Technology-based interventions to promote healthy behavior have been an emerging field of research for the past 10 to 20 years [1,2]. Among the healthy behaviors that are promoted, technology-based physical activity has brought to light new opportunities for public health interventions [3]. Several studies have evaluated the prospects of technologies such as exergames and active videogames [4], virtual reality [5], wearable physical activity trackers [6], website-delivered physical activity interventions [7], mobile phone apps [8,9], and video conferencing [10]. Electronic health (eHealth) physical activity promotion technologies have been designed not only for healthy adults [8], but also for vulnerable people in health care contexts, including cancer survivors [11,12], those in need of treatment for overweight and obesity [13,14] or cardiac rehabilitation [15], and older people [16,17]. All these technologies are popular (ie, positive assessment by many), with promising and reported positive outcomes [11-17]. However, the phenomena of usage cessation and losses to follow-up (ie, the law of attrition) are common problems [18]. Moreover, only 45% of the technology interventions are theoretically based, and the acceptability mechanisms have been insufficiently studied [1,19].

Acceptability and acceptance theories have provided interesting insights [20,21] into why some tools are chosen, accepted, and used more than others. The literature on acceptability and acceptance has emerged in different fields (eg, ergonomics, social psychology, management science) [22]. However, the acceptability and acceptance concepts have not been formally defined [23], and the distinction between the two has been based on the temporality of usage [21]. Acceptability refers to the a priori perceived use, whereas acceptance refers to the actual use [22]. Based on the proposed definitions [23], we define acceptability in this paper as the psychological antecedents of the behavioral intention to use technology without experience of the system. In the field of social psychology, the theory of reasoned action and the theory of planned behavior (TPB) [24,25] hold that attitudes and representations determine behavioral intention and real behavior [22]. These theories are the foundation for the technology acceptance model (TAM) [26], the most frequently used model in health informatics [27]. Nevertheless, several extensions have been proposed—TAM2 [28] and TAM3 [29]—revealing that the original TAM was not optimal in eHealth [27]. The unified theory of acceptance and use of technology (UTAUT), particularly its extension, UTAUT2, is today the most complete model, as it combines theory of reasoned action, TAM, a motivational model, TPB, a combined TPB and TAM, a model of personal computer use, diffusion of innovations theory, and social cognitive theory [29-31]. The UTAUT2 comprises 26 items divided into 8 constructs: Performance Expectancy (PE, 3 items), Effort Expectancy (EE, 4 items), Social Influence (SI, 3 items), Facilitating Conditions (FC, 4 items), Hedonic Motivation (HM, 3 items), Price Value (PV, 3 items), Habit (HT, 3 items), and Behavioral Intention (BI, 3 items).

Acceptability assessments in several studies in eHealth contexts have been based on tools without or with only partially demonstrated psychometric qualities [32,33]. However, to ensure the quality of future research, it is necessary to have scales with validated psychometric qualities [34]. The UTAUT model can be considered as a relevant framework for assessing the acceptability of eHealth, particularly for patient-centered assessment [20]. Yet, for theoretically based technologies in the health and wellness field, only 2 studies have been based on the UTAUT model [1], and 2 were conducted in France [27]. The scarcity French studies [1,27] may be due to the absence of validated scales in French to evaluate acceptability. To our knowledge, validated scales in the French language have been based on the TAM model [35] or on other definitions of acceptability in which the concept of acceptability is merged with the definition of usability [22,36]. The UTAUT2 [31] has already been translated into other languages (eg, German [37], Turkish [38], and Portuguese [39]) and has proven its validity; however, the psychometric qualities of the scales have been only partially demonstrated.

The aims of this study were to adapt the UTAUT2 [31] to the eHealth context and to validate this version, which we called the eHealth acceptability scale, in French-speaking samples. This validated tool would allow for the development of further studies in this field.

In line with the guidelines for scale validation from Vallerand et al [40] and Boateng et al [41], we conducted successive stages: translation and adaptation, dimensionality tests, reliability tests, and construct validity tests.

We managed the administration of the scale using LimeSurvey CE, version 2.06+ (LimeSurvey CE). We distributed the scale link by email or face-to-face at the end of students’ courses. We also distributed the link by email to health professionals and adults registered for adapted physical activity. In addition, we posted the link online via social media networks.

