Identifying middle school students’ challenges in computational thinking-based science learning

Design as a core focus of learning using computational programming and modeling

From a pedagogical perspective, this means that engaging students in developing design-based computational representational practices, such as the ones discussed above, can be closely aligned with the development of students’ CT skills.

Several scholars have pointed out that computing can be used successfully as a medium for teaching and learning other subjects and that this can facilitate learning in both the subject and computing domains. For example, Papert (1991) stated that programming is reflexive with other domains; that is, learning programming in concert with concepts from another domain (such as math and science) can be easier than learning them separately. Kay and Goldberg (1977) showed that object-oriented programming using SmallTalk is useful for learning math, science, and art. Emile, a scaffolded graphical programming interface designed and used to help students learn physics, represents another example of synergistic learning (Guzdial 1994). Redish and Wilson (1993), Soloway (1993), and Kafai et al. (1997) also demonstrated that reorganizing scientific and mathematical concepts around computational mechanisms lowered the learning threshold, especially in domains like physics and biology. More recently, some researchers have exploited the synergy between CT and science to develop CT-based science curricular units for K-12 classrooms (Sengupta et al., 2015; Basu, Kinnebrew & Biswas 2014; Allan et al. 2010; Repenning et al. 2010).

In each of the environments discussed above, students learn through an iterative model building process. Previous studies have shown that middle school and elementary children can successfully use programming as a mode of inquiry to develop models of scientific phenomena, which in turn helps them develop a deep understanding of the relevant scientific concepts (diSessa et al. 1991; Sengupta & Farris 2012). CTSiM adopts this learning-by-design pedagogical approach (Kolodner et al. 2003), and students iteratively design, test, and revise computational models of physics and ecology.

Agent-based modeling can leverage students’ prior knowledge

CTSiM is an agent-based modeling environment. The term “agent” here indicates an individual computational object or actor (for example, a rollercoaster car or a fish in a fish tank), which performs actions (for example, moving forward, changing directions) based on simple rules, and these rules can be designed and controlled by the user. Several researchers have shown that agent-based modeling can leverage K-12 students’ pre-instructional intuitions and support their learning of (a) complex and emergent phenomena in biology, such as population dynamics in ecological systems (Basu et al. (2014); Dickes and Sengupta 2012; Wilensky and Reisman 2006), and (b) phenomena in the domain of Newtonian mechanics that require students to develop an understanding of the relations between position, speed, and acceleration as aggregation of continuous change in these variables over time (Basu et al. 2012; diSessa et al. 1991).

The advantages of visual programming

In a visual programming (VP) environment, students construct programs using graphical objects and a drag-and-drop interface, thus making the programming more intuitive and accessible to the novice programmer (Kelleher and Pausch 2005). Visual constructs significantly reduce issues with program syntax and understanding textual structures making it easier for students to focus on the semantic meaning of the constructs (Soloway 1993). For example, visual interfaces make it easier to interpret and use flow of control constructs, such as loops and conditionals (Parsons and Haden 2007a, b).

CTSiM provides a library of visual constructs that students can choose from and arrange spatially to generate their computational models. If students try to drag and drop a programming construct incorrectly, the system disallows the action and indicates the error by explicitly displaying an “x” sign. Therefore, CTSiM eliminates the possibility of generating programs (that is, models) with syntax errors. Examples of other agent-based VP environments include AgentSheets (Repenning 1993), StarLogo TNG (Klopfer et al. 2005), Scratch (Maloney et al. 2004), ViMAP (Sengupta et al., 2015), and Alice (Conway 1997). They have been used successfully in teaching children CT through game design, storytelling, and modeling activities.

Integration of domain-specific primitives and domain-general abstractions

Previous research suggests that learning a domain-general programming language and then using it for domain-specific scientific modeling involves a significant pedagogical challenge (Guzdial 1994; Sherin et al. 1993). To address this issue, CTSiM combines domain-general and domain-specific primitives. Domain-general primitives are computational constructs (for example, “when-do-otherwise do” and “repeat” representing conditionals and loops). Domain-specific primitives are designed specifically to support modeling of particular aspects of the topic of study, for example, kinematics or ecology. Imposing domain-specific names on the constructs creates semantically meaningful structures for modeling actions in the particular domain. For example, “forward,” “speed up,” and “slow down” represent movement, acceleration, and deceleration actions in kinematics, respectively, and “create new” and “die” imply birth and death of agents in ecology, respectively. Students develop more complex agent behaviors by combining computational and domain-specific primitives. Examples include “model car speed” behavior in kinematics and “breathe” and “eat” behaviors in ecology. Previous studies suggest that such an approach that combines domain-general and domain-specific computational primitives can effectively support the development of children’s scientific models and conceptual understanding of the domain, as well as support the development of their programming concepts and skills.

