Skip to main content
SearchLoginLogin or Signup

Young Children’s Mathematics Learning From Same-Gender and Other-Gender Intelligent Character Prototypes

Volume 3, issue 2: Summer 2022. DOI: 10.1037/tmb0000069

Published onApr 30, 2022
Young Children’s Mathematics Learning From Same-Gender and Other-Gender Intelligent Character Prototypes
·

Abstract

An experiment examined how U.S. preschool-aged children’s (N = 113; 53% girls; M age = 4.44 years, SD = .35) math-related parasocial interactions (e.g., math talk) with a same-or other-gender intelligent character prototype impacted their performance on add-1 math problems in a virtual math game and subsequent transfer task with physical objects. The interactive intelligent character prototype was controlled with a Wizard of Oz approach and was not autonomous. Children who engaged in more math talk, particularly with a same-gender intelligent character prototype, demonstrated better performance on the math task during the virtual game and in the transfer task compared to children who engaged in less math talk. Children also reported that they liked same-gender media characters more than children that reported how much they liked other-gender media characters. The results suggest that engaging in contingent interactions about math and sharing a salient aspect of identity with intelligent characters, in this case gender, can facilitate children’s learning of foundational math skills. Intelligent characters who teach science, technology, engineering, and mathematics skills through digital media to young children hold promise as remote education tools that can provide individualized educational support to children.

Keywords: remote learning, children’s math learning, intelligent characters, parasocial interactions, gender

Editors: Rachel Flynn and Fran Blumberg were the Special Collection Editors. Fran Blumberg was the action editor for this article.

Change of Affiliations: Marisa M. Putnam is now working at MEF Associates. Joy Nissen is now affiliated with Widener University.

Acknowledgements: The authors thank the children, parents, and schools who made this research possible. The authors also thank Nickelodeon at Viacom, Inc. and Mariana Diaz Wionczek for permission to use the Dora the Explorer characters in the intelligent character game. Thank you to Ian Lyons for his constructive feedback throughout this project. The authors also thank the Children’s Digital Media Center members who contributed to this research, particularly Naomi Aguiar, Kaitlin Brunick, Stevie Chancellor, Melissa Richards, Marie Frolich, Angella Liu, Lauren Seibel, Nancy Perez, Elizabeth Fantini, Emily Zuckerman, Molly Biedermann, Zita DeZamaroczy, Christina Duval, and Sophia Farfan.

Funding: This research was funded by a grant from the National Science Foundation Graduate Research Fellowship Program Grant (No. DGE-1444316) to Marisa M. Putnam and a National Science Foundation (DRL Grant No. 1,252,113) to Sandra L. Calvert. Any opinions, findings, and conclusions or recommendations expressed in these materials are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation.

Disclosures: We have no conflicts of interests to disclose. This research was approved by Georgetown University’s Institutional Review Board. If interested in the data and measures in the present study, please contact the authors.

Correspondence concerning this article should be addressed to Marisa M. Putnam, Department of Psychology, Georgetown University, 37th and O Streets, N.W., Washington, DC 20007, United States mp1265@georgetown.edu


Science, technology, engineering, and mathematics (STEM) domains drive a nation’s innovation, solve modern problems, and are central to a nation’s 21st century economic success (Noonan, 2017a). Building children’s STEM skills are recognized as a national educational priority in the U.S. that supports entry into these domains (National Science Teachers’ Association [NSTA], 2014). However, many early childhood educators and parents report feeling unsure about how to support children’s STEM learning before kindergarten (McClure et al., 2017) which may lead to a less developed STEM foundation for children when they begin school. In addition, children in the U.S. fall behind their international peers in math achievement before they reach adolescence (Organisation for Economic Co-operation and Development, 2019).

Building evidence about best practices in remote education emerged as a new and additional U.S. national educational priority (United States Department of Education, 2020) during the coronavirus disease 2019 (COVID-19) pandemic when many teachers and children could not participate in traditional, in-person, schooling during mass stay-at-home orders (Sandford, 2020). This new priority presented an opportunity to blend lessons learned from research on educational gaming approaches, that can be delivered remotely, into the everyday landscape of more traditional classroom learning.

Educational experiences began to be conducted remotely in 2020 through curricula delivered on virtual interfaces like Zoom; prior to the pandemic U.S. children under age eight spent only about 5 daily minutes of screen time on school activities (Rideout & Robb, 2020). Novel teaching approaches blended preschool curriculum with educational digital media, sometimes featuring popular media characters, to engage children within their evolving home-based educational settings (Levine et al., 2021). While new, these approaches are anchored in previous literature that argues that gaming approaches that facilitate cognitive development are a promising avenue for designing engaging informal educational practices (Blumberg et al., 2019) and that digital STEM interventions may encourage preschool-age children to extend digital math experiences into other learning contexts (Gulz, Kjällander, et al., 2020).

The purpose of the present study was to examine the effects of young children’s interactions with same- or other-gender intelligent character prototypes on their fluency with early math skills in virtual and physical contexts. The math skill of interest was the add-1 rule, knowing that adding one to a number will increase that number by a single unit (Baroody, 1985).

The Add-1 Rule

Early math skills are foundational for, and predictive of, children’s academic achievement (Duncan et al., 2007). STEM is often stereotyped as a male domain (Leaper, 2015), and boys express more interest in STEM domains compared to girls (National Science Board [NSB], 2018; National Science Foundation, 2017; Noonan, 2017b). There is also a gender gap in the STEM workforce in the U.S., with more men participating compared to women (NSB, 2018; Noonan, 2017b). However, empirical evidence indicates that girls and boys demonstrate similar foundational numerical competencies (Hutchison et al., 2019). Increasing interest and skills in STEM for both boys and girls in childhood may contribute to building a competent and diverse STEM workforce that includes more participation by women (NSTA, 2014). Providing both boys and girls with new opportunities to practice math skills at a young age, then, may be a viable pathway for building STEM skills by increasing girls’ STEM interest in STEM pathways.

