Structural differences in the semantic networks of younger and older adults
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Structural differences in the semantic networks of younger and older adults Dirk U. Wulff1,2 , Thomas T. Hills3 , and & Rui Mata1,2 1 University of Basel 2 Max Planck Institute for Human Development 3 University of Warwick Cognitive science invokes semantic networks to explain diverse phenomena, from memory retrieval to creativity. Research in these areas often assumes a single underlying semantic network that is shared across individuals. Yet, recent evidence suggests that content, size, and connectivity of semantic networks are experience-dependent, implying sizable individual and age-related differences. Here, we investigate individual and age differences in the semantic networks of younger and older adults by deriving semantic networks from both fluency and similarity rating tasks. Crucially, we use a mega-study approach to obtain thousands of sim- ilarity ratings per individual to allow us to capture the characteristics of individual semantic networks. We find that older adults possess lexical networks with smaller average degree and longer path lengths relative to those of younger adults, with older adults showing less inter- individual agreement and thus more unique lexical representations relative to younger adults. Furthermore, this approach shows that individual and age differences are not evenly distributed but, rather, are related to weakly-connected, peripheral parts of the networks. All in all, these results reveal the inter-individual differences in both the content and structure of semantic net- works which may accumulate across the life span as a function of idiosyncratic experiences. Keywords: semantic networks, cognitive aging, mental lexicon Introduction making (Bhatia, 2019) using large-scale word vector spaces and free-association networks. However, general theories of Semantic networks are the representational basis of our learning and development (Ramscar et al., 2014; Ramscar cognitive system (Baronchelli et al., 2013; Beer, 2000; et al., 2017), as well as empirical findings (Benedek et al., Borge-Holthoefer & Arenas, 2010) and an integral part of 2017; Dubossarsky et al., 2017; Morais et al., 2013), sug- prominent models of memory (Anderson, 1983), reasoning gest that semantic networks could vary considerably between (Collins & Loftus, 1975), and creativity (Beaty et al., 2018; individuals and across the life span. Crucially, researchers Kenett et al., 2018). Past work has often make the simpli- now seem to agree that understanding experience-dependent fying assumption that a common semantic network can be changes and individual variation in cognition is an important used to understand human semantic cognition (Anderson, frontier for the science of aging (Lindenberger, 2014). 1983; Collins & Loftus, 1975; Hills et al., 2012; Jones et Aging research has made significant progress in the past al., 2015). This assumptions is implicit, for instance, in ef- decades quantifying age-related changes in semantic cogni- forts to model retrieval from memory (Wulff et al., 2021), tion, including large increases in the size of the knowledge- judgments of relatedness (Kraemer et al., 2021), or decision store across adult development, perhaps best documented in the large differences in vocabulary size in older relative to younger adults (Verhaeghen, 2003). More recently, however, Dirk U. Wulff https://orcid.org/0000-0002-4008-8022 research suggests that individual learning and life span de- Thomas T. Hills https://orcid.org/0000-0003-3842-2076 Rui velopment can also lead to changes in the structure of human Mata https://orcid.org/0000-0002-1679-906X knowledge (Cosgrove et al., 2021; Nation, 2017; Wulff et We are grateful to Laura Wiles for editing the manuscript. This al., 2019). For example, recent efforts have used data from work was supported by a grant from the Swiss Science Foundation large scale free-association studies to show that older adults’ (100015_197315) to Dirk U. Wulff. semantic networks are less connected and efficient relative Correspondence concerning this article should be addressed to those of younger adults (Dubossarsky et al., 2017; Wulff to Dirk U. Wulff, Department of Psychology, University of Basel, Missionsstrasse 60-62, 4055 Basel, Switzerland. E-mail: et al., 2021). dirk.wulff@gmail.com Quantifying individual and age differences in the size and
2 WULFF, HILLS, AND MATA structure of human knowledge is important because this may represent a missing link in understanding age-related decline in several aspects of cognitive functioning. Older adults tend to perform worse on a broad set of cognitive tasks, and such findings are commonly attributed to a decline in fluid cog- nitive abilities (Healey & Kahana, 2016; Salthouse, 2010). However, some have argued that changes in the underlying size and structure of representations can contribute to age differences in cognitive performance, for example, due to activation-spreading across many targets in memory (fan ef- fect; Buchler & Reder, 2007)) or difficulties in discrimina- tion learning between many similar items (Ramscar et al., 2017). One first step needed to understand the contribution of se- mantic networks to age differences in cognitive performance is to document the changes in the size and structure of seman- tic networks across the life span. In the present study, we Figure 1 seek to describe potential life span differences in the struc- ture of semantic networks by making two novel empirical Methodological approach. Panel A illustrates the two contributions (see Figure 1). First, we investigate age dif- steps, edge inclusion and filtering, involved in inferring ferences in the size and semantic network structure for ag- networks from verbal fluency sequences. For details see gregates of younger and older groups obtained from verbal Materials and Methods. The resulting network is based fluency data (e.g., “Name all animals you can”) under dif- on 142 sequences of the older adults’ group of study 1. ferent conditions: Our analyses of the semantic structure of To simplify the visualization more conservative inferences verbal fluency productions are the first to include age com- settings were employed than used in the analyses reported parisons using different categories (animal vs. country) to below. Panel B illustrates the creation of networks from document whether structural differences generalize across similarity ratings by normalizing individual’s responses to domains; further, we investigate age differences in verbal the