Trust and Contextual Readiness in Electric Vehicle Acceptance in Germany
Abstract:
Electric vehicle (EV) acceptance involves evaluations of vehicles, charging, and support, but intention, confidence, and use measures may overlap. This study examines these relationships in a quota-based online panel of 485 adults in Germany. The baseline partial least squares (PLS) structural equation modeling (SEM) contains 17 constructs, 45 indicators, and 27 direct paths. Adjusted $R^2$ values were 0.783 for behavioural intention (BI), 0.688 for trusting intention (TI), and 0.651 for use behaviour. Use behaviour denotes endorsement of three use statements, not verified adoption: only 166 respondents confirmed EV driving experience. BI ($\beta$ = 0.594) and facilitating conditions ($\beta$ = 0.235) were the strongest baseline correlates of this outcome. Reliability ($\beta$ = 0.473) and contextual acquaintance ($\beta$ = 0.377) were the strongest correlates of TI. Five statistically significant paths opposed their stated hypotheses and do not support those directional predictions. Additional analyses removed overlapping habit and reliability constructs, specified costs formatively, and restricted estimation to experienced respondents, each with 5,000 bootstrap samples. The intention-use and facilitating-conditions-use associations remained positive, whereas the small negative trusting-intention-use association was not statistically supported in the reduced model or experienced subset. A new analysis of 528 interaction terms retained multi-item moderators continuously; 77 passed Benjamini-Hochberg adjustment, and 46 passed the more conservative Benjamini-Yekutieli adjustment. Motivation, system confidence, and use-item endorsement provide a conceptual organisation of the results, but empirical separation, actual adoption, and out-of-sample predictive capability remain unestablished.1. Introduction
The decarbonisation of private motorised transport depends not only on the availability of electric vehicles (EVs) but on their adoption and repeated use. The EV acceptance literature identifies technological, economic, contextual, social, attitudinal and experience-related determinants, yet the relative influence of these determinants varies substantially by context and study design [1]. For transport planning, this creates a practical problem: investment in vehicles and charging facilities can establish formal availability without ensuring that households regard the overall system as reliable and compatible with everyday mobility.
Technology-acceptance research provides a strong starting point. The Unified Theory of Acceptance and Use of Technology (UTAUT) links performance expectancy (PE), effort expectancy (EE), social influence (SI), facilitating conditions (FC), and intention to use [2]. Its consumer extension adds hedonic motivation (HM), price value (PV), and habit (Ha) [3], and later synthesis has challenged the stability of individual UTAUT relationships across technologies and populations [4]. Electric mobility is unusually system-dependent. Vehicle use involves charging access, range planning, energy and maintenance costs, market availability, institutional support and service arrangements. A comprehensive explanation therefore needs to retain conventional acceptance determinants while also modelling confidence in the wider EV system.
Trust in technology is the willingness to depend on a technology despite possible adverse consequences [5]. Trust-based extensions emphasise structural assurance (SA), reliability (Re) and contextual familiarity [5], [6], [7], [8]. Contextual acquaintance (CA) is particularly relevant for EVs because prospective users evaluate charging, range, costs, and vehicle fit within concrete household and travel routines. Diffusion theory and EV field studies likewise indicate that knowledge and direct experience can change evaluations [9], [10], [11], [12], [13]. These mechanisms motivate separate theoretical consideration of behavioural intention (BI), trusting intention (TI) in electromobility, and responses to use statements; whether these measures are empirically distinct requires examination.
This article asks how established acceptance determinants, EV-specific evaluations, TI, and CA are conditionally associated with BI and responses to EV-use statements in Germany. Its contribution is an integrated comparison of motivational and system-related correlates, together with tests of how measurement overlap, cost specification, and respondents’ experience affect the interpretation. The complete 17-construct baseline is reported alongside additional reduced, formative-cost and experience-restricted analyses. This comparison identifies which findings persist across specifications and which depend on measurement or sample composition. The proposed readiness layers are a conceptual organising framework, not an established sequence of adoption stages.
2. Theoretical Framework and Hypotheses
BI is a central precursor of use in reasoned-action and technology-acceptance models [2], [14], [15]. FC describes the resources and support necessary to enact an intention [2], [3]. The comprehensive model therefore links BI and FC directly to use behaviour (UB), while also testing Ha, compatibility (Com), personal innovativeness (Pin) and costs (Co) as direct correlates of use. In this study, the variable labelled UB is an agreement scale; its interpretation as actual enactment is examined rather than assumed.
The motivational layer includes PE, EE, SI, PV, HM, FC, Ha, Com, attitude towards environment and technology (AtET), Pin, Co, Re, CA and TI. These variables integrate UTAUT-based, attitudinal, economic, contextual, and trust-related explanations [1], [2], [3], [4]. In a multivariate model, each coefficient represents the association remaining after accounting for the other specified predictors. Co requires particular care: acquisition expense, running costs, and total-cost acceptability are different components, motivating a formative specification. After reversing Co1 and Co2, higher item scores indicate more favourable cost evaluations rather than greater financial burden.
