PreprintPDF Available

Affect across adulthood: Evidence from English, Dutch and Spanish

Authors:
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Emotions play a fundamental role in language learning, use and processing. Words denoting positivity account for a larger part of the lexicon than words denoting negativity, and they also tend to be used more frequently, a phenomenon known as positivity bias. However, language experience changes over an individual’s lifetime making the examination of the emotion-laden lexicon an important topic not only across the life span but also across languages. Furthermore, existing theories predict a range of different age-related trajectories in processing valenced words. The present study pits all of these predictions against written productions (Facebook status updates from over 20,000 users) and behavioral data from three publicly available mega-studies on different languages, namely English, Dutch, and Spanish across adulthood. The production data demonstrated an increase in positive word types and tokens with advancing age. In terms of comprehension, the results showed a uniform and consistent effect of valence across languages and cohorts based on data from a visual word recognition task. The difference in RTs to very positive and very negative words declined with age, with responses to positive words slowing down more strongly with age than responses to negative words. We argue that the results stem from lifelong learning and emotion regulation: advancing age is accompanied by an increased type frequency of positive words in language production, which is mirrored as a discrimination penalty in comprehension. To our knowledge, this is the first study to simultaneously target both language production and comprehension across adulthood and in a cross-linguistic perspective.
Content may be subject to copyright.
Running head: AFFECT ACROSS ADULTHOOD 1
Affect across adulthood: Evidence from English, Dutch and Spanish
Aki-Juhani Kyröläinen
McMaster University and Brock University
Emmanuel Keuleers
Tilburg University
Paweł Mandera
Ghent University
Marc Brysbaert
Ghent University
Victor Kuperman
McMaster University
Author Note
DRAFT
AFFECT ACROSS ADULTHOOD 2
Correspondence concerning this article should be addressed to Aki-Juhani
Kyröläinen, Department of Linguistics and Languages, McMaster University, Togo Salmon
Hall 513, 1280 Main Street West, Hamilton, Ontario, Canada L8S 4M2. Tel: 905-525-9140, x.
24388, E-mail: akkyro@gmail.com. The first author’s contribution was supported by the
Social Sciences and Humanities Research Council of Canada Insight Development Grant,
430-2019-00851, (Dr. Kyröläinen, PI) and the first and the fifth authors’ contributions by
the Social Sciences and Humanities Research Council of Canada Partnered Research
Training Grant, 895-2016-1008, (Dr. Gary Libben, PI). The fifth author’s contribution was
also partially supported by the Ontario Early Researcher award (Kuperman, PI), the
Canada Research Chair (Tier 2; Kuperman, PI), and the CFI Leaders Opportunity Fund
(Kuperman, PI). We would also like to thank the two anonymous reviewers for their
comments and suggestions on our work.
AFFECT ACROSS ADULTHOOD 3
Abstract
Emotions play a fundamental role in language learning, use and processing. Words
denoting positivity account for a larger part of the lexicon than words denoting negativity,
and they also tend to be used more frequently, a phenomenon known as positivity bias.
However, language experience changes over an individual’s lifetime making the
examination of the emotion-laden lexicon an important topic not only across the life span
but also across languages. Furthermore, existing theories predict a range of different
age-related trajectories in processing valenced words. The present study pits all of these
predictions against written productions (Facebook status updates from over 20,000 users)
and behavioral data from three publicly available mega-studies on different languages,
namely English, Dutch, and Spanish across adulthood. The production data demonstrated
an increase in positive word types and tokens with advancing age. In terms of
comprehension, the results showed a uniform and consistent effect of valence across
languages and cohorts based on data from a visual word recognition task. The difference
in RTs to very positive and very negative words declined with age, with responses to
positive words slowing down more strongly with age than responses to negative words.
We argue that the results stem from lifelong learning and emotion regulation: advancing
age is accompanied by an increased type frequency of positive words in language
production, which is mirrored as a discrimination penalty in comprehension. To our
knowledge, this is the first study to simultaneously target both language production and
comprehension across adulthood and in a cross-linguistic perspective.
Keywords: aging, valence, emotion, language production, language comprehension,
word recognition
AFFECT ACROSS ADULTHOOD 4
Affect across adulthood: Evidence from English, Dutch and Spanish
Introduction
Emotions leave a noticeable footprint on language learning, use and processing (for
reviews, see e.g., Borghi et al., 2017; Citron, 2012; Ochsner, 2000). To give a few examples,
emotionally loaded – i.e., positive or negative – words (e.g., vacation,illness) tend to be
learned better and recognized or produced faster than their neutral counterparts (e.g.,
table,street) (Kuperman, Estes, Brysbaert, & Warriner, 2014, and references therein). Also,
psychological positivity or valence is a primary vehicle of learning abstract words, i.e.,
words that do not have tangible referents in the material world are learned and recognized
faster if they have strongly valenced meanings (Kousta, Vigliocco, Vinson, Andrews, &
Campo, 2011; Ponari, Norbury, & Vigliocco, 2018; Vigliocco, Ponari, & Norbury, 2018).
An influential view explains the substantial impact of affect on language by arguing
that subjective valence of a stimulus engages one of two motivational subsystems: one
geared towards objects that facilitate survival and another geared towards responding to
threat and danger (e.g., Bradley & Lang, 2000; Wurm, 2007). Yet the affective biases, the
relative need for survival and avoidance of danger, emotion regulation and language
experience change over an individual’s lifetime. The dynamic nature of human emotion
and cognition suggests that age modulates how language reflects affect and how affect is
perceived through language. As we discuss below, the current state of knowledge about
the nature of the interaction between aging, affect and language use is incomplete. This
paper harnesses large-scale collections of behavioral data across adulthood to examine the
distribution of affect in language production over age and its consequences for written
word recognition. Below we identify four empirical findings that are both robust in the
literature and pertinent to the topic of our interest and a broader context. We then briefly
summarize our (non-exhaustive) examination of relevant theoretical accounts, formulate
their predictions for production and comprehension of affective lexicon over age, and
outline how our studies test these predictions. In short, this study examines the
AFFECT ACROSS ADULTHOOD 5
production and comprehension of affective words in adulthood.
Key empirical findings
The question of aging and affective language can only be studied in a broader
context of the interaction between age and language in general (see Burke & Shafto, 2008;
Shafto & Tyler, 2014, for discussion). Four relevant findings appear to have strong
empirical support in the current literature and serve as explicanda for existing theoretical
accounts and our own theorizing below. Thus, (i) vocabulary knowledge displays a steady
growth across adulthood (Brysbaert, Stevens, Mandera, & Keuleers, 2016; Hartshorne &
Germine, 2015; Keuleers, Stevens, Mandera, & Brysbaert, 2015; Ramscar, Hendrix, Shaoul,
Milin, & Baayen, 2014; Verhaeghen, 2003). Also, (ii) older individuals display longer
response latencies in tasks related to language production and comprehension (see
Ratcliff, Thapar, Gomez, & McKoon, 2004; Rayner, Reichle, Stroud, Williams, & Pollatsek,
2006; Salthouse, 1996) and lower reading speed (see Kemper & Liu, 2007; Rayner et al.,
2006; Solan, Feldman, & Tujak, 1995; Whitford & Titone, 2017). A third finding (iii) is
that even the structure of lexicon is shaped by emotion, as evidenced by the
cross-linguistic positivity bias. There are more distinctly positive words (i.e., words
scoring above the middle of the valence scale) in vocabularies of multiple languages, and
these words are used more often than negative ones (Boucher & Osgood, 1969; Dodds
et al., 2015; Warriner & Kuperman, 2015).
Finally, (iv) aging is robustly associated with a changes in cognitive selectivity,
specifically an increasing preference for positive rather than negative stimuli (see
Carstensen & DeLiema, 2018; Reed, Chan, & Mikels, 2014, for an overview). This effect
has been found in a wide array of different cognitive domains ranging from attention,
working memory, autobiographical memory and decision making, among others (see
Carstensen & Turk-Charles, 1994; Carstensen, 2006; Mather & Carstensen, 2005; Reed
et al., 2014; Ruffman, Henry, Livingstone, & Phillips, 2008; Samanez-Larkin & Knutson,
AFFECT ACROSS ADULTHOOD 6
2015). As reviewed in Mather and Carstensen (2005), older individuals experience or
express fewer negative emotions. They tend to have more positive memories about their
past in autobiographical memory task; in memory tasks, they remember positive verbal or
pictorial stimuli better and they show a higher rate of forgetting the negative ones; and in
attention tasks, they focus on positive stimuli more than their younger counterparts. In
terms of emotion-recognition ability, however, recent evidence suggests that the peak
performance of this ability remains fairly stable between ages 40 and 60 years (Hartshorne
& Germine, 2015).
Theoretical accounts
To our knowledge, no single relevant account has targeted both language production
and comprehension over the lifespan, nor have been findings (i)-(iv) considered jointly.
Our review below focuses then on formulating predictions that existing accounts would
generate had they had findings (i)-(iv) as an input: we also mention theoretical proposals
that are by design confined to only some facets of the interaction between language, aging
and affect.
The “emotion regulation” account proposes that for the increased positivity in aging
individuals (finding iii) is due to the accumulating improvement in control over emotions,
which parallels biological maturation (see reviews by Charles & Carstensen, 2007; Urry,
2016). Older individuals are less prone to feeling negative emotions, engage in more
efficient coping strategies in the face of distress, and are more strategic in selecting
partners and experiences that lead to positive emotional states. Under this account, a
combination of findings (i), (iii) and (iv) leads to a prediction that the growing
vocabulary of older individuals will contain an ever-increasing proportion of positive
words, leading to age-related amplification of positivity bias. To our knowledge, this
prediction has not been tested yet: our Study 1 fills this lacuna.
As will become important for Studies 1 and 2 below, the amplification of the
AFFECT ACROSS ADULTHOOD 7
positivity bias predicted under the “emotion regulation” account may take different forms.
A useful distinction for operationalizing this change is one between word tokens (the
number of times a given word occurs in language production of an individual or a group)
and word types (the number of distinct words in one’s vocabulary). Metaphorically
speaking, word types represent the number of different tools in one’s linguistic tool-kit,
and their tokens represent how often each tool is used. As observed in Warriner and
Kuperman (2015) in a study of generic (non-age specific) corpus data, the positivity bias
emerges both in the type frequency (there are more positive words than negative ones)
and token frequency (positive words are used more often). If the preference for positive
stimuli in older participants in memory and attention tasks (e.g., Carstensen &
Turk-Charles, 1994; Kalenzaga, Lamidey, Ergis, Clarys, & Piolino, 2016; Reed &
Carstensen, 2012) does translate into an age-driven increase in positivity bias, it may
emerge in type frequency (older individuals use a larger number of distinct positive
words), in token frequency (older individuals use an average positive word more often),
or both. We demonstrate below that different scenarios for the change in positivity bias in
language production (showing in type-frequency only, token-frequency only, or both) give
rise to radically different predictions regarding recognition of words connotating affect:
these are examined in Study 2.
A second account that we consider can be labeled a “use it or lose it” account (see
Aschwanden et al., 2019; Hampshire, Sandrone, & Hellyer, 2019; Shimamura, Berry,
Mangels, Rusting, & Jurica, 1995). The prime example of this account is the very
well-documented frequency effect (see Brysbaert et al., 2011, for overview; Brysbaert,
Mandera, & Keuleers, 2018): words that occur more frequently in language production are
easier (faster) to recognize during language comprehension. Conversely, a decrease in the
use of a word would be paralleled by a greater effort of recognizing that word.
Importantly, this account predicts that language comprehension behavior will closely
mirror changes in the frequency distribution of emotion-laden words. If aging comes with
AFFECT ACROSS ADULTHOOD 8
a more frequent use of positive words, we expect to observe a faster recognition of positive
rather than neutral or especially negative words. This prediction for language
comprehension is expected to hold both for an increase in positive word types or tokens,
and regardless of the reasons for that increase (e.g., due to a better emotion regulation or
due to a prevalence of positive experiences in the older adults’ environments).
A third usage-based account, which we refer to as “lifelong learning”, generates a
different set of predictions. This account proposes that – as a result of knowledge
accumulation over time – older individuals have encountered a larger number of words
and, on average, they have encountered each word more often (see Baayen, Tomaschek,
Gahl, & Ramscar, 2017; Ramscar et al., 2014). Because vocabulary size increases with age
(finding i), discrimination among words becomes more difficult, resulting in increased
reaction times in a number of lexical tasks (see Milin, Feldman, Ramscar, Hendrix, &
Baayen, 2017; Ramscar, Dye, & McCauley, 2013; Ramscar, Hendrix, Love, & Baayen, 2013;
Ramscar et al., 2014; Ramscar, Dye, Blevins, & Baayen, 2018, for example) and in other
cognitive tasks such as decision making (see Blanco et al., 2016). Thus, under this account,
no cognitive deficit is required to explain the observed slow-down in recognition times
(finding ii) (see Ramscar, Dye, & McCauley, 2013; Ramscar, Hendrix, Love, & Baayen,
2013; Ramscar et al., 2014).
The lifelong learning account also has implications for the comprehension of
affective words. Specifically, the increase in the number of distinct positive word types
with age would lead to a slow-down in the recognition of positive words relative to negative
ones, see above. Under this account, the ease of recognizing a word is proportional to the
ease of discriminating this word from all other competitors; this ease is in turn
proportional to how informative are the cues associated with this word (Ramscar,
Hendrix, Love, & Baayen, 2013; Ramscar et al., 2014; Rescorla & Wagner, 1972; Rescorla,
1988). An increase in the number of positive word types with age is predicted to make a
word’s positive connotation a less informative cue. A similar proposal and empirical
AFFECT ACROSS ADULTHOOD 9
evidence that positive words are more numerous, more alike one another, and less
informative than negative words is also made in Garcia, Garas, and Schweitzer (2012) and
in Alves, Koch, and Unkelbach (2017). The implication of these findings is that if the
lexical space for positive words becomes more crowded with age, it is more difficult to tell
positive words apart from each other, predicting slower responses to positive words. This
is opposite to the prediction under the “use it or lose it” account. We note that our
treatment of the “lifelong learning” account makes an assumption that affective
connotations (e.g., positivity or negativity) can serve as cues to the word’s meaning: we
justify this assumption in the General discussion.
Several additional accounts are relevant for some but not all findings (i-iv) and only
generate predictions for language comprehension: we consider them below. One common
explanation of the slow-down of lexical processing in older individuals is cognitive decline
that comes with age. A number of studies on age-related cognitive functioning suggest
that aging can compromise some cognitive abilities related to fluid intelligence, e.g.,
attention, excecutive control and working memory (Bopp & Verhaeghen, 2005; Kausler,
1991; Salthouse, 1991; but see Verhaeghen, 2011), whereas abilities associated with
crystallized intelligence, e.g., domain-related knowledge and experience, remain
unaffected (Verhaeghen, 2003). These deficits are argued to lead to more effortful lexical
processing while lexical knowledge is assumed to remain intact (see Burke & MacKay,
1997; Thornton & Light, 2006, among others), see finding (ii). While this “cognitive
decline” account can explain changes in the overall language comprehension behavior, to
our knowledge, it makes no specific prediction that the decline is particularly pronounced
for, say, positive versus negative words. Furthermore, the account makes no predictions
for age-driven changes in the production of affective language or the structure of the
affective lexicon.
Moreover, a proposal exists that, with age, the engagement of motivational systems
associating valence with survival or danger weakens (Al-Shawaf, Conroy-Beam, Asao, &
AFFECT ACROSS ADULTHOOD 10
Buss, 2015; LeDoux, 2012). We label this a “flattening affect” proposal. If this is the case,
then differences in response latencies to positive and negative words should become
smaller over time, because the affective differences underlying this behavioral contrast
dwindle.
Finally, there is a possibility of a null effect of emotional information on language
comprehension: i.e., affect and its implications for production patterns have no
independent influence on age-related changes in language production or comprehension.
Indeed, some literature advocates a view that aging does not lead to either flattening nor
amplification of the role that affect plays in memory tasks (see Comblain, D’Argembeau,
Van der Linden, & Aldenhoff, 2004; Denburg, Buchanan, Tranel, & Adolphs, 2003; Eich &
Castel, 2016; Leshikar, Dulas, & Duarte, 2015, for example).
