The research responds to a persistent gap in the literature on computational thinking (CT) integration in science education. Although a common rationale for weaving CT into science curricula has been the promise of boosting learning in both domains simultaneously, few empirical studies have actually demonstrated equivalent gains in science learning. Across two systematic reviews of CT integration literature published between 2020 and 2022, only 19 empirical studies assessed both science and CT learning, and just 10 were impact studies that compared students’ knowledge and skills before and after an intervention. Several of those studies found statistically significant gains in both domains, but one found no significant gains, and there has been little consensus on how or why CT integration might deepen science understanding rather than simply compound domain-related challenges.

The theoretical foundation of the Decoding Approach is mechanistic reasoning, a construct that science educators have long sought to cultivate. Following frameworks developed by Russ and colleagues and later consolidated by Schwarz and colleagues, mechanistic reasoning spans three levels: Level 1 describes only the surface characteristics of what happened; Level 2 addresses how a process unfolded over time or space through co-occurrence and sequencing; and Level 3 addresses why the phenomenon occurred, attending to entities, their actions, and cause-and-effect chains. Because modern science standards such as the Next Generation Science Standards emphasize understanding how and why phenomena occur — not merely retaining facts — the researchers reasoned that decoding, which examines causality in both the real-world phenomenon and the modeled world, could push students toward Level 3 reasoning and enable them to transfer their understanding of causal mechanisms to new systems.

Decoding is rooted in established ideas from cognitive science. The real world and the modeled world serve as “contrasting cases” that focus learners’ attention on what to notice, and decoding adds the explicit identification of similarities between the two contexts, making it a form of analogical reasoning. The approach also draws on the professional practice of model validation, though the researchers distinguish it from expert-level validation: the middle school participants in decoding were just beginning to learn the phenomena being modeled, and the goal was not to produce expert-caliber predictive models but to help learners inch toward fuller understanding by bridging code, simulation, and scientific information. The approach sits between a “programming-first” model of CT integration, in which students create models of their own, and a “simulation-only” approach, in which students manipulate a model through an interface without ever seeing the code that generates the simulation.

The intervention took place over three cohort years from 2021 to 2023, initially planned as an in-person three-week summer program followed six months later by a two-day continuation workshop. Implementation varied considerably due to the COVID-19 pandemic: the first cohort totaled 40 contact hours conducted entirely virtually, the second was hybrid with 60 contact hours, and the third totaled 70 contact hours conducted fully in person in museum classrooms and exhibit halls. Participants were New York City youth between the ages of 11 and 13, recruited with strategies designed to reflect the demographics of the city’s school-age population, including partnerships with public schools and organizations serving under-resourced students from groups underrepresented in STEM and computing. Applicants were bucketed by gender, ethnicity or race, and school or program affiliation, then randomly drawn into treatment and comparison groups; students selected for the comparison group were offered spots in the following year’s full program.

The curriculum focused on the authentic scientific problem of rising Lyme disease rates in New York City, combining real-world investigation, data collection and analysis, and computational modeling with StarLogo Nova, a block-based modeling environment developed at MIT. Students began with instruction on ecosystem concepts — food webs, carrying capacity, ecological relationships — and explored systems through physical models and embodied simulations in which they acted out the roles of individuals in a system. The first focal model depicted population dynamics and energy flow among acorns, mice, and foxes. In early “Use” activities, students manipulated population sizes with slider variables; by day four, code was introduced and students began decoding, mapping mechanisms in the real world to those in the code. Later “Modify” activities required students to change code so foxes could eat both acorns and mice, and to compare mouse code across two different models and justify the differences. In the final week, students created new code, adding agents such as wild turkeys, dogs, and opossums to the tick-mouse-deer Lyme disease model, using decoding in both directions as they decided what behaviors to include and how to represent them.

The continuation workshop was designed partly as a research vehicle for studying transfer. After a three-hour refresher with increasingly difficult coding challenges, student teams took part in a five-hour “Hackathon” centered on a new ecological system — zebra mussels in New York waterways — modifying an existing model to represent the invasive species. Because the zebra mussel scenario was new to students, this activity allowed the research team to observe whether students transferred mechanisms such as predation, energy transfer, and disease transmission from earlier models to an unfamiliar context. Teams then presented their modified models, including adaptations of mechanisms from past models.

Quantitative results rested on the Survey of Knowledge and Skills in Computational Thinking (KSCT), a 20-item instrument combining an ecosystems scale, decoding “triplets,” and difficult computer science items, administered before the summer program, immediately after it, and again at the start of the continuation workshop. Baseline comparisons revealed that the treatment and comparison groups were not equivalent — the comparison group skewed toward higher-scoring individuals, which the researchers attributed to attrition driven by the pandemic. Nevertheless, treatment students (n = 46) improved significantly from a pre-test mean of 9.435 to a post-test mean of 12.174, with p < .01 and a large effect size of d = 0.96, while comparison students showed no significant change. Because the groups were non-equivalent, the researchers also conducted an analysis of covariance, which confirmed that group assignment had a significant effect on score change while pre-test scores did not.

Scores by subscale added nuance. On the ecosystems scale, treatment students gained significantly (d = 0.586), though they entered the program with relatively strong understanding — 80 percent could correctly predict the impact of an ecosystem change at pre-test, compared with a typical 54 percent found in prior research. The largest gains appeared on items about ecosystem equilibrium and identifying the logistic growth curve, which rose from 33 percent to 52 percent correct. Understanding of energy loss across food-chain levels changed little. Most striking were gains on the decoding scale, where students matched coded mechanisms to scientific processes: means rose from 1.913 to 3.391 with a large effect size of d = 1.157. A durability analysis of the 27 treatment students who took all three surveys found no significant difference between post and continuation scores, suggesting that learning persisted over the six-month interval, though the researchers note that self-selection into the continuation workshop may have influenced this result.

The qualitative strand of the study centered on artifact-based interviews, in which students thought aloud while investigating agent-based models, plus student work collected during the program. Two reviewers coded transcripts for mechanistic reasoning levels, comparisons between real-world and simulated contexts, and decoding behaviors, working through rounds of coding and discussion until reaching agreement. The researchers distinguished "basic" decoding — identifying a mapping between a code mechanism and a science process without further elaboration — from "advanced" decoding, in which students applied the mapping to make predictions or assess a model's validity. Interview transcripts show students spontaneously identifying that collision code represented infection, or that a "wiggle" procedure controlled movement, even when those words did not appear in the code itself.

Because only four focal students completed the full set of three interviews and three surveys across all cohort years — two cases in year one and two in year three — the qualitative analysis is deliberately limited in scope and was not treated as quantitative evidence. The researchers also acknowledge that the treatment and comparison groups were non-equivalent at baseline, that attrition shaped both quantitative and qualitative samples, and that the durability findings may reflect sampling effects. The implementation itself varied across cohorts in format and contact hours, spanning pandemic-era virtual instruction to fully in-person museum programming, which complicates direct comparisons between years.

Even with these limitations, the study's implications are noteworthy. The evidence from the four case studies shows students using coded mechanisms as an "active" and "executable" representation with which to reason about scientific phenomena, and in some cases transferring that reasoning to entirely new systems such as the lionfish overpopulation scenario posed in the final interview, where students were asked to imagine the code they would write. The authors argue that the results reveal untapped opportunities for deepening students' understanding of ecosystems through CT integration with an explicit decoding emphasis — and, more broadly, that reading, interpreting, and critiquing models built by others constitutes a valuable learning experience in its own right, one that does not require every student to become a model-builder for the synergy between computational thinking and science learning to take hold.