An important early math skill for all children is fluency with the add-1 rule (Baroody, 1985). Knowledge of the add-1 rule frees children’s cognitive resources and allows them to learn more advanced math skills (Baroody, 1985; Baroody et al., 2013). Fluency with the add-1 rule can be measured by the appropriateness of sums generated (e.g., 1 + 2 = 3), efficiency (i.e., quick accuracy), and adaptiveness, i.e., transferring knowledge to new add-1 problems Baroody et al., 2013). Previous research finds that for preschool children, age predicts children’s average latency, or efficiency, on the add-1 rule which reflects that this is a developmental time frame when they are becoming fluent with the add-1 rule (Calvert et al., 2020). Media interventions have been one route to developing add-1 fluency (Baroody et al., 2013; Calvert et al., 2020) as well as other early math skills (Gulz, Londos, et al., 2020). Media interventions designed to include elements that are grounded in early math research and include meaningful feedback and scaffolded support as well as a focus on motivational aspects relevant to the target age group (e.g., media characters) may offer preschoolers effective remote learning opportunities (Husain et al., 2015).

Educational Media as a Teacher for Young Children

Educational digital media are a promising venue for delivering remote STEM education, in part because children under 8 years old spend an average of two and a half hours daily using screen media (Rideout & Robb, 2020). Research on children’s experiences with digital media, including evidence about impacts of social contingency and social meaningfulness, can be incorporated into remote education and delivered on digital screens in children’s homes.

Social Contingency

Children learn about the world through socially contingent interactions and feedback from their environment. These interactions and feedback occur in both off-screen physical contexts and in onscreen virtual contexts (Calvert, 2017).

Social contingency involves social partners engaging in ways that are immediate, reliable, and accurate (Roseberry et al., 2014). However, learning from screens is more difficult than learning from live interactions (Anderson & Kirkorian, 2015). Transferring learning to a different context, such as from an onscreen to a live situation, is particularly difficult for young children (Barr, 2010; Barr et al., 2019).

Interactive media can reduce barriers to learning (Aladé et al., 2016; Kirkorian, 2018) and facilitate children’s transfer across contexts when children interact contingently with people, through interfaces like Facetime (Barr, 2019; Kirkorian, 2018; Roseberry et al., 2014), or media characters (Calvert, 2017; Calvert et al., 2007; Lauricella et al., 2010; Richards & Calvert, 2017b). Children’s socially contingent parasocial interactions (PSIs) with media characters can facilitate their learning from a screen (Calvert, 2017; Liebers & Schramm, 2019). For example, 4-year-old children who participated more “with” a character by answering her questions during a television episode performed better on a comprehension task than children who observed the episode (Calvert et al., 2007).

Recent developments in artificial intelligence are particularly relevant to remote learning and to advancements in PSIs with intelligent media characters, computer-generated agents embodied as media characters that can respond contingently to children (Brunick et al., 2016). PSIs with intelligent character prototypes have moved social contingency experiences from a general statement (e.g., sharing a favorite color) to a specific response (e.g., solution to a math problem; Calvert, 2021). These specific responses can increase the efficacy of educational media (Calvert et al., 2020), including the potential efficacy of remote delivery systems.

Although still in prototype form, embodied onscreen intelligent media characters are a digital resource that can be programmed to bolster children’s STEM learning using social contingency and social meaningfulness (Calvert, 2017; Calvert et al., 2020). For example, intelligent characters can respond to children through PSIs where the character asks a question, pauses for a reply, and responds as if they heard what the child said (Brunick et al., 2016). Currently, STEM lessons using intelligent characters are delivered through prototypes that use a Wizard of Oz approach. In this approach, a hidden adult “Wizard” is behind a screen using a virtual agent to prompt children and to contingently respond to their math (Calvert et al., 2020) and science (Finkelstein, 2018) replies. This approach has facilitated children’s effective STEM learning, making intelligent characters and intelligent peers a potential resource for remote learning if the agent’s socially contingent replies are implemented by software rather than controlled by a Wizard. Children engage with media characters similarly to other social partners (Richert et al., 2011) and may experience interactions with these intelligent character prototypes as if they are authentic social exchanges with an autonomous social partner.

Literature on pedagogical agents facilitating instruction onscreen provides evidence about the efficacy of socially contingent interactions related to STEM content. College students remember and transfer information about science topics when an agent engages in a personalized conversation, which contributes to feelings of social presence between a student and agent (Moreno & Mayer, 2004). When learners feel they are socially present with pedagogical agents, it may encourage cognitive engagement with the material presented (Moreno & Mayer, 2004). Meta-analytic findings support the premise that pedagogical agents facilitate onscreen math and science instruction effectively for kindergarteners through college students (Schroeder et al., 2013). Importantly, children in kindergarten through 12th grade benefited more from interacting with a pedagogical agent, compared to college students (Schroeder et al., 2013). Children may learn more from pedagogical agents because they are motivated by social interactions with the agents that they perceive similarly to social interactions with people (Schroeder et al., 2013) and with media characters (Richert et al., 2011).