The present study
Our literature review above identified predictions regarding different facets of
production and comprehension patterns for emotion-laden words. These predictions are
often conflicting and no account is designed to paint a full picture of how age effects on
production propagate to the distribution of affective language and shape comprehension
behavior. We test these predictions in two studies and offer a rich and novel empirical
material for further theorizing on this topic.
Study 1 reports analyses of the large dataset of written productions and answers the
question of whether and how the positivity bias presents itself in naturally occuring
free-form language over five decades of age. While extensive, the current literature on the
topic leaves some questions unanswered. First, most prior research has compared a group
of older participants with one control group of younger individuals on a given task
(Carstensen & Turk-Charles, 1994; Isaacowitz, Charles, & Carstensen, 2000; Leclerc &
Kensinger, 2011; Reed et al., 2014; Ruffman et al., 2008; Scheibe & Carstensen, 2010). This
experimental design does not shed light on whether the change in affect is gradually
AFFECT ACROSS ADULTHOOD 11
developing across the timeline of aging or is confined to a certain age group. Second, most
tasks involving language exercise tight experimental control and thus elicit verbal
productions in response to highly constrained sets of stimuli. It is unclear whether and to
what degree the patterns elicited experimentally generalize over natural language
behavior, produced under conditions immune to a reactivity bias. Third, much of prior
work uses relatively small samples of individuals, often using university students as
younger controls (see Verhaeghen, 2011, for discussion). To address these gaps, our Study
1 investigates the structure of the affective lexicon in over 20,000 adults from 24 to 85 years
of age based on their status updates on the social networking website Facebook. We
analyze these texts as ecologically valid linguistic indicators of the distribution of affect
over adulthood. While affective computing in general has received a wide attention in
machine learning (see Calvo & D’Mello, 2010; Hussain & Cambria, 2018; Poria, Cambria,
Bajpai, & Hussain, 2017, for example), to the best of our knowledge, Study 1 is the first
systematic inquiry of age-related effects on the emotional content in natural written
language. Yet for studies investigating the relationship between language and cognition,
naturally occurring data are critical as language is inherently societal and grounded in
time and place. Indeed, natural language productions, including an individual’s social
media behavior, have been fruitfully used to predict a number of traits such as depression
(Eichstaedt et al., 2018), age and gender (Sap et al., 2014), personality (Guntuku et al.,
2017), and stress (Guntuku, Buffone, Jaidka, Eichstaedt, & Ungar, 2019). For a similar Big
data approach addressing other dimensions of language and age see Schwartz et al. (2013).
Adjudicating between possible instantiations of the age ×affect interaction in
language comprehension (listed above) requires a high-power empirical examination. Our
Study 2 addresses this by examining patterns of responses to words varying in their
psychological positivity, in three languages (Dutch, English, and Spanish) and in
participants spanning four decades of adulthood (from 24 to over 60). The empirical
sources are three mega-studies reporting lexical decision accuracy and reaction times for
AFFECT ACROSS ADULTHOOD 12
thousands of participants. The analysis is presented in two parts. Given the discrepancy in
availability of valence norms across the three languages, we present the analysis in two
parts. In Study 2a, a descriptive analysis of the data originating from the three languages
is presented and in 2b a more detailed regression analysis of the English data is provided.
Additionally, Study 2 pits these data against a wealth of theoretical accounts of
cognitive and emotional change to identify those accounts that are both compatible with
distributional patterns in data production and find cross-linguistic support in
comprehension data from a visual lexical decision task. Evidently, theorizing based on
concurrent consideration of language production and comprehension data is only
meaningful under an assumption that preferences in one of these communicative
functions are relevant and material for the other. Indeed, multiple theoretical accounts –
while varying in detail – advocate this linking hypothesis, see MacDonald (2013),
Pickering and Garrod (2013) and Ramscar and Baayen (2013) as topic papers as well as
the commentaries.
Put in a simplified way, distributional biases in favor of some linguistic units or
structure over others arise as a result of multiple, often conflicting, physiological, cognitive
and communicative demands for efficiency coming from both speakers and listeners
(Ferreira, 2008; Jaeger & Tily, 2011). Language production both reflects these
distributional biases and contributes to them. Yet speakers are also comprehenders and, as
much research demonstrates (e.g., Hale, 2006; Levy, 2008), they make predictions about
upcoming forms which are attuned to probabilities of those forms in the input. The
structures that speakers preferentially produce become more probable in the language,
and in turn those structures become more expected, and easier to process during
comprehension. Theories differ in their proposed mechanisms that enable the cross-talk
between production and comprehension, yet they generally agree that distributional
preferences expressed in production are learned by comprehenders and thus inform
comprehension behavior. For our purposes, this linking hypothesis suggests that a
AFFECT ACROSS ADULTHOOD 13
preference for, say, positive words in naturally occurring speech or writing of older
individuals will translate into an increased expectation of positive words by
comprehenders and will change how informative a word’s positive connotation is as a cue
to word recognition (e.g., Ramscar & Baayen, 2013). In the General discussion, we will
elaborate on the implications of our joint consideration of language production and
comprehension for the theoretical accounts linking these two cognitive abilities.
In sum, we pursue three goals. One is to conduct a novel investigation of whether
age has an effect on preferential use of negative or positive words, and how this effect is
expressed in the frequency distributions of those valenced words (Study 1). This study
contributes large-scale and ecologically valid observational evidence to a field which has
at times been constrained to under-powered experimentation. Second, we investigate
responses to valenced words in a lexical decision task across languages and age groups
(Study 2), with the goal of pitting these data against a wealth of theoretical accounts of
cognitive and emotional change and identifying those accounts that find cross-linguistic
support. Our third and final goal is to suggest at least some mechanisms that tie together
age-driven changes in both language production and comprehension. We elaborate on this
topic in the General discussion.
Study 1
The goal of this study was to determine age-related changes in frequency-trajectories
of the affective lexicon across adulthood. We present an analysis based on naturally
occurring written texts, i.e., Facebook status updates in English. We focused on the type
and token frequency distributions associated with the affective lexicon in order to test
whether and how positivity bias presents itself (Dodds et al., 2015; Warriner & Kuperman,
2015). The analysis allowed us to test whether the change in affect over age was gradual or
confined to a specific age group.
AFFECT ACROSS ADULTHOOD 14
Method
Participants, materials and procedure
The data consisted of 12,026,030 status updates posted on Facebook by 115,112 users
as part of the myPersonality Facebook application (see Kosinski, Stillwell, & Graepel, 2013,
for details). The secondary use of these data was approved by the McMaster Research
Ethics Board (certificate: 2018-089). We only included the status updates from those users
who self-reported to be between 24 and 85 years of age and reside either in the United
States, Canada or the United Kingdom. These restrictions allowed us to obtain comparable
results between the different studies presented in this article, see Study 2a, b. The final
data set contained status updates from 22,288 users.
Several pre-processing steps were taken to prepare the status updates for analysis
and were carried out in R, version 3.6.1 (R Core Team, 2019). As a pre-processing step, the
status updates were tokenized, parsed and lemmatized using UDPipe as implemented in
the R package udpipe, version 0.8.2 (Straka & Straková, 2017). After that, the affective
lexicon was established by matching the lemmata of the status updates to their
corresponding English valence norms (N= 13,915) collected in the previous mega-study
by Warriner, Kuperman, and Brysbaert (2013). Furthermore, all words that lacked the
valence information were removed. Finally, in order to keep the different data sets
comparable in terms of age, the users were divided into five age groups: 24–29, 30–39,
40–49, 50–59 and 60+. These same age groups were also used in Studies 2a,b. The
summary information of the variables associated with the users and the lexical properties
of the status updates is provided in Table 1.
AFFECT ACROSS ADULTHOOD 15
Table 1
Summary information of the variables in the status update data set.
User Status update Word Word type Length of status update
Age group N N N N M SD
24–29 9475 1106362 7850535 13170 7.1 6.18
30–39 6349 635842 5043796 12875 7.93 6.83
40–49 4033 384078 3200399 12520 8.33 7.18
50–59 1885 140301 1236494 11296 8.81 7.51
60+ 546 36571 318692 8585 8.71 7.53
Results
To test for the presence of the positivity bias based on the Facebook status update
data, we calculated the average valence score for each user based on the affective words
used in their status updates. First, word-based valence ratings were averaged for each
sentence. Second, the average valence rating of a given status update was calculated by
averaging over the sentence-level ratings in a given update. Third, the valence rating of the
status updates were averaged for each user. We will refer to this final measure as the
average user valence score.
The average user valence scores were then compared to the mid-point of the valence
scale, i.e., a rating of 5, to test for the presence of a positivity bias reported in
non-age-specific datasets, see Warriner and Kuperman (2015) and reference therein.
Valence scores averaged per user across the age groups were uniformly greater than the
mid-point of the valence scale, Table 2. Thus, the positivity bias is found in the Facebook
status updates, similar to many other text types, see Dodds et al. (2015).
Additionally, the positivity bias strengthened with advancing age as the average
valence score of the users was posititively correlated with age: r(22286) = .17, p= < .001,
AFFECT ACROSS ADULTHOOD 16
95% CI [.16, .18].1It is worth pointing out that the maginitude of the correlation aligned
with the experimental results reported by Augustine, Mehl, and Larsen (2011) in Study 2
for recordings of conversational English.
Table 2
Distribution of the average valence scores of the users in each age groups along with the results of a
Welch’s one sample upper-tailed t test comparing the average valence score of the users in a given
age to the mid-point of the valence scale.
Valence ttest
Age group M SD t value df p value
24–29 5.93 0.19 473.3 9474 < .001
30–39 5.98 0.2 387.54 6348 < .001
40–49 6.01 0.22 293.34 4032 < .001
50–59 6.04 0.24 185.93 1884 < .001
60+ 6.03 0.3 79.98 545 < .001
In order to identify whether the positivity bias emerged specifically in type
frequency, we calculated the number of word types produced in a given age group and
then averaged the valence ratings associated with the types. The distributional
information of the valence ratings among the word types is given in Table 3.
The summary information indicated that increasingly more positive word types were
produced with advancing age. A linear regression model was fitted to the data and the
results confirmed that the main effect of age group was a statistically significant predictor
1There are several ways of calculating the average valence score for a given user. However, the manner of
calculating the score did not alter the trend either when it was calculated simply over all the affective words
produced by a given user, r(22286) = .15, p= < .001, 95% CI [.13, .16], or when the valence score was
calculated based on individual status updates of a given user excluding the sentence-level averaging,
r(22286) = .16, p= < .001, 95% CI [.15, .17].
AFFECT ACROSS ADULTHOOD 17
Table 3
Summary information of the valence rating across the word types and the age groups in the status
update data.
Valence Type frequency
Age group M SD Mdn Lower quartile Upper quartile N
24–29 5.07 1.28 5.21 4.25 5.95 13170
30–39 5.08 1.28 5.21 4.25 5.95 12875
40–49 5.09 1.29 5.23 4.26 5.95 12520
50–59 5.13 1.3 5.26 4.32 6 11296
60+ 5.21 1.32 5.35 4.43 6.13 8585
of the valence rating: F(4, 58441) = 19.50, p= < .001. Furthermore, the age-related
difference in the valence ratings emerged as statistically significant between the youngest
adult group and the age group from 50 to 59 (b = 0.06, t(58441) = 3.37, p < .001, 95% CI =
0.02, 0.09) and the youngest and the age group 60+ (b = 0.14, t(58441) = 7.83, p < .001,
95% CI = 0.11, 0.18). To use our metaphor, as people age they consistently use an
increasing number of distinct words as tools recruited to convey a positive meaning. The
change is gradual and not confined to a specific age group across adulthood. However, it
is possible that this age-related positivity bias as reflected in word types arose due to a
confound of how type frequency in general is distributed in language. To rule out this
possibility, we provided two additional analyses.
The first potential source of the confound concerns the relation between type and
token frequency. The youngest adult group produced substantially more word tokens and
consequently more word types in our dataset than the oldest adult group—see the
summary information in Table 1and 3. This imbalance might skew the distribution of
frequency data because of the well-attested co-dependency of token and type frequency
(see Baayen, 2001, for example). To rule out this confound, we reanalyzed the data while
AFFECT ACROSS ADULTHOOD 18
holding the number of word types constant across the age groups. The increased number
of positive word types with advancing age was supported by this re-analysis, leading us to
conclude that this finding is not artifactual, see Supplementary materials S1 for the
analysis.
The second potential confound concerns the relation between type frequency and
lexical growth. It is conceivable that the increased type frequency of positive words
emerged solely due to the increasing size of the lexicon and it is not age-related. If this is
the case, we should observe a relative increase in type frequency of positive words in
larger corpora compared to smaller ones. To rule out this possibility, we extracted
documents of the American and British English subsections of the 1.9 billion-token Global
Web-based English (GloWbE) corpus. Written productions in this corpus are not specific
to any age. We incrementally increased the size of the samples we extracted from this
corpus and examined the positivity bias relative to the sample size. The results
demonstrated that the proportion of positive word types in the sample decreased when
the size of the sample increased. This runs counter to our observations of a stronger
positivity bias in written productions of older individuals. We conclude that the mere
growth in vocabulary that comes with age cannot explain away the stronger positivity bias
in aging language users, see Supplementary materials S2 for the analysis.
As our next step, we proceeded to identifying token frequency-trajectories in the
affective lexicon across adulthood. To this end, we first set three pairs of thresholds to
specify positive and negative word groups: the median split of the valence distribution
thus covering 100% of available words (the ties were assigned to the lower half); the top
and bottom 25% of the valence distribution, thus covering 50% of available words; and the
top and bottom 15% (covering 30%). This enabled us to study the effect of age on mildly
valenced words (e.g., the median split) and extremely valenced ones (the extreme 30%).
We will refer to them as lexical coverage groups. Second, the token frequency was
calculated for positive and negative words for each lexical coverage group: the counts
AFFECT ACROSS ADULTHOOD 19
were based on the lemmata used in the status updates separately for each age group. To
account for the differences in sample sizes across groups, the frequencies of the lemmata
were normalized per 100,000 and log transformed (base 2) with a constant of 1 added to
the values to move away from zero. Third, we estimated a token frequency-trajectory of
the positive and the negative words by calculating the median of their log normalized
token frequencies in a given age and lexical coverage group. These trajectories are
visualized in Figure 1.
30
50
100
24−29 30−39 40−49 50−59 60+ 24−29 30−39 40−49 50−59 60+ 24−29 30−39 40−49 50−59 60+
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
Age group
Median of log token frequency (normalized per 100,000)
Valence negative positive
Figure 1. Median frequency-trajectories for the positive and negative words across the age
groups and for the 30% (left), 50% (middle) and 100% (right) lexical coverage.
The shape of the token frequency-trajectories indicated that the positive words were
used more often than the negative ones with advancing age and this monotonic
age-related effect was effectively stable across the age groups and the lexical coverage
groups. Additionally, a marked increase in the frequency of use of the affective lexicon
occurred when comparing the token frequency-trajectories between the youngest and the
oldest adult groups. However, the strength of the age-related modulation of the valence
effect varied depending on the lexical coverage group. This modulation became larger
AFFECT ACROSS ADULTHOOD 20
when the lexical coverage group shifted towards the extremes of the valence distribution:
the largest modulation was observed in the 30% lexical coverage group. To confirm that
this type of modulation of the valence effect was not just limited to the median but also
characterized the token frequency distribution as a whole, the token frequency
distribution for the positive words in the youngest (24–29) and the oldest (60+) adult
group is visualized in Figure 2.
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.0 2.5 5.0 7.5 10.0
Log token frequency (normalized per 100,000)
Density
Age group 24−29 60+
Lexical coverage group: 30%
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 5 10
Log token frequency (normalized per 100,000)
Density
Age group 24−29 60+
Lexical coverage group: 50%
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 5 10
Log token frequency (normalized per 100,000)
Density
Age group 24−29 60+
Lexical coverage group: 100%
Figure 2. Distribution of the log normalized token frequency for the positive words in the
youngest and the oldest adults groups across the three lexical coverage groups.
The density plots showed a clear pattern of increased use of the positive word tokens
in the oldest adult group compared to the youngest adults. We used the permutation test
of equality between two density distributions as implemented in the R package sm,
version 2.2.5.6, to formally test whether the densities differed (Bowman & Azzalini, 1997).
