Development and Preliminary Psychometric Evaluation of the Exercise and Nutrition Consistency Scale (ExNCS): A Cross-Sectional Study of Perceived Exercise-Nutrition Integration in Young, Physically Active Turkish Adults

Highlights

Public health relevance—How does this work relate to a public health issue?

• Aligning and sustaining physical activity and diet as one lifestyle pattern is central to preventing noncommunicable disease, yet the existing instruments it identified assess exercise and nutrition separately, or as co-occurring components, rather than capturing their behavioral consistency as a unified construct.
• The ExNCS is a brief Turkish self-report measure of perceived exercise-nutrition integration intended for descriptive, group-level research; it is not a screening instrument.

Public health significance—Why is this work of significance to public health?

• In 583 predominantly young adults who exercised at least weekly, the 16-item total score showed high internal consistency (α = 0.913) and a strong but exploratory general dimension (bifactor-(S-1) ω_h = 0.84–0.88), but no factor model met conventional fit criteria under ordinal estimation, so the dimensional structure remains provisional.
• Total scores were higher among more frequent exercisers (η² = 0.153, 95% CI 0.102–0.204); because several items refer to exercise, this is preliminary known-groups evidence rather than independent validation.

Public health implications—What are the key implications or messages for practitioners, policy makers, and/or researchers in public health?

• Researchers may use the total score only as a preliminary, descriptive indicator in group-level studies; subscale scores, particularly CHR, should not be scored or interpreted independently.
• Independent replication, response-process, convergent and criterion evidence, test–retest reliability, and measurement invariance are required before between-group, clinical, or English-language use.

Abstract

Background: Instruments assessing physical activity and diet usually measure each behavior separately or sum them as independent lifestyle components. This cross-sectional study developed the Exercise and Nutrition Consistency Scale (ExNCS), a Turkish self-report measure of perceived exercise-nutrition integration, and reports a preliminary psychometric evaluation. Methods: Adults who exercised at least weekly (N = 583; 58.7% male; 93.7% aged 18–25) completed 20 expert-reviewed items. Using a seed-documented split, item screening and exploratory analyses used the development half (n = 291), and six measurement models were compared in the holdout half (n = 292) using robust maximum-likelihood and ordinal (WLSMV) estimation. Results: Four reverse-worded items were removed, leaving 16 items. Retention criteria disagreed (two versus four factors), and no model met conventional WLSMV fit criteria (best: correlated four-factor, CFI = 0.865, RMSEA = 0.122). The symmetrical bifactor model was inadmissible; a bifactor-(S-1) model, estimated despite poor fit, indicated a strong general dimension (ECV = 0.73; ω_h = 0.88) that is best read as an exploratory composite rather than a confirmed construct. Total-score α was 0.913, scores increased with exercise frequency (η² = 0.153, 95% CI 0.102–0.204), sex invariance was inconclusive, and acquiescence was detected. Conclusions: The ExNCS is a candidate item set, but its dimensional structure and score interpretation require independent replication.

1. Introduction

In recent years, integrating multiple health-promoting behaviors has emerged as a central concern in public health, health psychology, and behavioral science [1,2,3]. Among these behaviors, regular physical activity and healthy dietary practices play vital roles in preventing chronic disease, enhancing mental well-being, improving physical performance, and promoting quality of life [4,5]. Both behaviors have been independently associated with reduced cardiovascular risk, improved metabolic health, enhanced cognitive functioning, and better emotional regulation [1,6,7,8]. Despite widespread awareness of their benefits, many individuals struggle not only to initiate these behaviors but also to maintain them as part of a stable, intentional lifestyle routine.

Research on multiple health behavior change (MHBC) has increasingly shown that health behaviors co-occur and influence one another through shared psychological and behavioral mechanisms [1,9,10]. This phenomenon, behavioral clustering or behavior bundling, suggests that individuals who adopt one positive health behavior are more likely to adopt complementary behaviors such as healthier eating [11,12,13,14]. Self-regulation, intrinsic motivation, behavioral identity, and perceived goal congruence mediate these interactions, but existing tools do not adequately capture the interplay or consistency between physical activity and nutrition behaviors [8,15].

For example, the Exercise Motivation Inventory-2 (EMI-2) assesses motives behind physical activity [16], and related Self-Determination Theory (SDT) work has extended this measurement tradition to integrated regulation [17]; the EMI-2 has since been adapted and shortened in several languages [18,19]. The International Physical Activity Questionnaire (IPAQ) evaluates self-reported physical activity [20,21,22,23], and the Three-Factor Eating Questionnaire (TFEQ) examines cognitive restraint, uncontrolled eating, and emotional eating [24,25,26,27,28,29]. Broader multiple-health-behavior and health-promoting-lifestyle instruments do combine the two domains; for example, the Health-Promoting Lifestyle Profile (HPLP) [30] and the FANTASTIC Lifestyle Assessment [31] each contain distinct physical-activity and nutrition subscales, yet they index the level of each behavior as an independent component summed into an overall lifestyle score, rather than the consistency or mutual alignment between them. Across both approaches, existing instruments either assess the behaviors in isolation or aggregate them as separate components, and, among the instruments identified in our targeted search, none was designed to evaluate how individuals align and sustain physical activity and nutrition as a single, mutually reinforcing pattern. Because that search was not exhaustive, this gap is stated relative to the nearest-neighbor instruments reviewed rather than as a claim that no such instrument exists.

Exercise-nutrition consistency must also be distinguished from several neighboring constructs; health consciousness denotes a general readiness to attend to and act on one’s health [32,33]; it is a dispositional orientation that need not translate into coordinated behavior in any specific domain. Exercise identity captures the extent to which exercise is part of the self-concept [17,34,35]; it concerns one behavior and says nothing about how eating is organized around it. Habit strength indexes the automaticity of a single behavior [36,37], and general lifestyle adherence, as indexed by multi-domain profiles such as the HPLP [30] and the FANTASTIC Lifestyle Assessment [31], sums the level of each behavior without reference to their relation. By contrast, the construct targeted here concerns the cross-behavior relation itself: whether a person perceives eating and exercise as coordinated and maintained together across days and circumstances. This relational focus is anchored in multiple-behavior-change research on cross-behavior transfer, in which autonomous motivation developed in the exercise domain spills over into eating self-regulation [38,39], and cognitions and stage transitions for physical activity and fruit-and-vegetable intake are positively interrelated over time [12,40]; meta-analytic evidence likewise indicates that multiple-behavior interventions operate partly through processes of change shared across behaviors [3,9]. Without a valid, reliable, integrated tool, researchers, clinicians, and health-promotion professionals struggle to assess holistic lifestyle behavior. Behavioral inconsistencies, healthy eating without regular physical activity, or exercise combined with poor nutrition may compromise the effectiveness of interventions or signal emotional regulation difficulties [41,42,43]. A systems approach is needed, one that views nutrition and physical activity not as independent elements but as interconnected components of a cohesive lifestyle pattern [44,45,46,47].

