A Quantitative Research Manuscript · PLS-SEM

The Offloading Effect

AI Dependency, Cognitive Offloading, and Academic Performance in Engineering Education

Akshat Panchasara · MDes, Intelligent User Experience Design · Dhirubhai Ambani University · July 2026

0Respondents

0Hypotheses

0Analysis steps

0Robustness checks

0R², performance

0Interactive figures

Structured Abstract

Background. As engineering students integrate AI assistants into daily coursework, a common fear is that dependency erodes the metacognitive habits that make learning durable. This study tests that fear directly.

Method. A cross-sectional survey of 302 engineering students and scholars measured AI Usage Intensity, AI Dependency, Cognitive Offloading, Problem-Solving Confidence, Metacognitive Awareness, and self-reported Academic Performance on validated 5-point scales. An eight-hypothesis serial-parallel mediation model was estimated with PLS-SEM (Lohmöller algorithm, Mode A), with 5,000-resample bootstrapping, MICOM, multi-group analysis, Gaussian-copula endogeneity checks, curvilinear tests, and Necessary Condition Analysis. Two of the eight hypotheses (H4, H8) were data-informed revisions made after the original a priori pair failed; this is disclosed in Section 3 and should be weighted accordingly.

Results. Seven of eight hypotheses were supported. AI Dependency drove both Cognitive Offloading (β=0.46) and Problem-Solving Confidence (β=0.53), and Confidence in turn fed Offloading directly (β=0.26), a short causal chain rather than a flat parallel structure. Offloading fed Metacognitive Awareness (β=0.33), giving Awareness a genuine predictor for the first time, yet Awareness still did not predict performance (β=0.05, n.s.), the one hypothesis that failed in both the original and the revised design. Confidence (β=0.40) and Offloading (β=0.35) remained the two strongest routes to performance, together carrying roughly three-quarters of Dependency's total effect, alongside a smaller, endogeneity-flagged direct effect. The confidence pathway showed a significant curvilinear bend, and Necessary Condition Analysis found AI Dependency helpful at the margin but never necessary.

Conclusion. In this sample, AI reliance operated through a confidence-to-offloading chain, not through metacognitive awareness, which remained inert as a performance lever even once properly supplied with its own predictor. A self-reported, AI-attributed performance measure means the direct effect should be read with stated caution.

1 · Introduction

The fear, and the question

Every generation of tools invites the same worry: that convenience will hollow out competence.

Engineering students now write code, debug, and research with an AI assistant a keystroke away. The intuitive fear is specific: that growing AI Dependency quietly erodes Metacognitive Awareness, the habit of checking one's own understanding, and that eroded awareness eventually shows up as weaker academic performance. This study was designed to test exactly that chain, and it fixed its hypotheses before a single response was inspected.

The data told a partial version of that story, and a more interesting one. Two mechanisms, not the feared one, did nearly all the work: Problem-Solving Confidence (in plain terms: believing you can tackle unfamiliar problems, with AI as a safety net) and Cognitive Offloading (in plain terms: letting the tool carry mental effort you would otherwise spend), linked in a short chain where confidence itself feeds offloading. Metacognitive Awareness, given every chance across two versions of this model, including a direct predictor built specifically to test it fairly, never once moved the performance measure. The manuscript that follows walks through how that structure was established, stress-tested, and what it means for how AI tools in education should be designed.

In Plain Terms We expected: more AI reliance → less self-reflection → worse grades. We found: more AI reliance → more offloading and more confidence → better self-reported performance, while self-reflection stayed intact and, if anything, rose slightly. The interesting part is the fine print, and this paper is mostly fine print.

2 · Theoretical Framework & Hypotheses

A serial-parallel mediation model

One exogenous driver, a chain running through two mediators into a third, plus direct routes to the outcome, eight hypotheses with a verdict each.

AI Dependency (in plain terms: feeling uncomfortable working without AI) is the exogenous driver. Problem-Solving Confidence and Cognitive Offloading each receive from Dependency directly, and Confidence also feeds Offloading (H4), forming a short causal chain. Offloading, in turn, feeds Metacognitive Awareness (H8), which had no viable predictor in the original design once its direct link from Dependency failed to hold up. All four constructs carry a path to Academic Performance. Verdicts below reflect the full analysis; each is unpacked in Sections 5–7.