All the estimated pvalues were less than .001 indicating that the density distributions of
the token frequencies were not equal between the youngest and the oldest adult groups
across the lexical coverage groups for the positive words, even after correcting for multiple
comparisons. In sum, the results demonstrated that not only did the token
AFFECT ACROSS ADULTHOOD 21
frequency-trajectory of the positive words become more pronounced with advancing age,
but the distribution as a whole shifted towards the increased use of the positive words.
Discussion
In Study 1, we investigated the use of the affective lexicon across adulthood in a large
data set of naturally occurring productions, namely Facebook status updates. The analysis
specifically focused on the token and type frequency distributions associated with the
affective lexicon. Type frequency captures the diversity of affective words used, whereas
token frequency informs us about how often affective words are used. The results of the
production data supported the “emotion regulation” account. The analysis demonstrated
that the positivity bias manifested itself in both token and type frequencies. Not only did
the Facebook status updates consist of more positively valenced lexical types, but positive
words were also used more frequently with advancing age. While the frequency of use of
both the positive and negative words increased as age advanced, the differences in
frequency distributions of positive words became more pronounced in the oldest adult
group compared to the youngest adults. This trend is compatible with a notion that age
correlates with better emotion regulation. Another partial contributor may be emotional
contagion, a tendency of social media users to generate more positive posts if they see
more positive posts in their newsfeed (Kramer, Guillory, & Hancock, 2014).
2
If social users
predominantly communicate with peers of their own age, then emotional contagion may
amplify effects of emotion regulation. A tendency towards positivity in some individuals,
reflected in their written texts, may induce a similar affective preference in their age group.
Together these results replicated and supported the positivity bias reported in
previous studies (Boucher & Osgood, 1969; Dodds et al., 2015; Warriner & Kuperman,
2015) and added important evidence on the specific role of age. The observed trend
towards positivity is not a domain of old age, nor is it a finding limited to tightly contolled
2We thank an anonymous reviewer for bringing this to our attention.
AFFECT ACROSS ADULTHOOD 22
experiments, but rather a gradual process that characterizes the entire adulthood.
Finally, the observed strengthening of the positivity bias in linguistic productions
with advancing age has important implications for models of language comprehension.
Were we to find an increasing positivity bias in word types or word tokens only, some of
the theoretical possibilities for the age-driven change in comprehension of affective words
would have been ruled out. The present language production data supports the possibility
of at least two frequency-related accounts of changes in word recognition, namely the “use
it or lose it“ account and “lifelong learning” (see Introduction). Study 2 below examines
predictions of these and additional accounts against cross-linguistic data sets of word
recognition reaction times.
Study 2
Study 2 presents two analyses of word recognition data. In Study 2a we investigated
the role of aging and affect cross-linguistically, by analyzing RTs in a visual lexical decision
task from English, Dutch and Spanish. This allowed us to compare the predictions from
the different existing theoretical accounts and to validate the results in three high-power
datasets. As the results of Study 1 showed that both the type and token frequencies of
words with positive connotation increased with advancing age, we focused on the relative
difference between recognition of the positive and the negative words in Study 2a. This
allowed us to specifically test the following hypotheses in a cross-linguistic setting: 1) “use
it or lose it”: relatively fast RTs to positive than negative words, 2) “lifelong learning”
account: a relatively stronger reduction in RTs to positive than to negative words, 3)
engagement of motivational systems: reduction in RTs to both positive and negative
words, and 4) preserved emotional memory: no age-related change in RTs to either
negative or positive words. For a detailed exposition of these accounts, see the
Introduction. Study 2b zooms in on English and addresses in depth a number of potential
confounds that can explain away some findings of Study 2a.
AFFECT ACROSS ADULTHOOD 23
Study 2a
Method
Participants, materials and procedure
The data analyzed here came from three publicly available mega-studies of visual
lexical decision in English (for accuracy, see Brysbaert, Mandera, McCormick, & Keuleers,
2019; for RTs, see Mandera, Keuleers, & Brysbaert, 2019), Dutch (for accuracy, see Keuleers
et al., 2015; for RTs, see Brysbaert, Keuleers, & Mandera, 2019) and Spanish (Aguasvivas
et al., 2018). The English data set was restricted to native speakers of English who resided
either in the UK or the USA. The English data set consisted of 61,858 words in total and
represented approximately 650,000 sessions (in each session a participant responded to
100 stimuli). Similarly, the Dutch data set was restricted to native speakers of Dutch or
Flemish and contained data on 54,319 words and approximately 450,000 sessions. While
the Spanish data set covered data from approximately 169,000 sessions across
Spanish-speaking countries, we restricted the data set to speakers of Peninsular Spanish,
approximately 49% of the data. This data set includes 45,389 words. Finally, the analysis
presented here only included the RTs for correct responses.
The studies were conducted online using crowdsourcing. All three mega-studies
made use of the same visual lexical decision task in which participants were asked to
indicate whether they knew the stimulus or not. Both accuracy and reaction time data
were collected along with basic demographic information. Each session consisted of 100
stimuli with the following approximate word-to-non-word ratio: 70/30 in English, 67/33 in
Dutch and 70/30 in Spanish. The RTs of the English and the Dutch data sets have been
compared to the results reported in previous studies. For example, Mandera et al. (2019)
demonstrated that these new norms were highly comparable to the RTs in the English
Lexicon Project with a correlation of .75 for the shared words. Similarly, Brysbaert,
Mandera, McCormick, and Keuleers (2019) showed that the words shared between these
AFFECT ACROSS ADULTHOOD 24
new Dutch norms and the norms in the Dutch Lexicon Project (Keuleers, Diependaele, &
Brysbaert, 2010) were strongly correlated (.7). Thus, these studies have demonstrated that
the new response norms are of high quality, while simultaneously representing a far more
diverse population.
In order to investigate the role of affect on word recognition across adulthood, the
words covered in the three mega-studies of visual lexical decision were matched with the
valence norms collected in the previous mega-studies for English (Warriner et al., 2013),
Dutch (Moors et al., 2013) and Spanish (Stadthagen-Gonzalez, Imbault, Sánchez, &
Brysbaert, 2017). To avoid undesirable variability in the levels of lexical knowledge across
age, we only considered those words that were very well known. We used the average
accuracy as a proxy for lexical knowledge with a cut-off point of 90% correct in each age
group (see Lemhöfer & Broersma, 2012; Stubbe, 2012, for justification). For valence, we
used the same three lexical coverage (30%, 50%, and 100% of the words) groups enabling
us to study the effect of age on mildly valenced words and extremely valenced ones.
Finally, the five age groups were considered in this study as in Study 1. The
language-specific summary of the data is provided in Table 4.
We tracked the relative changes in reaction times associated with valence across
adulthood in each of the languages in the following manner. First, we calculated the
average RTs weighted by the number of observations in each age group: this was done
separately for the positive and negative words in each lexical coverage group. Second, this
averaging was carried out separately for each of the languages. Finally, to estimate a
relative change related to valence, we calculated a RT (average negative RT - average
positive RT) in each dataset defined by language, age group, and lexical coverage group.
AFFECT ACROSS ADULTHOOD 25
Table 4
Language-specific summary of the valence data for each language in the three different lexical
coverage groups.
Variable
Word Valence rating Reaction time (ms) Number of observations (NOBs)
Valence Valence Valence Valence
Negative Positive Negative Positive Negative Positive Negative Positive
Language Lexical coverage N N M SD M SD M SD M SD M SD M SD
Dutch 30% 621 642 2.24 0.32 5.53 0.36 1062.05 119.78 1028.09 104.94 73.62 12.74 70.68 12.85
Dutch 50% 1019 1048 2.5 0.42 5.24 0.46 1060.51 114.3 1025.42 103.24 73.68 12.68 70.36 12.74
Dutch 100% 2047 2063 3.1 0.69 4.79 0.58 1056.7 110.86 1029.16 104.57 73.27 12.76 70.76 12.88
English 30% 1984 2037 2.92 0.47 6.91 0.42 969.3 121.48 915.31 103.98 99.04 46.8 94.07 47.12
English 50% 3241 3360 3.31 0.63 6.6 0.51 972.14 121.17 923.06 107.32 99.18 45.93 94.65 45.48
English 100% 6326 6470 4.03 0.89 6.1 0.66 967.17 119.37 932.77 110.34 99.06 48.39 95.62 45.11
Spanish 30% 1256 1296 2.77 0.57 7.22 0.49 975.11 268.88 923.55 267.17 31.94 33.04 32.19 28.15
Spanish 50% 2066 2109 3.31 0.82 6.88 0.58 982.72 278.18 935.62 275.11 31.79 28.62 32.25 29.26
Spanish 100% 4096 4095 4.15 1.04 6.32 0.73 985.72 277.61 949.51 277.28 32.14 28.07 32.26 28.36
Results
Figure 3(top row panels) visualizes the age- and valence-related trajectories of the
RTs across the three languages and lexical coverage groups. The bottom row panels
visualize average RTs for the positive and the negative words across age groups (the 100%
lexical coverage only).
Across the three languages, the trajectory of valence change over age was highly
comparable, even when considering mildly to extremely valenced words.
3
First, all values
of RT were above zero, indicating that responses to negative words were consistently
slower than those to positive words (for supporting and conflicting evidence see
Kuperman et al., 2014; Kousta et al., 2011, respectively). Second, the application of the
different lexical coverage groups showed that more extremely valenced groups came with
a stronger difference between positive and negative words. This is similar to amplification
3The only deviation from this trend was displayed in the Spanish data in which the youngest adults had a
substantially longer RTs than the older ones; this is discussed in the Limitations and future prospects.
However, this deviation does not affect RT which forms the basis of the analysis in this section.
AFFECT ACROSS ADULTHOOD 26
of the age effect on extremely valenced words seen in type and token frequencies in Study
1. Third, values of RT decreased between the age of 30–39 to 50–59 (in English) or 60+
(in Dutch and Spanish). The change was subtle, ranging from 6 to 7 ms in English, from 9
to 15 ms in Dutch and from 18 to 23 ms in Spanish (for RT between 30–39 and 60+ in
each language, depending on the lexical coverage of valence). Importantly, the age-driven
decrease in RT was consistent and monotonic. The transition from the age group of
24–29 to 30–39 came with either a strong decrease of 19 ms in
RT (Spanish) or effectively
no change in either English or Dutch based on the lexical coverage of 100%. This
discrepancy may partly be due to differences in the number of observations per word (the
lowest in Spanish). Still, over at least three decades (age 30+) the trend indicates that the
difference between response latencies to positive and negative words diminished with
advancing age.
The mathematical change in RT cannot directly point to the cause of the observed
reduction. To address this, Figure 3(bottom row panels) summarizes mean RTs for
positive and negative words in the 100% lexical coverage group across the three languages.
It demonstrates that both positive and negative words came with longer RTs when age
advanced, but the increase in RTs was greater in positive words. We will return to this
effect in Study 2b.
AFFECT ACROSS ADULTHOOD 27
en
nl
sp
24−29 30−39 40−49 50−59 60+ 24−29 30−39 40−49 50−59 60+ 24−29 30−39 40−49 50−59 60+
20
25
30
35
40
45
50
55
60
65
70
75
80
Age group
(<− positive) −Reaction time (ms) − (negative −>)
Lexical coverage group 30 50 100
en
nl
sp
24−29 30−39 40−49 50−59 60+ 24−29 30−39 40−49 50−59 60+ 24−29 30−39 40−49 50−59 60+
900
920
940
960
980
1000
1020
1040
1060
1080
1100
Age group
Average reaction time (ms)
Valence negative positive
Figure 3. The top panels present the change in the delta RTs of the affective words across
adulthood in English (en), Dutch (nl) and Spanish (sp). The error bars correspond to one
standard deviation. The bottom panels present the age-related modulation of RTs for the
positive and negative words for each of the three languages in the 100% lexical coverage
group. The error bars correspond to standard error.
AFFECT ACROSS ADULTHOOD 28
Discussion
In Study 2a, we investigated the influence of age and valence on RTs to English,
Dutch and Spanish words in a visual lexical decision task. We used RT as a measure of
relative change in the affective lexicon between positive and negative words across
adulthood. The results indicated a similar cross-linguistic modulation of RT by age.
First, RT displayed a overall reduction with advancing age. Second, the change in RT
was a result of the RTs to positive words being more affected by age than those to negative
words. Specifically, RTs to positive words increased over time at a higher rate than those to
negative words. We will pit these findings against existing theoretical accounts in the
General discussion.
The descriptive statistics presented in this section indicated a small effect size
aligning with previous studies on the valence effect in visual word recognition (Kuperman
et al., 2014). Additionally, our analyses in Study 2a did not account for many lexical
variables that are known to influence lexical decision latencies. The reasons for excluding
lexical control variables were as follows. First, not all relevant lexical predictors are
publicly available as norms for all the three languages. Second, once the lexical norms
available for all languages are matched with the norms for valence, some languages
(notably, Spanish) show drastic data sparseness. Study 2b, however, re-analyzes English
lexical decision data while taking into account a number of relevant lexical variables (see
Baayen, Feldman, & Schreuder, 2006; Balota, Cortese, Sergent-Marshall, Spieler, & Yap,
2004; Yap & Balota, 2009, for example): this is done in order to verify that the observed
age-driven changes in recognition of emotional words were not spurious.
Study 2b
This study re-analyzed the RTs to the English positive and negative words with
additional controls. First, we conducted a formal statistical test of whether the critical
patterns reported in Study 2a for English were reliable: namely, the age ×valence
AFFECT ACROSS ADULTHOOD 29
interaction where both positive and negative words show slower RTs across age groups,
and the slow-down is especially pronounced in positive words. Importantly, this test –
implemented as a regression model – included a large number of other lexical predictors
known to influence RTs in lexical decision (see Baayen et al., 2006; Balota et al., 2004;
Brysbaert et al., 2011; Keuleers et al., 2010; Yap & Balota, 2009). Thus, we sought to
validate the critical interaction over and above the influence of these predictors. A better
performance of the model with the critical age ×valence interaction over a model with
main effects of age and valence, while other predictors are contolled for, would be
evidence in favor of the interaction.
Second, we explored the possibility that the age
×
valence interaction itself could be
further modulated by another lexical property associated with the affective lexicon. A
better performance of the model with the critical age ×valence interaction over a model
where age and valence are two of the three terms in a three-way interaction would indicate
that the critical interaction is indeed valid across the entire lexical space. A vast majority of
lexical properties are likely to change their values across the lifespan. For instance, the
number of orthographic neighbors depends on one’s orthographic lexicon size, which
changes with age. We also demonstrated changes in word frequency in Study 1. Yet to our
knowledge, reliable age-locked estimates are not currently available for these lexical
properties. As a result, for this test, we chose word length (in letters) because it is a word
property that is not influenced by accumulation of knowledge over time – a
five-letter-word remains as a five-letter-word regardless of a person’s experience with such
words with advancing age – allowing us to keep it constant across the age groups. There
are reasons to believe, however, that the well-established effect of word length on RTs may
change as a function of age. Age has been shown to be associated with a reduction in
perceptual span (Rayner, Castelhano, & Yang, 2009) and visual acuity (Akutsu, Legge,
Ross, & Schuebel, 1991). Other changes in oculomotor behavior, such as longer (Rayner
et al., 2006) but slower saccadic eye movements (Peltsch, Hemraj, Garcia, & Munoz, 2011)
AFFECT ACROSS ADULTHOOD 30
in older readers can also affect reading performance and subsequently RTs in a visual
lexical decision task, especially in long words. In sum, we tested whether our critical
interaction of age ×valence is reliable on its own or whether it is further qualified by a
modulating influence of word length.
Participants, materials and procedure
The RTs came from the same mega-study on English as in Study 2a. The lexical
properties that we considered included affect as a critical independent variables and 10
other lexical predictors. We only considered variables that 1) were publicly available, 2)
had a large coverage among the affective words to avoid large-scale data loss, and 3) had
been previously shown to influence RTs in a visual lexical decision task (see Baayen et al.,
2006; Balota et al., 2004; Brysbaert et al., 2011; Keuleers et al., 2010; Yap & Balota, 2009).