The system-confidence layer is represented by TI, specifically confidence in arriving reliably and on time. Re reflects expectations that EVs function dependably; SA captures protections and institutional or infrastructural support; CA reflects familiarity and perceived fit with everyday use. EE, FC, and PE were also linked to TI. The model relates TI to BI and UB. These paths test conditional associations beyond the conventional intention pathway; the labels do not establish that Re and TI are empirically separable.
The original model formulated H1–H27 as directional relationships and H28–H45 as moderator families. Figure 1 presents the conceptual outcome grouping; Table 1 defines the constructs. Supplementary Table S9 reports all direct paths and their original predicted directions. The hypotheses are not rewritten after observing the results. Additional robustness and interaction analyses are explicitly exploratory and do not constitute new prespecified hypothesis tests.

| Code | Construct | Primary Role |
|---|---|---|
| AtET | Attitude towards environment and technology | Exogenous predictor |
| BI | Behavioural intention | Endogenous outcome |
| CA | Contextual acquaintance | Exogenous predictor |
| Co | Costs (fayourable evaluations) | Exogenous predictor |
| Com | Compatibility | Exogenous predictor |
| EE | Effort expectancy | Exogenous predictor |
| FC | Facilitating conditions | Exogenous predictor |
| Ha | Habit | Exogenous predictor |
| HM | Hedonic motivation | Exogenous predictor |
| PE | Performance expectancy | Exogenous predictor |
| Pin | Personal innovativeness | Exogenous predictor |
| PV | Price value | Exogenous predictor |
| Re | Reliability | Exogenous predictor |
| SA | Structural assurance | Exogenous predictor |
| SI | Social influence | Exogenous predictor |
| TI | Trusting intention in electromobility | Endogenous outcome |
| UB | Use behaviour (use-item endorsement) | Endogenous outcome |
3. Methodology
A quantitative cross-sectional online survey was used. The target population comprised adults whose main residence was in Germany and who held a Class B driving licence. Participants were recruited through a professional online panel. Age, gender, and monthly net household income were controlled through quotas; a minimum of 150 participants with EV driving experience were also sought. The design supports controlled coverage of relevant population segments but does not constitute probability sampling, so claims of national representativeness are avoided [16]. Data collection ran from 3 April to 3 May 2025.
The survey yielded 548 completed cases before quality screening. Response speed and flatlining or other implausible response patterns were assessed. Sixty-three cases were excluded, leaving $n$ = 485. All 485 retained participants provided affirmative consent. Four missing income responses were imputed into underrepresented quota categories-two into EUR 2,001–2,500 and two into EUR 3,501–5,000-before processing in $R$ and SmartPLS, following the documented procedure [17]. The final sample comprised 238 men and 247 women (Table 2). Age categories ranged from 18–24 years ($n$ = 41) to over 75 years ($n$ = 65), and 166 respondents reported EV driving experience.
All 485 respondents supplied answers to the three mandatory UB statements: using an EV for most daily journeys, EV use as a regular part of life, and an EV as the first choice for commuting and errands (Supplementary Table S3). The retained answers were not filtered to EV-experienced respondents. The printed questionnaire uses present-tense wording and does not document an instruction asking non-users to imagine future use. Of the retained sample, 166 answered yes to EV driving experience, 306 answered no, and 13 were unsure. Their intended interpretation cannot be recovered retrospectively; UB is therefore described as use-item endorsement rather than verified current use or intended adoption. The experienced-sample analysis includes only the 166 affirmative responses, not the 13 uncertain responses.
Construct indicators were assessed on seven-point Likert-type scales [18]. The initial Model I contained 50 indicators. Five reverse-worded items-AtET3, Co2, EE1, PE3 and PV1-were excluded during the original indicator assessment. The resulting baseline contained 45 indicators across 17 constructs. Five additional items were considered borderline (Co1, Ha2, HM1, Re3 and TI2) and retained on theoretical or item-to-total-correlation grounds. These selection decisions define the archived baseline; they are not evidence that a formative indicator should be removed for a low reflective loading. Accordingly, the formative-cost analysis restores Co2. Full wording, original selection decisions, and loadings are provided in Supplementary Tables S3–S5.
