older women at higher education levels; see Table 1), and some caution is warranted regarding the causal direction between early-life education and later-life declarative memory abilities (see above). Nevertheless, the findings suggest that early-life education may help counter nonverbal declarative memory declines in aging, especially for women. The observed pattern may provide an argument for further efforts to increase access to education, cognitive stimulation, and workforce participation, particularly for women in societies where education between the sexes is still unequal.

Thus, education may not only delay the onset of dementia/probable Alzheimer’s disease (Hall et al., 2009; Mortimer, Snowdon, & Markesbery, 2003; Reuser, Willekens, & Bonneux, 2011), in which declarative memory declines are a hallmark symptom, but also help counter declarative memory declines in healthy aging. The findings also suggest the possibility that onset delays in the diagnosis of Alzheimer’s disease that have been associated with higher education might in fact be partly due to improved declarative memory (and other cognitive functions, such as naming) specifically for real/verbal items, which are not only commonly encountered in everyday life, but are also common in neuropsychological assessments of cognitive impairment. Although this hypothesis is speculative, the findings from the current study suggest that it may warrant further examination.

In conclusion, this study elucidates various aspects of nonverbal declarative memory in aging. The study underscores the importance of examining education and sex, as well as the real/novel distinction and other moderating variables, in studies of declarative memory. It also shows the utility of investigating non-Western populations. The findings may lead to new lines of research regarding psychological, neurobiological, and social factors that could affect declarative memory in aging. Finally, the results have potentially important real-world and translational implications.