The concreteness norms were obtained from Brysbaert, Warriner, and Kuperman (2014)
and the age-of-acquisition norms were obtained from Kuperman, Stadthagen-Gonzalez,
and Brysbaert (2012). The remaining variables were extracted from the English Lexicon
Project (Balota et al., 2007). As a frequency norm, we used the subtitle frequency provided
in the English Lexicon Project (see Brysbaert & New, 2009, for details). It is important to
note that the analysis presented for Study 2b contained only the affective words that had
lexical information available in all the publicly available data bases: thus, the sample size
in Study 2b is somewhat smaller than in 2a. We used the same three pairs of thresholds to
specify positive and negative word groups: the median split of the valence distrib ution
thus covering 100% of available words (the ties were assigned to the lower half); the top
and bottom 25% of the valence distribution, thus covering 50% of available words; and the
top and bottom 15% (covering 30%). The number of the affective words attested in each
lexical coverage group was the following: 30%: n= 3544, 50%: n= 5905 and 100%: n=
11873. Finally, the words were divided into positive and negative ones based on the same
three lexical coverage groups. The lexical variables and their distribution in each of the
AFFECT ACROSS ADULTHOOD 31
lexical coverage group are presented in Table 5
Table 5
Summary information of the lexical variables and their distribution in each of the lexical coverage
group for the negative and positive words.
Data set
Lexical coverage: 30% Lexical coverage: 50% Lexical coverage: 100%
Valence Valence Valence
Negative Positive Negative Positive Negative Positive
Variable M SD M SD M SD M SD M SD M SD
Frequency (pm) 11.12 48.56 47.88 237.96 9.5 41.8 45.88 274.87 10.28 57.16 36.28 242.64
Length 7.6 2.39 7.35 2.32 7.49 2.39 7.33 2.35 7.31 2.37 7.22 2.35
Orthographic neighborhood 2.76 0.97 2.64 0.92 2.71 0.96 2.65 0.93 2.63 0.95 2.6 0.93
Phonological neighborhood 2.7 1.15 2.6 1.08 2.65 1.13 2.59 1.09 2.59 1.12 2.55 1.09
Bigram frequency (average, log) 8.16 0.45 8.14 0.48 8.14 0.45 8.14 0.47 8.15 0.46 8.15 0.47
Number of phonemes 6.4 2.22 6.11 2.14 6.31 2.21 6.1 2.16 6.16 2.19 6.02 2.15
Number of syllables 2.49 1.06 2.4 1.02 2.44 1.06 2.4 1.03 2.37 1.04 2.34 1.03
Number of morphemes 1.85 0.8 1.71 0.73 1.82 0.8 1.7 0.75 1.77 0.78 1.69 0.75
Age of acquisition 9.67 2.39 7.87 2.53 9.81 2.39 8.15 2.57 9.72 2.47 8.57 2.6
Concreteness 2.95 0.93 3.17 1.14 3.01 0.95 3.28 1.14 3.23 1 3.39 1.1
Although the posititive and the negative words displayed differences in terms of
lexical properties as reflected by the summary information in Table 5, the valence of a
particular word, i.e., whether positive or negative, was not highly predictable from the
other lexical properties associated with the word. This indicates that valence provides a
unique contribution to the semantic structuring of a word. The details of this analysis is
provided in Supplementary materials S4.
The summary information of the RTs (ms) is provided in Table 6separately for the
positive and the negative words in each of the lexical coverage group.
Statistical considerations
To offer evidence for the critical interaction—age and valence—we fitted linear
regression models to the data, and the Akaike information criterion (AIC) was used to
determine the contribution of a given predictor in the model (Akaike, 1974). Specifically,
AFFECT ACROSS ADULTHOOD 32
Table 6
Summary information of the English reaction time (ms) data for the negative and positive words
broken down by the age group and the three lexical coverage group.
English data set
Lexical coverage: 30% Lexical coverage: 50% Lexical coverage: 100%
RT (ms) RT (ms) RT (ms)
Valence Valence Valence
Negative Positive Negative Positive Negative Positive
Age group M SD M SD M SD M SD M SD M SD
24–29 920.76 105.55 868.54 83.77 925.81 106.19 876.14 88.61 924.52 105.8 888.43 94.24
30–39 935.77 103.68 878.39 80.76 938.44 102.75 887.03 85.8 935.47 103.03 899.04 90.14
40–49 945.31 104.33 892.52 85.5 946.4 102.85 897.92 87.62 942.01 102.62 909.04 92.45
50–59 971.06 112.15 922.92 91.63 973.71 110.77 928.34 95.1 969.05 110.11 937.71 98.79
60+ 1039.05 122.88 989.3 106.52 1040.63 123.51 994.01 109.82 1033.61 121.24 1003.29 113.63
we used AIC, the difference in AIC between two models. The logic of this method is as
follows. We calculated the difference in AIC between a model with a given predictor and a
model without that predictor. If the removal of the predictor resulted in an increased AIC
value, this less complex model was highly unlikely given the data, thus presenting
evidence for the inclusion of the test predictor. As a rule of thumb, a difference in AIC
between two models of 2 or less suggests substantial evidence for a particular model,
values between 3 and 7 indicate considerably less support, and 10 or greater indicates that
the model in question is very unlikely (Burnham & Anderson, 2002). Importantly, this
type of model comparison was only carried out for predictors of interest and other lexical
control variables were not removed from the models even if they were not statistically
significant to avoid reporting anti-conservative results (see Harrell, 2001, for discussion).
If the best fitting model showed that the critical age ×valence interaction was significant,
we considered this as evidence that this critical effect was not spurious and survived even
when other major lexical predictors were accounted for.
For the purposes of statistical inference when comparing the reaction times of the
AFFECT ACROSS ADULTHOOD 33
positive and negative words between the age and the lexical coverage groups, we
estimated marginal means for these contrasts (Searle, Speed, & Milliken, 1980) as
implemented in the R package emmeans, version 1.4.1. The estimated marginal means are
based on the predictions of the fitted model allowing us to carry out planned comparisons
to test the following valence effects: 1) overall slower RTs to the negative than to the
positive words, 2) relatively greater slow-down in RTs to the positive than to the negative
words with advancing age, and 3) decrease in the positivity bias effect with advancing age.
The estimated pvalues of the effect were adjusted for multiple comparsions using the
using the Tukey method (Tukey, 1994). Finally, as some of these comparisons are best
represented as relative effects, we reported these effects as percentages of the estimated
marginal means, i.e., ratios of the RTs.
Results
We carried out two backward stepwise model fitting procedures to test for the
age-related valence effect on RTs in English as recommended in Zuur, Ieno, Walker,
Saveliev, and Smith (2009). In the first, we tested whether the age-related valence effect
was further modulated by length (in letters). The first model specification contained log
transformed RT as the response variable along with the three-way interaction between age,
valence and length. The second model specification contained the same response variable
but separate interaction terms, namely age by valence and valence by length. Both of these
model specifications included the same set of lexical control variables: log-transformed
frequency, orthographic neighborhood, phonological neighborhood, log-transformed
average bigram frequency, number of phonemes, number of syllables, number of
morphemes, age-of-acquisition, and concreteness. These models were fitted to the three
different lexical coverage groups. The results of these fitted models are provided in
Supplementary materials S3. The model comparison supported the inclusion of separate
interactions terms indicating that the age ×valence interaction was not modulated by
AFFECT ACROSS ADULTHOOD 34
length and this support was consistent across the three lexical coverage groups.
In the second model fitting procedure, we tested whether the modulation of the
valence effect on RTs by age was supported by the data. We used a third model
specification that only differed by the absence of the interaction between age and valence.
The model comparison offered evidence for the inclusion of the age-related valence effect
and the results given in Supplementary materials S3. In sum, the model comparisons
across the lexical coverage groups fully supported the hypothesis that age modulated the
valence effect on RTs. The summary information of the best fitting model to English
affective words in the three lexical coverage groups is provided in Table 7.
Table 7
Summary information of the best fitting linear regression models of English RTs with an interaction
of valence and age in the three different lexical coverage groups.
Lexical coverage: 30% Lexical coverage: 50% Lexical coverage: 100%
Intercept 6.66 (0.01)∗∗∗ 6.66 (0.01)∗∗∗ 6.65 (0.01)∗∗∗
Frequency (log) 0.02 (0.00)∗∗∗
0.02 (0.00)∗∗∗
0.02 (0.00)∗∗∗
Length (in letters) 0.01 (0.00)∗∗∗ 0.01 (0.00)∗∗∗ 0.01 (0.00)∗∗∗
Valence: positive 0.01 (0.00) 0.01 (0.00) 0.00 (0.00)
Age: 30–39 0.02 (0.00)∗∗∗ 0.01 (0.00)∗∗∗ 0.01 (0.00)∗∗∗
Age: 40–49 0.03 (0.00)∗∗∗ 0.02 (0.00)∗∗∗ 0.02 (0.00)∗∗∗
Age: 50–59 0.05 (0.00)∗∗∗ 0.05 (0.00)∗∗∗ 0.05 (0.00)∗∗∗
Age: 60+ 0.12 (0.00)∗∗∗ 0.12 (0.00)∗∗∗ 0.11 (0.00)∗∗∗
Orthographic neighborhood 0.00 (0.00) 0.00 (0.00)0.01 (0.00)∗∗∗
Phonological neighborhood 0.00 (0.00)0.00 (0.00)0.00 (0.00)
Bigram frequency (average, log) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00)∗∗∗
Number of phonemes 0.00 (0.00) 0.00 (0.00) 0.00 (0.00)∗∗∗
Number of syllables 0.02 (0.00)∗∗∗ 0.02 (0.00)∗∗∗ 0.02 (0.00)∗∗∗
Number of morphemes 0.01 (0.00)∗∗∗
0.01 (0.00)∗∗∗
0.01 (0.00)∗∗∗
Age-of-acquisition 0.01 (0.00)∗∗∗ 0.01 (0.00)∗∗∗ 0.01 (0.00)∗∗∗
Concreteness 0.00 (0.00)∗∗∗ 0.00 (0.00)∗∗∗ 0.00 (0.00)
Length (in letters): Valence: positive 0.01 (0.00)∗∗∗
0.00 (0.00)∗∗∗
0.00 (0.00)∗∗∗
Valence: positive, age: 30–39 0.00 (0.00) 0.00 (0.00) 0.00 (0.00)
Valence: positive, age: 40–49 0.00 (0.00) 0.00 (0.00) 0.00 (0.00)
Valence: positive, age: 50–59 0.01 (0.00) 0.01 (0.00)0.01 (0.00)∗∗
Valence: positive, age: 60+ 0.01 (0.00)0.01 (0.00)∗∗ 0.01 (0.00)∗∗∗
R20.47 0.45 0.42
Adj. R20.47 0.45 0.42
RMSE 0.08 0.08 0.09
∗∗∗p < 0.001,∗∗p < 0.01,p < 0.05
AFFECT ACROSS ADULTHOOD 35
As a global pattern, all affective words elicited an increase in RTs driven by
advancing age, i.e., a simple main effect of age, and this effect was also consistent across
the three lexical coverage groups. This follows the well-attested pattern reported in
previous studies (see Ratcliff et al., 2004; Rayner et al., 2006; Salthouse, 1996, among
others). It is worth pointing out that the simple main effect of valence was not statistically
significant suggesting that the effect of valence appeared to be strongly modulated by
age-related processes. As the main focus of this study is on the age-related valence effect,
we will concentrate on the estimated interaction effect of age ×valence. The estimated
effect of this interaction on RTs is visualized in Figure 4. The results present three
age-related modulations of the valence effect that are relevant for the purposes of the
present study and are discussed separately below.
The first age-related valence effect discussed here is concerned with the effect itself
across adulthood; namely it displayed differentials patterns for the negative and positive
words. The negative words were consistently processed slower than the positive ones and
this valence effect was fairly consistent across the three different lexical coverage groups
(see Figure 4). For example, the estimated average RT (ms) was 881.34 (95% CI [879.09,
883.59]) for the negative words and 862.39 (95% CI [860.27, 864.51]) for the positive ones
among the youngest adult group in the 100% lexical coverage group, corresponding to a
facilitatory effect of 2.2%. In the case of the oldest adult group, the estimated average RT
for the negative words was 984.98 (95% CI [982.46, 987.5]) and 973.12 (95% CI [970.73,
975.51]) for the positive ones, respectively. This corresponded to a facilitatory effect of
1.22%.
The second effect was concerned with the modulation of RTs by the age-related
valence effect. The results indicated that the increase in RTs was greater with the positive
words and this effect was especially noticeable in the 100% lexical coverage group among
the youngest and the oldest adults. For example, in the case of the negative words, the
ratio of the RTs between the youngest and the oldest adult group was estimated to be 0.89
AFFECT ACROSS ADULTHOOD 36
(SE = 0.001) corresponding to a reduction of 11% in the oldest adult group. For the
positive words, the ratio was estimated to be 0.8862 (SE = 0.14) translating into a 11.38%
reduction in the oldest adult group. The difference in these age-related valence effects was
also statistically significant for the negative words (t(59344) = -69.33, p< .001) and for the
positive words (t(59344) = -76.91, p< .001). Thus, the results offered support for the
modulation of the age-related valence effect where the increase in RTs was greater for the
positive words with advancing age.
The third age-related valence effect concerned the relative modulation of the
age-related valence effects themselves. It is related to positivity bias and its modulation by
age; namely, the facilitatory effect of positivity bias was reduced as age advanced. This
relative modulation of the valence effect is clearly visible in Figure 4, when we contrast the
youngest and the oldest adult groups, and it was strongest in the 100% lexical coverage
group. For example, the observed facilitatory effect of the positivity bias was estimated to
be 1.22% in the oldest adult group and 2.2% in the youngest. The reduction in the
facilitatory effect of the positivity bias, albeit small, was estimated to be 0.97% and was
statistically significant (t(59344) = 4.29, p< .001). This reduction in positivity bias
between the youngest and the oldest adult group was also statistically significant in the
50% lexical coverage group with an estimated reduction of 0.89% (t(29504) = 2.82, p<
.001). Finally, in the 30% lexical coverage group, the reduction was estimated to be 0.87%
and was also statistically significant (t(17699) = 2.17, p= .03).
AFFECT ACROSS ADULTHOOD 37
30
50
100
24−29 30−39 40−49 50−59 60+
840
850
860
870
880
890
900
910
920
930
940
950
960
970
980
990
1000
840
850
860
870
880
890
900
910
920
930
940
950
960
970
980
990
1000
840
850
860
870
880
890
900
910
920
930
940
950
960
970
980
990
1000
Age group
Reaction time (ms)
Valence negative positive
Figure 4. The estimated partial effect of the age and valence interaction on RTs across the
three different lexical coverage groups in English. The RTs were backtransformed from log
scale to ms.
AFFECT ACROSS ADULTHOOD 38
Discussion
In Study 2b, we investigated the role of age-related valence effect on RTs to English
words in a visual lexical decision task while controlling for other lexical properties
associated with the affective words. The results of the analysis provided converging
evidence for the cross-linguistic age-related valence effects reported in Study 2a. We
carried out model comparisons that confirmed the support for the age-related valence
effect in RTs. Additionally, the model comparisons also offered evidence that the critical
interaction between age and valence was unlikely to be further modulated by length (in
letters) given the data. We used length as it can serve as an excellent control variable in
aging studies for two reasons. First, it is not influenced by accumulation of knowledge
across the life span compared to frequency of use, for example. Second, its influence in
language comprehension is connected to vision and aging is known be associated with
physiological changes in vision that can affect reading performance such as reduction in
perceptual span.
The best fitting model also provided evidence for the following age-related valence
effects on RTs in English. First, the positive words were consistently processed faster than
the negative ones, regardless of how extreme the valence response was. Second, the
positive words were more affected by age-related reduction in RTs than the negative words.
Thus, the positive words were estimated to display a greater increase in RTs with
advancing age. Third, the facilitatory effect of the positivity bias reduced with advancing
age.
General discussion
This paper examined production and comprehension of affective language across
adulthood (from 24 years of age). The overall goal was to identify whether and how
strongly age modulates the distribution of affective information in naturally occurring
language production and how these distributional shifts influence comprehension
AFFECT ACROSS ADULTHOOD 39
behavior. To this end, we recruited massive datasets of written productions (Facebook
status updates from over 20,000 users) and collections of lexical decision data in English,
Dutch and Spanish (responses from thousands of users to over 40,000 words in each
corpus). In each data set and task, we focused on the critical interaction of age by valence:
namely, we were interested whether language production or comprehension behavior
differed as a function of valence and whether these affective differences varied across age.