Known challenges for programming and learning-by-modeling in science

Developing a scaffolding framework for an environment like CTSiM which is intended to be used in classroom settings warrants an in-depth understanding of the different types of difficulties students at different levels of understanding face in the environment (Puntambekar and Hubscher 2005). Previous research has separately documented challenges associated with science learning, programming challenges faced by students and challenges faced with inquiry learning using modeling and simulation. However, when science learning and learning programming skills are combined in a modeling- and simulation-based learning environment, the types of challenges that emerge have not been explored. In this section, we explore the known challenges in each of these areas.

Instructional approaches in science emphasize learning by engaging in knowledge construction practices, investigation, and argumentation. These approaches to learning through inquiry not only provide the potential to connect knowledge more effectively to real-world contexts but also pose particular challenges for learners (Reiser 2004). For example, learners may not be familiar with general strategies for designing empirical tests of hypotheses and in using specific domain knowledge to plan and guide investigations (Schauble et al. 1991). They also tend to focus on achieving desired results rather than on understanding the principles behind the results (Perkins 1998) and find it difficult to generalize appropriately from their work on specific problem scenarios. Further, students tend to have difficulty mapping their intuitive understandings to formal representations and evaluating alternate representations (Sherin 2001). In addition, students may face social interaction and collaboration challenges or linguistic and discourse challenges (Reiser 2004).

The challenges faced in learning science through investigative processes or discovery learning can be grouped in a number of ways. Quintana et al. (2004) categorizes the challenges into three categories, those related to sense making, process management, and articulation and reflection. Sense making entails constructing and interpreting empirical tests of hypotheses. Students need to coordinate their reasoning about experiments or data comparisons with the implications of the findings for an explanation of the scientific phenomena. This coordination and mapping task is complex and requires rich subject matter knowledge to design data comparisons and interpret findings in light of the hypotheses. Process management involves the iterative processes of designing an investigation, collecting data, constructing and revising explanations based on data, evaluating explanations, and communicating arguments. These require both discipline-specific processes and content knowledge that may be new to learners. Finally, scientific investigations require the complementary processes of reflection and articulation as students monitor and evaluate their progress, reconsider and refine their plans, and articulate their understanding as they proceed. Thus, in learning science through inquiry or investigative process, students face challenges at several levels. They face challenges with the content knowledge, as well as the cognitive complexity of discipline-specific strategies for sense making and process management, and the metacognitive processes for social interaction and discourse association with scientific practices (Reiser 2004).

On the other hand, de Jong and van Joolingen (1998) identify a number of characteristic problems that learners may encounter in discovery learning with computer simulations and classify them according to the main discovery learning processes: hypothesis generation, design of experiments, interpretation of data, and regulation of learning. These challenges hold good for discovery learning using computer simulations in any domain including science. Generating hypotheses and adapting or rejecting hypotheses based on collected data seem to be common challenges. Also, students display confirmation bias—the tendency to seek for information confirming hypotheses or construct experiments that are not intended to test a hypothesis. They tend to design inconclusive experiments and show inefficient experimentation behavior. In addition, interpretation of data is often directed by the hypothesis and the tendency to find data confirming the hypotheses. In particular, students find interpretation of graphs extremely difficult. Finally, planning experiments and working in a systematic fashion are processes students find challenging.

Students’ challenges with studying scientific phenomena using a complex systems framework have also been studied extensively. In such systems, the collective, global behavior emerges from the properties of individual elements and their interactions with each other. The global or macro behaviors—known as emergent phenomena—are often, not easily explained by the properties of the individual elements. For example, in chemistry and physics, gas molecules’ elastic collisions at the micro level produce the macro-level properties of pressure and temperature. In biology, animals interact with others of the same and different species and the environment to survive, grow, and reproduce at the individual level that leads to phenomena such as evolution, natural selection, and population dynamics at the ecosystem level. Students find the behaviors of individual elements intuitive but struggle to understand their relations with the aggregate behavior (Wilensky & Resnick 1999; Chi 2005).

Studies have not yet been conducted for studying students’ challenges with learning CT skills, but several studies have documented the challenges students face while writing programs. Most of these challenges are, however, in the context of undergraduate programming with text-based programming languages. For example, students are found to have difficulties with assembling programs and writing syntactically correct programs. Programming languages tend to have only a few components which are combined in many different ways, and learning to understand the semantic results of different combinations is considered complex. Also, understanding how to combine programs to achieve particular goals is known to be a challenge (Spohrer 1989). When students try to assemble programs by combining elements, they often get confused with syntax problems as they struggle to understand semantic ones. Another known programming challenge in the literature is students’ lack of understanding of computational processes. Many students do not understand how interpretation of traditional computer languages works, e.g., where does control flow and how do variables get updated (DuBoulay 1989).