The empirical design and baseline estimates originated in the underlying dissertation. Partial least squares (PLS) structural equation modeling (SEM) used path weighting, a maximum of 3,000 iterations, a stop criterion of 10$^{-7}$, and initial weights of 1 [19], [20], [21], [22]. The archived SmartPLS outer-model matrix specifies all 17 baseline blocks as reflective, including Co; its automatic weighting setting therefore corresponds to Mode A. Original bootstrapping used 5,000 samples, two-tailed tests, percentile confidence intervals (Cls), and a fixed seed [23]. Blindfolding used an omission distance of 7. Reflective assessment included indicator reliability, Cronbach's alpha, composite reliability (CR), average variance extracted (AVE), and Fornell-Larcker comparisons [24], [25], [26], [27]. Heterotrait-monotrait (HTMT) ratios and variance inflation factors (VIFs) were audited. Harman's single-factor statistic is reported only as a limited diagnostic [28].
| Characteristic | Category | $\boldsymbol{n}$ | Share |
|---|---|---|---|
| Gender | Men | 238 | 49\% |
| Gender | Women | 247 | 51\% |
| Age | 18–24 | 41 | 8\% |
| Age | 25–34 | 69 | 14\% |
| Age | 35–44 | 74 | 15\% |
| Age | 45–54 | 74 | 15\% |
| Age | 55–64 | 94 | 19\% |
| Age | 65–74 | 68 | 14\% |
| Age | $>$75 | 65 | 13\% |
| Income | $<$EUR 1,250 | 64 | 13\% |
| Income | EUR 1,251–2,000 | 87 | 18\% |
| Income | EUR 2,001–2,500 | 63 | 13\% |
| Income | EUR 2,501–3,500 | 93 | 19\% |
| Income | EUR 3,501–5,000 | 91 | 19\% |
| Income | $>$EUR 5,000 | 87 | 18\% |
| EV experience | Yes | 166 | 34\% |
Journal hypothesis assessment requires both $p <$ 0.05 in the original two-tailed bootstrap test and agreement with the originally predicted direction. Coefficient magnitude and $f^2$ describe the size of an association separately; the original $|\beta| \geq$ 0.10 convention is not used to override direction or significance. The direct-path results retain the historical dissertation decision in a separate column and add the journal assessment. This makes the distinction between an unchanged numerical result and its directional interpretation explicit.
Additional analyses were implemented in Python using the same PLS pathweighting algorithm. The 485 $\times$ 45 encoded indicator matrix matched the archived SmartPLS input exactly; all 27 baseline point estimates were reproduced to a maximum absolute difference below 3 $\times$ 10$^{-14}$. Four sensitivity analyses were then specified: (a) removal of Ha and Re and their incident paths, retaining 15 constructs and 39 indicators; (b) Mode B formative Co using all four original cost indicators, with the other 16 blocks reflective, giving 46 indicators; (c) the combined reduced/formative specification; and (d) the original full model restricted to 166 respondents confirming EV driving experience. Each used 5,000 nonparametric case-bootstrap resamples, with the full measurement and structural model re-estimated in every resample. The new analyses used a reproducible seed of 20260911, 95% percentile intervals, and two-sided normal-reference $p$-values derived from the bootstrap standard error [21], [23]. They are sensitivity analyses selected after inspection of the original results.
Removing Ha and Re addresses the two highest HTMT pairs without merging intended action with confidence. Ha contains current or anticipated routine wording close to BI, whereas Re and TI both concern dependable performance. BI and TI remain the theoretically targeted outcomes; removing their overlapping predictors tests coefficient dependence on those blocks. This reduction is not presented as an optimal or independently validated model. For formative Co, acquisition expense, running costs, and total-cost acceptability are treated as components rather than interchangeable manifestations. Co1 and Co2 are reverse-coded as 8 minus the seven-point response, as in the original coding scheme, and Co3–Co4 are retained in their positive orientation. Formative assessment uses content coverage, indicator VIFs, weights, loadings, and bootstrap intervals, rather than alpha, CR, AVE, or reflective HTMT. No independent global cost measure was available for a redundancy analysis.
The new moderation analysis uses a two-stage score approach [22]. PLS construct scores from the complete reflective baseline enter each outcome equation with all of that outcome’s original predictors, a moderator main effect, and one predictor-by-moderator product. The 24 exogenous-to-outcome paths are each tested against all 22 moderators, defining one family of 528 interaction tests. Eleven multi-item moderators are unweighted means of their correctly oriented original items, standardised continuously rather than median-split; eleven categorical moderators retain explicit group contrasts. Binary moderators are centred but not variance-standardised. The full first-stage PLS model and interaction regressions are re-estimated in each of 5,000 case-bootstrap resamples (seed 20260912). Scores use the resample's standardisation. Two-sided $p$-values use the interaction coefficient divided by its bootstrap standard error. Benjamini-Hochberg adjusted $p$-values are the monotone, capped values of 528 $p$($i$)/$i$; Benjamini-Yekutieli additionally multiplies by the harmonic sum of 1/$j$ for $j$ = 1, $\ldots$, 528. The latter is included because these tests are dependent. Main-text examples pass both adjustments at 0.05. New coding excludes uncertain experience responses, unresolved education/employment categories, and four unresolved income responses for the respective moderator only ($n$ = 472–485); Supplementary Table S41 documents every contrast. This analysis provides new estimates, not reconstructed $p$-values for the historical median-split results.