We recruited participants in various categories of the general population: students (studying sports, psychology, management, and computer science at a university in the South of France), health professionals (in the field of obesity), and adults with health conditions (ie, diabetes, cardiovascular disease, and obesity) registered for adapted physical activity sessions. To conduct the successive stages of validation, we divided the participants into 5 sample groups.

This study was approved by the French National Commission for Information Technology and Civil Liberties (authorization no: UCA-E18-00), and all participants gave their electronic consent before participation.

We performed all statistical analyses with IBM SPSS version 23 (IBM Corporation) and IBM SPSS Amos version 23 (IBM Corporation). We examined the missing data trends. The cutoff for an acceptable percentage of missing data has not been well established in the literature [46]. However, 5% is considered inconsequential [47], and the risk of statistical bias is considered when the rate is higher than 10% [48]. In our global sample, the missing rate was under 10%. For structural equation modeling, the maximum likelihood estimation and the multiple imputation for handling missing data presented close to equivalent good properties [49]. We applied the maximum likelihood estimations (considered the standard for structural equation models [46]) to be used in Amos v23.

To conduct the successive stages of validation, we divided the participants into 5 samples. Samples 1 (n=20), 2 (n=10), 3 (n=227), and 4 (n=319) were independent groups, and sample 5 (n=61) was a subgroup of sample 4. The global sample included 576 volunteers, mainly students (n=349, 60.6%), with 53.5% men (n=303) and a mean age of 26.8 (SD 10.9) years. We excluded 18 volunteers because they had never used eHealth technology. Table 1 presents detailed participant characteristics for each sample.

We performed the first content clarity analysis on sample 1 (n=20), which revealed an acceptable clarity score (mean range from 4.40 to 7.00; mean 6.22, SD 0.71). Only 3 items (ie, EE2, SI2, and SI3) obtained a score of less than 5, which we rephrased according to participants’ suggestions. We performed a second content clarity analysis on sample 2 (n=10) regarding the 3 rephrased items. The clarity score increased for 2 items (SI2: mean range 4.40 to 6.20; SI3: mean range 4.90 to 6.20) but decreased for the third (EE2: mean range 4.65 to 3.30). We retained the 2 items with increased clarity scores in their rephrased form and the item with a decreased clarity score in its original translated wording. Multimedia Appendix 1 shows the preliminary pool of 23 items with their mean clarity scores.

We conducted a first maximum likelihood CFA on sample 3 (n=227) with the 23-item and 7-factor model. Standardized factor loadings were all higher than the recommended value of .50 [62], except for item FC4, for which the factor loading was .27. As a result, we removed item FC4. We conducted a second CFA using sample 4 (n=319) with the 22-item (ie, without FC4 item) and 7-factor correlated model (χ²₁₈₈=471.80, P<.001). Fit indexes were as follows: χ²/df=2.51, CFI=.94, TLI=.91, and RMSEA=.069 (90% CI 0.061-0.077), revealing an acceptable model fit, with good standardized factor loadings for all items (ie, ≥0.63).

Based on the recommendations of Myers et al [50], we examined several models to assess the dimensionality of the scale, using samples 3 and 4 merged (n=546). Table 2 presents model fit indexes for each model. First, the unidimensional model did not present good fit indexes. Second, the first-order all-factor correlated model presented good fit indexes, as previously demonstrated. Third, the hierarchical second-order model presented acceptable fit indexes. Fourth, the bifactor confirmatory model presented the best fit indexes: χ²₁₇₃=434.86 (P<.001), χ²/df=2.51, CFI=.97, TLI=.95, and RMSEA=.053 (90% CI .047-.059). These results sustained the possibility of extracting a global acceptability score from the scale.

We tested the invariance of the scale factorial structure following Gregorich’s recommendations [55], with samples 3 and 4 merged (n=535; 11 without sex information). The invariance tests were based on multigroup comparisons: female group (n=249) and male group (n=286). Each group presented good fit indexes for the CFA model (Table 3). We tested invariance in the 22-item 7-factor correlated model. Dimensional, metric, strong, and strict models presented good fit indexes (ie, CFI, TLI, and RMSEA) with significant chi-square P values (ie, P<.001). No significant differences between models were reported, except for the strict model (Table 3). A partial strict model, unconstrained for error of measurement for items EE2 and HM1, showed good fit indexes with no significant difference from the strong model.

Cronbach alphas ranged from .77 to .95 in samples 3 and 4 (n=546) for the 7 eHealth acceptability factors (ie, α_PE=.84; α_EE=.88; α_SI=.95; α_FC=.78; α_HM=.92; α_PV=.86; α_HT=.77) and were .93 for BI and .60 for usage.