We expect to see some of these known science and programming challenges with CTSiM as well. Since CTSiM tries to leverage the synergy between computational thinking and science education by making students computationally model a scientific phenomenon, we also anticipated situations where students’ programming challenges might be compounded by challenges with the science domain content, or vice versa, and were prepared to interleave scaffolds for the science content and the programming task.

Scaffolds in existing CT and science learning environments

The term scaffolding, as it relates to education, was introduced by Wood et al. (1976) as a metaphor describing how teachers and tutors assist learners in completing learning tasks that, without assistance, the learners would be unable to complete. Additionally, the authors list six scaffolding functions that tutors may employ: recruitment, reduction in degrees of freedom, direction maintenance, marking critical features, frustration control, and demonstration. This definition of the scaffolding process focuses on a relationship between two people and their interactions; it highlights the difficult but important task of continually diagnosing and adapting to the needs of the learner, whether that involves providing additional support, in the case that the learner is struggling, or removing support, in the case that the learner is excelling (Puntambekar and Hubscher 2005). Since this metaphor was introduced, researchers have expanded and generalized it to different aspects of computer-based learning environments. Some researchers define scaffolds as interface features (e.g., explanation construction tools; Reiser 2004). Others define scaffolds as activity sequencing within the learning environment (e.g., requiring students to answer questions before starting an invention task; Roll et al. 2012). Still others define scaffolds as supportive actions taken for the purpose of supporting learners in completing their tasks (e.g., providing hints; Azevedo and Jacobson 2008; Basu & Biswas 2016).

In this section, we discuss scaffolding mechanisms documented in the literature for helping students overcome the science and programming challenges discussed in the previous section. Reiser (2004) proposes two complementary mechanisms of scaffolding in software tools to help students with their science inquiry challenges related to sense making, process management, articulation, and reflection. He proposes (i) structuring problem-solving tasks to make them more tractable and to shape tasks for learners in ways that makes their problem-solving more productive and (ii) problematizing subject matter to provoke learners to devote resources to issues they might not otherwise address. Students’ learning tasks can be structured by providing structured work spaces to help decompose a task and organize work to help recognize important goals to pursue. Explicit structures such as prompts, agendas, or graphical organizers can help learners monitor their progress and keep track of what goals have been addressed and what aspects of the task are pending. Also, restricting the problem space by narrowing options, preselecting data, or offloading more routine parts of the task can help learners focus resources on the aspects of the task more productive for learning. The second proposed mechanism for scaffolding is to make some aspects of students’ work more “problematic” in a way that increases the utility of the problem-solving experience for learning. Rather than simplifying the task, the software leads students to encounter and grapple with important ideas or processes. This may actually add difficulty in the short term, but in a way that is productive for learning. For example, eliciting articulation or collaboration can help counter the tendency toward superficial and non-reflective work. Similarly, eliciting arguments and decisions can force students to think deeply about the content and the relations between the evidence and their arguments.

On the other hand, de Jong and van Joolingen (1998) describe different ways to support learners’ challenges with discovery learning using modeling and simulation. They suggest providing the learner with direct access to domain information and then providing support for specific discovery processes. Insufficient prior knowledge might be the cause that learners do not know which hypothesis to state, cannot make a good interpretation of data, and move to unsystematic experimentation behavior; hence, providing access to domain information comprises the first level of support. Then, students can be supported by providing them with a hypothesis menu or scratchpad, experimentation hints and strategies, tools for making predictions, and planning and monitoring tools. Decomposing and structuring the discovery process can also be useful scaffolds.

With respect to supporting learning of emergent science phenomena, agent-based modeling holds immense potential. As discussed earlier in the “Agent-based modeling can leverage students’ prior knowledge” section, it provides the means to build on students’ intuitive understandings about individual agents acting at the micro-level to grasp the mechanisms of emergence at the aggregate macro-level.

Scaffolds for programming challenges are limited and have focused on pointing out syntax errors in students’ programs or providing tools to help debug programs. Alleviating syntax problems is believed to help students focus on the semantic ones (Soloway 1993). In fact, research comparing learning in a more and a less syntactically strict language, Java and Python respectively, attribute the greater success of students in Python to be a result of reduced syntactic complexity (Mannila et al. 2006). Alleviating syntactic complexity is something we achieve in CTSiM by using a visual programming paradigm. Thus, bugs in CTSiM are always semantic errors and never the result of a typing error or a misremembered detail of the language syntax. Some research also claims that visual programming languages can make understanding the algorithmic flow of control more accessible by making complex elements of flow of control, such as loops and conditionals, more natural (Parsons and Haden 2007a, b).