Social Meaningfulness

Children’s virtual learning and transfer of knowledge from digital media, particularly in STEM domains, can also be supported through socially meaningful, one-sided, emotionally tinged parasocial relationships (PSRs) with media characters (Calvert & Richards, 2014; Calvert et al., 2014; Howard Gola et al., 2013; Lauricella et al., 2011; Schlesinger et al., 2016). PSRs can be developed over time with novel characters (Howard Gola et al., 2013), have similar components as children’s relationships with humans (e.g., trust), and occur most often with favorite same-gender characters (Bond & Calvert, 2014; Richards & Calvert, 2017a).

Media characters travel with children off-screen, depicted on clothing and embodied as toys, which enables characters to be friends and teachers to children across multiple contexts (Calvert, 2017; Richards & Calvert, 2017b). Characters who are culturally popular (e.g., Elmo; Lauricella et al., 2011) and similar to toddlers (e.g., same gender; Calvert et al., 2014) are particularly effective teachers when children are familiar with, or familiarized with, the character.

Gender is a particularly salient social group (Brown et al., 2020), traditionally focused on the dichotomy of male and female categories in the U.S. (Bem, 1981). Although U.S. culture is shifting away from a strictly binary conceptualization of gender, young children continue to more often be categorized by themselves and others as either male or female (Blakemore et al., 2013). Especially during the preschool years, children are interested in what it means to be a boy or a girl and favor their own gender (Bem, 1981; Brown et al., 2020).

Gender schemas can guide perception, learning, and memory (Bem, 1981; Martin & Halverson, 1983). Children prefer engaging with same-gender peers (Brown et al., 2020), which is mirrored in children’s preferences for same-gender media characters (Richards & Calvert, 2017a). Children better remember information related to their gender, compared to other genders (Signorella et al., 1993). Information consistent with children’s existing gender schemas, or considered “for me,” may be more engaging and more readily cognitively processed by children (Calvert & Huston, 1987).

Gender-stereotypical content may be processed more quickly compared to nonstereotypical content because gender stereotypes are simple gender schemas, often based on superficial qualities like hair length (Calvert & Huston, 1987). Information that is inconsistent with a child’s gender may result in slower processing, as it violates gender schematic expectations (Calvert & Huston, 1987). An implication for children’s STEM learning is that children may be more motivated to engage with and remember STEM content if it is presented by a same-gender character who may make the STEM content more meaningful for children. When a child and media character share socially meaningful characteristics, like gender, this similarity may encourage children to consider the character as a meaningful friend and/or teacher. Children may process STEM content quicker and more deeply if their gender schemas are activated (Calvert & Huston, 1987).

Among adult learners, the gender of a pedagogical agent elicits gender-stereotypical perceptions of the agent (Kim & Baylor, 2006). The designed physical gender expression of a pedagogical agent quickly provides learners with information that may or may not align with gender stereotypes (Moreno & Flowerday, 2006). However, results are mixed on the impact of pedagogical agent gender on adult learner cognitive outcomes. For example, the gender of pedagogical peer-agent’s teaching about learning theories did not influence cognitive outcomes for male and female college students on measures of free recall, multiple choice, and transfer (Schroeder & Adesope, 2015). Adult learners scored better on a math task when they engaged with an opposite-gender peer pedagogical agent that provided feedback on their performance (Krämer et al., 2016).

Combining Social Meaningfulness and Social Contingency

Intelligent agents may be particularly effective 21st century STEM educators if they are embodied as socially meaningful and socially contingent intelligent characters (Brunick et al., 2016; Calvert et al., 2020). Pedagogical agents that are designed as peers, yet with more advanced knowledge than the learner, can provide feedback that scaffolds learners to acquire skills or knowledge (Kim & Baylor, 2006). A series of studies suggest that intelligent character prototypes can support children’s add-1 rule learning in a socially contingent interactive digital math game (Calvert et al., 2020). The intelligent character prototype, Dora the Explorer, was popular among preschoolers and engaged children in math-related PSIs, or math talk, through scaffolds that incrementally supported children’s math solutions (Calvert et al., 2020). Scaffolding and conversational turn-taking also approximate how children learn in nonmedia environments (Jordan & Vaala, 2019).

The results revealed that PSIs and PSRs both uniquely contribute to children’s add-1 rule fluency in virtual and transfer tasks, with no evidence of gender differences (Calvert et al., 2020). Overall, children who engaged in more math talk and who had stronger PSRs with Dora correctly answered add-1 problems quicker compared to children that engaged in less math talk during the virtual game and who had weaker PSRs with Dora. Transfer performance was best for children: (a) who saw and heard Dora rather than only hearing a non-Dora voiceover during game play, and (b) when Dora responded contingently rather than noncontingently to children’s math replies. An implication is that socially contingent PSIs with intelligent character prototypes facilitate children’s virtual math learning, with or without a strong PSR. However, a socially meaningful character may be key to anchoring children’s learning and creating strong mental representations of the add-1 rule that facilitate transfer to a physical context.

Intelligent character prototypes can be designed to be socially meaningful to children in a variety of ways, such as being a child’s favorite character or perhaps by being the same gender. Being a favorite character is associated with better math learning from an intelligent character prototype (Calvert et al., 2020) and being the same gender as a character is associated with improved learning of STEM content (Calvert et al., 2014). The role of the alignment between child gender and character gender for math learning is of interest in the present study.

The Present Study

The purpose of this study was to examine the impact of young children’s PSIs with same- or other-gender intelligent character prototypes on children’s fluency with the add-1 rule in virtual and physical contexts. The hypotheses were:

Hypothesis 1: Children would report liking a same-gender character more than an other-gender character, prior to game play.

Hypothesis 2: Children engaging in more, versus less, math talk with the character during virtual game play would demonstrate reduced add-1 average latency.

Hypothesis 3: Children engaging in more math talk during virtual game play with a same-gender character, versus an other-gender character, would demonstrate reduced add-1 average latency.