Table 1 *Hypotheses and verdicts*
ID	Hypothesis	Verdict
H1	AI Dependency positively affects Cognitive Offloading	Supported
H2	AI Dependency positively affects Problem-Solving Confidence	Supported · curvilinear
H3	AI Dependency directly affects Academic Performance	Supported · endogeneity flag
H4	Problem-Solving Confidence positively influences Cognitive Offloading	Supported
H5	Problem-Solving Confidence positively affects Academic Performance	Supported
H6	Cognitive Offloading positively affects Academic Performance	Supported
H7	Metacognitive Awareness positively affects Academic Performance	Not supported
H8	Cognitive Offloading positively influences Metacognitive Awareness	Supported

In Plain Terms Seven of eight hypotheses held. The one that still didn't, H7, is a familiar face: it's the same Metacognitive Awareness → Performance link that failed in the original design too, now re-tested with a better-supplied predecessor (H8 gives it a real predictor for the first time) and it still comes up short. That's not a coding artifact, it's the same finding twice: whatever awareness students report, it doesn't move the performance measure used here.

3 · Method

Participants, measures, and analytical strategy

A confirmatory design, with one honest exception disclosed below, and the estimation machinery stated openly.

Participants and procedure

302 engineering students and scholars (after listwise deletion from 308 responses; 6 rows removed) completed a structured questionnaire: 175 undergraduates, 94 postgraduates, and 33 PhD scholars. The original eight hypotheses were registered before data inspection. A non-response bias check comparing early and late respondents found two constructs (Problem-Solving Confidence, Academic Performance) significantly higher among late respondents, a generalizability caveat carried into Section 12.

Measures

Six constructs on 5-point Likert scales: AI Usage Intensity (4 items, one reverse-coded), AI Dependency (3), Cognitive Offloading (4), Problem-Solving Confidence (4), Metacognitive Awareness (3), and Academic Performance (3, self-reported and AI-attributed, a stated limitation). A tertile split on Usage Intensity produced Low (n=103), Moderate (n=159), and Heavy (n=40) user groups for multi-group analysis.

Model revision, disclosed

The original design (Section 2's a priori set) hypothesized AI Dependency → Metacognitive Awareness and Metacognitive Awareness → Academic Performance directly; neither held. Both were then replaced with two new paths, Problem-Solving Confidence → Cognitive Offloading and Cognitive Offloading → Metacognitive Awareness, chosen after seeing that the originals failed. This is an honest exploratory-confirmatory hybrid, not a fully a priori design: H4 and H8 (Table 1) were data-informed revisions, re-estimated on the same 302 respondents rather than a fresh holdout sample. The other six hypotheses remain a priori throughout. Readers should weight H4 and H8 accordingly, as strong post-hoc findings warranting replication, not as pre-registered confirmations.

Analytical strategy

PLS-SEM was estimated with an in-house implementation of the Lohmöller iterative algorithm (Mode A reflective measurement, path weighting, convergence at Δ<10⁻⁷), built and validated directly because the available Python PLS package is incompatible with current dependencies, a deliberate methodological disclosure rather than a silent substitution. Reporting follows Hair et al. (2019) and Richter et al. (2016). Significance testing used 5,000 bootstrap resamples; robustness checks comprised Harman's single-factor and unmeasured-latent-method CMB tests, Gaussian-copula endogeneity analysis (Park & Gupta, 2012), a quadratic specification test, MICOM (Henseler et al., 2016), PLSpredict, and Necessary Condition Analysis (Dul, 2016).

4 · Measurement Model

Can the constructs be trusted?

Before any structural claim, the instruments themselves go on trial.

Reliability and convergent validity clear conventional thresholds for five of six constructs (α and CR ≥ 0.70, AVE ≥ 0.50). The exception is AI Usage Intensity (α = 0.43), whose weakness is examined honestly below rather than patched over. CFA fit is acceptable: CFI = 0.916, TLI = 0.899, RMSEA = 0.063, SRMR = 0.063. KMO = 0.925 and Bartlett's test (p < .001) confirm factorability; Harman's single factor explains 37.0% of variance, below the 50% concern threshold.