While several existing computer-based learning environments include scaffolds like explanation construction tools, guiding questions, argumentation interfaces, workspaces for structuring tasks, data comparison tools, and tools for observing effects of plans made or models built, several of these scaffolds are part of the environment design and are not adaptive. As Puntambekar and Hubscher (2005) point out, such tools now described as scaffolds provide us with novel techniques to support student learning, but they neglect important features of scaffolding such as ongoing diagnosis, calibrated support, and fading. Adaptive scaffolding involves responding to individual learner challenges. In computer-based environments, it involves tracking and interpreting learner actions. For example, in Ecolab (Luckin and du Boulay 1999)—a modeling- and simulation-based environment, the scaffolding agent intervenes whenever students specify an incorrect relationship in their models and provides a progression of five hints, each more specific than the previous one, with the final hint providing the answer. Co-Lab (Duque et al. 2012), on the other hand, tracks student actions to provide feedback about both students’ solutions (the models built by students) and work processes, but is still limited to reminding students about specific actions they have not taken or should employ more frequently for model building and testing. AgentSheets (Repenning et al. 2010) is an example of one of the very few CT-based environments which include adaptive scaffolds. Students are provided an automatic assessment of the computational artifacts they build (games or science simulations). The CT patterns present in students’ artifacts are compared against desired CT patterns for the artifacts and represented in terms of what is known as the Computational Thinking Pattern graph.

Materials

CTSiM curricular units

Kinematics (physics) and ecology (biology) were chosen as the curricular topics for synergistic learning of science and CT using CTSiM. They are common and important curricular topics in the middle school curriculum, and as Sengupta et al. (2013) argued, researchers have shown that K-12 students have difficulties in understanding and interpreting concepts in these domains (Chi et al. 1994). Furthermore, it has been argued that students’ difficulties in both the domains have similar epistemological origins, in that both kinematics phenomena (e.g., change of speed over time in an acceleration field) and system-level behaviors in an ecosystem (e.g., population dynamics) involve understanding aggregation of interactions over time (Reiner et al. 2000; Chi 2005). Also, agent-based modeling is well suited for representing such phenomena, as it enables learners to invoke their intuitions about agent-level behaviors and organize them through design-based learning activities, in order to explain aggregate-level outcomes. Studies have shown that pedagogical approaches based on agent-based models and modeling can allow novice learners to develop a deep understanding of dynamic, aggregate-level phenomena—both in kinematics and ecological systems by bootstrapping, rather than discarding their agent-level intuitions (Farris & Sengupta, 2016; Dickes & Sengupta, 2013; Dickes, Sengupta, Farris & Basu, 2016; Wilensky and Reisman 2006; Levy and Wilensky 2008). Student activities in kinematics and ecology are explained in greater detail below.

Kinematics unit: We extended previous research by Sengupta, Farris & Wright (2012) to design the kinematics unit in three phases.

Kinematics phase 1: This covered activities 1 and 2, where students used turtle graphics to construct geometric shapes that represented: (1) constant speed and (2) constant acceleration. In activity 1, students were introduced to programming primitives such as “forward,” “right turn,” and “left turn” that dealt with the kinematics of motion, primitives like “repeat” which corresponded to a computational construct (independent of a domain construct), and primitives like “pen down” and “pen up” which were Netlogo-specific drawing primitives. The students were given the task of generating procedures that described the movement of a turtle for drawing n-sided regular shapes, such as squares and hexagons. Each segment of the regular shape was walked by the turtle in unit time indicating constant speed. Therefore, activity 1 focused on students learning the relationship between speed, time, and distance for constant speed motion. In activity 2, students were given the task of extending the turtle behavior to generate shapes that represented increasing and decreasing spirals. In this unit, segments walked by the turtle, i.e., its speed per unit time, increased (or decreased) by a constant amount, which represented a positive (or negative) acceleration. Activity 2 thus introduced students to the relations between acceleration, speed, and distance using the “speed up” and “slow down” commands to command the motion of the turtle.

Kinematics phase 2 corresponded to activity 3, where students interpreted a speed-time graph to construct a representative turtle trajectory. Starting from the speed-time graph shown in Fig. 3, students developed a procedure where the length of segments the turtle traveled during a time interval corresponded to the speed value on the graph for that time interval. For example, it was expected that students would recognize and model the initial segment of increasing speed by a growing spiral, followed by the decrease in speed by a shrinking spiral, whose initial segment length equaled the final segment length of the last spiral. Students were given the freedom to choose the shapes associated with the increasing and decreasing spirals. We hypothesized this reverse engineering problem would help students gain a deeper understanding of the relations between acceleration, speed, distance, and time.