In response, the present study introduces the Exercise and Nutrition Consistency Scale (ExNCS), a self-report instrument developed for adults aged 18 and older following established scale-development guidance [48], through a targeted literature review, item generation, expert content review, and factor-analytic evaluation. For this study, exercise-nutrition consistency is defined as the self-perceived degree to which a person’s eating and exercise are coordinated and sustained together across days, contexts, and emotional states. Because it is assessed by self-report on a single occasion, the ExNCS captures perceived integration, respondents’ appraisal of their habitual pattern, rather than observed behavioral alignment or its stability over time. Its items cover four kinds of content: reported cross-behavior coordination (e.g., adjusting diet on exercise days); reported maintenance across contexts and emotional states (e.g., keeping a dietary routine on non-exercise days); beliefs about the interdependence of the two behaviors; and confidence in sustaining both. The last two are not consistency itself but cognitions that theory treats as proximal to it, so the total score reflects perceived integration together with supporting cognitions rather than favorable attitudes or self-efficacy alone. Three complementary frameworks, addressing the motivation, planning, and execution phases of health behavior change, inform this definition and the item content: SDT [49,50,51], whose concept of integrated regulation, the most autonomous form of extrinsic motivation, maps onto the consistency construct, because individuals who internalize both behaviors as expressions of personal values regulate them in a coordinated way; the Transtheoretical Model (TTM) [52,53,54], which specifies the temporal stages through which such integration develops; and the Health Action Process Approach (HAPA) [45,55], which bridges the motivation-behavior gap through planning, coping, and action-control processes. Items that may appear to pull in different directions, attending more to diet on exercise days and maintaining a dietary routine on non-exercise days, index the coupling and stability components, respectively, and are conceptually compatible; the items carry no fixed recall period and therefore refer to the respondent’s typical pattern. The measurement model is treated as reflective: variation in a common tendency toward integration is assumed to produce covariation among the items [56]. This assumption is only approximate, because the items are not strictly interchangeable; they span reported behavior, beliefs, self-appraisal, and confidence, and two items concern a single behavior (maintaining exercise under low mood; keeping eating patterns steady under emotional fluctuation) rather than the cross-behavior relation. The interpretation is therefore narrowed to perceived integration and its closely related cognitions rather than to a set of perfectly exchangeable indicators of one behavior; claims about actual behavioral coordination are kept outside the conclusions. A formative specification was not adopted, because the items were written as parallel indicators within each facet rather than as non-redundant causal components.

The four subdimensions operationalize specific aspects of these integrated theories. Exercise-Nutrition Awareness (ENA) captures TTM’s consciousness-raising and self-evaluation together with SDT’s awareness of the interdependence of behaviors [49,52,53]; the nutrition-performance content domain from which these items were drawn is documented in applied work [57]. Commitment to Healthy Routines (CHR) reflects HAPA’s planning processes and SDT’s movement toward integrated regulation [45,50]; sustained lifestyle-change programs illustrate these processes in applied settings [58,59]. Emotional Consistency and Discipline (ECD) addresses HAPA’s emotion-regulation and coping-planning mechanisms, critical for maintaining behavior under stress [45,55]; applied exercise-nutrition interventions in populations under psychological strain illustrate this content domain [60]. Performance and Efficiency Orientation (PEO) operationalizes the goal-driven aspects of all three theories: TTM’s action stage, SDT’s autonomous motivation toward performance goals, and HAPA’s action control [50,52,55]; applied reviews document the performance-nutrition content domain [61,62]. Together, the four dimensions assess how well individuals have integrated exercise and nutrition into a unified, self-regulated lifestyle. Because the present cross-sectional design cannot test the stage-transition (TTM) or phase-transition (HAPA) processes directly, these two models guide item generation here rather than serving as empirically tested frameworks; SDT’s integrated-regulation construct is the framework most directly examined by the structural analyses. This theoretical premise also anticipates that the four facets will be strongly intercorrelated and may be best represented as components of a superordinate construct rather than as fully independent factors.

The present study pursued four objectives: (a) to generate items and obtain expert content ratings through a targeted literature review and expert-panel evaluation; (b) to explore dimensionality in a randomly selected development half-sample; (c) to evaluate competing one-factor, two-factor, correlated four-factor, second-order, and bifactor specifications in the remaining holdout half-sample; and (d) to provide preliminary evidence on reliability and known-groups differences. Three expectations followed from the theoretical model and the study objectives: H1, the items would reflect four strongly intercorrelated facets consistent with a superordinate integration factor; H2, the total score and each facet would show internal consistency of at least 0.70; and H3, scores would increase with self-reported exercise frequency. Analyses of measurement invariance across sex, estimator sensitivity, the alternative one- and two-factor structures, the bifactor-(S-1) specification, careless responding, acquiescence, covariate-adjusted group comparisons, and the seed-documented re-split were not part of the original analysis plan and are reported as subsequent (non-prespecified) analyses.

2. Materials and Methods

2.1. Participants

A sample of N = 583 adults was recruited via convenience and snowball sampling through Google Forms (Google LLC, Mountain View, CA, USA). Inclusion criteria required that participants be ≥18 years old, engage in intentional exercise at least once per week, and provide informed consent; in this study, “physically active” refers only to this inclusion criterion and does not imply that participants met recommended physical activity guidelines. Data were collected online across Türkiye between January and March 2026. Because the link was distributed through open social-media announcements, institutional e-mail lists, and onward snowball referral, the number of people who saw the invitation cannot be determined. Of the 604 forms submitted, 21 were excluded as ineligible (age < 18 years or no intentional exercise at least once per week), leaving 583 complete responses, all of which were analyzed; these were partitioned into a development half (n = 291) and a holdout half (n = 292) (Section 2.4; Supplementary Figure S1). All participants completed all 20 items retained after content validation. The sample was 58.7% male (n = 342) and 41.3% female (n = 241), with a mean age of 20.90 years (SD = 4.92, range 18–58); 93.7% of participants were aged 18–25. Educational attainment was undergraduate (58.0%), high school (31.2%), associate (7.4%), master’s (2.7%), and doctorate (0.7%). Exercise frequency was distributed as follows: 1–2 days/week (28.8%, n = 168), 3–4 days/week (46.7%, n = 272), and ≥5 days/week (24.5%, n = 143). No formal a priori sample-size calculation was performed; the recruitment target of at least 500 respondents followed Comrey and Lee’s [63] “very good” benchmark. As a post hoc precision rationale, a holdout sample of n = 292 gives power > 0.99 to reject close fit (RMSEA = 0.05) when population misfit is RMSEA = 0.08 at the degrees of freedom of the models compared (df = 92–104, α = 0.05) [64]; the corresponding minimum n for power of 0.80 is approximately 134. However, the ratio of cases to free parameters in the holdout half is only about 6:1 for the most complex models, below the commonly cited 10:1 guideline, so parameter estimates for those models are imprecise [65]. This is therefore a predominantly young-adult, university-based convenience sample rather than a general-population community sample; the survey form required a response to every item, so no item-level data were missing, and submissions were restricted to one response per account, limiting duplicate entries. Forced responding eliminates missing data but does not ensure response quality and may itself encourage careless answering. No attention-check or social-desirability items were embedded, a decision taken to keep the survey brief; careless responding was therefore examined post hoc with response-pattern indices, and the potential influence of socially desirable responding is addressed in Section 4.5. No responses were excluded from the primary analyses. The Istanbul Aydın University Social and Humanities Ethics Committee granted ethical approval prior to data collection (Meeting No: 2025/13; approval date: 26 December 2025); the study followed the Declaration of Helsinki, and all participants provided informed consent.