Reliability & convergent validity by construct

Cronbach's α · Composite Reliability · AVE, against 0.70 and 0.50 thresholds · hover bars

Figure 1. Five constructs clear every bar. AI Usage Intensity fails α badly, driven by its reverse-worded item, and is used only as a grouping variable, never as a structural predictor.

Construct correlations & discriminant validity

Left: latent-variable correlations · Right: HTMT ratios (0.90 threshold) · hover cells

Figure 2. One pair breaches HTMT: AI Usage Intensity ↔ Cognitive Offloading at 0.987. Empirically, how often students reach for AI and how much effort they offload to it are nearly the same signal, a finding in its own right, handled by keeping Usage Intensity out of the structural model.

Exploratory factor structure

Oblimin-rotated ML loadings (21 items × 6 factors) with scree · hover cells

Figure 3. The six-factor structure recovers the designed constructs. PSC3 ("I can explain the logic behind AI-generated solutions") cross-loads toward Metacognitive Awareness, flagged and retained as a stated judgment call, not silently dropped.

Measurement verdict The five structural constructs are measurement-sound. AI Usage Intensity is not, and its near-identity with Cognitive Offloading (HTMT 0.987) is itself informative: usage frequency and offloading may be one behavior wearing two names. It serves as the grouping variable only.

5 · Structural Model

The chain, and the one holdout

Eight paths estimated at once, with Role controlled, seven of them landing exactly where hypothesized.

The model explains R² = 0.603 of Academic Performance (Q² = 0.591; PLSpredict RMSE 0.65 vs. naïve 0.99, confirming out-of-sample relevance). The structure is now a short chain rather than a flat parallel design: Dependency feeds both Confidence and Offloading directly, Confidence also feeds Offloading (H4, β = 0.26), and Offloading feeds Awareness (H8, β = 0.33), Awareness' first significant predictor across either version of this model. Confidence (β = 0.40) and Offloading (β = 0.35) remain the two strongest routes to performance. The one path that still fails is H7, Awareness → Performance (β = 0.05, n.s.), even now that Awareness finally has a real predictor feeding it.

Figure 4 · The structural model, interactive

Node size = R² · edge width = |β| · garnet = significant · parchment-gray = n.s. · gold = significant with caveat · hover nodes and edges

Figure 4. Toggling "Significant only" removes exactly one element: Metacognitive Awareness→Performance, and the Role control. Seven of eight hypothesized arrows survive; watch the awareness pathway fade alone.

Note. Bootstrap 5,000 resamples; Role (UG/PG/PhD) entered as control on Academic Performance. * p<.05, ** p<.01, *** p<.001.

In Plain Terms Seven arrows out of eight pointed where we expected, and the sizes tell their own story: Confidence and Offloading are the heavy cables (β around 0.35–0.40 into Performance), the Dependency→Performance shortcut is a thinner wire worth distrusting (see Section 7), and the lone gray arrow, Awareness→Performance, is the one hypothesis that just didn't carry any weight, even after we gave it a real predictor to work with in H8. Two of these eight hypotheses (H4, H8) were added after the original design's H1 and H4 came up empty, so treat those two as strong leads worth re-testing on new data, not settled facts the way the other six are.

6 · Mediation Analysis

Where the effect actually travels

Five possible routes from Dependency to Performance. Three carry weight; two are negligible.

With Offloading now receiving from both Dependency and Confidence, and Awareness receiving only from Offloading, the total effect of AI Dependency on Academic Performance (0.580) decomposes into a direct effect (0.153, endogeneity-flagged) and five specific indirect routes. Three clear zero comfortably: straight through Offloading (0.159), straight through Confidence (0.210), and the two-step Confidence→Offloading chain (0.049). The two routes that pass through Metacognitive Awareness, whether reached via Offloading alone or via the full four-step chain, stay negligible, consistent with H7's null result. Altogether, 73.7% of Dependency's total effect on performance is mediated (total indirect 0.427 of 0.580).

Mediation decomposition

Sankey: total effect splitting across five specific indirect routes and one direct path · band width = effect size · hover bands

Figure 5. The two Awareness-involving bands are visibly threads beside three cables. The direct band (gold) survives but carries the endogeneity caveat from Section 7.