Kinematics phase 3, represented by activity 4, involved modeling the motion of a rollercoaster car along a pre-specified track with multiple segments. In more detail, students were asked to model a rollercoaster as it moved through different segments of a track: (1) up (pulled by a motor) at constant speed, (2) down (with gravitational pull), (3) flat (cruising), and then (4) up again (moving against gravity). The students had to build their own model of rollercoaster behavior to match the observed expert behavior for all of the segments.

The ecology unit was represented by activities 5, 6, and 7, where students modeled a closed fish tank system in two phases. In the first phase (activity 5), students constructed a macro-level, semi-stable model of the fish tank ecosystem by modeling the fish and duckweed species as two agent types. Activity 5 required students to model the food chain, respiration, locomotion, excretion, and the reproductive behaviors for the fish and duckweed. The inability to develop a sustained macro-model, where the fish and the duckweed could be kept alive for extended periods of time, even though all of the macro processes associated with the two agents were correctly modeled (that is, the behaviors generated by the students’ computational model matched the behaviors generated by the expert model), encouraged students to reflect on what may be missing from the macro-model. This led to the realization about the need to model the waste cycle and its entities, primarily the two forms of bacteria and their behaviors. This prompted the transition to the second phase (activity 6) where students identified the continuously increasing fish waste as the culprit for the lack of sustainability of the fish tank. Students then built the waste cycle model for the fish tank, with the Nitrosomonas bacteria that converts the toxic ammonia in fish waste into nitrites, which is also toxic, and the Nitrobacter bacteria that converts the nitrites into nitrates. Nitrates are consumed by the duckweed (as nutrients) thus preventing an excessive buildup of toxic chemicals in the fish tank environment. The combination of graphs from the micro- and macro-world visualizations was intended to help the students develop an aggregate-level understanding of the interdependence and balance among the different agents (fish, duckweed, and bacteria) in the system. After completing the ecology micro-unit, students worked on activity 7 where they discussed the combined micro-macro model with their assigned researcher and how the macro-micro model phenomena could be combined into an aggregated causal model describing the sustainability of the fish tank ecosystem.

The sequencing of curricular modules allowed students to tackle modeling and reasoning with a single agent in kinematics first and then build more complex computational models with multiple agents in ecology. This was an intentional design decision because studies in developmental psychology (for example, Lehrer et al. 2008) and agent-based modeling for education (for example, Goldstone & Wilensky 2008) show that individual agent-level reasoning occurs developmentally prior to understanding interactions among agents, and eventual aggregate-level reasoning with multiple agents and processes. Furthermore, within each unit, the sequencing of the activities implied increasing conceptual challenges that students would face in learning the relevant phenomena. For example, in the kinematics unit, when students modeled a single agent, the computational modeling tasks were presented in the order of increasing complexity, starting from constant shapes (squares to triangles to circles) to spirals of the same shapes (where speed became a function of the acceleration) to modeling real-world systems involving constant and variable speed segments.

Sample and procedure

Fifteen 6th grade students (age ranged between 11 and 13) from an ethnically diverse middle school in middle Tennessee worked on CTSiM in a pull-out study with one-on-one individualized verbal guidance from one of five members of our research team. The students were chosen by their classroom science teacher, who also happened to be their science teacher. The teacher ensured that the chosen students were representative of different genders, ethnicities, and performance levels based on their state-level test scores (Tennessee Comprehensive Assessment Program or TCAP). All the students in the class (those who were chosen for the pull-out study as well as those who were not) had provided their consent (student and parental consent) for working with CTSiM. Hence, while the majority of the class participated in the CTSiM pull-out study, the remainder of the class (nine students) was allowed to explore the CTSiM learning environment on their own without one-on-one guidance (with minimal guidance from the teacher and some other members of our research team) during the science period, and the teacher made sure they learnt the same science topics as covered in the CTSiM learning activities during this time. In this paper, we focus only on the 15 students who participated in the pull-out study, and the data and analysis we present are derived from their pre- and post-tests, their interactions with the CTSiM environment, and the conversations they had with their assigned researcher. Since this was our first study with the CTSiM system, our goal was to use the one-on-one interactions to determine the approaches the students used in constructing their models, the problems they faced during model building, how they discovered and responded to errors in their models, and scaffolds provided by the researchers that were effective in helping them deal with the challenges they faced when they lacked domain knowledge, or when they tried to correct errors in their models.