2.2. Scale Development and Item Generation

A 28-item initial pool was generated through a targeted (non-systematic) literature review grounded in SDT, HAPA, and TTM, following established scale-construction guidance [66,67,68]. Candidate content was drawn from theoretical sources on integrated regulation and self-regulatory planning [45,49,50,51,52,53,54,55], from existing exercise-motivation, physical-activity, and eating-behavior instruments [16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31], and from applied exercise–nutrition literature [57,58,59,60,61,62]; no formal database search protocol, search dates, or screening procedure were used, so the review is not described as systematic. Items were authored a priori to target the four theoretically specified subdimensions: seven items each for ENA, CHR, ECD, and PEO, with each item explicitly assigned to its intended subscale and reviewed for theoretical alignment before content validation. Items were authored in Turkish; the ExNCS was developed and validated exclusively as a Turkish-language instrument. An English rendering of each item is provided solely to allow non-Turkish-speaking readers to understand the item content; it is not a validated cross-cultural adaptation. The English version was not administered, formally back-translated, or psychometrically evaluated, and all evidence reported here pertains to the Turkish form; formal cross-cultural adaptation, comprising independent forward and back translation, expert-committee reconciliation, pretesting, and confirmation of measurement equivalence, would be required before the English wording could be used in research [69,70]. A five-point Likert-type scale was used (1 = Strongly Disagree to 5 = Strongly Agree). Four items (3, 8, 11, 20) were worded in the negative direction to reduce acquiescence bias and were reverse-scored (6-original) in IBM SPSS Statistics 28 (IBM Corp., Armonk, NY, USA) prior to all analyses. Supplementary Table S1 presents the final 16-item Turkish instrument alongside the English working translation, and Supplementary Table S2 maps each item to the construct definition given in Section 1.

2.3. Content Validation

Nine specialists in exercise and sport psychology, nutrition, and scale development (psychometrics) evaluated the item pool using Lawshe’s [71] Content Validity Ratio (CVR; minimum 0.78 for nine panelists at p < 0.05), item-level CVI (I-CVI ≥ 0.78), the modified kappa statistic (κ* ≥ 0.74 = excellent) [72], and scale-level CVI (S-CVI/Ave ≥ 0.90) [73,74]. Panelists rated each item independently, in a single round, for relevance to its intended facet; items falling below the criteria were removed by expert consensus. After content validation, 20 items were retained (five per subscale). The content-validation phase is summarized by the a priori thresholds applied (CVR ≥ 0.78, I-CVI ≥ 0.78, κ* ≥ 0.74, S-CVI/Ave ≥ 0.90), and a full item-level table will be reported with the planned re-evaluation of the revised item set. Retained items kept the wording that was rated, so no item carries ratings that no longer match its final form. Because the item set requires revision before further validation, a new content-validity study with full item-level reporting is planned for the revised form. No cognitive interviews or pilot testing with members of the target population were conducted [75,76,77]; therefore, no response-process evidence shows that respondents interpret terms such as healthy eating, discipline, and consistency as intended, and this is treated as a limitation.

2.4. Data Analysis

Descriptive, reliability, and analysis-of-variance procedures were conducted in IBM SPSS Statistics 28.0 [78]. All factor-analytic, confirmatory, and model-based analyses reported here (polychoric EFA, parallel analysis, confirmatory and bifactor models, invariance testing, and sensitivity analyses) were conducted in R version 4.3.3 (R Foundation for Statistical Computing, Vienna, Austria) with the psych (2.4.1), lavaan (0.6-17), and semTools packages; the complete script is provided as Supplementary File S1. Item analysis examined means, standard deviations, skewness, excess kurtosis [79], response-category frequencies, and corrected item-total correlations (CITC ≥ 0.30) [80].

SPSS generated an initial partition by random case selection without a recorded seed and therefore cannot be regenerated, and the item-total statistics that informed item removal were computed on the full sample, so the confirmatory subsample was not independent of item selection. All split-sample analyses were therefore repeated as a retrospective internal analysis, on a documented random partition generated in R (with a fixed random-number seed, set.seed = 2026; sampling without replacement) into a development half (n = 291) and a holdout half (n = 292), which did not differ in age (t = 1.14, p = 0.25), sex (χ² = 0.05, p = 0.83) or exercise frequency (χ² = 1.10, p = 0.58). Item screening (CITC on the 20-item pool), all exploratory analyses, and the specification of the alternative two-factor model used the development half only; all confirmatory model comparisons and the model-based reliability and discriminant-validity indices used the holdout half only. Analyses on the full sample combine observations that also informed development and are labeled as such; they are expected to be optimistic relative to performance in new data. Because the full dataset had already informed item selection and model development, re-partitioning the same participants does not remove that earlier influence or render the holdout half independent of item selection; this re-split is therefore a retrospective internal check, not independent confirmation. Independent confirmation in a new sample of participants remains a requirement for a future study (Section 4.5).

Because the items are five-point ordinal indicators with marked ceiling effects, the primary EFA used polychoric correlations [81,82], minimum-residual extraction, and oblique (oblimin) rotation. Factor retention was examined with parallel analysis on polychoric correlations (100 replications, 95th percentile) in its component-based [83] and factor-based [84] forms and with the minimum average partial (MAP) test [85]; sampling adequacy was evaluated with the Kaiser–Meyer–Olkin measure (KMO) and Bartlett’s test of sphericity, and a salient loading of 0.40 was applied. A principal axis factoring solution with promax rotation on Pearson correlations was also estimated for comparison [86]. For every solution, the complete pattern matrix, communalities, and factor correlations are reported. A solution was considered inadmissible only if a communality reached or exceeded 1 or a residual variance was negative [87], not merely because an oblique pattern coefficient exceeded 1.

Six measurement models were estimated in the holdout half: a one-factor model (the structure implied by total-score use); a two-factor model specified from the development-half EFA; the theoretically specified correlated four-factor model; a second-order model; a symmetrical bifactor model (one general factor plus four orthogonal group factors); and a bifactor-(S-1) model in which the CHR items serve as the reference facet and load only on the general factor [88]. CHR was chosen as the reference domain on content grounds: its items concern the joint, day-to-day maintenance of exercise and healthy eating as a combined routine (e.g., “both exercise and healthy eating are part of my daily routine”), which is the content closest to the integration the general factor is intended to represent. As a consequence, the general factor is defined in terms of commitment-to-routine content and is therefore more accurately read as that specific common dimension than as a content-balanced index of overall exercise–nutrition integration. The bifactor-(S-1) model was added after the symmetrical bifactor model proved inadmissible and its CHR group factor vanished, an anomaly that is expected for symmetrical G-factor models in single-level designs [88]. Ordinal weighted least squares-mean and variance-adjusted (WLSMV) estimation on polychoric correlations was treated as primary because it matches the item metric [89]; robust maximum likelihood (MLR) was reported alongside it because information criteria and observed-score reliability indices require a likelihood. Fit was assessed with χ², CFI, TLI, RMSEA (90% CI), and SRMR [90,91], with CFI/TLI ≥ 0.95 and RMSEA ≤ 0.06 indicating good fit and CFI/TLI ≥ 0.90 and RMSEA ≤ 0.08 acceptable fit, recognizing that these are heuristics whose appropriateness depends on the model and data [92]; robust versions of CFI, TLI, and RMSEA are reported. Non-nested models were compared with BIC under MLR. For the admissible bifactor solution, ECV, ω, ω_h, ω_hs, and H were computed [93,94]; because bifactor models can fit unsystematic or implausible response patterns well, a bifactor fit advantage was not taken as evidence of structure in itself [95,96]. The same six models were re-estimated on the full sample under ML, MLR, and WLSMV as an estimator-sensitivity analysis. Harman’s single-factor test was computed on the full sample as a diagnostic for common-method variance.