Our central finding was uniform across tasks, languages and cohorts. In both natural
written productions and word recognition data, we observed an age
×
valence interaction.
We viewed the behavioral data as a (partial) reflection of cognitive, emotional and other
changes that age progression engenders in a human adult. Additionally, we provided
evidence that the valence of a particular word is not strongly predictable from the other
lexical properties associated with the word as discussed in the Supplementary materials
S4. This further supports the role of valence as an important cue associated with semantic
structuring in language. Thus, below we discuss the direction and the functional form of
these interactions in light of the theoretical accounts of affective lexicon as outlined in the
Introduction.
Data from the free-form social media texts in English enabled quantification of the
structure of the affective lexicon in over 20,000 users from 24 to 85 years of age (Study 1).
A central observation was that older individuals use a larger number of distinct positive
words (i.e., an increase in positive word types) and also they used an average positive
word more often (i.e., an increase in positive word tokens). The amplification of the
positive bias was gradual rather than confined to a specific age group. This finding was
robustly replicated across multiple methods of data aggregation. It converges with
previous reports of a trend in older adults to experience or express fewer negative
emotions in a variety of tasks related to memory and emotion recognition (see Carstensen
& DeLiema, 2018; Reed et al., 2014, for an overview). As in prior studies (e.g., Carstensen
& Turk-Charles, 1994; Kalenzaga et al., 2016; Reed & Carstensen, 2012), we link the
AFFECT ACROSS ADULTHOOD 40
gradually increasing positivity bias in older adults with their improving mastery of
emotion regulation strategies, which highlight positive experiences over negative ones.
The novel contribution of these production data is three-fold. First, it uncovers
evidence that runs counter to theoretical accounts positing a null effect of age on emotion
regulation and on tasks involving expression or recognition of affect (Comblain et al., 2004;
Denburg et al., 2003; Eich & Castel, 2016; Leshikar et al., 2015). Second, it harnesses a
very large high-power dataset of ecologically valid, free-form data representing a broad
range of age, ability, personality, and other properties relevant for individual variability in
cognition and emotion. Thus, the present findings serve as an important replication and
validation of the results often obtained from relatively small cohorts of participants,
drastically different in age and exposed to a relatively small number of tighly controlled
experimental stimuli. Third, our findings in language production are useful for both
predicting and constraining the range of theoretical possibilities regarding the roles of
aging and affect in word recognition behavior: for an elaboration of a hypothesis linking
production and comprehension see below.
Analyses of word recognition data across four decades of adulthood (24 y.o. and
older) and three languages uncovered one and the same pattern (Study 2a). Finding (i)
was that all words elicited longer lexical decision RTs in older participants compared to
younger ones. This is fully in line with reports of an age-driven slow-down in the speed of
lexical processing. As such it is compatible with predictions of two radically opposing
accounts. In the “cognitive decline” account, the slow-down is due to a deficient or
suboptimal cognitive functioning caused by aging (e.g., Salthouse, 2009), while the
“lifelong learning” account the slow-down is not a deficit but an expected result of the
continuously growing vocabulary and a greater effort of discriminating one word from an
increasing number of competitors (see Ramscar et al., 2014). Finding (i) alone does not
provide direct evidence in favor of either account.
Finding (ii) was that the difference in reaction times to very positive and very
AFFECT ACROSS ADULTHOOD 41
negative words declined with age. Finding (iii) was that this was due to responses to
positive words slowing down more strongly with age than responses to negative words.
To rephrase, the overall reduction in response latencies observed across the board as
readers age was more pronounced in positive rather than negative words. Study 2b
further confirmed the validity of all findings in the English data set, over and above the
impact of other lexical predictors.
Findings (ii) and (iii) are theoretically important. Natural written production data
revealed that as they age people use an ever larger number of positive word types, and use
an average positive word more often than negative ones (more positive word tokens). As
argued in the Introduction, an increase in positive word tokens is predicted to facilitate
recognition of positive words: under this frequency-based or “use-it-or-lose-it” account,
the ease of word recognition is directly proportional to one’s familiarity with the word,
gauged as the amount of exposure to the word or its frequency of occurrence (Baayen,
2010; Brysbaert et al., 2018). Yet an increase in positive word types predicts the opposite
under the “lifelong learning” account. Namely, it predicts that psychological positivity
becomes a less salient cue to discriminating a word from other words in the mental lexicon
(Alves et al., 2017; Garcia et al., 2012). The crowding of the lexical space associated with
positive connotations in older individuals is expected to lead to a relative slow-down in
positive words over age. Finding (iii) offers a ready confirmation for the “lifelong learning”
account: with age, RTs to positive words grow slower at a higher than RTs to negative
words. Critically, Finding (iii) runs counter to the predictions of the frequency-based
account. It cannot be explained under the cognitive decline account either that posits a
general slow-down of all verbal processes (see Laver & Burke, 1993, for discussion).
Second, to our knowledge, the deficits that aging is proposed to bring to attention,
memory and other cognitive processes are not argued to be sensitive to any aspect of
lexical semantics, including affective lexical connotations (see Burke & MacKay, 1997;
Thornton & Light, 2006, among others). Importantly, the age ×valence interaction
AFFECT ACROSS ADULTHOOD 42
indicates that it is not just accumulation of semantic information over time that can
facilitate processing as discussed in Laver and Burke (1993), for example, but the context
of the accumulation itself must be taken into account in order to begin to model the
differential valence effect reported in this study. The crowding of the lexical space, and
specifically the affective lexicon, can also be used to motivate the shrinkage of the
difference in reaction times to the positive and negative words, as the informativity of a
cue is not only determined by the stimulis itself but also the relation of the cue to other
stimuli: for discussion see (Ramscar, Dye, & McCauley, 2013; Rescorla, 2001). Thus, the
shrinkage of the positivity bias suggests that as the use of the affective lexicon as a whole
increases with advancing age, the valence loses its informativity as a cue resulting in a
shrinkage of the positivity bias effect in comprehension.
Given the context of how language can accumulate over time, the results presented
here cannot rule out the possibility of aging influencing affective lexical processing in
multiple ways simultaneously. Aging comes with an increase in positivity bias, both in
words types and tokens. It is likely then that the counter-directed trends – a processing
advantage coming with a higher token frequency and a discrimination penalty from a
higher type frequency– are present in the structure of one’s vocabulary as it develops
across the life span. If so, the chronometric data on word recognition reflect a
juxtaposition of effects on lexical processing, some of which mitigate each other.
What we do establish in this study is that central findings cannot be explained
without involving one of the conflicting accounts, i.e., the lifelong learning account. Over
and above possible other influences, “lifelong learning” is a mechanism that can explain
both the general effect of aging on language processing found here and in much previous
research (Baayen et al., 2017; Milin et al., 2017; Ramscar, Hendrix, Love, & Baayen, 2013)
and the newly reported age ×valence interaction in language comprehension. This
account satisfies the principle of parsimony (see Baltes, 1987; Salthouse, 2009; Singer,
Verhaeghen, Ghisletta, Lindenberger, & Baltes, 2003, among others) as the account can be
AFFECT ACROSS ADULTHOOD 43
used to motivate the patterns reported for production and comprehenison. In this study,
we have not presented a computational model of lifelong learning but pointed to general
principles that a cognitively plausible model should be able to account for. We leave it to
future studies to implemented a precise model, for example in the framework of
discriminative learning (Baayen et al., 2017; Milin et al., 2017; Ramscar et al., 2014;
Ramscar, Hendrix, Love, & Baayen, 2013). At the same time, it is noteworthy that while
word semantics is in the focus of any account of language learning, to our knowledge,
connotations are rarely discussed in the literature on discriminative learning. This paper
provides evidence that emotional connotations, like positivity or negativity, can serve as
cues towards associating forms and meanings and – like with other cues – their
informativity can change as a result of age-related changes in the distributional patterns of
the mental lexicon as a whole (see Hinojosa, Moreno, & Ferré, 2019, for similar
observation).
The results reported above go beyond a novel large-scale empirical contribution to
understanding age effects in production and comprehension of affective language. As
stated in the Introduction, we agree with the view that production and comprehension
behaviors are shaped by the same set of (often conflicting) cognitive, physiological, and
communicative processes (MacDonald, 2013; Pickering & Garrod, 2013): one of these
processes is aging. While proposals on the actual mechanism providing the
production-comprehension link differ (see the discussion in the special issue by
MacDonald, 2013), they converge on the notion that comprehenders are sensitive to
statistical probabilities of linguistic units that emerge in production. Furthermore, these
probabilities guide comprehenders in forming expectations about the upcoming material
and allocating their attentional resources to these expectations proportionally to how
probable they are to occur (Kuperberg & Jaeger, 2016; Levy, 2008; Ramscar et al., 2014).
Thus, distributional biases in language that are shaped by, and revealed in, production
also shape the comprehension behavior. These distributional biases are not static, either in
AFFECT ACROSS ADULTHOOD 44
a language community, or in an individual: they change both at the macro-level – due to
constant changes in the extra-linguistic environment, in language, and in an individual
organism – and the micro-level – due to continuous exposure to the new linguistic input.
For the purposes of this paper, an important implication of the coordinated nature of
production and comprehension behaviors is that we can interpret the age ×valence
interaction observed as both a constraining and explanatory factor of the age ×valence
interaction that we report in word recognition studies. This is what underlies our ability to
use production and comprehension data jointly as a testbed for a number of existing and
logically possible theoretical accounts outlined in the Introduction. However, it is worth
pointing out that we have presented evidence for the modulation of the valence effect with
advancing age in this study. We have not presented evidence for a change in valence itself
across adulthood, i.e., for a valence trajectory associated with a given word. For example,
it is plausible that the perceived valence of a word also changes as a function of age but it
is unlikely that this type of valence trajectory, even if present, would substantially alter the
results presented here as previous studies have reported only small differences in valence
ratings, for example between younger and older Finnish adults (Söderholm, Häyry, Laine,
& Karrasch, 2013) and among German children and adults (Bahn, Kauschke, Vesker, &
Schwarzer, 2018; but see Monnier & Syssau, 2017, for French).
We present novel evidence on the role of affect on language production and word
recognition over four decades of adult age and in three languages representing multiple
national communities. This evidence tests the validity of existing theoretical proposals of
the interaction between language, emotion and age. It does so by identifying distributional
patterns of affective language in natural written productions and pitting them against
word recognition behavior, using large data-sets for both tasks. Under the hypothesis
linking production and comprehension via statistics of language, we were able to critically
evaluate a spectrum of competing accounts. The results of this study are most readily
compatible with the “lifelong learning” account (Baayen et al., 2017; Milin et al., 2017;
AFFECT ACROSS ADULTHOOD 45
Ramscar et al., 2014; Rescorla, 1988; Rescorla & Wagner, 1972; for a broader discussion see
Siegel & Allan, 1996). This account presents aging as a continuous accumulation of lexical,
semantic, and—by extension—affective knowledge. It construes behavioral differences in
processing positive and negative words as a direct consequence of growing
information-processing demands on word recognition in an expanding lexicon. Our
findings demonstrate that the cognitive demands of word recognition and discrimination
are co-determined by the structure of the affective lexicon. While the effect size of this
modulation was estimated to be small, it is in line with previous studies and resulting
from a task that did not require an overt emotional response. Even small effects can be
integral parts of a detailed understanding of cognition (Salthouse, 2012). Further work is
required to expand this effort over additional semantic variables and generate a
comprehensive account of word processing over the life span.
Limitations and future prospects
Our use of Facebook data as a source of textual and affective information requires a
discussion of how generalizable the present findings are to the general population. We
discuss two potentially relevant aspects of the Facebook data: demographics of Facebook
users vs population at large, and affective and stylistic demands of writing in Facebook.
Online data sources tend to provide access to both a larger and more diverse pool of
participants compared to traditional convenience pools. However, these welcome
characteristics do not automatically mean that the sampled data can be taken to represent
the target population as a whole (Goodman, Cryder, & Cheema, 2013; Paolacci &
Chandler, 2014). Self-selection bias of individuals choosing to use a certain social media
platform (or participate in any online experimental study) is an important factor to
consider. Our choice of Facebook is partly determined by the breadth of its user base: 79%
of Americans used Facebook in 2019 and it is also widely used globally as compared to,
say, Twitter which is used by 22% of US adults
AFFECT ACROSS ADULTHOOD 46
(https://www.pewresearch.org/internet/fact-sheet/social-media/) and is primarily
concentrated in the United States, Japan and Russia (https://www.statista.com/statistics/
242606/number-of-active-twitter-users-in-selected-countries/). Facebook is also widely
used across all age groups: 62% of online users aged 65+ are on Facebook as are 72% of
online users between age 50–64 (https://www.omnicoreagency.com/facebook-statistics/).
Still, several caveats are in order.
While fairly representative on the online community, on average US-based Facebook
users are younger, better educated and have higher incomes than the US population in
general. In our data, the median age of our Facebook cohort of 22,000 users is 31 y.o, while
the median age in the US in the year of data collection was 36.7 (https://www.census.gov/
data/tables/2012/demo/age-and-sex/2012-age-sex-composition.html). Reasons that are
most often cited as influential for the differences in age distribution between online users
and the population at large is different levels of digital literacy and access to required
technology (Correa, Hinsley, & De Zuniga, 2010; Jung, Walden, Johnson, & Sundar, 2017).
We have no access to information about the socio-economic status, educational or technical
proficiency levels of contributors to the present Facebook text base. Thus, it is logically
possible that self-selection across age groups of Facebook users leads to an observed
consistent increase in the positivity bias time-locked with age. While this explanation
cannot be ruled out based on the current data, we find it unlikely. This is because the
increase in positivity bias is observed across the entire age range, also in the age groups
which are very well represented in the Facebook cohort. Also, our observation of the
age-related increase in the positivity bias in naturally written texts dovetails perfectly with
similar observations obtained from experimental paradigms unaffected by the
self-selection bias (see above).
In our opinion, an acceptable solution for testing the validity of age-related findings
made on observational data from social media is to replicate the analyses in, say, 10 years.
At that time, all age groups will be represented by an even broader user base than today, as
AFFECT ACROSS ADULTHOOD 47
well as by individuals who are more technologically savvy and accustomed to the use of
social media than the present cohorts in the same age groups: the difference is expected to
be most drastic in the older age groups (65+) that will be occupied by the present 55+
year olds (see Charness & Boot, 2009). If similar patterns of affective language use are
replicated, they are not due to the self-selection bias. It is worth pointing out that sampling
biases are not limited to the Facebook data used but also extend to any online study such
as the data originating from the mega-studies analyzed in this study. For example, this
might be case with the Spanish data set, as the youngest adults displayed a different
pattern compared to ones present in the Dutch and the English data. While this did not
affect the results presented in this study, as they were based on RTs, this line of
investigation should receive closer examination in future studies to better understand
potential differences stemming from sampling biases and demographic information.
Another potential confound in the patterns of affective language use on Facebook
may stem from the linguistic register accepted in this social platform. Register, or a text
variety associated with a specific situational context, is one of the strongest predictors of
linguistic variation: Online texts are known to vary substantially in their composition and
language use (Biber & Egbert, 2018). For instance, Jaidka, Guntuku, Buffone, Schwartz,
and Ungar (2018) demonstrated that the same individuals were more likely to use
Facebook to convey personal concerns and emotions while their Twitter usage was
associated with personal needs and drives (Jaidka et al., 2018). It is thus possible that our
choice of Facebook as a data source has made salient some specific aspects of individual
world outlook and affective perception and downplayed others, due to stylistic and
linguistic register demands of this social media platform. However, we are not aware of
any evidence that status updates in Facebook especially amplify positive rather than
negative dimensions of affective language. Furthermore, there is no a priori reason to
expect that the register-driven features of Facebook would influence different age groups
in a systematic and gradient way, with older age groups using more positive words. A
AFFECT ACROSS ADULTHOOD 48
definitive answer to the role of register in age-related patterns of language use will require
further linguistic research, which will evaluate affective language in other large and
diverse sources of texts produced naturally by individuals of different ages.