Bootstrap distributions of the three significant indirect routes

5,000 resamples · dashed line at zero · hover bars

Figure 6. All three distributions sit clear of zero. The Confidence→Offloading chain (H4's contribution) is the newest and smallest of the three, but real.

7 · Robustness Checks

Wait, not so fast

The clean result from Section 5, deliberately re-interrogated four ways.

Endogeneity (Gaussian copula)

The copula term for AI Dependency's direct path (H3) is significant (p = .035): an omitted variable, plausibly general ability or motivation, contaminates that coefficient, which shifts substantially under correction. The Problem-Solving Confidence path shows no such flag. Practical reading: trust the mediated pathways more than the direct one.

Curvilinearity

The dependency→confidence relationship (H2) bends (quadratic term p = .042). Confidence rises with dependency, then flattens at the top of the range: the boost is real but not limitless, and that plateau now also caps how much H4 (Confidence→Offloading) and H5 (Confidence→Performance) can receive downstream at high dependency.

The confidence curve

302 respondents · fitted quadratic · hover points for individual values

Figure 7. The fitted curve's flattening tail is the statistical shape of diminishing returns: past moderate dependency, additional reliance buys little additional confidence.

Necessity (NCA)

Is AI Dependency required for good performance? No. The ceiling-envelopment effect is trivial (d = 0.064): high performers exist at every dependency level. AI reliance helps at the margin; it is not a prerequisite.

Necessary Condition Analysis

CE-FDH ceiling envelope over the dependency–performance space · hover points

Figure 8. A necessity relationship would show a large empty region below-left of the ceiling. The space is nearly full: nothing about low dependency forbids high performance.

Measurement invariance (MICOM)

Five of six constructs achieve compositional invariance (c > 0.95) across every group pair; AI Usage Intensity fails only tautologically, since it defines the groups. The Section 8 comparisons therefore rest on invariant measures.

In Plain Terms Four stress tests, four honest outcomes: one result (the direct effect) is downgraded to "handle with care," one (the confidence path) gains a nuance (it plateaus), and two checks (necessity, invariance) pass cleanly. This is what the results look like after trying to break them.

8 · Multi-Group Analysis

Not the same story for everyone

Light and heavy AI users run on different engines, and the difference now starts one step earlier, in what feeds Offloading.

Comparing Low (n=103) and Heavy (n=40) usage groups on the Academic Performance model, two paths differ significantly: Confidence collapses from β=0.62 (Low) to β=0.00 (Heavy, p<.001), while Awareness turns from mildly negative to mildly positive (p=.034). The new H4 link (Confidence→Offloading) tells a matching story one step upstream: it is strong for light users (β=0.42) and effectively zero for heavy users (β=−0.06, p<.001), heavy users' offloading no longer draws on their confidence at all, it runs on Dependency alone. Role-based groups (UG vs. PG/PhD) show no significant differences, the split is about usage, not seniority.

Path coefficients, Low vs. Heavy AI users

Slope chart · Academic Performance predictors (left) and Cognitive Offloading predictors (right) · gold = significant group difference · hover lines

Figure 9. The Confidence line crosses in both panels. For heavy users, confidence stops mattering twice over, neither for performance directly, nor for feeding offloading upstream.

9 · Importance–Performance Map

Where intervention buys the most

Total effects against current standing, the prioritization view.

AI Dependency holds the highest total effect on Academic Performance (0.58, direct plus all five indirect routes) while sitting lowest on current performance (64/100), placing it squarely in the high-importance, low-performance quadrant that IPMA logic marks as the priority target. Problem-Solving Confidence follows closely (0.49 importance), reflecting its role as both a direct driver and the source of the H4 chain into Offloading. Metacognitive Awareness is the mirror image of Dependency: reasonably well-developed (71/100), yet nearly inert as a lever (0.05), unchanged by adding H8 upstream of it.

Importance–Performance Map

X = total effect on Academic Performance · Y = construct performance (0–100) · dashed lines at means · hover points

Figure 10. The bottom-right point (AI Dependency) is where design attention pays; the top-left point (Metacognitive Awareness) is already strong but moves nothing.

10 · Discussion

What survived the gauntlet

Three findings stand; each with its precise weight.