Reliability was assessed using Cronbach’s α and McDonald’s ω [97]; construct validity was assessed using average variance extracted (AVE ≥ 0.50), composite reliability (CR ≥ 0.70), the HTMT ratio (<0.85) [98], and the Fornell–Larcker criterion [99]. Measurement invariance across sex was tested by multi-group confirmatory factor analysis on the first-order four-factor model (N = 583; 342 men, 241 women), first under maximum likelihood with a mean structure (configural, metric, scalar) and then, because the items are ordinal, under WLSMV following the identification sequence of Wu and Estabrook [100,101] (configural; equal thresholds; equal thresholds and loadings; additionally equal intercepts). Invariance was judged by ΔCFI (<0.01) and ΔRMSEA (<0.015) [102,103], with the scaled Δχ² test reported, and conclusions were conditioned on adequate configural fit. Known-groups differences were examined with one-way analysis of variance across exercise-frequency strata for the total score and each subscale, with η² and its 95% confidence interval (obtained by inverting the noncentral F distribution) as the effect-size index [104]. Levene’s test indicated unequal variances for ENA, ECD, and PEO, so Welch’s F was used for the omnibus tests and the Games–Howell procedure for all pairwise comparisons [105], reported as mean differences with 95% confidence intervals and Hedges’ g. Two covariate-adjusted sensitivity analyses (ANCOVA) were conducted, adjusting first for sex, age, education, and body-mass index computed from self-reported height and weight and then additionally for years of exercise experience and usual session duration.

Careless responding was screened post hoc on the 20 items in their original (unrecoded) direction using the longest string of identical consecutive responses (flag ≥ 10, i.e., half the items), intra-individual response variability (flag = 0), and multivariate outlyingness (Mahalanobis D², p < 0.001) [106,107,108]; the principal analyses were repeated after excluding flagged cases. Because the 20-item pool contained four negatively keyed items, acquiescence was examined with a random-intercept item factor model [109] estimated under MLR, in which an acquiescence factor with loadings fixed to +1 for positively keyed items and −1 for (recoded) negatively keyed items was added, orthogonally, to a single substantive factor.

3. Results

3.1. Item Analysis and Item Retention

In the development half, the four reverse-worded items had CITC below the 0.30 criterion after recoding (ENA3R = 0.09; CHR3R = 0.23; ECD1R = 0.07; PEO5R = 0.15; Supplementary Table S4), whereas the 16 positively worded items ranged from 0.42 to 0.75; removing the four items raised α in that half from 0.882 to 0.920. Item removal was thus determined entirely within the development half and did not use holdout data. For reference, Table 1 reports descriptive statistics for the 16 retained items in the full sample; the corresponding full-sample values for the removed items (Item 3 = 0.06, Item 8 = 0.28, Item 11 = 0.11, and Item 20 = 0.13, i.e., ENA3R, CHR3R, ECD1R, and PEO5R) and the full-sample change in α (0.877 to 0.913) point to the same decision. All retained items showed |skewness| ≤ 1.77 and |excess kurtosis| ≤ 2.83, but response distributions were strongly ceilinged for ENA2 (63.6% at the scale maximum), ENA5 (67.1%), and PEO4 (54.4%), and the lowest categories of these items were sparse (1.7–4.6% of responses; as few as three observations in a category in the holdout half; Supplementary Table S3).

Table 1. Item analysis for the 16 retained items (N = 583).

Item	M	SD	Skew	Kurt	CITC
ENA1	3.63	1.17	−0.41	−0.67	0.63
ENA2	4.37	1.01	−1.66	2.17	0.45
ENA4	3.85	1.07	−0.71	−0.04	0.62
ENA5	4.46	0.90	−1.77	2.83	0.48
CHR1	3.48	1.14	−0.15	−0.82	0.71
CHR2	3.73	1.12	−0.44	−0.73	0.65
CHR4	3.25	1.14	0.00	−0.68	0.63
CHR5	3.79	1.13	−0.59	−0.46	0.64
ECD2	3.23	1.19	−0.09	−0.79	0.61
ECD3	3.60	1.26	−0.50	−0.75	0.59
ECD4	2.95	1.28	0.10	−0.96	0.46
ECD5	3.19	1.24	−0.11	−0.89	0.58
PEO1	3.46	1.17	−0.32	−0.65	0.65
PEO2	3.75	1.14	−0.62	−0.41	0.65
PEO3	4.02	1.01	−0.81	0.11	0.57
PEO4	4.17	1.11	−1.31	0.97	0.39

Note. N = 583. CITC = corrected item-total correlation; Kurt = excess kurtosis. Values are descriptive; item-retention decisions used the development half (n = 291).

3.2. Exploratory Factor Analysis

In the development half, sampling adequacy was high (20 items: KMO = 0.895, Bartlett χ² (190) = 2972.36, p < 0.001; 16 retained items: KMO = 0.919, χ² (120) = 2622.54, p < 0.001). The first four polychoric eigenvalues were 8.47, 1.94, 0.96, and 0.84; the first accounted for 52.9% of the total item variance, indicating a dominant general dimension. Retention criteria disagreed: component-based parallel analysis on polychoric and Pearson correlations and the MAP test supported two factors, whereas factor-based parallel analysis supported four. The data therefore did not unambiguously determine the number of factors, and we carried forward the four-factor specification to confirmatory testing because it was theoretically specified, not because the exploratory evidence favored it.

The two-factor polychoric solution (Supplementary Table S5) was admissible (communalities 0.27–0.77) and accounted for 60.5% of the common variance; the factors correlated r = 0.49. Factor 1 comprised all ENA and PEO items together with CHR1, CHR2, and CHR5 (pattern coefficients 0.45–0.88); factor 2 comprised all four ECD items (0.58–0.87) and CHR4 (0.44). CHR1 (0.54 and 0.41) and CHR4 (0.42 and 0.44) cross-loaded substantially, and PEO4 had a low communality (0.27). The solution therefore separates an awareness-performance-commitment dimension from an emotional-discipline dimension and distributes the commitment items across both. The same two-factor pattern emerged in the full sample (Supplementary Table S5), although that replication is not independent because the full sample contains the development half.

A four-factor solution estimated by principal axis factoring with promax rotation on Pearson correlations was admissible in the development half: all communalities were below 1 (maximum 0.86), and the single pattern coefficient above 1.0 (PEO2 = 1.02) occurred in an oblique solution with a communality of 0.86 and is therefore not a Heywood case [87]; the polychoric four-factor solution behaved similarly (maximum communality 0.92). An earlier description of this four-factor EFA as inadmissible, based on pattern coefficients above 1.0, could not be re-examined because the original partition cannot be regenerated, and it is withdrawn. More importantly, the four exploratory factors did not reproduce the intended facets: the CHR items grouped with ENA1 and ENA4; ENA2 and ENA5 formed a doublet (with PEO3); PEO1 and PEO2 formed a factor with ENA4; and only the ECD facet emerged as specified. Inter-item correlations within ECD were strong (r = 0.43–0.73) rather than near-perfect. The exploratory evidence thus supports a strong general dimension with one clearly separable emotional-discipline component and provides no independent support for the four-facet specification.

3.3. Confirmatory and Hierarchical Factor Analysis

Table 2 reports the six models in the holdout half; under the primary ordinal estimator (WLSMV), no model met conventional fit criteria. The one-factor model fitted very poorly (CFI = 0.680, RMSEA = 0.182), as did the EFA-derived two-factor model (CFI = 0.775, RMSEA = 0.153). The correlated four-factor model (CFI = 0.865, TLI = 0.834, RMSEA = 0.122 [0.109, 0.135], SRMR = 0.075) and the bifactor-(S-1) model (CFI = 0.875, TLI = 0.837, RMSEA = 0.121 [0.108, 0.134], SRMR = 0.075) fitted best but remained well short of benchmarks, the second-order model fitted worse than the correlated model (CFI = 0.847), and the symmetrical bifactor model did not converge. Under MLR, the ordering was the same, and fit was closer to, but still below, conventional thresholds (four-factor: CFI = 0.909, RMSEA = 0.084; bifactor-(S-1): CFI = 0.920, RMSEA = 0.081); BIC favored the correlated four-factor model (12,423.0) marginally over the bifactor-(S-1) model (12,424.7) and both clearly over the second-order (12,443.9), two-factor (12,527.5), and one-factor (12,748.7) models. Because the one-factor model, the structure implied by summing the items, was clearly rejected, the total score cannot be regarded as strictly unidimensional [110]; the question addressed by the bifactor-(S-1) model is whether a general factor nonetheless dominates the reliable variance of that sum.