range of 0 and 1. The weighted network is based on the fluency under different retrieval time allowances (1 minute average ratings of the older adults’ group of study 3. vs. 10 minutes), which allow us to assess the idea that older adults can catch-up or even outperform younger adults in ver- bal fluency when given the opportunity to search their poten- (e.g., Alzheimer’s disease; Kalbe et al., 2004). In recent tially larger semantic stores. Second, we adopt a megastudy years, however, research has begun to analyze verbal fluency approach (Keuleers & Balota, 2018) to provide the first com- data in novel ways to extract from them semantic networks parison of younger and older adults’ semantic networks at (Henry et al., 2004) and understand individual and age dif- the level of individuals in a study involving over 2,000 sim- ferences in semantic cognition (Cosgrove et al., 2021; Zemla ilarity ratings from each participant. This is crucial, given & Austerweil, 2018). Such approaches leverage the fact that that aggregate networks likely do not accurately reflect the the proximity of elements within the sequence of responses structure of individual level networks. Our work thus con- should reveal information on whether two elements are con- tributes to mapping the structural differences in the semantic nected in an underlying semantic network. Several algo- networks of younger and older adults by considering differ- rithms utilizing this principle have been proposed. To infer ent elicitation tasks (fluency, similarity) using both aggregate semantic networks of younger and older adults, we rely on and individual semantic networks. a random walk plus filtering algorithm, which was recently found to predict human behavior better than other algorithms Results (Zemla & Austerweil, 2018). In the first step, this algorithm adds to a single network for each age group edges for every Age-related differences in fluency networks pair of elements that occurred less than two positions apart from each other across all verbal fluency sequences of the Verbal fluency is a neuropsychological test that requires age group. Then, in a second step, all edges in the network participants to retrieve as many elements as possible from a that were added only once across all sequences or were less natural category (Bousfield, 1953), say, animals, within in frequent than expectation derived from random behavior are a given amount of time. Verbal fluency tasks are typically removed (see Figure 1A). Previous research has found this employed to measure fluid cognitive abilities, for instance, approach to produce plausible networks that predict human in screening instruments for age-related cognitive pathology
INSERT SHORTTITLE COMMAND IN PREAMBLE 3 behavior better than other network inference methods for flu- ency data (Goñi et al., 2011; Zemla & Austerweil, 2018). We compared semantic networks of younger and older Wulff et al. (2016) − Animals Younger adults Wulff et al. (2016) − Animals Older adults caterpillar sea_lion adults on the basis of four verbal fluency data sets, stemming fly puma leopard cheetah muskrat water_buffalo from published work (Wulff et al., 2016) and two new studies lizard toad butterfly seal dolphin mosquito polar_bear lion penguin hyena shark panther crocodile walrus frog cockroach bluebird octopus cougar turtle rhinoceros fish (see Methods for details). Table 1 provides an overview of animal gopher pony ant mule llama weasel gerbil worm hamster horse dog goldfish octopus squid blue_jay mole possum wolf elk tiger bear otter gazelle alligator whale gorilla ape orangutan puppy owl crow boar parrot bobcat the datasets. Following previous work (Dubossarsky et al., calf donkey dolphin rattlesnake robin kitten ram pig ostrich gerbil snake sheep cat eagle sparrow iguana lamb woodchuck snail beaver zebra beaver otter porpoise bull goat spider raccoon baboon turtle chicken cow insect gila_monster coyote monkey black_bear porcupine rooster tortoise tarantula bird 2017; Wulff et al., 2016; Zortea et al., 2014), we compared goose deer armadillo lizard elephant rattlesnake opossum fox ferret duck frog chimpanzee eel squirrel giraffe mosquito hen snake lobster chimp pigeon skunk hippopotamus fly bison bird emu rabbit antelope turkey crab eagle robin crow armadillo kangaroo moose whale mouse spider younger and older adults’ networks with respect to three squirrel rat guinea_pig rabbit reindeer elk ox buffalo antelope alligator fish rhinoceros orangutan camel shark zebra gazelle hawk owl dinosaur gopher hog buffalo chipmunk llama camel goat bison rat worm turkey human aardvark butterfly snail monkey chimpanzee cat macroscopic network measures: average degree (connectiv- chipmunk donkey elephant seal ostrich ant marmot coyote ape human ox pig goose beetle moose chicken bee crocodile chimp skunk lion sheep mole deer puma fox baboon duck porcupine aardvark lamb cockroach ity, hki), average local clustering coefficient (structuredness, muskrat gorilla emu hawk wolf bear cow lemur kangaroo hippopotamus possum mountain_lion meerkat pony horse tiger giraffe panda grizzly_bear penguin leopard mule dog anteater koala canary raccoon bull rooster badger hen polar_bear mouse mare bobcat hyena ram C), and average shortest path length (efficiency, L). These walrus cheetah cattle platypus wallaby jaguar cougar swan flamingo lemur panther calf philly colt metrics are frequently employed to characterize the structure Study 1 − Animals Study 1 − Animals Younger adults Kanarienvogel Taufliege Raupe Kakerlake Tausendfuesser Older adults Weisser Hai Tigerhai Wildkatze Grille Schnecke Seepferdchen of cognitive networks and have been successfully linked to Staffordshireterrier Laus Floh Kolibri Wildpferd Goldhamster Rebhuhn Raupe Seestern Nacktmull Zecke Fledermaus Heuschrecke Schmeissfliege Kellerassel Hausschwein Schwein Wanze Kartoffelkaefer Terrier Flughund Motte Pferdebremse Kakadu Schnake Laus Hummel Terrier Henne Ziege Libelle Mistkaefer Wiesel Marder Hase MueckeGottesanbeterin Stechmuecke Regenwurm Uhu Wespe Mops Fohlen Schmetterling Made Waschbaer Bremse Feuerkaefer Kartoffelkaefer Windhund Kaninchen Auerhuhn Nashorn Fliege Pekinese Spitz Insekt Junikaefer Streifenhoernchen Marienkaefer WurmJunikaefer Lamm Eber Ferkel Biber Dachs Ameise Warzenschwein Collie Hahn Meerschweinchen Blaumeise various measures of cognitive performance (for reviews, see Stachelschwein Wombat Spitzmaus Elch Nasenbaer Igel Mistkaefer Otter Bisamratte Fuchs Bueffel Wolf Koala Chamaeleon Kaenguru Faultier Biene Maulwurf Vogel Geier Eule Spatz Grashuepfer Blattlaus Insekt Maikaefer Hornisse Meise Termite Rabe