First, the operative mechanisms are Confidence and Offloading, and they interact. Confidence doesn't just act alone, it feeds Offloading directly (H4), meaning part of Dependency's benefit to Offloading is routed through how capable a student already feels. Together the two constructs carry roughly three-quarters of Dependency's total effect on performance.

Second, Metacognitive Awareness resists every version of this model. It failed as a direct outcome of Dependency in the original design, and now, even freshly supplied with a genuine predictor (Offloading, via H8, itself a real and sizable effect, β=0.33), its own link to performance (H7) still lands at β=0.05, n.s. Students who offload more do report noticing and checking their understanding more, that part moved. It simply doesn't convert into the performance measure used here, twice over.

Third, the direct effect remains the least trustworthy number in the model. Endogeneity correction shifts it substantially; the outcome measure is self-reported and AI-attributed. The defensible claim is mechanism-level (dependency → confidence → offloading → perceived performance), not a causal grade-improvement claim.

The one-sentence version AI reliance in this sample works through a short chain, what students believe they can do, feeding what they hand over to the tool, not through any decay, or any productive use, of how they think about their thinking.

11 · Design Directions

From coefficients to design decisions

Eight directions, each traceable to a specific statistical source, for AI-tool and curriculum designers.

DD-01 · H2 + curvilinear test

Design for the confidence plateau

Source: β=0.53 Dependency→Confidence; quadratic p=.042, flattening at high dependency

Confidence gains from AI taper. Past moderate reliance, more AI does not build more self-belief, and since Confidence now feeds Offloading directly (H4), that plateau caps a second pathway too.

Introduce graduated challenge: reduce AI scaffolding as competence signals rise
Surface "you solved this without help" moments explicitly
Treat maximum-assistance mode as a floor state, not the default

DD-02 · H1/H4/H6 chain

Make offloading a decision, not a drift

Source: Dependency→Offloading β=0.46 (H1); Confidence→Offloading β=0.26 (H4); Offloading→Performance β=0.35 (H6)

Offloading is fed from two directions and drives outcomes on its own, design it deliberately instead of letting it accumulate invisibly.

Show a per-task "what AI carried vs. what you carried" summary
Prompt an explain-back step before AI output can be submitted onward
Let students set offloading budgets per assignment type

DD-03 · H7 null, twice over

Stop assuming reflection transfers

Source: H7 n.s. (β=0.05) even after H8 gave Awareness a real predictor (β=0.33)

Students do reflect more as they offload more, but that reflection still moved nothing on performance, in the original design and in this revised one. Verification must be built into the workflow, not hoped for.

Embed test-before-accept gates in AI coding tools
Grade the verification artifact, not just the solution
Do not market "critical thinking prompts" as a performance fix; the data says that lever is weak, twice tested

DD-04 · Endogeneity check

Never read usage as ability

Source: copula p=.035 on H3 (Dependency→Performance direct path)

The dependency→performance link is confounded. Institutional dashboards must not treat AI usage as a performance signal in either direction.

Pair any usage analytics with objective assessment before policy decisions
Ban "AI-use score" as an input to grading or ranking
Report mechanism metrics (offloading, confidence) instead of raw usage

DD-05 · MGA

Two user modes, two designs

Source: confidence path β=0.62 (Low) vs 0.00 (Heavy), p<.001, on both the Offloading and Performance models

Light users need confidence support; heavy users need offloading governance, confidence stops mattering for them at both stages of the chain. One interface serving both serves neither.

Light-user mode: encouragement, worked examples, visible progress
Heavy-user mode: verification friction, offloading summaries, retention checks
Detect mode from behavior, not self-declaration

DD-06 · NCA

Keep the no-AI path first-class

Source: necessity effect d=0.064 (trivial); AI is not required for high performance

High performers exist at every dependency level. Mandating AI tooling has no empirical basis in this data.

Assessment design must remain completable without AI assistance
Offer AI-free tracks without grade penalty
Frame AI as an accelerant, never a prerequisite, in course policy language

DD-07 · HTMT 0.987

Measure dependency, not clicks

Source: Usage Intensity ↔ Offloading HTMT=0.987; Usage α=0.43

Usage frequency and offloading are empirically one signal. Analytics and future instruments should measure psychological reliance, not raw activity.