4. Results
Sections 4.1–4.2 report the complete archived baseline, with its original coefficients and bootstrap statistics. Section 4.3 separates historical moderation summaries from the new multiplicity-adjusted analysis. Section 4.4 reports additional estimates from the same retained sample and its experienced subset; these are new robustness results, not replacements for the archived numbers.
All retained baseline loadings were statistically different from zero, but their magnitudes varied. The lowest were Ha2 (0.402), Co1 (0.574), TI2 (0.616), and Re3 (0.670). Baseline CR exceeded 0.81, and AVE exceeded 0.62. Ha had a Cronbach’s $\alpha$ of 0.659. The Fornell-Larcker criterion was met for 16 of 17 constructs; Ha failed because its AVE of 0.622 was lower than its largest squared correlation, 0.68 with BI. Table 3 describes the reflective baseline only. Its Co reliability statistics do not validate the formative specification assessed in Section 4.4.
Construct | Cronbach’s $\boldsymbol{\alpha}$ | Composite Reliability | AVE | Assessment |
|---|---|---|---|---|
AtET | 0.869 | 0.938 | 0.883 | Meets CR and AVE criteria |
BI | 0.918 | 0.948 | 0.860 | Meets CR and AVE criteria |
CA | 0.906 | 0.941 | 0.842 | Meets CR and AVE criteria |
Co | 0.700 | 0.834 | 0.634 | Reflective baseline only; formative assessment in Section 4.4 |
Com | 0.902 | 0.953 | 0.911 | Meets CR and AVE criteria |
EE | 0.906 | 0.955 | 0.914 | Meets CR and AVE criteria |
FC | 0.800 | 0.882 | 0.714 | Meets CR and AVE criteria |
Ha | 0.659 | 0.817 | 0.622 | Caution |
HM | 0.899 | 0.952 | 0.908 | Meets CR and AVE criteria |
PE | 0.922 | 0.963 | 0.928 | Meets CR and AVE criteria |
Pin | 0.888 | 0.930 | 0.816 | Meets CR and AVE criteria |
PV | 0.935 | 0.968 | 0.939 | Meets CR and AVE criteria |
Re | 0.785 | 0.875 | 0.704 | Meets CR and AVE criteria |
SA | 0.882 | 0.927 | 0.809 | Meets CR and AVE criteria |
SI | 0.892 | 0.933 | 0.823 | Meets CR and AVE criteria |
TI | 0.782 | 0.871 | 0.700 | Meets CR and AVE criteria |
UB | 0.939 | 0.961 | 0.891 | Meets CR and AVE criteria |
Ten baseline HTMT ratios exceeded 0.90, including Ha-BI (1.012), TI-Re (0.992), Ha-CA (0.959), and PV-Co (0.954). The baseline inner VIF exceeded 5 for CA predicting BI (5.624); HM predicting BI was close to 5 (4.975). These results challenge empirical separation and motivate the additional analyses in Section 4.4. They prevent treating three theoretical outcome labels or high R² values as proof of distinct readiness dimensions.
The previously reported Harman first-factor share was 41\%, which cannot exclude common-method bias [28]. A supplementary collinearity diagnostic regressed each baseline construct score on all other construct scores. The resulting VIFs ranged from 1.671 to 5.961, including CA 5.961, BI 5.768, and HM 5.036 (Supplementary Table S40). These values flag shared variance but cannot distinguish substantive overlap from method variance; they are not a pass/fail test for common-method bias. Documented procedures included voluntary participation, assurances of anonymous and aggregated processing, and separate construct question blocks with both positive and negative wording. The survey used one respondent, one occasion, and the same seven-point agreement format for all focal constructs. No designated marker variable, temporal separation, or independent behavioural source was available. Method variance and wording effects therefore remain plausible contributors to overlap and unexpected signs.
Adjusted $R^2$ was 0.783 for BI, 0.688 for TI and 0.651 for UB. Reported blindfolding $Q^2$ values were 0.77, 0.68, and 0.57, respectively. The estimated-model standardised root mean square residual (SRMR) was 0.059; the bootstrap mean was 0.038, and the reported 95% upper value was 0.042. Although the SRMR is below a conventional 0.08 heuristic, it exceeds the reported bootstrap upper value. These diagnostics therefore do not jointly establish model fit [29], [30]. $R^2$ quantifies in-sample explained variance, and blindfolding $Q^2$ is not independent holdout validation [31], [32], [33]. No out-of-sample prediction, benchmark comparison, or validated decision-support capability is claimed.