Hypothesis 4: Children engaging in more, versus less, math talk with the character during virtual game play would correctly answer more add-1 transfer task problems.

Hypothesis 5: Children engaging in more math talk with a same-gender character, versus an other-gender character, in the virtual math game would correctly answer more transfer task problems.

Method

Participants

One-hundred and twenty-five 4- and 5-year-old children were recruited from preschools in the Washington D.C. metropolitan area.1 Parents/guardians provided written consent for children’s participation and completed a demographic questionnaire. Twelve children, six in each condition, quit the game and were dropped from the sample. The final sample was 113 children (M age = 4.44 years, SD = 0.35; 53% female).

Parents’ report of child race and ethnicity included White (67.26%), mixed race/other (11.50%), Asian (6.19%), Hispanic (6.19%), Black (2.65%), and no response (6.19%). Parents’ reported education level included high school diploma (0.88%), some college (1.77%), college degree (28.32%), graduate school/professional degree (63.72%), or no response (5.31%). The study was approved by Georgetown University’s IRB.

Materials

Intelligent Character Prototype Math Game

Two versions of an intelligent character prototype were utilized. The game in the Dora condition featured the female Hispanic media character Dora from the television program Dora the Explorer. The Dora version was developed by Calvert et al. (2020).2 The game in the Diego condition featured a male media Hispanic character Diego from the television program Go, Diego, Go! The games were identical aside from different central characters.

The math game was presented on a digital screen and simulated social interaction with children through contingent verbal interactions using a Wizard of Oz approach. At the beginning of the game, the featured character told children that they needed help shopping for three sets of supplies (e.g., balloons, hats, and goodie bags) for the other character’s birthday party. The character was shopping with Boots, a monkey from Dora’s show. These characters stood at a cash register and asked for help counting items as the objects moved down a conveyor belt. Swiper the Fox, from Dora’s show, stole the items if children did not answer the math problems correctly.

The game engaged children in math talk through the character’s prompts to solve 12 add-1 problems. For each set of birthday supplies, a group of one, two, three, or four items came down the conveyor belt and the character named the number of items present. After these items fell into the grocery bag, another single item appeared. The character asked children to solve a problem based on the number of total items that had come down the conveyor belt (e.g., “Here come 2 balloons. Oh! Here comes 1 more balloon. What does 2 and 1 make?”). The first round of the game presented add-1 math problems in sequential order using balloons (1 + 1, 2 + 1, 3 + 1, 4 + 1) with items moving across the conveyor belt in 5.5 s. In the second round, party hats moved down the conveyor belt in 2.5 s and add-1 problems were in sequential order. In the third round, the order of problems was randomized and goodie bags moved across the conveyor belt in 2.5 s. At the end of the game, children watched an animated birthday party scene and the character asked them to sing happy birthday.

The game provided prompts and personalized feedback within each round including scaffolds for children who did not solve the problems (e.g., answered incorrectly or did not answer). In the first scaffold level, items moved down the conveyor belt slower while flashing. If children did not solve a math problem, a second scaffold level was introduced. The character asked children to count with them as the items appeared and moved one-by-one on the conveyor belt. If children did not correctly solve the add-1 problem after counting with the character, a third level scaffold was introduced where Boots gave the correct answer to the add-1 problem and prompted children to say it. Children repeated the answer to move forward to the next problem.

Measures

Pregame Character Knowledge Child Survey

Children were shown a photo of the character featured in the game they were randomly assigned to play. Children were asked if they recognized the character and to say the character’s name. Responses were coded into a 3-point scale: (0) did not know character, (1) recognized character, and (2) knew character’s name.

Children were asked how much they liked the character, responding verbally or by pointing to a 5-point Likert smiley face scale. Experimenters recorded children’s responses with options ranging from 1 (did not like the character at all) to 5 (liked the character a whole lot).

Children’s Parasocial Relationship Attachment and Friendship Subscale

Consistent with Calvert et al. (2020) measurement of PSR, children answered the attachment and friendship subscale questions (i.e., trustworthy, safe, cute, friend) from the Children’s Parasocial Relationship Survey about the character that would be featured in the math game (Richards & Calvert, 2017a).

Math Talk Parasocial Interactions

Math talk measured the extent to which children were engaging with a character on task, through PSIs about math. Scores were calculated by summing the number of math talk prompts a child responded to on-task (any numerical reply, correct or incorrect) and dividing that sum by the number of math talk opportunities that were presented within the game.

The number of math talk prompts children received varied (range: 12–78) based on children’s performance. If the child required more scaffolds to correctly answer a math problem, they received more math talk prompts. Reliability was computed on 20% of the sample, Cronbach’s α = 1.00.

Average Latency

Consistent with prior studies (Calvert et al., 2020), latency was scored and analyzed as a composite average score of children’s average reaction time to provide correct answers to all add-1 problems in the game. Scores were calculated from experimental session videos using Noldus the Observer XT version 14. Latency was operationalized as the number of seconds between the character finishing each of the initial math questions (e.g., “What does 2 and 1 make?”) and when the child responded correctly to the math problem (“3”). If a child answered this initial math prompt incorrectly (e.g., “5”) and required scaffolds, that time was included in the latency score. Reliability was computed on 20% of the sample, Cronbach’s α = .96.