Table 2. Comparison of six measurement models in the holdout half-sample (n = 292; seed-documented partition).

Model	Estimator	χ2	df	CFI	TLI	RMSEA [90% CI]	SRMR	AIC	BIC
One-factor	MLR	521.21	104	0.750	0.712	0.134 [0.123, 0.146]	0.096	12,631.0	12,748.7
Two-factor (EFA)	MLR	356.37	103	0.852	0.827	0.104 [0.092, 0.116]	0.091	12,406.2	12,527.5
Correlated four-factor	MLR	255.62	98	0.909	0.888	0.084 [0.071, 0.096]	0.066	12,283.3	12,423.0
Second-order	MLR	280.99	100	0.895	0.874	0.089 [0.077, 0.101]	0.073	12,311.6	12,443.9
Bifactor (symmetrical)	MLR	n.c.	—	—	—	—	—	—	—
Bifactor-(S-1)	MLR	231.95	92	0.920	0.895	0.081 [0.068, 0.094]	0.064	12,262.9	12,424.7
One-factor	WLSMV	1073.83	104	0.680	0.631	0.182 [0.170, 0.193]	0.125	—	—
Two-factor (EFA)	WLSMV	733.19	103	0.775	0.737	0.153 [0.141, 0.165]	0.098	—	—
Correlated four-factor	WLSMV	426.97	98	0.865	0.834	0.122 [0.109, 0.135]	0.075	—	—
Second-order	WLSMV	506.62	100	0.847	0.816	0.128 [0.116, 0.141]	0.085	—	—
Bifactor (symmetrical)	WLSMV	n.c.	—	—	—	—	—	—	—
Bifactor-(S-1)	WLSMV	429.05	92	0.875	0.837	0.121 [0.108, 0.134]	0.075	—	—

Note. MLR = robust maximum likelihood; WLSMV = ordinal estimation on polychoric correlations (primary estimator). χ² is the scaled statistic; CFI, TLI, and RMSEA are robust versions; BIC is defined only under MLR. Two-factor (EFA) = model specified from the development-half EFA (Section 3.2); bifactor-(S-1) = CHR items load only on the general factor [88]. n.c. = did not converge.

The full-sample estimator-sensitivity analysis (Table 3) reproduced this ordering under ML, MLR, and WLSMV. It also showed that the symmetrical bifactor solution, which an ML/MLR analysis can appear to favor (χ² = 426.24, CFI = 0.929), converges only to an inadmissible solution in which CHR1 has a large negative residual variance and a CHR group loading far above 1. That apparent fit and the bifactor indices derived from the corresponding split-sample solution (ECV = 0.667; ω_h = 0.846), which already contained a boundary estimate for CHR1, are therefore not interpretable and are replaced below by the admissible bifactor-(S-1) solution. The symmetrical bifactor model did not converge under WLSMV in either sample. Taken together, these results do not confirm any of the specified structures; they indicate a strong general dimension with residual multidimensionality whose form remains to be established.

Table 3. Estimator-sensitivity analysis: the six measurement models re-estimated on the full sample (N = 583).

Model	Estimator	χ2	df	CFI	TLI	RMSEA	SRMR	AIC	BIC
One-factor	ML	1301.66	104	0.749	0.710	0.141	0.093	25,163.7	25,303.5
Two-factor (EFA)	ML	830.71	103	0.847	0.822	0.110	0.086	24,694.7	24,838.9
Correlated four-factor	ML	521.53	98	0.911	0.891	0.086	0.060	24,395.6	24,561.6
Second-order	ML	602.66	100	0.895	0.873	0.093	0.069	24,472.7	24,629.9
Bifactor (symmetrical)	ML	inadm. [426.24]	88	[0.929]	[0.903]	[0.081]	[0.056]	[24,320.3]	[24,529.9]
Bifactor-(S-1)	ML	453.57	92	0.924	0.901	0.082	0.057	24,339.6	24,531.8
One-factor	MLR	951.73	104	0.754	0.717	0.138	0.093	25,163.7	25,303.5
Two-factor (EFA)	MLR	619.17	103	0.853	0.829	0.107	0.086	24,694.7	24,838.9
Correlated four-factor	MLR	394.63	98	0.917	0.898	0.083	0.060	24,395.6	24,561.6
Second-order	MLR	456.60	100	0.900	0.880	0.090	0.069	24,472.7	24,629.9
Bifactor (symmetrical)	MLR	inadm. [322.09]	88	[0.934]	[0.910]	[0.078]	[0.056]	[24,320.3]	[24,529.9]
Bifactor-(S-1)	MLR	343.83	92	0.930	0.908	0.079	0.057	24,339.6	24,531.8
One-factor	WLSMV	2020.27	104	0.702	0.656	0.179	0.118	—	—
Two-factor (EFA)	WLSMV	1364.56	103	0.796	0.762	0.148	0.090	—	—
Correlated four-factor	WLSMV	729.49	98	0.884	0.858	0.115	0.064	—	—
Second-order	WLSMV	908.39	100	0.865	0.838	0.122	0.076	—	—
Bifactor (symmetrical)	WLSMV	n.c.	88	—	—	—	—	—	—
Bifactor-(S-1)	WLSMV	703.03	92	0.900	0.869	0.110	0.065	—	—

Note. Full sample (N = 583), which includes the development half, so estimates are optimistic. χ² is scaled under MLR and WLSMV, with robust CFI, TLI, and RMSEA. AIC and BIC are identical under ML and MLR and are undefined under WLSMV. inadm. = inadmissible solution (Section 3.3); bracketed values are shown for transparency only and are not interpretable. n.c. = did not converge.

In the admissible bifactor-(S-1) solution in the holdout half (WLSMV; Table 4), all items loaded substantially on the general factor (0.46–0.85), with the CHR items, which define the general factor in this parameterization, loading 0.72–0.85. The general factor accounted for most of the common variance (ECV = 0.734) and of the total-score variance on the latent-response scale (ω_h = 0.876 against ω = 0.950; ω_h/ω = 0.92) and was well defined (H = 0.939). On the observed-score scale (MLR), which corresponds more closely to the summed total score, the values were ECV = 0.689, ω_h = 0.839, and ω = 0.929 (full sample: ECV = 0.696, ω_h = 0.848, ω = 0.936). Specific-factor reliability beyond the general factor was low for ENA (ω_hs = 0.20) and PEO (0.14) and moderate only for ECD (0.49), in agreement with the EFA, in which ECD was the only facet that separated clearly. These indices are consistent with a total score whose reliable variance is dominated by a single general factor. Because they derive from a model that did not reach acceptable fit under the primary ordinal estimator, and because that general factor is anchored on CHR content (Section 2.4), they describe the composite rather than confirm the intended integration construct. They do not establish unidimensionality because the one-factor model was rejected, and they are conditional on a bifactor specification introduced after the symmetrical model failed. The CHR facet has no specific factor in this model because, as the reference domain, its group factor is omitted by design; this omission means that CHR-specific variance cannot be estimated here, not that it is absent, and the vanished CHR group factor in the symmetrical specification is itself the anomaly that motivated the (S-1) parameterization. Pending a model in which CHR-specific variance can be estimated, CHR should not be scored or interpreted as a separate subscale. The ENA-specific factor also requires caution: its two substantial specific loadings belong to ENA2 (0.62) and ENA5 (0.51), the two most ceilinged items (63.6% and 67.1% at the scale maximum), so it is at least as plausibly a response-distribution (doublet) artifact as a substantive awareness facet.