Survey instruments: drop frequency items, keep dependency/offloading scales
Product analytics: log what was delegated, not how often the tool opened
Treat "time in tool" as a vanity metric in education contexts

DD-08 · IPMA

Aim interventions at the dependency quadrant

Source: AI Dependency total effect 0.58, performance 64/100, the priority quadrant

The highest-leverage, lowest-performing construct is dependency itself: shaping how students depend is the intervention with the most headroom.

Curriculum: teach delegation strategy as an explicit skill
Tools: make dependency visible to the student, trend lines, not judgments
Leave Metacognitive Awareness programs as culture, not as a performance lever

12 · Limitations

What this study cannot claim

Self-reported, AI-attributed outcome. Academic Performance items name AI in their wording ("improved due to AI"), structurally inviting inflated association with AI Dependency; the mediated pathways are the more behaviorally grounded results.
Cross-sectional design. Directional hypothesis language reflects theorized order, not demonstrated causality; the endogeneity finding underlines this concretely.
Single institution. Generalization beyond this population is unestablished; the non-response check (two constructs higher among late respondents) adds a further sampling caveat.
One weak instrument. AI Usage Intensity (α=0.43) failed reliability and was demoted to grouping duty; a higher-order AI Reliance construct was tested and did not rescue it.
Unstable strong CMB test. The unmeasured-latent-method model showed Heywood-like loadings for two blocks; Harman's test and HTMT/AVE remain the stable common-method evidence.
Estimation software. PLS-SEM ran on a purpose-built, validated implementation of the Lohmöller algorithm rather than a maintained commercial package, disclosed openly; replication in SmartPLS or seminr is a natural next step.

13 · Conclusion

Dependent, not diminished

This study set out to catch AI dependency eroding the reflective habits of 302 engineering students. The revised design gave that hypothesis every advantage, a real upstream predictor, a full re-estimation, and it still failed to catch it, twice, in the informative way. Confidence and Offloading, linked in a short causal chain, emerged as the genuine machinery connecting reliance to perceived performance, together carrying roughly three-quarters of the effect, surviving bootstrap, invariance, and necessity testing, and arriving with honest fine print: the confidence gains plateau, that plateau limits a second downstream pathway now too, heavy users run on offloading alone, and the direct effect wears an endogeneity flag it cannot remove. For designers of educational AI, the mandate is specific: govern the offloading, scaffold the confidence, keep the no-AI path open, and stop waiting for reflection alone to move the numbers.

† · References

Bibliography

Dul, J. (2016). Necessary Condition Analysis (NCA): Logic and methodology of "necessary but not sufficient" causality. Organizational Research Methods, 19(1), 10–52.
Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18(1), 39–50.
Hair, J. F., Risher, J. J., Sarstedt, M., & Ringle, C. M. (2019). When to use and how to report the results of PLS-SEM. European Business Review, 31(1), 2–24.
Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. Journal of the Academy of Marketing Science, 43(1), 115–135.
Henseler, J., Ringle, C. M., & Sarstedt, M. (2016). Testing measurement invariance of composites using partial least squares. International Marketing Review, 33(3), 405–431.
Lohmöller, J.-B. (1989). Latent variable path modeling with partial least squares. Physica-Verlag.
Paas, F. G. W. C. (1992). Training strategies for attaining transfer of problem-solving skill in statistics: A cognitive-load approach. Journal of Educational Psychology, 84(4), 429–434.
Park, S., & Gupta, S. (2012). Handling endogenous regressors by joint estimation using copulas. Marketing Science, 31(4), 567–586.
Podsakoff, P. M., MacKenzie, S. B., Lee, J.-Y., & Podsakoff, N. P. (2003). Common method biases in behavioral research: A critical review of the literature and recommended remedies. Journal of Applied Psychology, 88(5), 879–903.
Richter, N. F., Cepeda, G., Roldán, J. L., & Ringle, C. M. (2016). European management research using partial least squares structural equation modeling (PLS-SEM). European Management Journal, 34(6), 589–597.
Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.
Shmueli, G., Sarstedt, M., Hair, J. F., Cheah, J.-H., Ting, H., Vaithilingam, S., & Ringle, C. M. (2019). Predictive model assessment in PLS-SEM: Guidelines for using PLSpredict. European Journal of Marketing, 53(11), 2322–2347.