Kuckuck Kohlmeise Kraehe Rehkitz WasserbueffelKalb Steinbock Schaf Murmeltier Pudel Boxer Bernhardiner Dackel Pferd Kuh Maus Dogge Schaeferhund Huhn Fuchs Hirsch Luchs Wildschwein Maulwurf Fischotter Biber Bisamratte Kaefer Kueken Igel Katze Floh Hornisse Truthahn Maikaefer Kreuzspinne Vogelspinne Stieglitz Spinne Bison Eichhoernchen Libelle Kojote Gaemse Lama Rind Kenett et al., 2020; Siew et al., 2019; Wulff et al., 2019). Schakal Blaumeise Dobermann Fliege Ratte Graugans Kojote Ameisenbaer Kaefer Golden Retriever Tsetsefliege Tapir Wisent Ara Flusspferd Wisent Marienkaefer Wellensittich Dogge Chinchilla Wuehlmaus Grizzlybaer Elster Storch Maultier Tarantel Rennmaus Stinktier Reh Baer Drossel Piranha Hase Wolfshund Kakerlake Golden Retriever Hund Nerz Zobel Hyaene Ratte Loewe Specht Rhesusaffe Dromedar Bison Dachs Flamingo Nacktschnecke Ozelot Meerschweinchen Braunbaer Star Panther Ochse Zebra Alpaka Ente Eichhoernchen For instance, degree has been linked to speed of retrieving Kaninchen Kreuzspinne Koi Wasserbueffel Panther Adler Leopard Nilpferd Kamel Tiger Papagei Erpel Weisskopfseeadler Regenwurm Hamster Erdmaennchen Luchs Gnu Panda Mammut Stier Wolf Krokodil Muecke Heuschrecke Bussard Schmetterling Wildschwein Hamster Schwein Affe Damhirsch Gepard Esel Elefant Strauss Termite Specht Stute Falke Papagei Warzenschwein Schwarzbaer Schimpanse Gnu BeutelrattePuma Loewe Biene Ara Erdmaennchen Fink Amsel Spinne Ameise Natter Gepard Maus Hyaene Reh Giraffe Nashorn Wespe Kanarienvogel Hengst Schwalbe Gans Weinbergschnecke Jaguar Antilope Uhu Spatz Kuckuck words in lexical decison tasks (De Deyne et al., 2013). To Schaeferhund Giraffe Gorilla Kauz Taube Nilpferd Rothirsch Berner Sennenhund Gorilla Meerkatze Echse Alligator Hummel Wuehlmaus Alligator Reiher Hausschwein Antilope Salamander Schimpanse Wellensittich Dingo Muschel Frosch Kroete Pfau Eisbaer Frosch Orang−Utan StarRotkehlchen Fisch Salamander Komodowaran Fennek Gazelle Lurch Schwalbe Nachtigall Moewe Wal Maultier Molch Bernhardiner Orang−Utan Kroete Buntspecht Eidechse Unke Skorpion avoid any confounding influences of network size, younger Hund Bonobo Dackel Rentier Lemur Hauskatze Kamel Kuh Dromedar Hirsch Anakonda Leopard Katze Krokodil Hai Aal Tintenfisch Miesmuschel Auster Lurch Schnecke Schwan Krebs Vogelspinne Gecko Kaulquappe Stier Krabbe Springbock Schildkroete Blindschleiche Waran Eidechse Kaenguru Fisch Adler Elster Eisvogel Lerche Reiher Rabe Fink Kolibri Eule Dohle Flughund Rotschwanz Perserkatze Siamkatze Puma Ringelnatter Robbe Made Affe Amsel Schwan and older adults’ networks were compared on the largest con- Seestern Blauwal Ziege Hummer Drossel Tiger Blindschleiche Flusspferd Baer Pottwal Karpfen Grauwal Pelikan Wal Geier MeiseEichelhaeher Henne Jaguar Auerhuhn Schlange Pavian Braunbaer Elefant Hammerhai Languste Kreuzotter Bussard Milan Vogel Storch Kueken Emu Esel Kobra Blauwal Alpaka Schwertwal Rotbauchunke Brillenschlange Falke Moewe Steinadler Klapperschlange Huhn Eisbaer Lama Schwertfisch Gazelle Boa Habicht Seeadler Kraehe Fledermaus Strauss Kreuzotter Hecht Orca Koala Zebra Katzenhai Otter Kobra Kondor Taube Gans Kranich Iltis nected, common sub-graph, containing only words that were Fasan Pferd Python Forelle Buckelwal Panda Ente Python Frettchen Hai Pinguin Wiesel Emu Auster Kormoran Pinguin Lachs Thunfisch Weisser Hai Schlange Hering Boa Natter Wombat Albatros Kakadu Truthahn Seehund Delfin Wels Finnwal Walhai Lachs Aal Muschel Bulle Leguan Viper Ameisenbaer Waschbaer Pavian Leguan Pottwal Tigerhai Sprotte Flamingo Goldfisch Ringelnatter Brillenbaer Krebs Sperling Seeelefant Rind Clownfisch Piranha Grizzlybaer Hecht Marabu Schaf Qualle Scholle Walross Pfau Pelikan Walross Rotbarsch Kragenbaer Bueffel Karpfen produced by both younger and older adults. Figure 2 shows Hahn Koi Schildkroete Barsch Sandotter Fasan Hummer Seeelefant Rotfeder Kabeljau Schwarzbaer Marder Zander Robbe SeepferdchenOchse Kiwi Zander Flusskrebs Languste Krake Stoer Thunfisch Pirol Delfin Seehund Seeloewe Mantarochen Koralle Pangasius Barsch Schleie Steinbutt Krabbe Makrele Goldfisch Rentier Flunder Rochen Plankton Seekuh Tintenfisch Skalar Guppy Polarfuchs Elch Wels Garnele Seeigel Scholle Forelle Walfisch Seeloewe the networks estimated for younger and older adults in each Study 1 − Countries Study 1 − Countries Younger adults Older adults of the four data sets analyzed. Antarktis Arktis Simbabwe Kamerun Suedpol Seychellen Tonga Mauritius Malediven Bermuda Kuba Reunion Namibia Mosambik San Marino Haiti The results are presented in Figure 3. Compared to Slowakei Schweiz Niederlande Tschechische RepublikSlowenien Daenemark Spanien Wales Nordirland England Faeroeer−Inseln Italien Schweden Schottland Vereinigtes Koenigreich Grossbritannien und Nordirland Irland Puerto Rico Tahiti Kamerun Panama Haiti Guatemala Guyana Suriname Guatemala Kolumbien Dominikanische Republik Mexiko Honduras Costa Rica Belize Vereinigte Staaten von Amerika Kenia Mali Eritrea Tschad Hawaii Togo Elfenbeinkueste Lettland Jamaika Argentinien Kongo Niger older adults, the networks of younger adults showed con- Deutschland Portugal Costa RicaKuba Ghana Madagaskar Uganda Island Norwegen Franzoesisch−Guayana Neuseeland Polen Frankreich Brasilien Ecuador Vereinigte Republik Tansania Belarus Estland Mexiko Venezuela Angola Indonesien Liechtenstein Panama Sudan Mauretanien Vereinigte Staaten von AmerikaArgentinien Aethiopien El Salvador Taiwan Kambodscha Luxemburg Andorra Ecuador Chile Nicaragua Honduras Australien Dominikanische Republik Brasilien Afrika Arktis Monaco Zypern Trinidad und Tobago Peru Vietnam sistently higher average degrees and lower average shortest Bolivien Paraguay Oesterreich Paraguay Nicaragua Somalia Thailand Venezuela Antarktis Finnland Groenland Kolumbien Philippinen Demokratische Volksrepublik Laos Senegal Burundi Belgien Niger Bolivien Suedafrika Myanmar Uruguay Kongo Botswana Chile Aegypten NepalRepublik Korea Griechenland Alaska Alaska Suedafrika Botswana Lappland Israel Bangladesch Kanada Uruguay Peru Zypern path lengths. However, results for the average clustering co- Kroatien Marokko Mali Sri Lanka Tunesien Sambia Japan Ghana Liechtenstein Malta Demokratische Volksrepublik Korea Ukraine Vatikanstadt Libyen Nigeria Madagaskar Kuwait Nigeria Groenland Kanada Algerien Korea Russische Foederation Algerien Vereinigte Republik Tansania Indien Arabische Republik Syrien Portugal Andorra Marokko Serbien Zentralafrikanische Republik Libyen Palaestina Libanon Aegypten Spanien Schweiz Tunesien Irak Montenegro Ruanda Mauritius Arabien Jordanien Vereinigtes Koenigreich Grossbritannien und Nordirland Arabische