Add-1 Rule Transfer Task With Physical Objects

The transfer task was adapted from Calvert et al. (2020). Children answered 10 add-1 rule math problems, presented in sequential order (1 + 1 … 10 + 1), by counting physical silver-colored star stickers. The experimenter showed children an initial number of stickers and said, “I have [number] stickers,” and placed the group of stickers into a small bag. The experimenter showed children an additional sticker and said “Oh! I have one more sticker. How many stickers do I have?” while holding the single sticker over the top of the bag until the children answered. Children responded verbally or with gestures (e.g., holding up fingers). The number of transfer task problems answered correctly by children was summed to create a transfer task score (range: 0–10).

Procedure

Children participated at their preschool or childcare centers and provided verbal assent. Within gender groups, children were randomly assigned to play either the Dora or Diego math game; children played the math game with a same- or other-gender character. The experimenter administered the pregame character knowledge survey and attachment and friendship subscale questions to children. Next, children played the math game on a video screen by responding verbally or with gestures. One experimenter sat next to the child while they played to assist them as needed. A second experimenter operated the character in the game Wizard of Oz style by hiding behind a room divider and using a menu of preset response options. This experimenter viewed a live feed of the child playing the game and could hear children’s verbal replies. After the math game, children completed the add-1 transfer task. Children received a small toy as a token of appreciation for participating in the study. Data and measures used for the present study can be obtained by contacting the authors.

Results

One-hundred and thirteen children finished the math game. Twenty-three of these children answered all problems correctly on the first trial in the game, demonstrating ceiling level accuracy and indicating that they already knew the add-1 rule. Following the analytic approach in Calvert et al. (2020), these children were excluded from analyses. The ninety children (54% girls; M age = 4.42 years) who completed the game and answered at least one add-1 problem incorrectly on the first trial comprised the analytic sample.

PSR Attachment and Friendship

The attachment and friendship subscale yielded low internal consistency overall (α = .61), which compromised the use of the subscale as a predictor variable in this study. Thus, it was not used in subsequent analyses.

Knowledge of the Character

Before the game, 33.7% of the children knew the character’s name, 18.0% recognized the character but did not know their name, and 48.3% reported that they did not know the character. Character knowledge was included as a control variable in subsequent analyses due to the range of children’s character knowledge.

Liking the Character

An ordinary least squares (OLS) regression estimated how much children liked the character after seeing a picture of the character and before playing the game, predicted by child-character gender match and controlling for character knowledge and child age. Child-character gender match was indicated as 0 for girls who played the Diego game and boys that played the Dora game. Child-character gender match was indicated as 1 for girls who played the Dora game and boys that played the Diego game.

Children liked same-gender characters significantly more than other-gender characters, n = 88, adjusted R 2 = 0.07, F(3, 84), p = .03.3 Children liked same-gender characters .71 points higher than children liked other-gender characters (B = 0.71, SE = 0.33, p = .03).

Game Play

The average duration of game play was 12.43 min (SD = 4.41, range: 6.07–25.9) and did not differ by condition, t(88) = –.78, p = .44; M SameGender = 12.79, SD = 4.70; M OtherGender = 12.06, SD = 4.10. Children correctly answered an average of 8.19 add-1 problems on their first try (SD = 2.81), which did not differ by condition, t(88) = .65, p = .52; M SameGender = 8.00, SD = 2.90; M OtherGender = 8.39, SD = 2.74.

Average Latency

Children’s average latency to answer the add-1 problems correctly was 24.84 s (SD = 21.59) and did not differ by condition, t(88) = –.48, p = .63; M SameGender = 25.91, SD = 22.88; M OtherGender = 23.73, SD = 20.37.

Math Talk

Children replied to 85.55% (SD = 18.19%, range: 21.62–100) of math talk prompts, on average. Children engaged in similar amounts of math talk with same- and other-gender characters, t(88) = .16, p = .87; M SameGender = 85.25, SD = 17.86; M OtherGender = 85.86, SD = 18.73.

An OLS regression predicted average latency by math talk, controlling for child-character gender match, character knowledge, and child age (Table 1, model 1). Children who engaged in more math talk answered add-1 problems faster, compared to children who engaged in less math talk. For every percentage point higher children scored on math talk, they correctly answered .81 s faster (p < .001, Cohen’s d = .04).

Table 1
OLS Regression Predicting Average Response Latency in Seconds for Add-1 Problems

Variable

Model 1

Model 2

Math talk

–.81** (.11)

–.57** (.14)

Child-character gender match * Math talk

–.51** (.18)

Child-character gender match (0 = other-gender, 1 = same-gender)

2.13 (3.51)

2.06 (3.40)

Character knowledge

–1.18 (1.79)

–1.27 (1.72)

Age

.003 (.01)

.01 (.01)

Constant

20.24 (24.44)

14.02 (24.07)

Regression statistics

 R 2

.45

.49

 Adjusted R 2

.42

.46

 F

14.01**

19.89**

 df

4, 84

5, 83

 N

89

89

Note. Sample excludes a child who did not answer character knowledge questions. Math Talk: Calculated by dividing the number of math talk prompts a child replied to by the number of math talk prompts that were available to a child. That number was multiplied by 100, and math talk was mean centered at 85.55. Controls: Child-character gender match: 1 = child and character are the same gender, 0 = child and character are different genders; character knowledge (0 = no knowledge, 1 = recognize character, 2 = know character’s name); Age (days). Table A1 in the Appendix displays the OLS regression without control variables. OLS = ordinary least squares. * p ≤ .05. ** p ≤ .01. Robust standard errors.

Interactions Between Math Talk and Child-Character Gender Match

An OLS regression predicted average latency scores from math talk and the interaction between math talk and child-character gender match, controlling for character knowledge and child age (Table 1, model 2). While engaging in more math talk with either character resulted in quicker correct answers to add-1 problems, compared to children who engaged in less math talk, the significant and negative interaction term indicated that the effect of math talk was greater for children interacting with a same-gender character (p < .001, Cohen’s d = .02; Figure 1). The significant and negative math talk coefficient in model 2 can be interpreted as: The more math talk children engaged in with a same-gender character, the quicker they correctly answered the math problems compared to children engaging in more math talk with different-gender characters.