Table 4 reports all 27 baseline direct paths. The strongest association with use-item endorsement was BI ($\beta$ = 0.594, $p <$ 0.001, $f^2$ = 0.243), followed by FC ($\beta$ = 0.235, $p <$ 0.001, $f^2$ = 0.073). Co had a positive conditional coefficient ($\beta$ = 0.157), while Com($\beta$ = -0.185) and TI ($\beta$ = -0.076) had negative coefficients. These three significant results do not support H19, H14, and H26 as originally worded. The Co score is oriented towards favourable evaluations, which also exposes a mismatch between the negative cost-burden hypothesis and its operationalisation. Ha to UB was positive but not significant ($p$ = 0.081). Pin was unrelated to UB after the competing predictors were included.
| H | Path | Expected | $\boldsymbol{\beta}$ | $\boldsymbol{t}$ | $\boldsymbol{p}$ | $\boldsymbol{f^2}$ | Historical Decision | Directional Assessment |
|---|---|---|---|---|---|---|---|---|
| H1 | PE $\rightarrow$ BI | Positive | +0.084 | 1.493 | 0.136 | 0.008 | Rejected | Not sig. |
| H2 | PE $\rightarrow$ TI | Positive | -0.019 | 0.406 | 0.685 | 0.000 | Rejected | Not sig. |
| H3 | EE $\rightarrow$ BI | Positive | -0.068 | 2.098 | 0.036 | 0.013 | Confirmed | Opposite |
| H4 | EE $\rightarrow$ TI | Positive | +0.048 | 1.265 | 0.206 | 0.005 | Rejected | Not sig. |
| H5 | SI $\rightarrow$ BI | Positive | +0.101 | 2.550 | 0.011 | 0.020 | Confirmed | Supported |
| H6 | PV $\rightarrow$ BI | Positive | +0.094 | 1.683 | 0.092 | 0.014 | Rejected | Not sig. |
| H7 | HM $\rightarrow$ BI | Positive | +0.067 | 1.084 | 0.278 | 0.004 | Rejected | Not sig. |
| H8 | FC $\rightarrow$ BI | Positive | -0.039 | 0.914 | 0.361 | 0.003 | Rejected | Not sig. |
| H9 | FC $\rightarrow$ UB | Positive | +0.235 | 6.136 | $<$0.001 | 0.073 | Confirmed | Supported |
| H10 | FC $\rightarrow$ TI | Positive | -0.109 | 2.769 | 0.006 | 0.017 | Confirmed | Opposite |
| H11 | Ha $\rightarrow$ BI | Positive | +0.320 | 6.094 | $<$0.001 | 0.126 | Confirmed | Supported |
| H12 | Ha $\rightarrow$ UB | Positive | +0.134 | 1.746 | 0.081 | 0.013 | Rejected | Not sig. |
| H13 | Com $\rightarrow$ BI | Positive | +0.068 | 1.448 | 0.148 | 0.006 | Rejected | Not sig. |
| H14 | Com $\rightarrow$ UB | Positive | -0.185 | 3.488 | $<$0.001 | 0.033 | Confirmed | Opposite |
| H15 | AtET $\rightarrow$ BI | Positive | +0.012 | 0.282 | 0.778 | 0.000 | Rejected | Not sig. |
| H16 | Pln $\rightarrow$ BI | Positive | +0.199 | 4.583 | $<$0.001 | 0.063 | Confirmed | Supported |
| H17 | Pln $\rightarrow$ UB | Positive | -0.004 | 0.094 | 0.925 | 0.000 | Rejected | Not sig. |
| H18 | Co $\rightarrow$ BI | Negative | +0.017 | 0.312 | 0.755 | 0.000 | Rejected | Not sig. |
| H19 | Co $\rightarrow$ UB | Negative | +0.157 | 3.430 | $<$0.001 | 0.033 | Confirmed | Opposite |
| H20 | Re $\rightarrow$ BI | Positive | +0.068 | 1.504 | 0.133 | 0.006 | Rejected | Not sig. |
| H21 | Re $\rightarrow$ TI | Positive | +0.473 | 9.559 | $<$0.001 | 0.257 | Confirmed | Supported |
| H22 | CA $\rightarrow$ BI | Positive | +0.118 | 2.051 | 0.040 | 0.012 | Confirmed | Supported |
| H23 | CA $\rightarrow$ TI | Positive | +0.377 | 6.320 | $<$0.001 | 0.106 | Confirmed | Supported |
| H24 | SA $\rightarrow$ TI | Positive | +0.101 | 2.052 | 0.040 | 0.009 | Confirmed | Supported |
| H25 | TI $\rightarrow$ BI | Positive | -0.048 | 0.967 | 0.334 | 0.003 | Rejected | Not sig. |
| H26 | TI $\rightarrow$ UB | Positive | -0.076 | 2.048 | 0.041 | 0.009 | Confirmed | Opposite |
| H27 | BI $\rightarrow$ UB | Positive | +0.594 | 9.200 | $<$0.001 | 0.243 | Confirmed | Supported |
BI was most strongly associated with Ha ($\beta$ = 0.320), Pin ($\beta$ = 0.199), CA ($\beta$ = 0.118) and SI ($\beta$ = 0.101). EE had a statistically significant negative coefficient ($\beta$ = -0.068, $p$ = 0.036); this does not support the positive H3, irrespective of its small magnitude. PV was positive but not significant ($\beta$ = 0.094, $p$ = 0.092), so H6 was not supported. PE, HM, FC, Com, AtET, Co, Re, and TI were not significant positive direct correlates in the complete intention equation.