Republik Syrien Litauen Angola Seychellen China Oman efficient were mixed. A multiverse analysis (Steegen et al., Armenien Ungarn Bosnien und Herzegowina Kosovo Albanien Turkmenistan Georgien Aserbaidschan Tuerkei Iran Israel Irak Aethiopien Elfenbeinkueste Namibia Kenia Mosambik Demokratische Republik Kongo Somalia Tschad San Marino Benelux Island Vatikanstadt Luxemburg Monaco Frankreich Belgien Bulgarien Albanien Griechenland Jordanien Saudi−Arabien Afghanistan Iran Vereinigte Arabische Emirate Pakistan Bahrain Tibet Hongkong Nordamerika Tadschikistan Sudan Simbabwe Bosnien und Herzegowina 2016) evaluating the results under various implementations Schottland Italien Mongolei Kirgisistan Vereinigte Arabische Emirate Tuerkei Jemen Deutschland Usbekistan Uganda Suedsudan Aserbaidschan Afghanistan Niederlande Ungarn Suedamerika Rumaenien Neuseeland Togo Eritrea Libanon Tschetschenien Bangladesch Daenemark Bulgarien Fidschi Malediven Georgien Pakistan Nordirland Ehemalige Jugoslawische Republik Mazedonien Australien Afrika Tschechoslowakei Tadschikistan Kasachstan Indien of our inference method suggests that inference is only ro- Malta Oman Irland Ehemalige Jugoslawische Republik Mazedonien Thailand Bali Bahrain Nordpol Finnland Polen Serbien Jugoslawien Kirgisistan Sri Lanka China Indonesien Dubai Vietnam Saudi−Arabien England Slowakei Rumaenien Kroatien Belarus Republik Moldau Mongolei Myanmar Singapur Nepal Japan Jemen Suedpol Schweden Russische FoederationMontenegro Kasachstan Philippinen Malaysia Hongkong Bhutan Kuwait Oesterreich Lettland Estland Litauen Slowenien Republik Korea Abu Dhabi Palaestina Taiwan Tschechische Republik Ukraine Usbekistan Andalusien Katar bust for degree and shortest path length, but not for clus- Demokratische Volksrepublik Korea Demokratische Volksrepublik Tibet Laos Korea Kambodscha Papua−Neuguinea Wales Norwegen Moskau Baltikum Republik Moldau Sowjetunion Abchasien tering, providing an explanation for the mixed results in the Study 2 − Animals Younger adults Study 2 − Animals Older adults tarantel springmaus rennmaus bison made latter (see Supplementary Material). These results corrob- seerobbe rotfuchs ringelnatter made lurch schwarze witwe kakerlake feuerqualle wurm kreuzspinne spinne erdmännchen gottesanbeterin kaulquappe kröte vogelspinne maikäfer libelle junikäfer blattlaus raupe weinbergschnecke schnecke kartoffelkäfer seeigel seepferdchen qualle tintenfisch wattwurm wisent siamkatze krebs regenwurm dinosaurier kellerassel mammut zitronenfalter pfauenauge ponyweißstorch kohlweißling wanze salamander wespe hummel marienkäfer nacktschnecke garnele bandwurm schwarzstorch orate the existence of systematic structural differences be- perserkatze nutria krabbe eidechse faultier mücke hornisse zwergkaninchen meerschweinchen rennpferd kaiserpinguin kabeljau goldfisch klapperschlange frosch käfer iltis wurm hamster ochse kreuzotter kanarienvogel koi flunder esel anakonda fliege robbe hund lamm stier skorpion leguan regenwurm laus krake katze sibirischer tiger antilope qualle biene sprotte maulwurf emu gepard dogge pinguin grashüpfer waran seehund skorpion känguru nilpferd alligator wattwurm raupe dromedar bulldogge rind python schmetterling salamander schäferhund tween younger and older adults’ semantic networks in terms milbe floh gecko flamingo kranich echse lachs kobra nasenbär schildkröte karpfen bulle chow−chow pudel zecke schlange papagei steinbock scholle ameisenbär kröte thunfisch ratte pferd steinbock schildkröte emu laus nashorn mistkäfer wellensittich forelle lurch erdhörnchen biber puma bernhardiner wüstenfuchs luchs elefant storch amsel meise widder barsch eidechse hecht elefant maus seelöwe katze elch boa bär specht drossel kakadu kaulquappe blauwal boxer aal zebra vogel eisbär dackel koalabär gepard fink pelikan frosch delfin tiger bergziege gämse of connectivity and efficiency, but not clustering. gnu hai wal terrier krokodil elster schwarzbär maulwurf fisch schlange strauss giraffe spatz blattlaus floh büffel fisch bussard star eichhörnchen makrele cockerspaniel pandabär rabe nymphensittich kamel puma ameise marienkäfer wanze hering siamkatze braunbär kaninchen maikäfer rentier termite strauß zebra eichelhäher käfer alpaka kuh ferkel grizzlybär löwe krähe möwe schmetterling wolf huhn kaninchen krokodil orang−utan habicht rentier ziege elch kalb spinne antilope Two additional findings concerning younger and older waschbär stinktier eisbär leopard kapuzineraffe gorilla clownfisch adler eule mäusebussard falke rotkehlchen meerschweinchen milan buntspecht fledermaus blaumeise kohlmeise wespe libelle mücke hornisse hummel ameisenbär lama wellensittich otter affe pinguin braunbär schwein wildkatze pandabär wiesel dachs schwarzbär schaf bonoboaffe taube wildschwein seeelefant pfau fasan adults’ verbal fluency data are worth noting. First, in the schimpanse thunfisch forelle dorsch hering hyäne marder wolf eichhörnchen fuchs chinchilla stier frettchen hase wiesel bulle schwein schwalbe ochse moskito tsetsefliege kreuzspinne biene fliege ameise kreuzotter kakerlake jaguar flusspferd boa giraffe nashorn papagei amsel adler hyäne marder löwe wildschwein bär ente hase hahn bache flamingo pelikan motte robbe pferd küken grizzlybär gans two studies that gave participants 10 minutes to retrieve items kuh schaf chamäleon ringelnatter orang−utan goldfisch delfin vogel gans henne fink gorilla schwan pute lama reh dachs hamster kobra uhu ara krebs hecht affe tiger rind schimpanse storch wildgans dackel ratte maus geier kakadu geier igel husky aal ziege blindschleiche star reiher dromedar huhn hirsch hängebauchschwein hahn pavian gnu karpfen barsch biber panda hausschwein alpaka hund ente lamm kohlmeise meise seeadler panda muschel esel reh fuchs luchs eule walross from semantic memory, there were no differences in the num- jaguar zander lachs strauss gazelle yorkshire terrier möwe dingo fischreiher nerz blaumeise drossel hummer schwertwal orca hai waschbär stachelschwein dalmatiner kanarienvogel leopard spatz albatros krabbe zwergwal blauwal labrador bussard leguan bernhardiner chihuahua kuckuck anakonda taube elster kranich scholle stachelrochen seepferdchen seekuh falke kamel bison golden retriever rhesusaffe rotkehlchen nilpferd schwalbe seelöwe tintenfisch buckelwal büffel mops bachstelze otter wal nymphensittich pavian känguru hirsch widder walhai seestern igel killerwal truthahn pudel buchfink habicht milan tigerhai schäferhund rabe koalabär nachtigall ber of items produced by younger and older adults (Table walross weisser hai seeigel oktopus schwertfisch rochen schwerthai pottwal weißer hai hammerhai katzenhai schwan uhu pitbull pute krake zeisig eichelhäher panther kondor specht brieftaube krähe gazelle gibbon meerkatze lachmöwe 1). Compared with the shorter retrieval periods of previous Figure 2 studies (cf. Hills et al., 2013; Rosen, 1980; Tombaugh et al., 1999), the longer retrieval period of 10 minutes seems Fluency networks. Figure shows the networks estimated to eliminate older adults’ disadvantage of slower memory for younger (left column) and older adults (right column) retrieval. Second, as a group, older adults produced more in each of the four data sets analyzed. Labels are not unique category elements across all four data sets (Table 1), displayed on top of their nodes to not obscure the structural which is supportive of the notion that older adults possess characteristics of the network. For details on the network a larger mental lexicon than younger adults (Verhaeghen, inference mechanism, see Methods. 