Table 4. Standardized loadings in the holdout half (n = 292, WLSMV): correlated four-factor model and bifactor-(S-1) decomposition.

Item	Four-Factor λ	Bifactor-(S-1) General (G)	Bifactor-(S-1) Specific	Subscale
ENA1	0.776	0.716	0.188	ENA
ENA2	0.663	0.553	0.617	ENA
ENA4	0.815	0.760	0.118	ENA
ENA5	0.702	0.593	0.514	ENA
CHR1	0.862	0.854	—	CHR
CHR2	0.767	0.765	—	CHR
CHR4	0.781	0.775	—	CHR
CHR5	0.721	0.717	—	CHR
ECD2	0.847	0.593	0.572	ECD
ECD3	0.765	0.573	0.422	ECD
ECD4	0.791	0.457	0.706	ECD
ECD5	0.915	0.588	0.727	ECD
PEO1	0.807	0.755	0.197	PEO
PEO2	0.801	0.742	0.288	PEO
PEO3	0.674	0.620	0.349	PEO
PEO4	0.506	0.455	0.326	PEO

Note. All loadings are significant (p < 0.001). CHR is the reference facet and has no specific factor. General factor: ECV = 0.734, ω = 0.950, ω_h = 0.876, H = 0.939; ω_hs: ENA = 0.20, ECD = 0.49, PEO = 0.14.

3.4. Reliability and Construct Validity

In the holdout half, Cronbach’s α ranged from 0.733 to 0.854 across subscales and composite reliability (CR, MLR) from 0.729 to 0.871; total-scale α was 0.906 (full sample: subscale α = 0.764–0.862; total α = 0.913). AVE ranged from 0.440 to 0.614: CHR (0.542) and ECD (0.614) met the ≥0.50 criterion, whereas ENA and PEO (both 0.440) did not, indicating that more than half of the variance in these items is error or item-specific variance (Table 5). HTMT satisfied the <0.85 criterion for five of six pairs; ENA-PEO (HTMT = 0.897; latent r = 0.90) exceeded it, and the Fornell–Larcker criterion was not met for three of six pairs (ENA-CHR, ENA-PEO, and CHR-PEO; Table 6). ENA, CHR, and PEO are therefore not empirically distinguishable in this sample. Together with the EFA and bifactor (S-1) results, this pattern suggests that these three facets may not be distinct at all, rather than distinct facets that merely share a general factor; only ECD showed discriminant evidence. Consequently, subscale scores are not recommended for separate interpretation, and the four-facet structure is regarded as unconfirmed.

Table 5. Reliability and construct-validity indices (holdout half, n = 292).

Subscale	α	ω	CR	AVE	√AVE
ENA	0.745	0.729	0.729	0.440	0.663
CHR	0.819	0.830	0.830	0.542	0.736
ECD	0.854	0.871	0.871	0.614	0.784
PEO	0.733	0.738	0.738	0.440	0.663
Total ExNCS	0.906	0.929	—	—	—

Note. α = Cronbach’s alpha; ω = McDonald’s omega; CR = composite reliability; AVE = average variance extracted; √AVE = square root of AVE. Values are from the holdout half; total ω is from the bifactor-(S-1) model (ω_h = 0.839).

Table 6. Discriminant-validity matrix (√AVE on diagonal; factor correlations below diagonal; HTMT above diagonal; holdout half, n = 292).

	ENA	CHR	ECD	PEO
ENA	(0.663)	0.818	0.374	0.897
CHR	0.884	(0.736)	0.723	0.800
ECD	0.452	0.681	(0.784)	0.526
PEO	0.901	0.846	0.591	(0.663)

Note. Diagonal (in parentheses) = √AVE; below diagonal = latent correlations (correlated four-factor model, MLR); above diagonal = HTMT ratios. ENA-PEO exceeds the HTMT criterion of 0.85, and three of six Fornell–Larcker comparisons are not satisfied.

3.5. Preliminary Known-Groups Comparison

Total and subscale scores increased monotonically across exercise-frequency strata, with all omnibus tests significant at p < 0.001 (Table 7). The total-score effect was large by conventional benchmarks (η² = 0.153, 95% CI [0.102, 0.204]); ECD showed a similar effect (η² = 0.156 [0.104, 0.207]), and ENA showed the smallest (η² = 0.064 [0.029, 0.104]). The confidence intervals for ECD, CHR, and PEO overlap substantially, so differences among subscale effect sizes should not be interpreted as differences in underlying mechanisms. Levene’s test indicated unequal variances for ENA (W = 4.88, p = 0.008), ECD (W = 3.12, p = 0.045), and PEO (W = 2.92, p = 0.055; p = 0.040 with median centering), so Welch’s F and Games–Howell comparisons were used for all scales. For the total score, mean differences were 0.46 (95% CI 0.29–0.62; g = 0.66) between 1–2 and 3–4 days/week, 0.34 (0.18–0.50; g = 0.51) between 3–4 and ≥ 5 days/week, and 0.79 (0.60–0.99; g = 1.08) between the extreme groups; all adjacent differences were significant on every scale, including ENA between 3–4 and ≥5 days/week (0.20, 95% CI 0.02–0.37, p = 0.022). Adjusting for sex, age, education, and body-mass index left the total-score effect essentially unchanged (partial η² = 0.141), whereas additional adjustment for years of exercise experience and usual session duration reduced it to partial η² = 0.053 (F (2, 563) = 15.80, p < 0.001), indicating that the frequency gradient is shared with other indicators of exercise engagement. Because several ExNCS items explicitly reference exercise and the grouping variable is a self-reported exercise behavior, these comparisons provide only preliminary known-groups evidence; they are not criterion validity and do not show that the scale captures coordination between physical activity and dietary behavior.

Table 7. Preliminary known-groups comparison: ExNCS scores by exercise frequency (N = 583).

Scale	1–2 d/wk	3–4 d/wk	≥5 d/wk	Welch’s F	η2 [95% CI]
ENA	3.78 ± 0.89	4.13 ± 0.74	4.32 ± 0.69	18.16	0.064 [0.029, 0.104]
CHR	3.13 ± 0.95	3.63 ± 0.83	3.95 ± 0.88	31.24	0.106 [0.062, 0.153]
ECD	2.68 ± 1.08	3.30 ± 0.91	3.80 ± 0.91	49.36	0.156 [0.104, 0.207]
PEO	3.51 ± 0.92	3.88 ± 0.76	4.21 ± 0.76	26.96	0.091 [0.050, 0.136]
Total	3.28 ± 0.77	3.73 ± 0.64	4.07 ± 0.69	46.58	0.153 [0.102, 0.204]

Note. Values are M ± SD. Welch’s F is reported because variances were unequal for ENA, ECD, and PEO (Section 3.5); Welch degrees of freedom ranged from (2, 316.2) to (2, 326.9), and all tests were significant at p < 0.001. η²: small = 0.01, medium = 0.06, large = 0.14 [104]. Pairwise comparisons, Hedges’ g, and covariate-adjusted results are reported in Section 3.5 and Supplementary Table S7.

