What the K–12 Evidence Actually Shows
This chapter reads the K–12 evidence about E-homework as a body of studies, not as a verdict attached to a platform feature. The goal is to understand what the pooled result supports, how much the studies differ, and why an independent meta-analysis can find no general online-format advantage without disproving the more specific K–12 pattern. That distinction gives the reader a way to judge both the strength and the reach of the claim.
The question is not only whether the average points upward. It is also what was averaged, how the studies were combined, and whether the same result is plausible in the next classroom. Those details are not technical decoration. They determine whether “E-homework helps” means a broad expectation, a context-bound finding, or an inference that the evidence cannot carry.
A pooled result starts with a defined evidence set
The focal study is a systematic review and meta-analysis of E-homework for K–12 students. Its search covered research published between 2000 and 2025 and used predefined inclusion criteria. Eligible studies had to examine K–12 students, compare an E-homework intervention with a relevant alternative, report an academic outcome, and provide enough statistical information to calculate an effect size. The review also required that outcomes could be classified as assignment performance or exam performance and that at least one moderator, such as school level, subject, feedback, or duration, could be coded.
After screening and citation tracking, 26 empirical studies remained. They included studies published in Chinese and English and represented journal articles, dissertations, and conference papers. The review extracted one effect size per study. That last choice matters because a single study may report several classes, subjects, tests, or time points. Counting all of those estimates as if they came from unrelated studies could give a study with many measurements too much influence. The review instead used a predefined hierarchy to select one result that best represented each study’s intervention and outcome.
The resulting number is a standardized mean difference, reported as Hedges’ g. In accessible terms, it expresses the average separation between the E-homework and comparison groups in standard-deviation units, with a correction intended to reduce small-sample bias. It is not a number of extra marks that every student should expect. A value can summarize studies that used different tests and score scales, but that comparability comes at a cost: the reader must keep the underlying populations, tasks, and implementation conditions in view.
The review used a random-effects model. A fixed-effect model would treat all studies as estimating one common underlying effect and would attribute differences between their estimates largely to sampling error. A random-effects model allows the underlying effect to vary across settings while estimating an average distribution of effects. That choice fits the subject. An online mathematics assignment in a primary classroom is not the same educational event as an interactive language task in a high school, even if both are placed under the label E-homework.
The composition of the evidence set reinforces the need for that caution. The 26 studies were not 26 identical classroom trials: the review included research published in Chinese and English, along with journal articles, dissertations, and conference papers. The largest share came from English-language studies, and the publication categories were unevenly sized. That mix broadens the evidence beyond one publication channel, but it also means that the average combines settings with different reporting conventions, samples, and levels of detail. A result supported by a broad collection is not automatically a result that applies equally across every subgroup represented by only a few studies.
The one-effect-per-study rule is another guard against a misleading denominator. Suppose one investigation reports an exam result, a homework result, and several follow-up measurements. Treating every number as a separate independent study would make that investigation count repeatedly and could make the combined estimate appear more certain than it is. Selecting one representative estimate does not solve every dependence problem, but it makes the unit of synthesis clearer: the review is combining one selected contribution from each of 26 studies, not claiming to have 26 independent tests of every possible outcome.
This structure also clarifies what the comparison is. The result is not a before-and-after score for one group of students. It is an average difference between an E-homework condition and comparison conditions across studies that met the review’s criteria. A positive difference can reflect the way the intervention was implemented relative to its comparator, including the feedback and practice opportunities available to students. It should not be rewritten as a fixed amount of learning produced by a device. The design tells us what was compared; the variation tells us how cautiously the comparison should travel.
The focal random-effects estimate is Hedges’ g = 0.309, with a 95% confidence interval from 0.199 to 0.419. The interval remains above zero, and the reported statistical test is significant. Within this evidence set, students who received E-homework performed better on the included academic outcomes on average than students in the comparison conditions.
That is a meaningful result, but “on average” does important work in the sentence. The estimate does not say that every implementation improved scores. It does not identify one platform as responsible. It does not tell a teacher how many points a particular class will gain. It also does not prove that the digital features, rather than differences in teaching, task selection, student participation, or other contextual conditions, caused the entire separation between groups.
A meta-analysis can strengthen an inference by combining evidence, but it cannot turn unlike studies into identical ones. The strength of the result comes partly from the number and design of the studies; its limit comes from the variation that remains after they are combined. That is why the next question is not whether the pooled estimate is positive, but how much it moves from one setting to another.
Why the average is conditional
The focal review reports substantial overall heterogeneity: the study estimates vary more than would be expected from sampling error alone. Its I² value is 89.34%, a high level of inconsistency by conventional interpretation. I² does not tell us which classroom feature caused the variation, and it should not be read as a percentage of classrooms that will fail. It indicates that the studies are not all estimating the same practical effect.