2003). Despite such differences, the age-related patterns in macroscopic network structure generalize across the different domains and conditions, which speaks to the generality of
4 WULFF, HILLS, AND MATA 5 2 direct estimates of the connection strength between words, 2.5 1 sidestepping the need to infer edges using complex algo- ∆ 〈k 〉 0 0 rithms. Fourth, similarity ratings deliver graded responses −2.5 −1 permitting the construction of networks with weighted edges. −5 −2 Finally, because network statistics are available for each in- dividual, the comparison between younger and older adults’ 0.1 0.1 0.06 0.06 networks can be carried out using standard methods of sta- tistical inference. ∆C 0.02 0.02 −0.02 −0.02 −0.06 −0.06 In our study, each of 36 younger and 36 older participants −0.1 −0.1 provided a total of 2,253 similarity ratings, of which 1,953 were given to all possible pairs of 63 common animals and 0.2 0.6 0.1 0.3 the remaining 300 to a set of repeat pairs, for which we found ∆L 0 0 reliability to be high (older adults: r = .76, younger adults: −0.1 −0.3 −0.2 −0.6 r = .74). We constructed networks by, first, mapping an individual’s ratings from the original scale of 1 (extremely Zortea et al. (2014) Dubossarsky et al. (2017) Wulff et al. (2016) Study 1 Animal Study 1 Country Study 2 Animal dissimilar) to 20 (extremely similar) to the scale of 0 (mini- mum rating) to 1 (maximum rating), in order to account for differences in scale use. Second, we placed edges between all 63 animal nodes with weights equal to the transformed Figure 3 ratings. Finally, we eliminated edges with weights below a threshold wmin = [0, .1, .2, .3, .4]. This last step was nec- Differences in the macroscopic structure of younger essary to be able to determine the average local clustering and older adults fluency networks. Gray bars show the coefficient, which is not defined for completely connected the difference between the younger and older adults’ age networks, while also providing us with a means to assess the group in Zortea et al. (2014) and that of age 30 and 70 robustness of our results to the choice of threshold. Figure in Dubossarsky et al. (2017), respectively. Yellow bars 4 shows the 72 networks obtained from younger and older show differences in networks inferred from the four fluency adults under wmin = .1. data sets. Error bars show 95% bootstrapped confidence intervals. Across all values of wmin , compared to older adults, the networks of younger adults showed consistently higher aver- age degrees (hki) and lower average shortest path lengths (L), and also higher local clustering coefficients (C) (see Figure these findings across elicitation procedures (cf. Dubossarsky 5). We found the same pattern of results when the networks et al., 2017; Wulff et al., 2021). were analyzed as unweighted networks. For small values of wmin , where more than 50% of all edges were retained, i.e., Age-related differences in individual-level similarity net- wmin ∈ (0, .1), moderate to large effects were observed that works consistently reached statistical significance. Effects for more A potential criticism of extant comparisons of younger restrictive values of wmin , i.e., wmin > .1 pointed in the same and older adults’ networks is that they lump together the direction, but they were smaller in size and, due to larger data of many individuals to form aggregate networks, thus variance, did not consistently reach significance. These re- obscuring individual and group differences. To address this sults corroborate the structural differences found for aggre- limitation, we conducted a comparison of younger and older gate networks and demonstrate, for the first time, systematic adults’ semantic networks at the level of the individual. age-related differences in the structure of semantic networks Specifically, we elicited a large number of similarity ratings at the level of the individual. and constructed networks directly from each individual’s re- Moreover, analyses reported in the Supplementary Ma- sponses. Aside from avoiding problems of aggregation, this terial confirm the existence of aggregation biases. For the approach had five additional advantages: First, similarity rat- average degree, the clustering coefficient, and the average ings likely recruit different memory retrieval processes and shortest path length, but not the average strength, estimates may overall be less affected by such processes than verbal based on aggregate networks, which we derived by averaging fluency, permitting an independent and, potentially, cleaner networks within age groups, were considerably higher than assessment of network structure. Second, by requiring par- the majority of estimates for individual-level networks. Ag- ticipants to rate a common set of words, similarity ratings gregate networks, however, still revealed group differences likely are less affected by vocabulary differences between consistent with those observed on the individual level, sug- younger and older adults. Third, similarity ratings deliver gesting some level of robustness for comparisons of groups
INSERT SHORTTITLE COMMAND IN PREAMBLE 5 Figure 4 Similarity rating networks. Each individual plot shows the network of one individual under wmin = .1. The first four columns show, ordered by network strength, the networks of younger adults. The second four columns those of older adults. Edges weights have been scaled according to w2 to increase visibility. Nodes are ordered and colored according to ten animal categories. These are, starting at 0°, African animals (plus kangaroo), large apes, birds, farm animals, fish, forest animals, pets, reptiles, and rodents. Animals names were translated from German.