3.6. Measurement Invariance Across Sex

Under maximum likelihood (Table 8, upper panel), the configural model fitted only marginally (CFI = 0.900, RMSEA = 0.064); constraining loadings produced a negligible change (ΔCFI = −0.003, ΔRMSEA = −0.001), and adding intercept constraints produced ΔCFI = −0.010, at the boundary of the criterion, with a significant Δχ² (12) = 57.51, p < 0.001. Because the items are ordinal, the analysis was repeated under WLSMV (Table 8, lower panel). The configural model fitted poorly (robust CFI = 0.865, robust RMSEA = 0.123; scaled CFI = 0.936, RMSEA = 0.105). Successive equality constraints on thresholds, loadings, and intercepts produced negligible changes in scaled CFI and RMSEA (|ΔCFI| ≤ 0.004), and the scaled Δχ² tests were non-significant for thresholds (Δχ² (32) = 23.68, p = 0.86) and loadings (Δχ² (12) = 15.82, p = 0.20) but significant for intercepts (Δχ² (12) = 58.72, p < 0.001). Because change-in-fit criteria presuppose an adequately fitting configural model, which was not obtained under either estimator, and because the intercept step was significant under both, measurement invariance across sex is regarded as inconclusive. These analyses concern the first-order model; invariance of the total score (general factor) was not tested, and neither latent nor observed mean comparisons between men and women are warranted on the present evidence.

Table 8. Measurement invariance of the first-order four-factor model across sex (N = 583; 342 men, 241 women) under ML and ordinal (WLSMV) estimation.

Model	χ2	df	CFI	TLI	RMSEA	Δχ2 (Δdf)	ΔCFI	ΔRMSEA
ML: Configural	662.98	196	0.900	0.878	0.064	—	—	—
ML: Metric	688.37	208	0.897	0.882	0.063	25.39 (12) *	−0.003	−0.001
ML: Scalar	745.88	220	0.888	0.878	0.064	57.51 (12) **	−0.010	+0001
WLSMV: Configural	824.03	196	0.936	0.921	0.105	—	—	—
WLSMV: Thresholds	859.19	228	0.935	0.932	0.098	23.68 (32)	−0.001	−0.007
WLSMV: +Loadings	828.79	240	0.940	0.940	0.092	15.82 (12)	+0004	−0.006
WLSMV: +Intercepts	864.33	252	0.937	0.940	0.091	58.72 (12) **	−0.002	−0.001

Note. ML = maximum likelihood with a mean structure and marker-variable identification; WLSMV = ordinal estimation with delta parameterization and the identification sequence of Wu and Estabrook [100]. Under WLSMV, χ² and the fit indices are scaled versions (robust configural values: CFI = 0.865, RMSEA = 0.123). Δ values are relative to the preceding, less constrained model. * p < 0.05; ** p < 0.001.

3.7. Sensitivity Analyses: Careless Responding and Acquiescence

Twenty-one respondents (3.6%) gave the identical response to all 20 items in their original direction, including the four negatively keyed items, and therefore showed zero response variability; 69 (11.8%) had a longest identical string of ten or more items, and 31 (5.3%) were multivariate outliers (Mahalanobis D², p < 0.001). In total, 97 respondents (16.6%) met at least one criterion (Supplementary Table S6). Excluding them left the principal results essentially unchanged: total α = 0.903; correlated four-factor model (MLR) CFI = 0.907, RMSEA = 0.085; bifactor-(S-1) model CFI = 0.923, RMSEA = 0.080; and a total-score known-groups effect of η² = 0.169 (n = 486). The corresponding WLSMV fit after exclusion was likewise essentially unchanged (correlated four-factor CFI = 0.876, RMSEA = 0.116; bifactor-(S-1) CFI = 0.896, RMSEA = 0.109), so no model reached acceptable ordinal fit in the reduced sample either. In the random-intercept model on the 20-item pool, the acquiescence factor had significant variance (0.157, SE = 0.029, p < 0.001; standardized loadings of absolute value 0.29–0.44) and improved fit over the single-factor model (scaled Δχ² (1) = 13.78, p < 0.001), although neither model fitted well (CFI = 0.674 and 0.695). Because both models fit poorly, these results suggest a possible acquiescent response-style contribution rather than its magnitude. Because the 16 retained items are all positively keyed, acquiescence cannot be separated from the substantive general factor in the final form and is likely to inflate it to an unknown degree; response style therefore remains a plausible, but not quantified, alternative explanation for part of the general-factor variance reported above.

4. Discussion

The present study developed the ExNCS and reports a preliminary psychometric evaluation of it as a candidate Turkish item set for perceived exercise-nutrition integration in young adults who exercise. Four findings summarize the evidence. First, the 16 retained items share a strong general dimension and yield an internally consistent total score (H2 supported for the total score and all facets). Second, no measurement model, including the theoretically specified four-facet model, met conventional fit criteria under the estimator appropriate for the item format, and the exploratory evidence favored two rather than four factors, so the dimensional structure remains provisional (H1 not supported). Third, total scores increased with exercise frequency (H3 supported), but this preliminary known-groups evidence is partly content-driven and was attenuated after adjustment for other exercise characteristics. Fourth, acquiescence, careless responding and the absence of response-process and external-validity evidence limit what the scores can be taken to mean. These points, rather than any claim of established validity, frame the discussion below.

4.1. Dimensional Structure: A Strong General Factor Within a Provisional Structure

The central structural question was whether the four subscales represent distinct constructs or facets of a superordinate construct. Exploratory retention criteria disagreed (two versus four factors), the four exploratory factors did not reproduce the intended facets, the one-factor, two-factor, correlated four-factor, and second-order models all fell short of conventional fit criteria under WLSMV, and the symmetrical bifactor model was inadmissible or non-convergent in every sample and under every estimator. The bifactor-(S-1) model, introduced after that failure, was admissible and indicated that a general factor dominates the reliable variance of the total score (ω_h = 0.84–0.88 depending on scale and sample), but its fit was no better than that of the correlated four-factor model. The present study therefore does not confirm a hierarchical or bifactor structure; it shows that a strong but exploratory general dimension underlies the item set and that residual multidimensionality, most clearly an emotional-discipline component, remains unexplained. Interpreting the general factor as integrated regulation, consistent with SDT [50], is one possible reading. Two alternative explanations are at least as plausible for the present evidence. First, the general factor may partly reflect response style: the random-intercept model on the 20-item pool revealed a significant acquiescence factor, the final 16-item form is uniformly positively keyed and so contains no internal control for acquiescence [111], and Harman’s single-factor test indicated that one unrotated component accounted for 44.3% of the total item variance, a diagnostic rather than a protective result [112]. Second, the general factor may reflect a broad health orientation or wording shared across items (several refer to “healthy habits” or “these habits”) rather than cross-behavior integration specifically. Symmetrical bifactor models also reward modeling unsystematic or implausible response patterns [95], which is one reason the earlier bifactor fit advantage should not have been read as structural support. Distinguishing substantive integration from these alternatives requires balanced keying, multi-method data, and external correlates.

At the facet level, the evidence points to a potential lack of distinction between facets rather than to distinct facets that merely share a general factor. ENA, CHR, and PEO correlated 0.85–0.90 at the latent level; ENA-PEO exceeded the HTMT criterion, three of six Fornell–Larcker comparisons failed, and in the EFA, these items formed a single factor; only ECD separated consistently (latent correlations 0.45–0.68; ω_hs = 0.49). The CHR group factor vanished in the symmetrical bifactor model, and CHR should therefore not be scored or interpreted independently. The ENA-specific factor is carried by two ceilinged items and may be a response-distribution artifact. Based on the current evidence, the total score is the only unit for which an interpretation is offered, and that interpretation is preliminary and pending independent validation; the four facets are best regarded as content groupings used in item writing rather than as psychometrically supported subscales. Whether a two-dimensional (integration and emotional discipline) structure, a unidimensional short form, or a revised four-facet structure is most defensible should be decided in an independent, more heterogeneous sample using ordinal methods and exploratory structural equation modeling (ESEM) [113], which accommodates the cross-loadings observed here. Consistent with the coupling-stability distinction drawn in Section 1, the two items that appear to pull in opposite directions (ENA1, adjusting diet on exercise days; CHR4, maintaining a dietary routine on non-exercise days) were positively correlated (r = 0.55), indicating that respondents who adjust their eating around exercise also tend to keep a baseline dietary routine.