This changes how the headline should be used. If heterogeneity were low, a teacher might treat the pooled estimate as a reasonably stable expectation across similar settings. With high heterogeneity, the average is better understood as a center of a spread. The relevant decision becomes conditional: How similar is the intended use to the interventions in the evidence? What kind of outcome will be measured? Does the task reward the same behavior that the platform supports? Which learner supports and teacher practices are present?
The review’s design reflects this concern through moderator analyses. It examines publication type, intervention duration, feedback type, school level, and subject domain. Some of those moderators show a statistically detectable pattern; others do not. But even a statistically significant moderator does not automatically reveal a mechanism. It may mark a cluster of related differences—student age, curriculum, task structure, teacher involvement, or assessment form—that travel together in the available studies.
Publication-bias checks add another layer of calibration. The focal review used funnel plots and Egger’s regression test and did not find a statistically significant signal of publication bias. That lowers concern that the pooled result is explained entirely by missing negative studies. It does not prove that no unpublished or selectively reported findings exist. A test with a modest number of studies has limited ability to detect every form of selection, and a non-significant result is not evidence that bias is absent.
The independent 2024 synthesis makes the conditional nature of the claim impossible to ignore. Bishop, Best, and Horsley examined 18 experimental and quasi-experimental studies comparing online homework with traditional pen-and-paper homework. Eleven studies favored online homework and seven favored paper homework, but their random-effects estimate was 0.06, with a 95% confidence interval from −0.19 to 0.32. Because that interval includes zero, the authors did not find a statistically significant overall achievement difference between the two formats in the published studies they examined.
At first glance, that looks like a contradiction. One synthesis reports a positive average and another reports no reliable overall format advantage. But the two questions are not identical. The focal review isolates K–12 students and examines E-homework as a defined digital intervention, including outcome and moderator distinctions. The 2024 synthesis uses a different study set and a broader context that is largely situated in higher education or general online-homework comparisons. It does not establish a K–12-specific null result.
The broader synthesis nevertheless matters. It tests the tempting explanation that putting homework online is itself the active ingredient. If a delivery label were enough, a general format advantage should appear consistently across populations and settings. Instead, the independent result is close to zero and highly variable. That pattern supports a more careful interpretation of the K–12 finding: the benefit may be associated with the way E-homework organizes feedback, practice, and participation under particular conditions, not with the digital medium in isolation.
The two syntheses also differ in how much their studies disagree. The 2024 paper reports very high residual heterogeneity, with I² above 97%. Its random-effects estimate therefore sits inside a wide range of study experiences. That does not make the review useless; it tells us that a single average is a poor substitute for asking which conditions were present in each comparison. The same principle applies to the K–12 review, especially when its exam outcomes vary widely.
There is also a difference between evidence of no average advantage and evidence that the medium has no value. The 2024 analysis did not find a statistically reliable overall difference in its sample. That is not the same as proving that online homework never helps. Some included studies favored online work, some favored paper, and the variation may reflect task, learner, or implementation differences. Absence of a general format effect is not evidence that every digital feedback loop is educationally inert.
Nor should the positive K–12 estimate be treated as proof of an inherent digital advantage. The focal result can be true because the included K–12 interventions, on average, created useful opportunities for practice and correction. It can also coexist with settings in which paper work, teacher explanation, or a mixed approach better matches the learning goal. The independent comparison reminds us to ask what was actually changed beyond the medium.
A useful reading rule follows. Carry five pieces of information with every claim about E-homework: the population, the comparison, the outcome, the variation across studies, and the conditions that might explain it. “K–12 students performed better on average” is a defensible statement about the focal synthesis. “Digital homework is better than paper” drops the population, outcome, and heterogeneity and therefore says more than the evidence supports.
This rule also clarifies what new evidence would change the conclusion. A large, well-controlled study in a new school level could test whether the K–12 pattern travels. A study that measures the quality and use of feedback could distinguish a useful explanation from a bare correctness signal. A longitudinal design could show whether assignment gains persist when students later work without the platform. A direct comparison of matched tasks could reveal whether the apparent difference comes from the assessment rather than the homework format.
The evidence currently supports a positive but conditional K–12 judgment. E-homework is associated with better academic outcomes on average in the focal review, yet the high heterogeneity means the average is not a universal forecast. An independent broader synthesis finds no significant overall online-versus-paper advantage, which weakens any claim that digital delivery alone explains the benefit. To understand the most important part of the title, we now need to separate the kinds of performance being measured: the assignments that sit inside the feedback loop and the exams that ask what remains when the loop is gone.