6 WULFF, HILLS, AND MATA Table 1 An Overview of Fluency Data and their Inferred Macroscopic Network Structure u Dataset Age N t n̄ r̄ Σn |V| hki C L a c c Wulff et al. (2016) 29-65 142 1 min 21.2 .74 .09 8 7.69 .46 2.53 66-94 142 1 min 17.9 .75 .11a 84 6.83c .38c 2.78 Study 1 - Animal 18-34 41 10 min 90.7 2.71 .15a 209.1b 5.44c .18 3.51c 66-81 71 10 min 89.6 12.6 .18a 209.1b 4.29c .13 4.07c Study 1 - Country 18-34 41 10 min 75.3 2.41 .08a 150.9b 7.17c .19 3c 66-81 71 10 min 69.7 10.9 .11a 150.9b 5.88c .21 3.36c Study 2 - Animal 18-32 36 10 min 92.6 7.03 .17 105 3.56 .3c 4.4 65-78 36 10 min 88.6 11.4 .19 105 3.33 .35c 4.72 Note. a Proportions were found to be significantly different between younger and older adults according to permutation tests. b Bootstrap estimates. c Significant (p < .05) group difference according to bootstrap test. Legend: n - number of non-duplicate, valid responses; r - number of duplicate responses; u - Number of unique responses across the a group’s retrieval sequences. relative to one another. edge weights and the proportion of triangles were consis- tently lower for older than younger adults, whereas path Locating Age-Related Differences in Semantic Network lengths were consistently larger. Crucially, we observed that Structure the differences between older and younger adults were con- siderably larger for the lower half of node pairs. Thus, the Past work on the development of semantic knowledge sug- differences between younger and older adults’ networks ap- gests that cumulative linguistic experience and general learn- pear to be mainly due to peripheral regions in the network, ing process combine to create specific semantic structures where edge weights are small, triangles rare, and shortest that allow efficient discrimination learning (Ramscar et al., path lengths long. 2017). Crucially, that work proposes that such learning pro- We should note that the results above do not seem to be cesses involve the strengthening of some associations while explained by age differences in use of the scale. We ob- weakening others to allow differentiating between meaning- served the judgments of younger and older adults not to dif- ful and meaningless pairs of items in memory. One impor- fer in terms of the judged minimum (d = 0, p = 1) or the tant consequence of this process is that age differences in judged maximum (d = .26, p = .277). However, we did network structure may not be homogeneous across pairs of find younger and older adults to differ in terms of the rat- associations due to the interaction of learning and cumulative ings‘ average (d = .56, p = .019) and, crucially, the ratings’ experience. skewness (d = −.51, p = .032), with older adults’ ratings To shed light on the differences between younger and being lower on average and more right skewed. This sug- older adults’ networks, we compared their networks on the gests that younger and older adults interpreted and used the level of node pairs with respect to three metrics that directly end points of the scales in the same way, and differed only underlie the macroscopic results in Figure 5 and allow us in how they distributed the word pairs in between the end to assess homogeneity of age differences across node pairs. points, as would be expected from different perceptions of Specifically, for each of the 1,953 node pairs, we compare similarity between judged pairs. the edge weight w under wmin = 0 (corresponding to hsi and hki), the proportion with which the pair forms triangles with Assessing Age-Related Differences in the Similarity of other nodes (C pair ) under wmin = .1, and the path length con- Network Structure necting the pair (L pair ) also under wmin = .1. Figure 6 dis- plays these results separately for younger and older adults One corollary of the idea that cumulative experience is with node pairs ordered by the average edge weight w across responsible for structural differences in semantic memory is both age groups. Ordering edges in this way allows direct not only that younger and older adults’ semantic networks inference-by-eye to reveal whether age-differences emerge differ in key respects but, also, that older adults differ more uniformly across the network. from each other as a function of their different accumulated We observed consistent differences between younger and experiences (Wulff et al., 2019). We tested this principle older adults in terms of all three metrics. Specifically, the by evaluating within age-group agreement in terms of edge
INSERT SHORTTITLE COMMAND IN PREAMBLE 7 1 1 1 −2.44 −3.23 −3.77 −3 −2.61 −1.9 −1.37 −0.93 −1.12 −0.62 Older adults 0.8 Younger adults 0.8 0.8 0.6 0.6 Older adults |E| |E| 0.6 Younger adults 0.4 0.4 w 0.2 0.2 0.4 0 0 0.2 10 25 0 1 500 1000 1500 1953 8 20 6 15 Node pairs ∆ 〈k 〉 ∆ 〈s〉 4 10 1 −2.55 −4.6 −5.31 −5.27 −5.06 −4.64 −3.47 −2.44 −2.05 −1.61 2 5 0 0 0.8 −2 −5 C pai r 0.6 0.25 weighted 0.25 0.4 0.2 unweighted 0.2 0.15 0.15 0.2 ∆ Cw ∆C 0.1 0.1 0 0.05 0.05 1 500 1000 1500 1953 0 0 Node pairs −0.05 −0.05 1.4 2.25 1.94 2.01 2.22 2.25 2.53 1.94 1.48 1.34 0.86 0.2 0.2 1.2 0 0 1 ∆ Lw L pair ∆L 0.8 −0.2 −0.2 0.6 −0.4 −0.4 0.4 −0.6 −0.6 0.2 0 0.1 0.2 0.3 0.4 0 1 500 1000 1500 1953 w min Node pairs Figure 5 Figure 6 Differences in the macroscopic structure of younger Comparisons between younger and older adults’ networks and older adults’ similarity rating networks. Blue and across all 1,953 node pairs. The panels show separately yellow circles, in panel 1, correspond to younger and older for younger (blue) and older (younger) adults the average adults, respectively. In panels 2 to 5, light blue circles and edge weights under wmin = 0 (upper panel), the proportion dark blue circles correspond to differences between the of triangles that existing edges form with other edges under younger and older adults’ networks derived from weighted wmin = .1 (middle panel), and the shortest paths between the and unweighted networks, respectively. Error bars show nodes wmin = .1. The numbers on top of each panel show 95% bootstrapped confidence intervals. the Cohen’s d (younger - older adults) for bins of 200 node pairs. weights. Specifically, we compared all pairs of individual networks using a weighted Jaccard index (JI). We found works of the two age groups did not systematically differ older adults’ networks to be considerably less similar to each in their average clustering coefficients. Importantly, we ex- other (JI = .33) than younger adults’ networks (JI = .45; tend past work by showing that these age patterns generalize d = .97). This result is compatible with the idea that cu- across categories (animals, countries) and time constraints (1 mulative exposure to linguistic and other information con- vs. 10 minutes), suggesting that such age-related differences tributes to individual differences in the structure of semantic are not a function of specific elicitation choices and general- networks. ize across domains. Discussion In addition, analyses of individual networks estimated from a similarity-judgment task involving thousands of judg- We investigated differences in the networks of younger ments from the same individuals ruled out potential problems and older adults at both the group and the individual level. of aggregation and confirmed the differences in average de- Our group-level analyses using verbal fluency data repli- grees and lower average shortest path lengths, while addi- cate previously observed differences between networks of tionally revealing systematic differences in terms of average younger and older adults (e.g., Cosgrove et al., 2021; Du- clustering coefficients, in the direction of lower clustering in bossarsky et al., 2017; Zortea et al., 2014): The aggre- older adults’ semantic networks. We found age differences gate older adults’ networks based on verbal fluency exhib- were especially pronounced for weakly-related, peripheral ited larger average degrees and lower average shortest path regions of the network. Further, older adults’ networks were lengths than younger adults’ networks, although the net- shown to be considerably less similar to each other than