We measured test-retest reliability in sample 5 (n=61) twice with an acceptable interval of 4 weeks [59]. Table 4 presents the results of the ICC and t tests. The ICCs for each construct ranged from .62 to .88. Thus, there were no significant differences in the t tests from time 1 to time 2.

We assessed convergent validity using Pearson correlation coefficients in sample 4 (n=319). BI was related to the eHealth acceptability subscales in the expected directions, even though the effect sizes were small for EE, FC, and PV. Usage was related to HT as expected, but not with FC. PEOU, SN, and PBC were related to the eHealth acceptability subscales in the expected directions. We observed additional significant correlation coefficients between constructs. Table 5 presents the complete matrix.

This study aimed to fill a gap in the acceptability literature by providing a validated scale based on the UTAUT2 model [31] that would be suitable for eHealth contexts in French-speaking samples. The eHealth acceptability scale comprised 22 items divided into 7 subscales: PE (3 items), EE (4 items), SI (3 items), FC (3 items), HM (3 items), PV (3 items), and HT (3 items).

The dimensionality tests showed that the first-order all-factor correlated model and the bifactor confirmatory model had good fit indexes. The results confirmed the possibility of both using the subscales individually and extracting a global score of acceptability. The internal consistency evaluated by Cronbach alphas was considered satisfactory [57] and thus was confirmed. The ICCs for each subscale were above .60 and there were no significant differences in the t test over a 4-week period. These results demonstrated the temporal stability of the eHealth acceptability scale. Although it might seem important to attain strict factorial invariance, practical experience suggests that this is almost unachievable [55]. The partial strict factorial invariance pointed to the sex invariance in our analysis. This conclusion was one of the major findings, as it confirms that the eHealth acceptability scale can be used in male and female French-speaking samples.

Convergent validity assessments showed that subscales of the eHealth acceptability scale were significantly positively related to BI, usage, and the PEOU construct from the TAM [26], and significantly positively related to the SN and PCB constructs from the TPB [44]. These preliminary results need to be confirmed in future studies.

The strength of this scale validation was that it followed all the steps recommended by Boateng [41].

Some limitations must nevertheless be acknowledged. One of these limitations, as in all rating scales, is the self-reported nature of the responses, which can be biased based on social desirability [63]. Another limitation is the homogeneity of the samples we used. Most of the participants were young and students. Few participants with low socioeconomic status were included, which limited generalizability. In populations that are not familiar with eHealth tools, it may be necessary to deliver specific education, notably by providing a description of the terms used. In addition, given the age distribution of our sample (ie, centered on ages 18-34 years), we could not test the age invariance. Furthermore, the simultaneous modification of the language (ie, into French) and the context (ie, adaptation to eHealth) may have led to potential interactions and is a limitation. The study would probably have been stronger if we had validated a French-language instrument before changing the context.

Compared with the English-language UTAUT2 model [31], the French eHealth acceptability scale comprised 22 items divided into 7 subscales: PE (3 items), EE (4 items), SI (3 items), FC (3 items), HM (3 items), PV (3 items), and HT (3 items), according to our analyses. We removed item FC4 for its inconsistency; the low loading was also observed to a lesser extent in the German translation [37], although not removed. The sex invariance demonstrated in our analysis was not provided in the original version [31], nor in the other translations [37,38].

In future studies, it will be necessary to test the constructs of the eHealth acceptability scale, which was based on the UTAUT2 model, in French samples and to estimate the explained variance in BI and usage. In addition, evaluation of age invariance will be necessary. The suggested adaptation to the eHealth context could also be replicated in other languages. Specifically, an English validation of the eHealth acceptability scale would be of interest in order to provide a common tool across French- and English-speaking samples. This scale could be used in future research to identify acceptability correlates in different contexts. It could also be used in clinical practice before implementing a new technology in health care or in the field of marketing as new technologies are developed.

We designed a 22-item French-language eHealth acceptability scale, divided into 7 subscales. The scale demonstrated good psychometric qualities (ie, reliability, dimensionality, validity). With this preliminary validation, the scale can be used with men and women to assess the acceptability of eHealth technology in French-speaking samples and offers promising avenues in research, clinical practice, and marketing.

Multimedia Appendix 1

Preliminary version of the eHealth acceptability scale and adapted items of the UTAUT2.