Context
This study emerged as part of an ongoing research project by Aki-Juhani Kyröläinen
and Victor Kuperman on age-related changes of distributional biases in language and how
they affect language production and comprehension. This research naturally connects to
the prior work by Emmanuel Keuleers, Paweł Mandera and Marc Brysbaert on aging and
visual word recognition and, importantly, to their behavioral studies conducted on a
massive scale. These studies have made it possible to test the theoretical predictions
stemming from a number of different accounts discussed in this study. Finally, the line of
research fostered through this collaboration is currently being expanded to other aspects
of visual word processing across languages in adulthood.
AFFECT ACROSS ADULTHOOD 49
References
Aguasvivas, J., Carreiras, M., Brysbaert, M., Mandera, P., Keuleers, E., & Duñabeitia, J. A.
(2018). Spalex: A Spanish lexical decision database from a massive online data
collection. Frontiers in Psychology,9, 2156. doi:10.3389/fpsyg.2018.02156
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on
Automatic Control,19(6), 716–723. doi:10.1109/TAC.1974.1100705
Akutsu, H., Legge, G. E., Ross, J. A., & Schuebel, K. J. (1991). Psychophysics of reading—X.
Effects of age-related changes in vision. Journal of Gerontology,46(6), P325–P331.
doi:10.1093/geronj/46.6.P325
Alves, H., Koch, A., & Unkelbach, C. (2017). Why good is more alike than bad: Processing
implications. Trends in Cognitive Sciences,21(2), 69–79. doi:10.1016/j.tics.2016.12.006
Aschwanden, D., Schumacher, V., Zimmermann, K., Werner, C., Allemand, M.,
Zimprich, D., & Martin, M. (2019). Do professors better maintain cognitive
functioning in older age? GeroPsych,32, 5–17. doi:10.1024/1662-9647/a000201
Augustine, A. A., Mehl, M. R., & Larsen, R. J. (2011). A positivity bias in written and
spoken English and its moderation by personality and gender. Social Psychological
and Personality Science,2(5), 508–515. doi:10.1177/1948550611399154
Baayen, R. H. (2001). Word frequency distributions. Dordrecht: Kluwer Academic Publishers.
Baayen, R. H. (2010). Demythologizing the word frequency effect: A discriminative
learning perspective. The Mental Lexicon,5(3), 436–561. doi:10.1075/ml.5.3.10baa
Baayen, R. H., Feldman, L. B., & Schreuder, R. (2006). Morphological influences on the
recognition of monosyllabic monomorphemic words. Journal of Memory and Language,
55(2), 290–313. doi:10.1016/j.jml.2006.03.008
Baayen, R. H., Tomaschek, F., Gahl, S., & Ramscar, M. (2017). The ecclesiastes principle in
language change. In M. Hundt, S. Mollin, & S. E. Pfenninger (Eds.), The changing
English language (pp. 21–48). Cambridge: Cambridge University Press.
AFFECT ACROSS ADULTHOOD 50
Bahn, D., Kauschke, C., Vesker, M., & Schwarzer, G. (2018). Perception of valence and
arousal in German emotion terms: A comparison between 9-year-old children and
adults. Applied Psycholinguistics,39(3), 463–481.
Balota, D. A., Cortese, M. J., Sergent-Marshall, S. D., Spieler, D. H., & Yap, M. J. (2004).
Visual word recognition of single-syllable words. Journal of Experimental Psychology:
General,133(2), 283–316. doi:10.1037/0096-3445.133.2.283
Balota, D. A., Yap, M. J., Cortese, M. J., Hutchison, K. A., Kessler, B., Loftis, B., . . .
Treiman, R. (2007). The English Lexicon Project. Behavior Research Methods,39(3),
445–459. doi:10.3758/BF03193014
Baltes, P. B. (1987). Theoretical propositions of life-span developmental psychology: On
the dynamics between growth and decline. Developmental Psychology,23(5), 611.
Biber, D., & Egbert, J. (2018). Register variation online. Cambridge: Cambridge University
Press.
Blanco, N. J., Love, B. C., Ramscar, M., Otto, A. R., Smayda, K., & Maddox, W. T. (2016).
Exploratory decision-making as a function of lifelong experience, not cognitive
decline. Journal of Experimental Psychology: General,145(3), 284.
doi:10.1037/xge0000133
Bopp, K. L., & Verhaeghen, P. (2005). Aging and verbal memory span: A meta-analysis.
The Journals of Gerontology Series B: Psychological Sciences and Social Sciences,60(5),
P223–P233. doi:10.1093/geronb/60.5.P223
Borghi, A. M., Binkofski, F., Castelfranchi, C., Cimatti, F., Scorolli, C., & Tummolini, L.
(2017). The challenge of abstract concepts. Psychological Bulletin,143(3), 263–292.
doi:10.1037/bul0000089
Boucher, J., & Osgood, C. E. (1969). The Pollyanna hypothesis. Journal of Memory and
Language,8(1), 1. doi:10.1016/S0022-5371(69)80002-2
Bowman, A. W., & Azzalini, A. (1997). Applied smoothing techniques for data analysis: The
kernel approach with S-Plus illustrations. Oxford: Oxford University Press.
AFFECT ACROSS ADULTHOOD 51
Bradley, M. M., & Lang, P. J. (2000). Affective reactions to acoustic stimuli.
Psychophysiology,37(2), 204–215.
Breiman, L. (2001a). Random forests. Machine Learning,45(1), 5–32.
Breiman, L. (2001b). Statistical modeling: The two cultures. Statistical Science,16(3),
199–231.
Brysbaert, M., Buchmeier, M., Conrad, M., Jacobs, A. M., Bölte, J., & Böhl, A. (2011). The
word frequency effect: A review of recent developments and implications for the
choice of frequency estimates in German. Experimental Psychology,58(5), 412–424.
doi:10.1027/1618-3169/a000123
Brysbaert, M., Keuleers, E., & Mandera, P. (2019). Recognition times for 54 thousand
Dutch words: Data from the Dutch Crowdsourcing Project. Psychologica Belgica,
59(1), 281–300. doi:10.5334/pb.491
Brysbaert, M., Mandera, P., & Keuleers, E. (2018). The word frequency effect in word
processing: An updated review. Current Directions in Psychological Science,27(1),
45–50. doi:10.1177/0963721417727521
Brysbaert, M., Mandera, P., McCormick, S. F., & Keuleers, E. (2019). Word prevalence
norms for 62,000 English lemmas. Behavior Research Methods,51(2), 467–479.
doi:10.3758/s13428-018-1077-9
Brysbaert, M., & New, B. (2009). Moving beyond kučera and francis: A critical evaluation
of current word frequency norms and the introduction of a new and improved word
frequency measure for american english. Behavior Research Methods,41(4), 977–990.
doi:10.3758/BRM.41.4.977
Brysbaert, M., Stevens, M., Mandera, P., & Keuleers, E. (2016). How many words do we
know? Practical estimates of vocabulary size dependent on word definition, the
degree of language input and the participant’s age. Frontiers in Psychology,7, 1116.
doi:10.3389/fpsyg.2017.01164
AFFECT ACROSS ADULTHOOD 52
Brysbaert, M., Warriner, A. B., & Kuperman, V. (2014). Concreteness ratings for 40
thousand generally known English word lemmas. Behavior Research Methods,46(3),
904–911. doi:10.3758/s13428-013-0403-5
Burke, D. M., & MacKay, D. G. (1997). Memory, language, and ageing. Philosophical
Transactions of the Royal Society B: Biological Sciences,352(1363), 1845–1856.
Burke, D. M., & Shafto, M. A. (2008). Language and aging. In F. I. M. Craik &
T. A. Salthouse (Eds.), The handbook of aging and cognition (pp. 373–443). New York:
Psychology Press.
Burnham, K. P., & Anderson, D. R. (2002). Model selection and multimodel inference: A
practical information-theoretic approach (2nd ed.). New York: Springer.
Calvo, R. A., & D’Mello, S. (2010). Affect detection: An interdisciplinary review of models,
methods, and their applications. IEEE Transactions on affective computing,1(1), 18–37.
doi:10.1109/T-AFFC.2010.1
Carstensen, L. L. (2006). The influence of a sense of time on human development. Science,
312(5782), 1913–1915. doi:10.1126/science.1127488
Carstensen, L. L., & DeLiema, M. (2018). The positivity effect: A negativity bias in youth
fades with age. Current Opinion in Behavioral Sciences,19, 7–12.
doi:10.1016/j.cobeha.2017.07.009
Carstensen, L. L., & Turk-Charles, S. (1994). The salience of emotion across the adult life
span. Psychology and aging,9(2), 259. doi:10.1037//0882-7974.9.2.259
Charles, S. T., & Carstensen, L. L. (2007). Emotion regulation and aging. In J. J. Gross (Ed.),
Handbook of emotion regulation (pp. 307–327). The Guilford Press.
Charness, N., & Boot, W. R. (2009). Aging and information technology use: Potential and
barriers. Current Directions in Psychological Science,18(5), 253–258.
Citron, F. M. M. (2012). Neural correlates of written emotion word processing: A review of
recent electrophysiological and hemodynamic neuroimaging studies. Brain and
Language,122(3), 211–226. doi:10.1016/j.bandl.2011.12.007
AFFECT ACROSS ADULTHOOD 53
Comblain, C., D’Argembeau, A., Van der Linden, M., & Aldenhoff, L. (2004). The effect of
ageing on the recollection of emotional and neutral pictures. Memory,12(6), 673–684.
doi:10.1080/09658210344000477
Correa, T., Hinsley, A. W., & De Zuniga, H. G. (2010). Who interacts on the Web?: The
intersection of users’ personality and social media use. Computers in Human Behavior,
26(2), 247–253. doi:10.1016/j.chb.2009.09.003
Denburg, N. L., Buchanan, T. W., Tranel, D., & Adolphs, R. (2003). Evidence for preserved
emotional memory in normal older persons. Emotion,3(3), 239–253.
doi:10.1037/1528-3542.3.3.239
Dodds, P. S., Clark, E. M., Desu, S., Frank, M. R., Reagan, A. J., Williams, J. R., . ..
Danforth, C. M. (2015). Human language reveals a universal positivity bias. In
Proceedings of the National Academy of Sciences (Vol. 112, pp. 2389–2394). National
Acad Sciences.
Eich, T. S., & Castel, A. D. (2016). The cognitive control of emotional versus value-based
information in younger and older adults. Psychology and Aging,31(5), 503–512.
doi:10.1037/pag0000106
Eichstaedt, J. C., Smith, R. J., Merchant, R. M., Ungar, L. H., Crutchley, P.,
Preoţiuc-Pietro, D., .. . Schwartz, H. A. (2018). Facebook language predicts
depression in medical records. Proceedings of the National Academy of Sciences,115(44),
11203–11208. doi:10.1073/pnas.1802331115
Fernández-Delgado, M., Cernadas, E., Barro, S., & Amorim, D. (2014). Do we need
hundreds of classifiers to solve real world classification problems? Journal of Machine
Learning Research,15(1), 3133–3181.
Ferreira, V. S. (2008). Ambiguity, accessibility, and a division of labor for communicative
success. Psychology of Learning and Motivation,49, 209–246.
doi:10.1016/S0079-7421(08)00006-6
AFFECT ACROSS ADULTHOOD 54
Garcia, D., Garas, A., & Schweitzer, F. (2012). Positive words carry less information than
negative words. EPJ Data Science,1(1), 3. doi:10.1140/epjds3
Goodman, J. K., Cryder, C. E., & Cheema, A. (2013). Data collection in a flat world: The
strengths and weaknesses of Mechanical Turk samples. Journal of Behavioral Decision
Making,26(3), 213–224. doi:10.1002/bdm.1753
Guntuku, S. C., Buffone, A., Jaidka, K., Eichstaedt, J. C., & Ungar, L. H. (2019).
Understanding and measuring psychological stress using social media. In Proceedings
of the international AAAI conference on web and social media (Vol. 13, pp. 214–225).
Guntuku, S. C., Lin, W., Carpenter, J., Ng, W. K., Ungar, L. H., & Preoţiuc-Pietro, D. (2017).
Studying personality through the content of posted and liked images on Twitter. In
Proceedings of the 2017 ACM on web science conference (pp. 223–227).
Hale, J. (2006). Uncertainty about the rest of the sentence. Cognitive Science,30(4), 643–672.
doi:10.1207/s15516709cog0000_64
Hampshire, A., Sandrone, S., & Hellyer, P. J. (2019). A large-scale, cross-sectional
investigation into the efficacy of brain training. Frontiers in human neuroscience,13, 221.
doi:10.3389/fnhum.2019.00221
Harrell, F. E., Jr. (2001). Regression modeling strategies. With applications to linear models,
logistic regression, and survival analysis. New York: Springer.
Hartshorne, J. K., & Germine, L. T. (2015). When does cognitive functioning peak? The
asynchronous rise and fall of different cognitive abilities across the life span.
Psychological Science,26(4), 433–443. doi:10.1177/0956797614567339
Hinojosa, J. A., Moreno, E. M., & Ferré, P. (2019). Affective neurolinguistics: Towards a
framework for reconciling language and emotion. Language, Cognition and
Neuroscience, 1–27. doi:10.1080/23273798.2019.1620957
Hussain, A., & Cambria, E. (2018). Semi-supervised learning for big social data analysis.
Neurocomputing,275, 1662–1673. doi:10.1016/j.neucom.2017.10.010
AFFECT ACROSS ADULTHOOD 55
Isaacowitz, D. M., Charles, S. T., & Carstensen, L. L. (2000). Emotion and cognition. In
F. I. M. Craik & T. A. Salthouse (Eds.), The handbook of aging and cognition (2nd ed.,
pp. 593–631). Mahwah, NJ: Lawrence Erlbaum.
Jaeger, T. F., & Tily, H. (2011). On language ‘utility’: Processing complexity and
communicative efficiency. Wiley Interdisciplinary Reviews: Cognitive Science,2(3),
323–335. doi:10.1002/wcs.126
Jaidka, K., Guntuku, S. C., Buffone, A., Schwartz, H. A., & Ungar, L. H. (2018). Facebook
vs. Twitter: Cross-platform differences in self-disclosure and trait prediction. In
Proceedings of the twelfth international AAAI conference on web and social media
(pp. 141–150).
Jung, E. H., Walden, J., Johnson, A. C., & Sundar, S. S. (2017). Social networking in the
aging context: Why older adults use or avoid Facebook. Telematics and Informatics,
34(7), 1071–1080. doi:10.1016/j.tele.2017.04.015
Kalenzaga, S., Lamidey, V., Ergis, A.
-
M., Clarys, D., & Piolino, P. (2016). The positivity bias
in aging: Motivation or degradation? Emotion,16(5), 602–610.
doi:10.1037/emo0000170
Kausler, D. H. (1991). Experimental psychology, cognition, and human aging (2nd ed.).
doi:10.1007/978-1-4613-9695-6
Kemper, S., & Liu, C.
-
J. (2007). Eye movements of young and older adults during reading.
Psychology and Aging,22(1), 84–93. doi:10.1037/0882-7974.22.1.84
Keuleers, E., Diependaele, K., & Brysbaert, M. (2010). Practice effects in large-scale visual
word recognition studies: A lexical decision study on 14,000 Dutch mono-and
disyllabic words and nonwords. Frontiers in Psychology,1, 174.
doi:10.3389/fpsyg.2010.00174
Keuleers, E., Stevens, M., Mandera, P., & Brysbaert, M. (2015). Word knowledge in the
crowd: Measuring vocabulary size and word prevalence in a massive online
AFFECT ACROSS ADULTHOOD 56
experiment. The Quarterly Journal of Experimental Psychology,68(8), 1665–1692.
doi:10.1080/17470218.2015.1022560
Kosinski, M., Stillwell, D., & Graepel, T. (2013). Private traits and attributes are predictable
from digital records of human behavior. Proceedings of the National Academy of
Sciences,110(15), 5802–5805.