Figure 1

Latency by Math Talk * Child-Character Gender Match

Add-1 Rule Transfer Task With Physical Objects

Children answered an average of 3.78 (SD = 3.27, range: 0–10) add-1 transfer task problems with physical objects, which did not differ by condition, t(87) = –1.57, p = .12; M SameGender = 4.31, SD = 3.47; M OtherGender = 3.23, SD = 3.00.

Math Talk

A negative binomial regression predicted the number of correctly answered transfer task problems, controlling for child-character gender match, character knowledge, and child age. Math talk did not predict children’s transfer performance (Table 2, model 1).4

Table 2
Negative Binomial Regression Predicting the Number of Add-1 Transfer Task Problems Answered Correctly

Variable

Model 1

Model 2

Math talk

1.00 (.01)

.99 (.01)

Child-character gender match * Math talk

1.02** (.01)

Child-character gender match
(0 = other-gender, 1 = same-gender)

1.35 (.25)

1.32 (.24)

Character knowledge

.98 (.10)

.98 (.10)

Age

1.00 (.001)

1.00 (.001)

Constant

.75 (.89)

.98 (1.14)

Regression statistics

 Pseudo R 2

.01

.02

 Wald X 2

5.41

10.40

 df

4

5

 N

88

88

 Likelihood ratio X 2 (1)

59.28

51.94

Note. Sample excludes a child who did not answer character knowledge questions, and a child who did not complete the transfer task. Math Talk: Calculated by dividing the number of math talk prompts a child replied to by the number of math talk prompts that were available to a child. That number was multiplied by 100, and math talk is mean centered at 85.55. Controls: Child-character gender match: 1 = child and character are the same gender, 0 = child and character are different genders; character knowledge (0 = no knowledge, 1 = recognize character, 2 = know character’s name); Age (days). Table A2 in the Appendix displays the Poisson regression models. IRR = incidence rate ratio. * p ≤ .05. ** p ≤ .01. Robust standard errors.

Interactions Between Math Talk and Child-Character Gender Match

A negative binomial regression predicted the number of correct transfer task problems by math talk and the interaction of math talk and child-character gender match, controlling for character knowledge and child age (Table 2, model 2).5 Children engaging in more math talk with a same-gender character, compared to an other-gender character, during the virtual game correctly answered more transfer task problems. The significant interaction of math talk and child-character gender match indicated that for each point higher children scored on math talk in the game, the incidence rate ratio (IRR) of correctly answering more transfer task problems was 1.02 times higher for children engaging with a same-gender character compared to an other-gender character (p = .01).

Discussion

The purpose of this study was to examine how preschool children’s math talk with a same- or other-gender intelligent character prototype influenced learning of the add-1 rule in both a virtual math game and in a transfer task with physical objects. Children who engaged in more math talk with a character correctly answered add-1 problems quicker than children who engaged in less math talk with a character; however, this effect was greater for children that engaged with a same-gender character compared to an other-gender character. Children who engaged in more PSIs with a same-gender character during the virtual game, compared to an other-gender character, also answered more transfer task problems correctly. Children reported that they liked a same-gender character more than children reporting how much they liked an other-gender character, before experiencing the characters in the game. This study provides new evidence that PSIs with a same-gender character may activate gender schemas and increase performance of early STEM skills.

Children’s math talk with a same-gender character may have streamlined children’s cognitive processing during the math game. Cognitive processing speed may have increased when children engaged with same-gender characters due to activation of gender schemas, which can serve as informational filters (Bem, 1981). Gender schematic processing may be particularly likely for children during the preschool years when own-gender group interest is high and rigid adherence to gender stereotypes is common (Brown et al., 2020). Gender schemas may have facilitated children’s processing and organization of information in the math game when the character was part of their gender group (i.e., “I see a girl character, I am a girl, this content is for me”; Calvert & Huston, 1987).

The character’s gender may have served as an implicit label and activated children’s gender schemas about whether the character was an appropriate social partner (Calvert & Huston, 1987) or increased feelings of social presence with the character (Moreno & Mayer, 2004). Consistent with this idea, children reported a preference for same-gender characters, even though the characters were unknown to many of the children before gameplay. This finding is consistent with children’s preference for same-gender peers and favorite media characters (Brown et al., 2020; Richards & Calvert, 2017a) and suggests that the physical design of the character’s gender conveyed meaningful social information to children (Moreno & Flowerday, 2006).

While relying on gender schemas to interact with characters did not result in a greater amount of math talk with a same-gender character, compared to an other-gender character, these socially meaningful interactions had a greater effect on math outcomes and may have enhanced the learning process. For children who were gender matched, the additional effect of a one percentage point increase in math talk is associated with a .02 standard deviation decrease in average latency. The 25th percentile of math talk, which is mean centered, is –4.15 and the 75th percentile of math talk is 14.45. For children that are gender matched to the character, going from the 25th to the 75th percentile in math talk is associated with a .93 standard deviation decrease in average latency. This large effect reinforces the importance, in this study and in previous studies (Calvert et al., 2020), of the role of math talk in supporting children in reducing their average latency to correctly answer add-1 problems. The implication is that STEM content associated with same-gender characters through PSIs activate gender schemas and lower cognitive burden, resulting in children’s quicker correct answers on average when they engage with same-gender characters compared to children that engage with different-gender characters. Beneficial effects may occur because children can expend more resources on fluency with, and automatic processing of, the add-1 rule rather than on the relevance of the character who delivered the content (Baroody et al., 2013).