4.2. Reverse-Worded Items

The four reverse-worded items underperformed in both the development half (CITC 0.07–0.23) and the full sample (0.06–0.28). This pattern is plausibly attributable to a wording-direction method effect; syntactically complex negative items can introduce a distinct method factor, particularly in samples with variable familiarity with Likert formats [111]. It may also partly reflect careless responding: 21 respondents gave the same answer to every item regardless of keying, which necessarily depresses the item-total correlations of negatively keyed items. Because the present design could not formally separate a method factor from substantive misfit, this removal is best understood as a data-driven decision pending confirmation. A consequence that must be stated plainly is that the final form contains no negatively keyed items and therefore no internal safeguard against acquiescence, which the random-intercept analysis showed to be present in this sample; the general factor reported in Section 4.1 is therefore likely to be partly inflated by it. Four concrete strategies are recommended for the next version: cognitive interviews with members of the target population to establish why the negative items failed (e.g., double negation or the ambiguity of “does not affect” in Turkish) and to rewrite them as clearly worded reversals or as balanced pairs of semantically opposite statements [75]; administration of a balanced form, with approximately equal numbers of positively and negatively keyed items per facet, so that acquiescence can be modeled with a random-intercept or method factor [109,114]; confirmatory comparison of substantive and method-factor specifications on that balanced form; and item response theory analysis of category functioning, given the sparse lower categories and ceiling effects observed here.

4.3. Reliability and Validity Evidence

Internal consistency was high for the total score (α = 0.906–0.913; ω = 0.929–0.936 depending on sample), and α and ω were similar at the subscale level [115]. However, high internal consistency in a uniformly keyed item set with a significant acquiescence component partly reflects shared response style and should not be read as evidence of validity. AVE for ENA and PEO (0.440) was below the 0.50 criterion, meaning that these items carry more error and item-specific variance than facet variance [90]. The exercise-frequency gradient (η² = 0.153, 95% CI 0.102–0.204) is useful preliminary known-groups evidence, but it is partly content-driven; it was substantially attenuated after adjustment for exercise experience and session duration (partial η² = 0.053), and it cannot show that higher scores reflect healthier behavior in both domains, because no independent dietary or physical-activity measure was included. Likewise, the cross-sectional association between ECD and exercise frequency does not show that frequent exercise develops emotional coping or discipline [45], and the overlapping confidence intervals of the subscale effects do not support ranking the facets by association strength.

4.4. Applications

The ExNCS is a brief (16-item, approximately 5-minute) Turkish-language item set. Based on current evidence, its total score may be used only provisionally as an exploratory, descriptive, group-level composite indexing perceived exercise-nutrition integration in research with samples similar to the present one (young Turkish adults who already exercise). It is not a screening instrument: no criterion evidence, cut-scores, or normative data exist, and it has not been administered to inactive adults. It should not be used for clinical or individual-level decisions, for individual profile interpretation or case conceptualization, or for comparing groups until measurement invariance has been established; subscale scores, CHR in particular, should not be scored or interpreted independently; and the English working translation should not be used to collect data. For researchers, the practical take-home points are to report the total score together with these limitations, to include balanced keying and attention-check items in future administrations, to collect external measures of physical activity and diet alongside the ExNCS, and to treat the present item set as a starting point for revision rather than a finished instrument.

4.5. Limitations and Future Directions

Two features of the analytic history are important limitations; item removal drew on full-sample statistics, so the confirmatory subsample was not independent of item selection, and the initial unseeded partition cannot be regenerated. The seed-documented re-split is a retrospective internal analysis that reduces, but does not remove, this dependence, because the same participants informed item selection and model development; it is not independent confirmation. The full-sample analyses (item descriptives, estimator sensitivity, known-groups comparisons, invariance and sensitivity analyses) reuse observations that informed development; several analyses were not part of the original analysis plan; and the bifactor-(S-1) specification was adopted only after the symmetrical model failed. Model performance reported here is therefore likely to be optimistic, and independent replication in a new sample is required before any structural or score-interpretation claim can be regarded as established [116].

The sample was predominantly young (mean age 20.90; 93.7% aged 18–25), largely university-based and educationally homogeneous (89% high-school or undergraduate), and recruited by convenience and snowball sampling; all participants exercised at least once per week. The findings therefore concern young Turkish adults who exercise and cannot be generalized to inactive adults, the group in which a lifestyle-integration measure would be most informative for public health, or to middle-aged and older adults, clinical populations, or elite athletes. The next validation study should use a stratified community sample spanning age bands (e.g., 18–29, 30–49, 50–64 and ≥65 years), include inactive and insufficiently active adults and people with cardiometabolic conditions, and be large enough to test invariance across age, sex and education.

Content validity rests on a single round of expert ratings that are reported only at the threshold level, without item-level indices; revised items were not re-rated, and no cognitive interviews or pilot testing with the target population were conducted, so there is no evidence that respondents interpret the items as intended. Several items assess beliefs, knowledge, or confidence rather than reported behavior, so the score reflects perceived integration and related cognitions, not behavioral alignment. Forced responding eliminated missing data but not careless responding: about one in six respondents met at least one careless-response criterion, and a significant acquiescence factor was present. No attention-check or social-desirability items were included, and all data were self-reported on a single occasion, so social desirability and common-method variance may have inflated the general factor and the internal-consistency estimates. Dimensionality criteria disagreed; no model met conventional fit criteria under ordinal estimation, the symmetrical bifactor model was inadmissible, and the bifactor indices come from a post hoc bifactor-(S-1) specification. The holdout half yielded only about six cases per free parameter for the most complex models, fewer than is generally recommended for bifactor models [117].

External, convergent, discriminant, and criterion validity were not examined, and the known-groups contrast is partly tautological. Future studies should correlate ExNCS scores with device-based physical activity, dietary intake, exercise motivation, eating behavior, habit strength, health consciousness, and exercise identity and should test predictive validity for intervention adherence. The cross-sectional design precludes conclusions about stability over time; test-retest reliability over 2–4 weeks (ICC ≥ 0.70) and responsiveness to intervention remain to be established. Sex invariance was inconclusive, and invariance across age, education, language, and culture is untested. The ExNCS was developed and evaluated solely in Turkish; the English wording is an unvalidated working translation that must undergo forward and back translation, expert-committee review, cognitive pretesting, and invariance testing before use [69,70].

5. Conclusions

This cross-sectional study offers a candidate Turkish item set, the 16-item ExNCS, intended to capture perceived integration of exercise and nutrition. In a predominantly young sample of adults who exercise, the items showed encouraging internal consistency, a strong but exploratory general dimension, and preliminary associations with exercise frequency. However, their dimensional structure was not confirmed: no model met conventional fit criteria under ordinal estimation, the four facets were not empirically distinct, and part of the general factor may reflect acquiescence. Because no model reached acceptable fit under the primary ordinal estimator, the total score is best treated as an exploratory composite whose interpretation requires further validation, and the high reliability and explained-common-variance estimates describe that composite rather than confirming the intended construct, and the subscales should not be used independently. Replication in an independent and more heterogeneous sample, revision toward balanced keying, cognitive interviewing, external and criterion validation, test–retest reliability, measurement invariance, and formal cross-cultural adaptation are required before the ExNCS can be recommended for wider research use.