Kousta, S.-T., Vigliocco, G., Vinson, D. P., Andrews, M., & Campo, E. D. (2011). The
representation of abstract words: Why emotion matters. Journal of Experimental
Psychology: General,140(1), 14–34. doi:10.1037/a0021446
Kramer, A. D. I., Guillory, J. E., & Hancock, J. T. (2014). Experimental evidence of
massive-scale emotional contagion through social networks. Proceedings of the
National Academy of Sciences,111(24), 8788–8790. doi:10.1073/pnas.1320040111
Kuperberg, G. R., & Jaeger, T. F. (2016). What do we mean by prediction in language
comprehension? Language, Cognition and Neuroscience,31(1), 32–59.
doi:10.1080/23273798.2015.1102299
Kuperman, V., Estes, Z., Brysbaert, M., & Warriner, A. B. (2014). Emotion and language:
Valence and arousal affect word recognition. Journal of Experimental Psychology:
General,143(3), 1065. doi:10.1037/a0035669
Kuperman, V., Stadthagen-Gonzalez, H., & Brysbaert, M. (2012). Age-of-acquisition
ratings for 30,000 English words. Behavior Research Methods,44(4), 978–990.
doi:10.3758/s13428-012-0210-4
Laver, G. D., & Burke, D. M. (1993). Why do semantic priming effects increase in old age?
A meta-analysis. Psychology and Aging,8(1), 34–43. doi:10.1037//0882-7974.8.1.34
Leclerc, C. M., & Kensinger, E. A. (2011). Neural processing of emotional pictures and
words: A comparison of young and older adults. Developmental Neuropsychology,
36(4), 519–538. doi:10.1080/87565641.2010.549864
LeDoux, J. E. (2012). Evolution of human emotion: A view through fear. Progress in Brain
Research,195, 431–442. doi:10.1016/B978-0-444-53860-4.00021-0
AFFECT ACROSS ADULTHOOD 57
Lemhöfer, K., & Broersma, M. (2012). Introducing LexTALE: A quick and valid lexical test
for advanced learners of English. Behavior Research Methods,44(2), 325–343.
doi:10.3758/s13428-011-0146-0
Leshikar, E. D., Dulas, M. R., & Duarte, A. (2015). Self-referencing enhances recollection in
both young and older adults. Aging, Neuropsychology, and Cognition,22(4), 388–412.
doi:10.1037/0882-7974.20.4.579
Levy, R. (2008). Expectation-based syntactic comprehension. Cognition,106(3), 1126–1177.
doi:10.1016/j.cognition.2007.05.006
MacDonald, M. C. (2013). How language production shapes language form and
comprehension. Frontiers in Psychology,4, 226. doi:10.3389/fpsyg.2013.00226
Mandera, P., Keuleers, E., & Brysbaert, M. (2019). Recognition times for 62 thousand
English words: Data from the English crowdsourcing project. Behavior Research
Methods, 1–20. doi:10.3758/s13428-019-01272-8
Mather, M., & Carstensen, L. L. (2005). Aging and motivated cognition: The positivity
effect in attention and memory. Trends in Cognitive Sciences,9(10), 496–502.
Milin, P., Feldman, L. B., Ramscar, M., Hendrix, P., & Baayen, R. H. (2017). Discrimination
in lexical decision. PLoS ONE,12(2), e0171935. doi:10.1371/journal.pone.0171935
Monnier, C., & Syssau, A. (2017). Affective norms for 720 French words rated by children
and adolescents (FANchild). Behavior Research Methods,49(5), 1882–1893.
doi:10.3758/s13428-016-0831-0
Moors, A., De Houwer, J., Hermans, D., Wanmaker, S., Van Schie, K., Van Harmelen, A.
-
L.,
.. . Brysbaert, M. (2013). Norms of valence, arousal, dominance, and age of
acquisition for 4,300 Dutch words. Behavior Research Methods,45(1), 169–177.
doi:10.3758/s13428-012-0243-8
Ochsner, K. N. (2000). Are affective events richly recollected or simply familiar? the
experience and process of recognizing feelings past. Journal of Experimental
Psychology: General,129(2), 242–261. doi:10.1037/0096-3445.129.2.242
AFFECT ACROSS ADULTHOOD 58
Paolacci, G., & Chandler, J. (2014). Inside the Turk: Understanding Mechanical Turk as a
participant pool. Current Directions in Psychological Science,23(3), 184–188.
doi:10.1177/0963721414531598
Peltsch, A., Hemraj, A., Garcia, A., & Munoz, D. P. (2011). Age-related trends in saccade
characteristics among the elderly. Neurobiology of Aging,32(4), 669–679.
doi:10.1016/j.neurobiolaging.2009.04.001
Pickering, M. J., & Garrod, S. (2013). How tightly are production and comprehension
interwoven? Frontiers in Psychology,4, 238. doi:10.3389/fpsyg.2013.00238
Ponari, M., Norbury, C. F., & Vigliocco, G. (2018). Acquisition of abstract concepts is
influenced by emotional valence. Developmental Science,21(2), e12549.
doi:10.1111/desc.12549
Poria, S., Cambria, E., Bajpai, R., & Hussain, A. (2017). A review of affective computing:
From unimodal analysis to multimodal fusion. Information Fusion,37, 98–125.
doi:10.1016/j.inffus.2017.02.003
R Core Team. (2019). R: A language and environment for statistical computing. R Foundation
for Statistical Computing. Vienna. Retrieved from https://www.R-project.org
Ramscar, M., & Baayen, R. H. (2013). Production, comprehension, and synthesis: A
communicative perspective on language. Frontiers in Psychology,4, 233.
doi:10.3389/fpsyg.2013.00233
Ramscar, M., Dye, M., Blevins, J. P., & Baayen, R. H. (2018). Morphological development.
In A. Bar On & D. Ravid (Eds.), Handbook of communication disorders (pp. 181–202).
doi:10.1515/9781614514909-010
Ramscar, M., Dye, M., & McCauley, S. M. (2013). Error and expectation in language
learning: The curious absence of mouses in adult speech. Language,89(4), 760–793.
doi:10.1353/lan.2013.0068
AFFECT ACROSS ADULTHOOD 59
Ramscar, M., Hendrix, P., Love, B. C., & Baayen, R. H. (2013). Learning is not decline: The
mental lexicon as a window into cognition across the lifespan. The Mental Lexicon,8,
450–481. doi:10.1075/ml.8.3.08ram
Ramscar, M., Hendrix, P., Shaoul, C., Milin, P., & Baayen, R. H. (2014). The myth of
cognitive decline: Non-linear dynamics of lifelong learning. Topics in Cognitive
Science,6(1), 5–42. doi:10.1111/tops.12078
Ratcliff, R., Thapar, A., Gomez, P., & McKoon, G. (2004). A diffusion model analysis of the
effects of aging in the lexical-decision task. Psychology and Aging,19(2), 278–289.
doi:10.1037/0882-7974.19.2.278
Rayner, K., Castelhano, M. S., & Yang, J. (2009). Eye movements and the perceptual span
in older and younger readers. Psychology and Aging,24(3), 755–760.
doi:10.1037/a0014300
Rayner, K., Reichle, E. D., Stroud, M. J., Williams, C. C., & Pollatsek, A. (2006). The effect of
word frequency, word predictability, and font difficulty on the eye movements of
young and older readers. Psychology and Aging,21(3), 448–465.
doi:10.1037/0882-7974.21.3.448
Reed, A. E., & Carstensen, L. L. (2012). The theory behind the age-related positivity effect.
Frontiers in Psychology,3. doi:10.3389/fpsyg.2012.00339
Reed, A. E., Chan, L., & Mikels, J. A. (2014). Meta-analysis of the age-related positivity
effect: Age differences in preferences for positive over negative information.
Psychology and Aging,29(1), 1–15. doi:10.1037/a0035194
Rescorla, R. A. (1988). Pavlovian conditioning: It’s not what you think it is. American
Psychologist,43(3), 151–160. doi:10.1037/0003-066X.43.3.151
Rescorla, R. A. (2001). Unequal associative changes when excitors and neutral stimuli are
conditioned in compound. The Quarterly Journal of Experimental Psychology Section B,
54(1b), 53–68. doi:10.1080/713932743
AFFECT ACROSS ADULTHOOD 60
Rescorla, R. A., & Wagner, A. R. (1972). A theory of pavlovian conditioning: Variations in
the effectiveness of reinforcement and nonreinforcement. In A. H. Black &
W. F. Prokasy (Eds.), Classical conditioning II: Current research and theory (Vol. 2,
pp. 64–99). New York: Appleton-Century-Crofts.
Ruffman, T., Henry, J. D., Livingstone, V., & Phillips, L. H. (2008). A meta-analytic review
of emotion recognition and aging: Implications for neuropsychological models of
aging. Neuroscience & Biobehavioral Reviews,32(4), 863–881.
doi:10.1016/j.neubiorev.2008.01.001
Salthouse, T. A. (1991). Theoretical perspectives on cognitive aging. Hillsdale, NJ: Lawrence
Erlbaum Associates, Inc.
Salthouse, T. A. (1996). The processing-speed theory of adult age differences in cognition.
Psychological Review,103(3), 403–428. doi:10.1037/0033-295X.103.3.403
Salthouse, T. A. (2009). When does age-related cognitive decline begin? Neurobiology of
Aging,30(4), 507–514. doi:10.1016/j.neurobiolaging.2008.09.02
Salthouse, T. A. (2012). Consequences of age-related cognitive declines. Annual Review of
Psychology,63, 201–226. doi:10.1146/annurev-psych-120710-100328
Samanez-Larkin, G. R., & Knutson, B. (2015). Decision making in the ageing brain:
Changes in affective and motivational circuits. Nature Reviews Neuroscience,16(5),
278–289. doi:10.1038/nrn3917
Sap, M., Park, G., Eichstaedt, J., Kern, M., Stillwell, D., Kosinski, M., . .. Schwartz, H. A.
(2014). Developing age and gender predictive lexica over social media. In Proceedings
of the 2014 conference on empirical methods in natural language processing (EMNLP)
(pp. 1146–1151).
Scheibe, S., & Carstensen, L. L. (2010). Emotional aging: Recent findings and future trends.
The Journals of Gerontology: Series B,65(2), 135–144. doi:10.1093/geronb/gbp132
Schwartz, H. A., Eichstaedt, J. C., Kern, M. L., Dziurzynski, L., Ramones, S. M.,
Agrawal, M., . . . Ungar, L. H. (2013). Personality, gender, and age in the language of
AFFECT ACROSS ADULTHOOD 61
social media: The open-vocabulary approach. PLoS ONE,8(9), e73791.
doi:10.1371/journal.pone.0073791
Searle, S. R., Speed, F. M., & Milliken, G. A. (1980). Population marginal means in the
linear model: An alternative to least squares means. The American Statistician,34(4),
216–221.
Shafto, M. A., & Tyler, L. K. (2014). Language in the aging brain: The network dynamics of
cognitive decline and preservation. Science,346(6209), 583–587.
doi:10.1126/science.1254404
Al-Shawaf, L., Conroy-Beam, D., Asao, K., & Buss, D. M. (2015). Human emotions: An
evolutionary psychological perspective. Emotion Review,8(2), 173–186.
doi:10.1177/1754073914565518
Shimamura, A. P., Berry, J. M., Mangels, J. A., Rusting, C. L., & Jurica, P. J. (1995). Memory
and cognitive abilities in university professors: Evidence for successful aging.
Psychological Science,6(5), 271–277.
Shmueli, G. (2010). To explain or to predict? Statistical Science,25(3), 289–310.
doi:10.1214/10-STS330
Siegel, S., & Allan, L. G. (1996). The widespread influence of the rescorla-wagner model.
Psychonomic Bulletin & Review,3(3), 314–321. doi:10.3758/BF03210755
Singer, T., Verhaeghen, P., Ghisletta, P., Lindenberger, U., & Baltes, P. B. (2003). The fate of
cognition in very old age: Six-year longitudinal findings in the Berlin Aging Study
(BASE). Psychology and Aging,18(2), 318. doi:10.1037/0882-7974.18.2.318
Söderholm, C., Häyry, E., Laine, M., & Karrasch, M. (2013). Valence and arousal ratings
for 420 Finnish nouns by age and gender. PLoS ONE,8(8), e72859.
doi:10.1371/journal.pone.0072859
Solan, H. A., Feldman, J., & Tujak, L. (1995). Developing visual and reading efficiency in
older adults. Optometry and Vision Science,72(2), 139–145.
doi:10.1097/00006324-199502000-00012
AFFECT ACROSS ADULTHOOD 62
Stadthagen-Gonzalez, H., Imbault, C., Sánchez, M. A. P., & Brysbaert, M. (2017). Norms of
valence and arousal for 14,031 Spanish words. Behavior Research Methods,49(1),
111–123. doi:10.3758/s13428-015-0700-2
Straka, M., & Straková, J. (2017). Tokenizing, POS tagging, lemmatizing and parsing UD
2.0 with udpipe. In Proceedings of the conll 2017 shared task: Multilingual parsing from
raw text to universal dependencies (pp. 88–99). doi:10.18653/v1/K17-3009
Stubbe, R. (2012). Do pseudoword false alarm rates and overestimation rates in Yes/No
vocabulary tests change with Japanese university students’ English ability levels?
Language Testing,29(4), 471–488. doi:10.1177/0265532211433033
Thornton, R., & Light, L. L. (2006). Language comprehension and production in normal
aging. In J. E. Birren & K. W. Schaie (Eds.), Handbook of the psychology of aging
(pp. 261–287). doi:10.1016/B978-012101264-9/50015-X
Tukey, J. W. (1994). The problem of multiple comparisons. In H. I. Braun (Ed.), The collected
works of John W. Tukey (Vol. 8, pp. 1–300). 1953. New York: Chapman and Hall.
Urry, H. L. (2016). Resources for emotion regulation in older age: Linking cognitive
resources with cognitive reappraisal. In A. D. Ong & C. E. Löckenhoff (Eds.),
Bronfenbrenner series on the ecology of human development. Emotion, aging, and health
(pp. 51–59). doi:10.1037/14857-004
Verhaeghen, P. (2003). Aging and vocabulary score: A meta-analysis. Psychology and Aging,
18(2), 332–339. doi:10.1037/0882-7974.18.2.332
Verhaeghen, P. (2011). Aging and executive control: Reports of a demise greatly
exaggerated. Current Directions in Psychological Science,20(3), 174–180.
doi:10.1177/0963721411408772
Vigliocco, G., Ponari, M., & Norbury, C. (2018). Learning and processing abstract words
and concepts: Insights from typical and atypical development. Topics in Cognitive
Science,10(3), 533–549. doi:10.1111/tops.12347
AFFECT ACROSS ADULTHOOD 63
Warriner, A. B., & Kuperman, V. (2015). Affective biases in English are bi-dimensional.
Cognition and Emotion,29(7), 1147–1167. doi:10.1080/02699931.2014.968098
Warriner, A. B., Kuperman, V., & Brysbaert, M. (2013). Norms of valence, arousal, and
dominance for 13,915 English lemmas. Behavior Research Methods,45(4), 1191–1207.
doi:10.3758/s13428-012-0314-x
Whitford, V., & Titone, D. (2017). The effects of word frequency and word predictability
during first-and second-language paragraph reading in bilingual older and younger
adults. Psychology and Aging,32(2), 158–177. doi:10.1037/pag0000151
Wurm, L. H. (2007). Danger and usefulness: An alternative framework for understanding
rapid evaluation effects in perception? Psychonomic Bulletin & Review,14(6),
1218–1225. doi:10.3758/BF03193116
Yap, M. J., & Balota, D. A. (2009). Visual word recognition of multisyllabic words. Journal
of Memory and Language,60(4), 502–529. doi:10.1016/j.jml.2009.02.001
Yarkoni, T., & Westfall, J. (2017). Choosing prediction over explanation in psychology:
Lessons from machine learning. Perspectives on Psychological Science,12(6), 1100–1122.
doi:10.1177/174569161769339
Zuur, A. F., Ieno, E. N., Walker, N. J., Saveliev, A. A., & Smith, G. M. (2009). Mixed effects
models and extensions in ecology with R. New York: Springer.
Running head: AFFECT ACROSS ADULTHOOD 1
Affect across adulthood: Supplementary materials
Aki-Juhani Kyröläinen
McMaster University and Brock University
Emmanuel Keuleers
Tilburg University
Paweł Mandera
Ghent University
Marc Brysbaert
Ghent University
Victor Kuperman
McMaster University
AFFECT ACROSS ADULTHOOD 2
Affect across adulthood: Supplementary materials
Supplementary materials S1: Bootstrap analysis of type frequency and valence
Logically, type frequency can be largely independent from token frequency in
language. However, based on an extensive literature on quantitative text analysis (see
Baayen, 2001, for example), it has been demonstrated that changes in token and type
frequency associated with words occur together in naturally produced text. Importantly,
this relationship between type and token frequency was also demonstrated in the status
update data set, as the youngest adult group produced the highest number of both types
and tokens. In order to control for this relationship, we carried out a bootstrap resampling
procedure where the affective word types were randomly sampled with replacement. The
sample size was held constant across the five age groups and the size of the affective
vocabulary was set to V= 6000, representing approximately 70% of the affective word
types produced among the oldest adult group. The results of this bootstrap procedure are
given in Table 1.