The 15 students were paired one-on-one with one of the five members of our research team. Thus, each researcher from our team worked with three students for the study with three 1-h sessions daily (9 am–10 am, 10 am–11 am, 12:30 pm–1:30 pm), one for each student assigned to them. On ay 1, all 15 students took the paper-based pre-tests for both the kinematics and ecology units. They took between 25 and 40 min to finish each test. Then, the students worked on the kinematics units (activities 1–4) from day 2 through day 4 and took the kinematics post-test on day 5. On days 6–8, they worked on the ecology units (activities 5–7) and then took the ecology post-test on day 9. The students worked in the CTSiM environment with their assigned researcher sitting next to them, interacting with them when needed. The entire study took place over a span of 2 weeks toward the end of the school year, after the students had completed their annual state-level assessments (TCAP).

All five members (one assistant professor and four graduate research assistants) of our research team who conducted the one-on-one interviews had prior experience with running similar studies. They met before the study and decided on a common framework for questioning the students and interacting with them as they worked with CTSiM. While the interviews were not strictly scripted since the conversations would depend on individual student actions and thought processes, a common flexible interview script was prepared and shared among the researchers. This ensured that all of the researchers’ interview formats and structures were similar (similar questions asked and similar examples to illustrate a concept) during each of the CTSiM learning activities. As part of the intervention, the researchers introduced the CTSiM system and its features to their students individually and introduced each of the learning activities before the student started them. However, the students were not told what to do; they had complete control over how they would go about their modeling and debugging tasks. But the researchers did intervene to help the students when they were stuck or frustrated by their own lack of progress. An important component of the researchers’ interactions with the students involved targeted prompts, where they got the students to focus on specific parts of the simulation results and verify the correctness of their model. When needed, the researchers also asked leading questions to direct the students to look for differences between the expert simulation results and their own results and then reflect on possible causes for observed differences. These questions often required the students to predict the outcome of changes they had made to their models and then check if their predictions were supported by the simulation results.

In addition, the researchers prompted the students periodically to make them think aloud and explain what they were currently doing on the system. They also provided pointers about how to decompose large complex modeling problems into smaller manageable parts and at appropriate times, reminded the students about how they had tackled similar situations in past work. All of the student and researcher conversations during the one-on-one interviews were recorded using the Camtasia software.¹ These videos also included recordings of the screen, so we could determine what actions the students performed in the environment and what the consequences of those actions were.

Analysis and coding

We scored students’ pre- and post-tests and also analyzed the Camtasia™-generated videos for all 15 students to characterize the types of challenges the students faced while working with CTSiM and the scaffolds that were provided to help them overcome these challenges. Two members of our research team came up with initial rubrics for grading the pre- and post-tests, which were then iteratively refined based on student responses. The initial rubric focused on correct answers for multiple-choice questions and keywords and important concepts for questions requiring short answer responses. A systematic grading scheme was developed after studying a subset of the short answer responses. The short answer grading scheme attempted to account for different ways a question could be answered correctly and was updated if we found a student response which could not be graded adequately using the current rubric. We have since used these pre-post grading rubrics in other studies (Basu et al 2014; Basu et al. 2015), and have found the rubrics to be reliable and valid with a variety of student responses from different studies.

The video data was coded along two dimensions: first, the type and frequency of challenges faced during each activity and second, the scaffolds that were used to help the students overcome the challenges. Initial codes were established using the constant comparison method by two researchers involved in the study. To do so, they chose data from two participants, whom we will call Sara and Jim (not their real names). They were selected as representative cases because they had the lowest and highest state standardized assessment (TCAP) scores in science among the 15 participants of the pull-out study (Basu et al. 2013). When the students voluntarily asked their interviewer/research member a question or mentioned they were not sure what to do next or asked for help with building and debugging models, these counted as challenges. Even if the students themselves did not ask for help, the students were frequently asked to explain what they were doing, why they were doing what they were doing, what they planned to do next and why, and predict the results of their actions. When the students could not correctly predict or explain their model behaviors, describe the semantics of programming blocks, explain their actions in the system, or how they planned to check and debug their model, these were also considered as student challenges. When we found instances of challenges, we documented the challenge using a brief description, its associated timestamp, and the scaffold provided.

Definitions and examples of the types of challenges (derived by initial analysis and repeated re-analysis to refine the definitions) are explained in detail in the “Challenges faced and scaffolds required” section. Fourteen challenge categories were identified and further grouped into four broad categories: (1) programming challenges, (2) modeling challenges, (3) domain challenges, and (4) agent-based reasoning challenges—to aid in the interpretation of the aggregate data set. Henceforth, we refer to the 14 initial categories as “subcategories” of these four broad categories.