TI was most strongly associated with Re ($\beta$ = 0.473), CA ($\beta$ = 0.377) and SA ($\beta$ = 0.101). FC had a significant negative coefficient ($\beta$ = -0.109, $p$ = 0.006), which does not support the positive H10. EE and PE were not significant. Across the 27 baseline paths, nine met both the significance and predicted direction criteria, five were significant in the opposite direction, and thirteen were not statistically significant.
Six specific indirect effects reached $p <$ 0.05: CA $\rightarrow$ BI $\rightarrow$ UB ($\beta$ = 0.070, $p$ = 0.041) ; EE $\rightarrow$ BI $\rightarrow$ UB ($\beta$ = -0.040, $p$ = 0.042) ; Ha $\rightarrow$ BI $\rightarrow$ UB ($\beta$ = 0.190, $p <$ 0.001) ; Pin $\rightarrow$ BI $\rightarrow$ UB ($\beta$ = 0.118, $p <$ 0.001) ; SI $\rightarrow$ BI $\rightarrow$ UB ($\beta$ = 0.060, $p$ = 0.014); and Re $\rightarrow$ TI $\rightarrow$ UB ($\beta$ = -0.036, $p$ = 0.046). Because the analysis is cross-sectional, these are statistical indirect associations, not temporal mediation effects.
The historical Model II tables contain 528 path-by-moderator records, using group coding or median splits. They report 58 nominally significant results as significance stars, without exact $p$-values. These records and coding are retained as an archival account in Supplementary Tables S11–S33; an exact retrospective false-discovery adjustment cannot be recovered from stars alone. The new interaction analysis provides unrounded numerical $p$-values for a separately specified family of 528 tests. Of these, 132 were nominally significant, 77 had Benjamini-Hochberg adjusted $p <$ 0.05, and 46 also passed Benjamini-Yekutieli adjustment. Selected results meeting the more conservative adjustment include experience $\times$ FC $\rightarrow$ UB (interaction $\beta$ = 0.324, adjusted $p <$ 0.001), charging-time evaluations $\times$ Pin $\rightarrow$ BI ($\beta$ = 0.114, adjusted $p <$ 0.001), and market availability $\times$ FC $\rightarrow$ UB ($\beta$ = 0.115, adjusted $p <$ 0.001). The complete 528 new estimates, bootstrap intervals, and both adjusted $p$-values are supplied in the accompanying comma-separated values (CSV) file. These post hoc associations depend on the specified measurement and interaction models and do not establish validated population segments.
Table 5 summarises selected path estimates across the baseline and additional models. Removing Ha and Re retained positive BI $\rightarrow$ UB ($\beta$ = 0.655,95% CI [0.548,0.757]) and FC $\rightarrow$ UB ($\beta$ = 0.249, [0.179,0.331]) associations. TI $\rightarrow$ UB was smaller, and its interval included zero ($\beta$ = -0.050, [-0.126, 0.029], $p$ = 0.213). CA $\rightarrow$ TI increased from 0.377 to 0.640, while TI $\rightarrow$ BI changed from -0.048 to 0.009 ($p$ = 0.847). Adjusted $R^2$ was 0.755 for BI, 0.611 for TI, and 0.647 for UB. Two reflective HTMT ratios remained above 0.90: Co-PV (0.954) and CA-SA (0.902). Thus, a stable intention/use association coexists with substantial redistribution of shared variance; the reduction does not validate three distinct layers.
| Path/Outcome | Baseline $\boldsymbol{n}$ = 485 | Reduced $\boldsymbol{n}$ = 485 | Formative Co $\boldsymbol{n}$ = 485 | Experienced $\boldsymbol{n}$ = 166 |
|---|---|---|---|---|
| BI $\rightarrow$ UB | 0.594 | 0.655 | 0.594 | 0.638 |
| FC $\rightarrow$ UB | 0.235 | 0.249 | 0.233 | 0.289 |
| Co $\rightarrow$ UB | 0.157 | 0.162 | 0.157 | 0.057 |
| Com $\rightarrow$ UB | -0.185 | -0.172 | -0.184 | -0.100 |
| TI $\rightarrow$ UB | -0.076 | -0.050 | -0.074 | -0.038 |
| CA $\rightarrow$ TI | 0.377 | 0.640 | 0.377 | 0.455 |
| Re $\rightarrow$ Tl | 0.473 | Not estimated | 0.473 | 0.250 |
| FC $\rightarrow$ TI | -0.109 | -0.135 | -0.109 | -0.171 |
| Tl $\rightarrow$ Bl | -0.048 | 0.009 | -0.048 | 0.153 |
| Adjusted $R^2$ BI | 0.783 | 0.755 | 0.783 | 0.763 |
| Adjusted $R^2 Tl$ | 0.688 | 0.611 | 0.688 | 0.703 |
| Adjusted $R^2$ UB | 0.651 | 0.647 | 0.651 | 0.607 |
The formative-cost model yielded Co $\rightarrow$ BI = 0.013 and Co $\rightarrow$ UB = 0.157 (95% CI for the latter [0.069, 0.250]), compared with baseline 0.017 and 0.157. Indicator VIFs were 1.205–2.005. Weights for Co1, Co3 and Co4 were 0.229, 0.489 and 0.510, with 95% intervals excluding zero. The Co2 weight was -0.057, with an interval spanning zero [-0.178, 0.060]; its loading was 0.311 [0.176, 0.438]. Co2 was retained for running-cost content coverage in this specified sensitivity model, despite its weak unique contribution, rather than selected for statistical significance. This remains a measurement limitation. The combined reduced/formative model also retained BI $\rightarrow$ UB = 0.654 and FC $\rightarrow$ UB = 0.247, while TI $\rightarrow$ UB = -0.048 had an interval spanning zero. Full results appear in Supplementary Tables S34–S38.