8 WULFF, HILLS, AND MATA younger adults’ networks. All in all, these results provide reported here. One promising proposal stems from models of converging evidence that the semantic networks of younger discriminative learning, whereby increasing experience leads and older adults differ systematically not only in content, as weakly and strongly related contents in memory to be driven has been amply suggested in past work (Verhaeghen, 2003), further apart from each other, resulting in a topological ex- but also in their structure (Wulff et al., 2019). Our results are pansion of the network. The nature of structural differences, particularly novel in pointing out the progressively idiosyn- the observations of amplified differences for more weakly re- cratic nature of semantic representations across the life span, lated words, as well as the lower similarity between older leading to more distinct semantic representations between in- adults’ compared to younger adults’ networks, seem to sup- dividuals over time. Further, our finding that individual and port this notion. However, so far, discriminative learning age differences may be strongest for peripheral parts of se- has only been successfully employed to account for age dif- mantic representations, emphasize the importance of investi- ferences in paired-associate learning (Ramscar et al., 2014; gating a large swath of individuals’ semantic representations Ramscar et al., 2017). Whether such a mechanism can be to understand the environmental and cognitive contributions expanded to account for the full set of results presented here to individual differences in semantic cognition. remains an open question. We should point out a number of limitations in our work. Third, and finally, our work made use of an extreme-group First and foremost, we must acknowledge that we cannot comparison design by comparing groups of younger relative definitively determine to what extent the age differences de- to older adults. This type of design is not optimal to study scribed above are due to age differences in representation the role of cumulative experience that is thought to underlie and/or control processes involved in searching and select- age differences in the content and structure of lexical and se- ing information from memory. The type of network models mantic networks. Ideally, estimates of cumulative experience we adopt here to describe lexical associations are, in princi- and associated semantic networks would be obtained longi- ple, compatible with mechanistic explanations based on both tudinally for large samples of individuals and across long representation and process and, therefore, cannot fully ar- spans of time involving years or decades. One major diffi- bitrate between the two (Castro & Siew, 2020). Our find- culty with such studies will be mapping semantic networks ing that results generalize across elicitation conditions (time for specific individuals but such efforts are under way (Wulff contraints), domains (animals, countries), and tasks (verbal et al., 2021). fluency, similarity judgement) could be indicative of age dif- Despite its limitations, our work has some important im- ferences being due to differences in the underlying represen- plications for understanding and modeling human cognition. tation, but only to the extent that one can confidently assume Both extant theories and some empirical evidence suggest different processes of search and comparison across the dif- sizable links between the structure of semantic networks and ferent conditions, domains, and tasks. It seems plausible cognitive performance in a wider range of tasks (see Wulff that the underlying cognitive processes are perhaps not iden- et al., 2019, for a review). In many of these tasks, older tical but, at least similar, as all share aspects of controlled adults are known to perform worse than younger adults (Salt- selection, involving the activation of concepts (e.g., "ani- house, 2010), which is often considered a consequence of mal") and their features (e.g., "has wings"). There are two declining fluid abilities (Healey & Kahana, 2016; Salthouse, main approaches that could be interesting to further address 2010). Our and similar findings of systematic differences in the role of representation and process in engendering age semantic networks open up an alternative route leading to age differences in the semantic networks estimated from lexical differences in cognitive performance, whereby older adults’ tasks. One approach involves using additional independent cognitive performance shows apparent decline because of the measures to statistically account for the contribution of con- consequences of learning for the size and the structure of se- trol processes using an individual differences approach (e.g., mantic networks. In turn, our finding that age differences Hoffman, 2018). Another approach involves making use of may be particularly pronounced in peripheral parts of seman- neuroimaging techniques to directly measure mechanisms of tic networks could have implications for future tests of the- control and memory retrieval. Past work suggests that repre- ories of individual and age differences in semantic cognition sentation and semantic control rely on distinct (but interact- that may, or may not, make predictions concerning different ing) brain regions (Ralph et al., 2017) and this information parts of semantic representations. could be potentially be leveraged to provide an estimate of Our results may have implications beyond our theoretical the role of control processes in semantic cognition. understanding of healthy cognitive aging. Lacking a cure, Second, on a related note, we do not detail a specific the best way to battle the “dementia epidemic” is timely di- mechanism to account for the interaction between cumula- agnosis and early treatment (Larson et al., 2013; Robinson et tive experience and network structure. Consequently, a key al., 2015). The diagnosis of mild cognitive impairments and challenge for future research lies in developing models for early dementia is, however, still predominately based on tests the age-related changes in the structure of semantic networks of cognitive performance (Robinson et al., 2015). Instru-
INSERT SHORTTITLE COMMAND IN PREAMBLE 9 ments such as the short dementia screener DemTect (Kalbe older adults and 41 younger adults. Responses were recorded et al., 2004) or the neuropsychological battery CERAD (Fil- using a microphone and transcribed by us. Participants were lenbaum et al., 2008) involve an individual undergoing a recruited through the internal participant database of the MPI series of standard cognitive tasks, including several of the for Human Development. The older adults’ age ranged from tasks listed above. Understanding the role of age-related 65 to 80 years with a median age of 70 years, the younger changes in the structure of semantic networks promises to adults’ age ranged from 17 to 33 with a median age of 25. improve our interpretation of current instruments for demen- Participants were paid 10€/hour for participation. The sec- tia screening and diagnosis. Further research in this direc- ond study was also collected at the Max Planck Institute for tion could lead to more personalized instruments that can de- Human Development using participants from the MPI’s in- tect changes in cognitive performance earlier and with higher ternal database. We collected 10-minute fluency data for ani- sensitivity by focusing on specific parts of semantic represen- mals from 36 older adults and 36 younger adults. Responses tations than is currently done. were recorded using a microphone and transcribed by us. In sum, we presented converging results from verbal flu- The older adults’ age ranged from 65 to 78 years with a me- ency and similarity judgment tasks concerning structural dian age of 70 years, the younger adults’ age ranged from differences in the semantic networks of younger and older 18 to 32 with a median age of 24. Participants were paid adults. Older adults seem to possess richer more idiosyn- 10€/hour for participation. Study 1, 2 and 3 were approved cratic networks, characterized by smaller average degree by the internal review board of the Max Planck Institute for and longer path lengths relative to those of younger adults. Human Development. Our results emphasize the importance of considering how Fluency data were subjected to minimal preprocessing. life span cognitive development and cumulative experience Responses were scrutinized for category membership and shape the content and structure of individuals’ semantic cog- spelling. A lenient criterion was used to assess category nition. membership to retain as much of the original data as possible. In the case of animals, all nonfictional entries that described Methods entire, nonhuman, and nonfictional animals were retained. This led us to exclude a few cases from the data, such as Fluency