Two researchers unaffiliated with the study coded the remaining video data from the other 13 participants, using our coding scheme described above. To establish reliability, they were first asked to determine the challenges and frequency counts for activities 3, 4, and 5 from Sara’s video data. Both coders reached good agreement with the researcher-developed codes (91.15 and 96.46 % agreement). Once reliability with the researcher codes was established, the coders were asked to code a different student to test their inter-rater reliability. The inter-rater reliability between coder 1 and coder 2 yielded a Cohen’s kappa of 0.895 (93.1 % agreement), implying a “very good” inter-rater reliability rating. Then, the coders divided up the work of coding the remaining 12 student videos. Once the challenges faced and scaffolds received for all 15 students were extracted from the video files (used to answer our first research question), we computed the average number of challenges of each type per activity (to answer research question 2).

Pre-post learning gains with CTSiM

Since these pre-to-post learning gains are clearly due to a combined effect of the use of the CTSiM environment and the verbal scaffolds provided to the students, we compared these gains against the pre-to-post gains for the other nine students in the class who explored CTSiM on their own without any external scaffolding. Unsurprisingly, we found that those nine students also showed learning gains, but the effect sizes (Cohen’s d) were much lower (0.05 versus 0.71 for kinematics and 1.09 versus 3.16 for ecology) compared to the students who received one-on-one scaffolding. We also computed repeated measures ANCOVA with TCAP science scores as a covariate of the pre-test scores to study the interaction between time and condition. Not surprisingly, there was a significant effect of condition (i.e., students who received one-on-one scaffolding and students who did not) on pre-post learning gains in kinematics (F(1,21) = 4.101, p < 0.06), as well as ecology (F(1,21) = 37.012, p < 0.001), indicating the scaffolding helped students learn science content better.

Challenges faced and scaffolds required

Number of challenges and their evolution over time

As further analysis beyond the different types of difficulties the students face when working with CTSiM and the scaffolds which can help them in such situations, we also studied how the frequency of challenges varied across learning activities in one domain and across domains. This helped understand the complexities associated with different learning activities and the variation in support required in these activities.

First, we ran an agglomerative complete-link hierarchical clustering algorithm to see how the students grouped based on their challenge frequency profiles per activity. The results showed that the students generally exhibited similar challenge profiles with the exception of one student (see Fig. 4).

Figure 4 shows the challenge profiles of the two clusters—the average challenge profile for the similar group of 14 students and one outlier, a single student who seemed to face many more challenges than the rest of the students. This student needed more scaffolding than the other students, and several challenges had to be scaffolded more than once before the student could overcome those difficulties. This student’s pre-test and standardized state-level test scores were much lower than those of the other students, which may explain why the student had a significantly higher number of challenges initially. Though this student had multiple challenges that persisted through multiple activities, the number of challenges the student had came closer to the number of challenges the others faced at the end of the kinematics (activity 4) and ecology units (activity 7). Similarly, the student’s post-test scores also matched that of the others, making this student’s pre-post gains higher than most of the students.

Next, we analyzed how the average number of challenges per student varied across the kinematics and ecology units and across the activities in each unit. The average number of challenges for an activity is calculated as the total number of challenges for all 15 students for an activity divided by 15. This number depends on new challenges that the students face in an activity, as well as the effectiveness of scaffolds received in previous activities. This is because whenever the students faced challenges in an activity, they were scaffolded. If the scaffolding was successful, the students were not likely to face the same challenges again in their model building and checking tasks. However, we did observe similar challenges resurfacing later in the same activity or in subsequent activities; therefore, the students were often provided with the same scaffolds more than once. Latter scaffolds often started with a reminder that this scaffold was provided before when the student faced the same challenge.

Figure 5 shows how the average number of challenges varied across the different activities. The number of challenges decreased across similar activities in the same domain. For example, the number of challenges decreased through the progression of shape drawing activities (activities 1–3); similarly, they decreased from activity 5 through activity 7 for the ecology units. The challenges increased in the transition between domains (activity 4 in kinematics to activity 5 in ecology) and between problem types in a domain (activity 3 to activity 4 in the kinematics domain). This was because activity 4 (the rollercoaster activity) introduced a number of new modeling and programming challenges. It required building a model of a real-world phenomenon by taking into account relevant variables such as steepness of the rollercoaster ramp. In addition, this was the first activity where the students’ simulation model behaviors had to match that of an expert model behavior. This required a better understanding of the simulation output, which was presented as a combination of an animation and graphs. Moreover, this activity was more challenging from a computational modeling viewpoint, because the model required the use of nested conditionals and variables. The students were experiencing these computational concepts for the first time, and this explained the increase in the difficulties they faced. Similarly, when students progressed from the kinematics domain to the ecology domain, activity 5 (the fish-tank macro-model) introduced additional complexities in a new domain. First, the students had to scale up from a single-agent to a multi-agent model. It also involved modeling multiple behaviors for each agent, and the students had to figure out how to modularize behaviors, for example, what to include in the fish “eat” behavior versus the fish “swim” behavior. (The two are related—a fish has to swim to its food before it can eat it).