Although the effect of math talk was greater for children who learned the add-1 rule through PSIs with a same-gender character, children who engaged in more PSIs with other-gender characters also produced quicker correct answers to add-1 math problems in the virtual game compared to children who engaged in fewer PSIs. A one percentage point increase in math talk is associated with a .04 standard deviation decrease in average latency. This implies that going from the 25th percentile to the 75th percentile in math talk is associated with a .74 standard deviation decrease in average latency. The other-gender characters were not socially meaningful in terms of gender but were socially contingent partners for children through PSIs, which support the body of work on the positive role of social contingency in children’s learning from educational digital media (Barr, 2019; Calvert, 2017; Calvert et al., 2020; Kirkorian, 2018; Roseberry et al., 2014). Additionally, these results contribute to research on young children’s early math learning from pedagogical agents (Gulz, Londos, et al., 2020) and support findings that peer pedagogical agents can drive learning experiences for children and facilitate the processing of on-screen content when they engage in more PSIs.

These results extend the line of research on children’s STEM learning from intelligent character prototypes and broader research on children’s learning from pedagogical agents (Calvert et al., 2020; Schroeder et al., 2013). The study provides support for the idea that pathways to learning can occur through PSIs without a PSR, when characters are like children in salient ways (Calvert et al., 2020). Previous research featured Dora as an intelligent character prototype teaching math, when she was a well-known character, and the results revealed that PSIs and PSRs each contributed to children’s quicker correct response times (Calvert et al., 2020). When children do not know, or do not share salient similarities with the character, like gender, it may be particularly important for characters to provide ample opportunities for a conversational back-and-forth interaction about STEM content, which has been found to support learning from peer agents (Kim & Baylor, 2006). Media characters may be effective and motivating for young children to learn math from pedagogical agents (Gulz, Kjällander, et al., 2020; Gulz, Londos, et al., 2020; Husain et al., 2015).

Children who engaged in more math talk with a same-gender character also correctly answered more add-1 transfer task problems with physical objects when compared to children who did more math talk with other-gender characters. The IRR of correctly answering more transfer task problems was 1.02 times higher for children engaging with a same-gender intelligent character compared to an other-gender intelligent character. In other words, there was a two percent increase in the probability that children gender matched to the character correctly answered another transfer task problem correctly, as they engaged in more math talk. While this effect size is small, this finding is practically important. Success on the transfer task involved children’s generalization of the add-1 rule to harder math problems than previously practiced in the game (5 + 1 to 10 + 1). Transferring learning from a virtual context to a physical context is challenging for young children (Barr, 2010; Barr et al, 2019), although evidence suggests that it may be supported when children interact contingently with a social partner in interactive media (Barr, 2019; Kirkorian, 2018). This study suggests that when children interacted about math with a gender-matched character for a short amount of time in a single virtual game session, they were able to transfer math learning to math problems in a physical context. A future direction for this work is to expose children to the game repeatedly and test whether this may increase the effect size of transfer task performance.

The transfer task results suggest that children’s memories or underlying representations of gender-relevant math information were strengthened through PSIs with same-gender characters, which supported children’s transfer of math skills to a physical setting. The creation of stronger mental representations during virtual game play were likely facilitated by gender schemas streamlining cognitive processing of the add-1 math content, similarly to how children better learn information about their own gender (Signorella et al., 1993). Same-gender characters may have served as contingent social meaningful anchors across virtual and physical contexts, through which children may be able to understand that the content was “for them” in the game. Previous research on intelligent character prototypes teaching the add-1 rule revealed that transfer was facilitated by children’s socially meaningful PSRs and socially contingent PSIs (Calvert et al., 2020). Being part of the same social group, gender in this case, may operate similarly to PSRs in this instance when other information is not known about the character.

Contrary to prediction, more math talk in the game with an other-gender character did not predict children’s performance in the transfer task with physical objects. Children may have experienced higher processing demands when they could not relate the add-1 content in a socially meaningful way to themselves through gender similarity. This is consistent with previous research that found that children demonstrated poorer add-1 performance with physical objects after interacting with an unknown female voice-only intelligent character prototype, compared to an intelligent character prototype with whom they had a strong PSR (Calvert et al., 2020). The results of the current experiment suggest that PSIs about STEM content may not independently facilitate add-1 transfer with physical objects. It may be particularly important for the character to be perceived as sharing characteristics salient to a child’s identity if the child does not have previous experience with the character.

An implication of these studies for remote education is that intelligent characters may hold promise as remote educators, particularly when they are gender-matched to children. This supports the idea that gender is a design feature of intelligent characters and peer pedagogical agents that impacts children’s experiences and learning outcomes (Kim & Baylor, 2006). A media character’s gender may operate as a strong indication to children of when media content is “for them” or “not for them” when children are first encountering the character. Although the present study examines math performance at a single time point in time with characters that children may not know, the math game is a venue in which children can work on repeatedly towards fluency with add-1 problems. Repeated exposure to the intelligent character math game prototype could be incorporated into remote instructional activities for young children, rather than relying on in-person verbal math drills or worksheets. The findings of this study specify properties of novel games that feature pedagogical agents—in this case being the same gender as the child—which may lead to better learning from electronic games (Blumberg et al., 2019) and potentially foster distance learning opportunities.