data Godzilla, cat eye, or animal trainer. Similarly, in the case of countries, we retained all existing and named territories such Four data sets from three studies were used to infer net- as Istrien, a region of Italy, Croatia and Slovenia, the desert works from fluency data. The first data set was obtained Sahara or cities, but not nonexisting, fictional territories such from (Wulff et al., 2016), who analyzed the data of two pub- as Middle-earth. Spelling was hand-corrected on the basis lished studies, i.e., from Hills et al. (2013) and the Midlife in of the Merriam-Webster online dictionary. Overall 96.8% to the United States (MIDUS3) longitudinal study. The data of 99% of responses were retained in the analysis. Hills et al. (2013) contains three waves of responses to one- minute animal fluency task collected at Stanford University, Measures of macroscopic network structure CA. At time point one, the data included a total of 201 par- ticipants aged 27 to 99 (Mdn = 68). To avoid practice effects The average degree of a network G = (V, E), with nodes and problems associated with participant attrition, we used (or vertices) V and edges E, is defined as hki = 2|E| |V| for unweighted networks and as hki = |V|(|V|−1) 2 P only the first wave. The MIDUS3 data contained one-minute i, j∈V;i, j ai j wi j , animal fluency data - recorded in phone interviews - from where ai, j denotes the presence of an edge between nodes 104 individuals aged 34 to 83. Audio recordings were tran- i and j and wi, j the according edge weight. The aver- scribed by us (see Supplementary Material). In order to ob- age degree or strength, as it is commonly referred to for tained a sufficient amount of data to infer fluency networks, weighted networks, describes the average connectivity in we joined the two data sets, but eliminated individuals with the network. The average local clustering coefficient for fewer than 10 fluency productions and mini-mental state val- unweighted networks is defined as C = |V| 1 P i∈V C i with Ci = |ki |(ki −1) j,h∈Ni a jh and ki being the degree of node i 2 P ues lower than 26, which is indicative of either low attention to the task or the onset of age-related disorders. Groups of and Ni the set of neighbors to i. For weighted networks, w +w Ciw = |si |(k1i −1) j,h∈Ni i j 2 ih ai j aih a jh with si = j∈Ni w j being P P younger and older adults were created by splitting the data at the median age. This resulted in groups of 142 individuals the strength of node i, the weighted analog to ki . The local each aged 29 to 65 years old and 66 to 94 years old, respec- clustering coefficient describes the degree of transitivity in tively. Our first study with original data was collected in the the network and is related to network modularity (Newman, context of another study on age-difference in decision mak- 2006). It is often conceived of as an indicator of the struc- ing running in the laboratories of the Max Planck Institute turedness of a network (Barrat et al., 2004). The average shortest path length is defined as L = |V|(|V|−1) 2 P (MPI) for Human Development, Berlin. We collected 10- i, j∈V;i, j Li j minute fluency data for both animals and countries from 71 where Li j is the length of shortest path between nodes i and
10 WULFF, HILLS, AND MATA j, also known as the geodesic distance. For weighted net- and a flat fee of 44.1€ for providing the similarity ratings. works, Li j is the sum of weights rather than the length. The average shortest path length describes the average distance References between nodes. Low average shortest path lengths have been Anderson, J. R. (1983). A spreading activation theory of associated with efficient information processes (Bullmore & memory. Journal of verbal learning and verbal be- Sporns, 2012; Latora & Marchiori, 2001). havior, 22(3), 261–295. Baronchelli, A., Ferrer-i-Cancho, R., Pastor-Satorras, R., Network inference approach Chater, N., & Christiansen, M. H. (2013). Networks Networks were inferred from verbal fluency data based on in cognitive science. Trends in Cognitive Sciences, the community model developed by Goñi et al. (2011) and 17(7), 348–360. studied by Zemla and Austerweil (2018). The model is based Barrat, A., Barthelemy, M., Pastor-Satorras, R., & Vespig- on a two-step procedure. First, nodes and edges are included nani, A. (2004). The architecture of complex for every pair of responses that occurred within a distance of l weighted networks. Proceedings of the National responses. For instance, for the response sequence “dog, cat, Academy of Sciences, 101(11), 3747–3752. mouse, rabbit” and a criterion of l = 2, edges would be in- Beaty, R. E., Kenett, Y. N., Christensen, A. P., Rosenberg, cluded for all pairs less than three responses apart, excluding M. D., Benedek, M., Chen, Q., Fink, A., Qiu, J., only the pair dog and rabbit, which are three responses apart. Kwapil, T. R., Kane, M. J., et al. (2018). Ro- Second, an edge is identified as a true edge if the frequency bust prediction of individual creative ability from of the connected words occurring with l or fewer steps apart brain functional connectivity. Proceedings of the exceeded a frequency threshold tmin reflecting the required National Academy of Sciences, 201713532. minimum frequency of co-occurring within l responses to be Beer, R. D. (2000). Dynamical approaches to cognitive sci- considered in the first place, as well as a frequency threshold ence. Trends in Cognitive Sciences, 4(3), 91–99. tchance . The latter is derived from the probability plinked of two Benedek, M., Kenett, Y. N., Umdasch, K., Anaki, D., Faust, ij words occurring within l responses by chance, which is cal- M., & Neubauer, A. C. (2017). How semantic mem- culated as plinked ij = pi jco−occur ∗ p≥l co−occur i j . Furthermore, pi j , ory structure and intelligence contribute to creative the probability of two words to co-occur within a fluency se- thought: A network science approach. Thinking & quence, and p≥l i j , the probability that two responses are no Reasoning, 23(2), 158–183. fi f j Bhatia, S. (2019). Predicting risk perception: New insights more than l responses apart, are calculated as pco−occur ij = MM from data science. Management Science, 65(8), i j = N(N−1) (−lN 2 ) with fi , f j denoting the number l(l+1) and p≥l 2 3800–3823. of times two responses occur across M sequence and N de- Borge-Holthoefer, J., & Arenas, A. (2010). Semantic net- notes the average number of productions per sequence. tchance works: Structure and dynamics. Entropy, 12(5), is then defined as the 1 − α quantile of the binomial distri- 1264–1302. bution B(M, plinked ij ). Consistent with prior literature, we set Bousfield, W. A. (1953). The occurrence of clustering in the l = 1, tmin = 1, and α = 1 (Goñi et al., 2011; Zemla & Auster- recall of randomly arranged associates. The Journal weil, 2018) for our main analyses. In addition, we evaluate of General Psychology, 49(2), 229–240. the robustness of the results in a multiverse analysis (Steegen Buchler, N. E., & Reder, L. M. (2007). Modeling age-related et al., 2016) presented in the Supplementary Material. memory deficits: A two-parameter solution. Psy- chology and Aging, 22(1), 104. Similarity ratings Bullmore, E., & Sporns, O. (2012). The economy of Similarity ratings were collected in the context of study 3 brain network organization. Nature Reviews Neuro- and prior to participants completing the verbal fluency task. science, 13(5), 336. Participants took home a tablet to provide, over the course of Castro, N., & Siew, C. S. Q. (2020). Contributions of modern roughly one week, on a scale from 1 to 20, similarity ratings network science to the cognitive sciences: revisit- for 2,268 pairs of animals, consisting of each possible pair ing research spirals of representation and process. of 63 frequently occurring animals and 315 repeated pairs. Proceedings of the Royal Society A: Mathemati- The 63 animals were selected on the basis of the verbal flu- cal, Physical and Engineering Sciences, 476(2238), ency responses of study 2 in a manner that equated word 20190825–25. frequency across younger and older adult age groups. See Collins, A. M., & Loftus, E. F. (1975). A spreading- Supplementary Material. Reliability was found to be high in activation theory of semantic processing. Psycho- both younger and older adults with correlations of r = .76, logical Review, 82(6), 407. r = .74 for younger and older adults, respectively. Partici- pants were paid 10€/hour for participation in the lab session
You can also read