Among 306 respondents denying EV experience, 181 selected the lowest response on all three UB items, whereas 54 selected 5–7 on at least one item. The 13 uncertain respondents were recorded separately. These answers demonstrate heterogeneity in use-item endorsement, not that non-experienced respondents actually used an EV. In the 166 confirmed-experience cases, BI $\rightarrow$ UB = 0.638 (95% CI [0.443, 0.843]) and FC $\rightarrow$ UB = 0.289 ([0.136, 0.484]) remained positive. Co $\rightarrow$ UB = 0.057, Com $\rightarrow$ UB = -0.100, and TI $\rightarrow$ UB = -0.038 each had intervals including zero. Adjusted $R^2$ values were 0.763, 0.703, and 0.607 for BI, TI, and UB. These estimates are subset sensitivities, not measurement-invariant multi-group comparisons; having driven an EV does not establish ownership or regular use. Supplementary Tables S37 and S39 report all subset paths and item diagnostics.
5. Discussion
The results provide a qualified conceptual account of motivation, system confidence and use-item endorsement. BI and FC remain positively associated with use responses across the baseline, reduced, formative-cost and experienced-sample models. In contrast, the evidence does not establish three empirically distinct readiness layers. Removing Ha and Re changes the allocation of shared variance: CA becomes more strongly associated with both intention and TI, and the small negative TI–UB coefficient is no longer statistically supported. The conceptual grouping remains useful for organising questions, while the claim of a validated three-stage adoption mechanism is unwarranted.
The positive BI–UB association is consistent with acceptance theory [2], [3], [14], 15], but intention and use-item endorsement were measured together. Its baseline $f^2$ of 0.243 is substantially larger than the FC–UB $f^2$ of 0.073. These conditional effect sizes help prioritise further investigation without expressing an intervention effect. The stability of both coefficients across the sensitivity analyses supports their relevance within these survey data. Ha and Pin are associated chiefly with intention; interpretation of Ha is particularly constrained by wording that includes both current and anticipated routine and by HTMT above one with intention.
Re and CA were the leading baseline correlates of TI, with $f^2$ values of 0.257 and 0.106. SA contributed more modestly ($f^2$ = 0.009). The association of CA with system confidence is compatible with the importance of experience and familiarity discussed in EV studies [10], [11], [12], [13], although those studies do not validate the present measures. TI did not add a positive baseline association with either intention or use. Its small negative use coefficient ($f^2$ = 0.009) became non-significant when the overlapping Ha and Re blocks were removed and in the experienced subset. This sensitivity is consistent with an unstable allocation of shared variance, but does not prove suppression or show that trust discourages EV use.
The positive Co coefficient concerns favourable cost evaluations, not evidence that greater expenditure encourages use. The formative-cost model retained the acquisition, running-cost and overall-cost components and yielded a very similar Co $\rightarrow$ UB coefficient. Consequently, changing reflective to formative weighting alone does not explain the observed positive coefficient. Com $\rightarrow$ UB and FC $\rightarrow$ TI remained negative in the reduced model. However, Com $\rightarrow$ UB was not statistically supported in the experienced subset. Because some respondents denying EV experience nevertheless endorsed current-use statements, mixed interpretations of the outcome, shared method variance, and correlated predictors remain plausible explanations. None can be identified as the unique cause from these data.
The independent assessment separates directional hypothesis support from detection of a conditional association. A significant opposite-signed coefficient is evidence against the stated direction, not confirmation of that hypothesis. The historical decisions remain visible in Table\ref{tab4} solely for provenance. Additional models assess sensitivity; their post hoc construction does not justify rewriting the original hypotheses.