The current research does not suggest that screen-based intelligent characters replace the physical, or remote, presence and expertise of a human. Rather, intelligent characters may be a useful tool in tandem with educators and parents when they can be designed to respond autonomously and do not need to be controlled via a Wizard of Oz approach. Screen-based intelligent characters can be flexibly utilized in remote education or to supplement more traditional in-school education. These contingent social partners can be programmed to be like children and encourage children’s active involvement with, and processing of, early STEM content. Teachers may be able to use intelligent characters to personalize support for children who are struggling (Brunick et al., 2016), as well as those who are rapidly mastering skills. Educators and parents may encourage their children’s learning if they support children’s engagement in more PSIs through toy play with the character, encouragement, and repeated exposure (Bond & Calvert, 2014). Through PSIs, children may form a deeper relationship with a character which may motivate them to return to the game for additional learning opportunities (Calvert, 2017).

An unanticipated methodological issue in the study was that the internal consistency of the attachment and friendship subscale of the Children’s PSR Survey was below acceptable thresholds (Richards & Calvert, 2017a). Previously, this subscale was successfully used to measure children’s attachment and friendship with well-known (Calvert et al., 2020) and favorite characters (Richards & Calvert, 2017a); however, in the present study, many children reported that they did not know the featured characters. Children who do not know a character cannot have a relationship with them, which is likely why internal consistency was low. Future research examining questions related to specific media characters may consider only utilizing the Child PSR Survey if children report character knowledge. Children’s lack of familiarity with characters highlights a challenge in studying children’s digital media and technology, which can change more rapidly than the pace of research. Notably, though, the characters in this prototype can be swapped out, which could make the game more useful to a broader base of children.

The generalizability of the findings is limited by the lack of an educationally diverse sample. Most parents of children in this sample held a postsecondary degree, and more than half held a graduate degree. The results suggest that children from households with higher socioeconomic statuses benefit educationally through engaging with an intelligent character prototype about STEM content, particularly when the character matches their gender. Future research should explore how a more socioeconomically diverse sample of children may learn from engagement with intelligent character prototypes, as children in lower income households engage in about 2 more daily hours of screen time compared to higher income households (Rideout & Robb, 2020). The characters examined here were Hispanic, which may provide familiarity to children who are the same ethnicity as Dora and Diego, a sample that may be of particular interest in future studies.

The intelligent character prototypes used in this study or in other studies (Calvert et al., 2020; Finkelstein, 2018) do not yet operate without a “Wizard” manipulating socially contingent replies to children. Thus, the promise of socially contingent autonomous intelligent characters or intelligent peers for remote education is not yet feasible. This study provides evidence of a principle needed for effective delivery of information, socially contingent replies, and can be used in the development of effective remote educational applications.

Conclusion

The necessity of educating young children online due to the COVID-19 pandemic, throughout the U.S. and most of the world, was unprecedented. However, given the heavy use of digital media and children’s access to the media characters that inhabit their virtual experiences (Rideout & Robb, 2020), educational digital media is well suited to bridge the space between school and home. This bridge between home and school through remote education can strengthen overall learning. Digital interfaces that include intelligent media characters can be harnessed as an educational tool that may be particularly well suited for remote education. This study teaches us that intelligent media character prototypes can foster children’s early STEM skills when children see the characters as “for them” and when characters engage children in conversational back-and-forth PSIs that optimize learning. Creating more opportunities for preschool-aged children to learn from intelligent digital media characters as their social partners may create a viable pathway for engaging children in 21st century STEM domains. Additionally, intelligent characters that support STEM learning can link home to school learning spaces, particularly during challenging times like the COVID-19 pandemic where remote education was necessary.

Appendix: Supplemental Tables

Table A1

OLS Regression Predicting Average Response Latency in Seconds for Add-1 Problems, No Control Variables

Variable

Model 1

Model 2

Math talk

–.79** (.11)

–.50** (.14)

Child-character gender match * Math talk

–.50** (.18)

Child-character gender match
(0 = other-gender, 1 = same-gender)

1.70 (3.40)

Constant

20.84 (1.71)

23.90 (2.70)

Regression statistics

 R 2

.44

.49

 Adjusted R 2

.43

.47

 F

50.34**

30.55**

 df

1, 88

3, 86

 N

90

90

Note. Math Talk: Calculated by dividing the number of math talk prompts a child replied to by the number of math talk prompts that were available to a child. That number was multiplied by 100, and math talk was mean centered at 85.55. Child-character gender match: 1 = child and character are the same gender, 0 = child and character are different genders. OLS = ordinary least squares.
* p ≤ .05. ** p ≤ .01. Robust standard errors.

Table A2

Poisson Regression Predicting the Number of Add-1 Transfer Task Problems Answered Correctly

Variable

Model 1

Model 2

Math talk

1.00 (.01)

.99 (.01)

Child-character gender match * Math talk

1.02** (.01)

Child-character gender match
(0 = other-gender, 1 = same-gender)

1.35 (.25)

1.32 (.24)

Character knowledge

.97 (.11)

.97 (.10)

Age

1.00 (.001)

1.00 (.001)

Constant

.90 (1.05)

1.16 (1.28)

Regression statistics

 Pseudo R 2

.03

.05

 Wald X 2

5.46

10.34

 df

4

5

 N

88

88

Note. Sample excludes a child who did not answer character knowledge questions and a child who did not complete the transfer task. Math Talk: Calculated by dividing the number of math talk prompts a child replied to by the number of math talk prompts that were available to a child. That number was multiplied by 100, and math talk is mean centered at 85.55. Controls: Child-Character Gender Match: 1 = child and character are the same gender, 0 = child and character are different genders; Character Knowledge (0 = no knowledge, 1 = recognize character, 2 = know character’s name); Age (days). IRR = incidence rate ratio.
* p ≤ .05. ** p ≤ .01. Robust standard errors.