Next, we looked at previously observed challenge categories which resurfaced and increased in activities 4 and 5. In activity 4, the only previously observed challenges which increased instead of going down with time were the programming challenge related to understanding the syntax and semantics of domain-specific primitives and the modeling challenge related to model validation. Facing challenges with respect to understanding domain-specific primitives seems understandable in the wake of new domain knowledge and related domain knowledge challenges. Also, activity 4 marked the first time the students had to perform model validation by comparing their model simulations against expert simulations and had to compare the two sets of animations and plots to assess the correctness of their models. Similarly, in activity 5, there were a few challenges previously observed in activity 4 which resurfaced and increased. For example, programming challenges related to the use of CT primitives increased, as did modeling challenges related to identifying relevant entities and their interactions, choosing correct preconditions, and specifying model parameters and component behaviors. A new domain, increase in domain complexity, and dealing with modeling multiple agents and multiple behaviors for each agent seem to have been the primary contributors. Further, the size (number of blocks contained) of the fish macro expert model was about thrice that of the expert rollercoaster model, increasing the probability of facing various difficulties in this activity (activity 5). Challenges with using CT constructs like conditionals resurfacing in the context of complex domain content emphasize the need for further practice and a more holistic understanding of the constructs. Unfortunately, we did not study computational learning gains using pre- and post-tests in this initial study, but they may have indicated that students needed repeated practice in different contexts to gain a deep understanding of the computational constructs. In other more recent studies with modified versions of CTSiM (modified based on challenges identified in this paper), we have shown synergistic learning of science and CT concepts (Basu et al. 2014; Basu et al. 2016). In Figs. 6, 7, 8, and 9, we investigate these issues further, by analyzing the data available from this study to study how the four primary categories of challenges individually varied across activities.

Figure 6 shows that students generally had fewer difficulties with domain knowledge in kinematics (activities 1–4) than in ecology (activities 5–7). For kinematics activities, the challenges did increase with the introduction of new domain-specific concepts like acceleration and the operation of a rollercoaster. But there was a sharp increase in the number of challenges when students had to deal with multiple agents and their interactions in the macro and micro fish tank activities. The difficulties were further compounded by students’ low prior knowledge in ecology as indicated by their low ecology pre-test scores.

Programming challenges show a similar trend as seen for average number of challenges in general in Fig. 5. Figure 7 shows that students initially had problems with understanding computational primitives, such as conditionals, loops, nesting, and variables, but these programming challenges decreased from activities 1 to 3. Activity 4 introduced a new type of programming challenge related to checking and debugging one’s model using the results from an expert simulation. Also, challenges with understanding primitives increased due to the number of new primitives (domain-based and computational) introduced in activity 4. Another big challenge in activity 4 was constructing nested conditionals to model rollercoaster behavior on different segments of the track. In activity 5, there were new types of programming challenges related to modularity and procedurality since the fish tank macro-model required students to specify component behaviors as separate procedures that were invoked from one main “Go” procedure. However, challenges with understanding conditionals, loops, nesting, and variables also increased, though they were not new to this activity. The reason for the resurfacing of old challenges may be explained by the increase in the complexity of the domain content in this activity (see Fig. 6), making it harder for the students to translate the domain content into computational structures. Overall, for both kinematics and ecology units, the programming challenges decreased over time across activities in the unit unless an activity introduced addition complexities.

Similarly, modeling challenges (see Fig. 8) increase in number in activity 4 for kinematics and activity 5 for ecology. Initial difficulties were related to systematicity, specifying component behaviors, identifying entities and interactions, and model validation. In activity 4, modeling a real-world system introduced new challenges related to choosing correct initial conditions. The students also had the additional task of verifying the correctness of their models by comparing against expert simulation behaviors. For activity 5, although the average number of challenges increased, there were no new types of modeling challenges. Existing modeling challenges resurfaced in light of the sharp increase in domain knowledge-related challenges. However, when the students switched to the fish tank micro-unit (activity 6), they had overcome most of these challenges.

For the agent-based thinking challenges (see Fig. 9), challenges went down with time in both the kinematics and ecology units. Since the kinematics models had single agents, the challenges related to agent-aggregate relationships did not occur in activities 1–4. Unlike the other three categories of challenges, the number of challenges did not increase in activity 4. This is possibly because activity 4 did not introduce any new agent-based-thinking-related challenges. However, the agent-based thinking challenges resurfaced in activity 5 when the students were required to model multiple new agents, and modeling multiple agents caused the number of challenges to increase sharply. Like other types of challenges, the students were also able to overcome most of these challenges by activities 6 and 7.