Multiplicity adjustment retains a subset of the new interaction associations, including the stronger FC–UB association among respondents confirming EV experience. The existing EV experience literature [10], [11], [12], [13] provides a reason to investigate such differences, but the cross-sectional interaction does not show that acquiring experience changes the slope. Continuous composite moderators preserve more information than median splits; nevertheless, equal item weighting, residual construct overlap, untested measurement invariance, and one-interaction-at-a-time estimation limit interpretation. Results surviving both adjustments are candidates for prespecified replication, not validated targeting rules.
$\bullet$ BI and FC were consistently associated with use-item endorsement. This suggests evaluating both motivational and resource-related measures in future transport studies. The present data do not establish that changing either will increase actual EV adoption.
$\bullet$ The Re and CA associations with TI identify potential evaluation topics, such as predictable charging and service information. Effects of Re interventions require independent behavioural outcomes and prospective comparison groups.
$\bullet$ Test use, route- and charging-planning support, and household-specific cost information are possible ways to investigate CA. Whether these interventions change confidence, intention, or actual use should be tested longitudinally or experimentally; effectiveness is not demonstrated here.
$\bullet$ Moderator findings that survived multiplicity adjustment may inform the design of replication studies. They should not determine population segmentation or resource allocation without external validation and assessment of measurement comparability.
$\bullet$ AtET was not a significant direct correlate of intention after the competing predictors were included. This conditional null result does not establish that environmental attitudes or related interventions are irrelevant.
The design is cross-sectional, self-reported, observational, and based on a non-probability quota panel. UB is an agreement scale with unresolved interpretation among respondents without confirmed EV experience, and the experienced subset is not a verified ownership or regular-use sample. Four income values were allocated to quota categories in the original preparation; the new income interaction analysis instead excludes the four unresolved responses. Original indicator selection and overlap remain concerns: the reduced model addresses the two most extreme HTMT pairs but does not establish discriminant validity for the remaining constructs. The formative Co block lacks an independent global criterion for redundancy analysis; its content coverage and weak Co2 weight require further validation. Collinearity diagnostics cannot separate method variance from substantive covariance. New interaction tests use multiplicity correction and continuous composite scores, but remain post hoc, assume the specified functional form, and do not establish measurement invariance. Significance differences across specifications or subsets are not themselves formal tests of differences between coefficients. Neither $R^2$, blindfolding $Q^2$, nor bootstrap resampling provides independent out-of-sample validation. Replication should use prospective adoption events or observed travel and charging behaviour, revised measures and prespecified interaction tests.
6. Conclusions
In this German quota sample, BI and FC were the most stable correlates of EV-use statement endorsement. Re and CA were associated with TI, while Ha and Pin were prominent baseline correlates of BI. The complete baseline, reduced specification, formative-cost model, and experienced-sample analysis identify a stable core of conditional associations alongside important sensitivity to measurement and respondent experience. They do not establish actual adoption effects or three empirically distinct readiness stages.
The unchanged baseline coefficients are reported together with a directional assessment that does not count significant opposite-signed paths as hypothesis support. Additional analyses qualify the small negative trust-use finding, clarify the favourable-cost orientation, and separate multiplicity-adjusted interaction evidence from historical exploratory summaries. Future research should test the proposed conceptual distinctions using improved measurement, prospective behavioural outcomes and independent predictive or intervention validation.
Informed consent to participate was obtained electronically from all participants included in the study. All 485 retained records contained an affirmative consent response.
Participant-level survey data are not publicly available because of the participant information and applicable data-protection restrictions. Complete de-identified data may be made available confidentially to editors and reviewers upon request, subject to institutional approval and appropriate data-protection safeguards. Aggregate results, item definitions and complete model outputs are provided in the article and Supplementary Material.
The author acknowledges the doctoral supervisors named in the dissertation and the panel provider’s support with participant recruitment.
The empirical study and a broader theoretical treatment formed part of the author’s doctoral dissertation. The present article uses the same retained sample and archived final model. The revision adds identified measurement audits, alternative-model sensitivity analyses, an experience-restricted analysis, and a multiplicity-adjusted interaction analysis. No new data were collected, and the supplementary estimates are distinguished from the original results. Bibliographic identifying details of the dissertation are withheld in this anonymous review version. The dissertation is titled “Not yet ready for electrification?! An empirical study of the barriers and drivers of electric vehicle adoption in Germany”.
The authors declare no conflicts of interest.
ChatGPT by OpenAI (GPT-5, accessed 8–17 September 2026) assisted with manuscript structuring, language editing, extraction scripts, and cross-file consistency checks. During the revision, it also assisted with Python code for supplementary PLS estimation, bootstrap sensitivity analyses, interaction testing, and multiplicity adjustment. The tool was instructed to use only literature sources contained in the dissertation and not to invent references. The original numerical results were checked against the dissertation, the retained raw-data export, and the archived SmartPLS report. The supplementary analyses and their aggregate outputs were independently reviewed by the author. The author remains fully responsible for the accuracy, originality and compliance of the manuscript.