Table 1
Distribution of the average valence score of the word types in English for the five age groups across
the bootstrap runs, s= 1000, with 95% percentile confidence intervals.
Valence
Age group MLower CI Upper CI
24–29 5.07 5.05 5.09
30–39 5.08 5.05 5.10
40–49 5.09 5.06 5.11
50–59 5.13 5.10 5.15
60+ 5.21 5.19 5.23
The results of the bootstrap procedure demonstrated that even when the vocabulary
AFFECT ACROSS ADULTHOOD 3
size was held constant among the five age groups the average valence rating increased
with advancing age. It is noteworthy that the age-related valence-trajectory displayed a
marked increase after the age of 50 and onwards as the 95% confidence intervals did not
overlap between the youngest and the two oldest adult groups. Together these results
provided strong evidence that advancing age goes hand in hand with the increased use of
more diverse affective word types.
Supplementary materials S2: Type frequency and base rate
To rule out the possibility that age-related type frequency of positivity bias arose due
to differences in the size of the lexicon, thus being unrelated to age, we extracted
documents form the Global Web-based English (GloWbE) corpus covering American and
British English. The documents were tokenized and lemmatized following the Universal
Dependency Schema. This collection of documents amounted to a non-age-specific corpus
consisting of approximately 790 million lemmata. These lemmata were matched with the
valence ratings to form an affective lexicon. The size of the corpus allowed us to
investigate the relation between the proportion of positive word types out of total word
types (i.e., positivity bias) and size of the corpus.
To mimic the growth of the lexicon, the corpus was divided into chunks of 10,000
word tokens each. We considered samples that included the first chunk, the first two
chunks, the first three chunks, etc. The word-type positivity bias was determined by
calculating the proportion of types in a sample that had a valence rating greater than five.
The results are visualized in Figure 1.
AFFECT ACROSS ADULTHOOD 4
0.50
0.55
0.60
0.65
0.70
0.75
0.80
10,000
20,000
30,000
40,000
50,000
60,000
70,000
80,000
90,000
100,000
Corpus size
Positivity bias (proportion)
Figure 1. Type frequency of positivity bias in the affective lexicon as a function of corpus
size.
The results replicated a well documented finding that positive types were used more
often overall than negative ones. Critically, the results also demonstrated that the
positivity bias, or the proportion of positive word types out of total, was reduced in a
sample when its size increased. In the smallest sample of 10,000 tokens, positive word
types (with valence ratings above 5) accounted for roughly 74% of all words that we have
valence ratings for, while in the largest sample of 100,000 tokens this proportion went
down to about 64%. This offered evidence that the age-related type frequency effect of
positivity bias reported in Study 1 was unlikely to be driven by the growth of the lexicon.
Supplementary materials S3: Model comparisons
The results of the model comparisons of the RTs in the English data set are provided
in Table 2.
AFFECT ACROSS ADULTHOOD 5
Table 2
Summary information of the model comparisons. Part A compares model fits between a three-way
interaction and two two-way interactions models reporting log-likelihood, number of parameters
(K), AIC and AIC. Part B compares model fits between a simple main effect of age and valence
and two two-way interaction models reporting log-likgehood, number of parameters (K), AIC and
AIC.
Lexical coverage: 30% Lexical coverage: 50% Lexical coverage: 100%
Part A: interaction of age and valence Log-likelihood KAIC AIC Log-likelihood KAIC AIC Log-likelihood KAIC AIC
Three-way interactionmodel: age * valence * length 1.878375 ×10428 3.7509502 ×104NA 3.091567 ×10428 6.1773332 ×104NA 6.111065 ×10428 1.2216331 ×105NA
Twotwo-way interaction model: age * valence + valence * length 1.879907 ×10421 3.7554141 ×10445 3.092843 ×10421 6.181286 ×10440 6.110881 ×10421 1.2217362 ×10510
Part B: main effect of age and valence
Main effect model: age + valence + valence * length 1.879109 ×10417 3.754618 ×104NA 3.092027 ×10417 6.1804546 ×104NA 6.109475×10417 1.2215349 ×105NA
Twotwo-way interaction model: age * valence + valence * length 1.879907 ×10421 3.7554141 ×10483.092843 ×10421 6.181286 ×10486.110881 ×10421 1.2217362 ×10520
Supplementary materials S4: Lexical properties of the affective words in English
One of the central aspects of lexical properties associated with words is that they
tend to co-vary. Even the concept of positivity bias builds on the idea that positive
connotation co-varies, at least partly, with frequency. The rationale of predicting whether a
given word is associated with negative or positive connotation from other lexical
properties builds on this background. This type of analysis will be informative regarding
the variables that co-vary with valence and also the extent to which these variables
discriminate between positive and negative words. If the affective connotation associated
with a particular word is highly predictable, it becomes difficult to disentangle the
contribution of affect from other lexical properties associated with a particular word.
Conversely, if this is not the case, this offers evidence that valence provides a unique
contribution to the semantics of a word.1
1
Models that provide high explanatory power do not necessarily provide high predictive power, especially,
if the data are characterized by highly complex and/or non-linear interactional relations (see Breiman, 2001b;
Shmueli, 2010; Yarkoni & Westfall, 2017, for discussion). Random forests, as implemented in R, were also
fitted to the data (Breiman, 2001a) because they have demonstrated high predictive performance in various
classification tasks (see Fernández-Delgado, Cernadas, Barro, & Amorim, 2014, for evaluation of 179
AFFECT ACROSS ADULTHOOD 6
We fitted a logistic regression model to the data using the following lexical variables
as predictors for valence: length, frequency (log), orthographic neighborhood,
phonological neighborhood, bigram frequency, number of phonemes, number of syllables,
number of morphemes, age of acquisition, and concreteness. The models were fitted
separately for each of the lexical coverage groups. The summary information of the fitted
models is reported in Table 3.
Table 3
Results of the logistic regression models fitted on the three different lexical coverage groups. The
estimated coefficients represent the log-odds for positive words.
Lexical coverage: 30% Lexical coverage: 50% Lexical coverage: 100%
Intercept 1.82 (0.77)
0.05 (0.57) 0.10 (0.40)
Length 0.14 (0.05)∗∗ 0.06 (0.04) 0.06 (0.03)
Frequency (log) 0.22 (0.04)∗∗∗ 0.23 (0.03)∗∗∗
0.17 (0.02)∗∗∗
Orthographic neighborhood 0.17 (0.11) 0.08 (0.09) 0.04 (0.06)
Phonological neighborhood 0.19 (0.09)0.13 (0.07) 0.10 (0.05)
Bigram frequency 0.08 (0.09) 0.09 (0.07) 0.07 (0.05)
Number of phonemes 0.09 (0.05) 0.08 (0.04)0.07 (0.03)∗∗
Number of syllables 0.35 (0.07)∗∗∗ 0.33 (0.06)∗∗∗
0.24 (0.04)∗∗∗
Number of morphemes 0.33 (0.07)∗∗∗
0.24 (0.05)∗∗∗ 0.21 (0.04)∗∗∗
Age of acquisition 0.28 (0.02)∗∗∗
0.25 (0.01)∗∗∗ 0.17 (0.01)∗∗∗
Concreteness 0.02 (0.04) 0.06 (0.03)
0.03 (0.02)
∗∗∗p < 0.001,∗∗ p < 0.01,p < 0.05
The results of the fitted models indicated that only a handful of lexical properties
associated with the affective words were estimated to have some degree of cue-validity in
discriminating between positive and negative words, namely frequency, number of
classifiers). However, for these data, Random forests effectively did not outperform the logistic regression
models; for example, the average AUC in the 30% lexical coverage group was .73 (95% CIs = .67, .78).
AFFECT ACROSS ADULTHOOD 7
syllables, number of morphemes, and age of acquisition. These lexical properties were
statistically significant across the three lexical coverage groups.
In order to test the predictibtability of valence, we used a 10-fold cross validation to
split the data into training and testing sets. Additionally, to evalute the predictive
performance of the logistic regression model we calculated Area Under The Curve (AUC).
AUC is one of the most commonly used measures to assess predictive performance of a
model in machine learning and medical research. In the case of valence, an AUC value of 1
indicates a perfect separation between the positive and negative words and a value of .5
indicates chance level performance. Additionally, AUC has a probabistilic interpretation;
namely, given a pair of words, what is the probability of ranking a random word with a
positive connotation higher than a negative one. This procedure was carried out separately
for each of the lexical coverage groups. The results are presented Table 4.
The results indicated that the predictibtability of valence was fairly low and became
worse as the size of the affective lexicon was increased from 30% to 100% coverage. While
certain lexical properties have cue-validity their ability to discriminate between positive
and negative words was fairly low.
Table 4
Average performance of the logistic regression models discriminating between the positive and
negative words in the three different lexical coverage groups.
AUC
Lexical coverage group MLower CI Upper CI
30% coverage .73 .67 .78
50% coverage .71 .67 .75
100% coverage .65 0.6 .66
Interestingly, the predictive performance of the models demonstrated that while
AFFECT ACROSS ADULTHOOD 8
lexical properties associated with the affective lexicon had a certain degree of
cue-validity—for example, frequency was a statistically significant predictor—the lexical
properties appeared to provide only a minimal amount of information in terms of
discriminating words into positive and negative. This finding indicates that the affective
connotation of a word carries its own unique fingerprint that is likely not to be masked by
other lexical properties that known to typically modulate language processing.
AFFECT ACROSS ADULTHOOD 9
References
Baayen, R. H. (2001). Word frequency distributions. Dordrecht: Kluwer Academic Publishers.
Breiman, L. (2001a). Random forests. Machine Learning,45(1), 5–32.
Breiman, L. (2001b). Statistical modeling: The two cultures. Statistical Science,16(3),
199–231.
Fernández-Delgado, M., Cernadas, E., Barro, S., & Amorim, D. (2014). Do we need
hundreds of classifiers to solve real world classification problems? Journal of Machine
Learning Research,15(1), 3133–3181.
Shmueli, G. (2010). To explain or to predict? Statistical Science,25(3), 289–310.
doi:10.1214/10-STS330
Yarkoni, T., & Westfall, J. (2017). Choosing prediction over explanation in psychology:
Lessons from machine learning. Perspectives on Psychological Science,12(6), 1100–1122.
doi:10.1177/174569161769339
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Brain training is a large and expanding industry, and yet there is a recurrent and ongoing debate concerning its scientific basis or evidence for efficacy. Much of evidence for the efficacy of brain training within this debate is from small-scale studies that do not assess the type of “brain training,” the specificity of transfer effects, or the length of training required to achieve a generalized effect. To explore these factors, we analyze cross-sectional data from two large Internet-cohort studies (total N = 60,222) to determine whether cognition differs at the population level for individuals who report that they brain train on different devices, and across different timeframes, with programs in common use circa 2010–2013. Examining scores for an assessment of working-memory, reasoning and verbal abilities shows no cognitive advantages for individuals who brain train. This contrasts unfavorably with significant advantages for individuals who regularly undertake other cognitive pursuits such as computer, board and card games. However, finer grained analyses reveal a more complex relationship between brain training and cognitive performance. Specifically, individuals who have just begun to brain train start from a low cognitive baseline compared to individuals who have never engaged in brain training, whereas those who have trained for a year or more have higher working-memory and verbal scores compared to those who have just started, thus suggesting an efficacy for brain training over an extended period of time. The advantages in global function, working memory, and verbal memory after several months of training are plausible and of clinically relevant scale. However, this relationship is not evident for reasoning performance or self-report measures of everyday function (e.g., employment status and problems with attention). These results accord with the view that although brain training programs can produce benefits, these might extend to tasks that are operationally similar to the training regime. Furthermore, the duration of training regime required for effective enhancement of cognitive performance is longer than that applied in most previous studies.
Article
Full-text available
We present a new database of Dutch word recognition times for a total of 54 thousand words, called the Dutch Crowdsourcing Project. The data were collected with an internet vocabulary test. The database is limited to native Dutch speakers. Participants were asked to indicate which words they knew. Their response times were registered, even though the participants were not asked to respond as fast as possible. Still, the response times correlate around .7 with the response times of the Dutch Lexicon Projects for shared words. Also results of virtual experiments indicate that the new response times are a valid addition to the Dutch Lexicon Projects. This not only means that we have useful response times for some 20 thousand extra words, but we now also have data on differences in response latencies as a function of education and age. The new data correspond better to word use in the Netherlands.
Article
Full-text available
We present a new dataset of English word recognition times for a total of 62 thousand words, called the English Crowdsourcing Project. The data were collected via an internet vocabulary test, in which more than one million people participated. The present dataset is limited to native English speakers. Participants were asked to indicate which words they knew. Their response times were registered, although at no point were the participants asked to respond as fast as possible. Still, the response times correlate around .75 with the response times of the English Lexicon Project for the shared words. Also results of virtual experiments indicate that the new response times are a valid addition to the English Lexicon Project. This not only means that we have useful response times for some 35 thousand extra words, but we now also have data on differences in response latencies as a function of education and age.
Article
Full-text available
Depression, the most prevalent mental illness, is underdiagnosed and undertreated, highlighting the need to extend the scope of current screening methods. Here, we use language from Facebook posts of consenting individuals to predict depression recorded in electronic medical records. We accessed the history of Facebook statuses posted by 683 patients visiting a large urban academic emergency department, 114 of whom had a diagnosis of depression in their medical records. Using only the language preceding their first documentation of a diagnosis of depression, we could identify depressed patients with fair accuracy [area under the curve (AUC) = 0.69], approximately matching the accuracy of screening surveys benchmarked against medical records. Restricting Facebook data to only the 6 months immediately preceding the first documented diagnosis of depression yielded a higher prediction accuracy (AUC = 0.72) for those users who had sufficient Facebook data. Significant prediction of future depression status was possible as far as 3 months before its first documentation. We found that language predictors of depression include emotional (sadness), interpersonal (loneliness, hostility), and cognitive (preoccupation with the self, rumination) processes. Unobtrusive depression assessment through social media of consenting individuals may become feasible as a scalable complement to existing screening and monitoring procedures.
Article
Full-text available
We present word prevalence data for 61,858 English words. Word prevalence refers to the number of people who know the word. The measure was obtained on the basis of an online crowdsourcing study involving over 220,000 people. Word prevalence data are useful for gauging the difficulty of words and, as such, for matching stimulus materials in experimental conditions or selecting stimulus materials for vocabulary tests. Word prevalence also predicts word processing times, over and above the effects of word frequency, word length, similarity to other words, and age of acquisition, in line with previous findings in the Dutch language.
Article
Standard neurocognitive models of language processing have tended to obviate the need for incorporating emotion processes, while affective neuroscience theories have typically been concerned with the way in which people communicate their emotions, and have often simply not addressed linguistic issues. Here, we summarise evidence from temporal and spatial brain imaging studies that have investigated emotion effects on lexical, semantic and morphosyntactic aspects of language during the comprehension of single words and sentences. The evidence reviewed suggests that emotion is represented in the brain as a set of semantic features in a distributed sensory, motor, language and affective network. Also, emotion interacts with a number of lexical, semantic and syntactic features in different brain regions and timings. This is in line with the proposals of interactive neurocognitive models of language processing, which assume the interplay between different representational levels during on-line language comprehension.
Article
Research on cognitive aging demonstrates age-related cognitive decline. Education is a protective factor against cognitive decline, but few studies have examined the cognitive development of highly educated individuals. This study compared the cognitive performance and intellectual engagement of retired professors (N = 47, M age = 72.9) and individuals with average education (N = 236, M age = 72.7) over 5 years. Although the highly educated sample showed better performance in perceptual speed and working memory, cognitive performance was rather stable over time in both samples. Interestingly, high intellectual engagement enabled individuals with average education to keep up with the performance of the highly educated sample on perceptual speed. These findings raise the question whether intellectual engagement is more beneficial than years of education in perceptual speed.
Book
Cambridge Core - Sociolinguistics - Register Variation Online - by Douglas Biber