**3. Methodology**

#### *3.1. Sample*

A survey was employed to collect data to test our hypotheses. We chose the customers in Chinese hospitality with service robots as our respondents. There are two reasons. First, the early outbreak of COVID-19 caused unprecedented damage to hospitality industries in China. Second, several service robots have been introduced into hotels and restaurants in China, which provide services, such as ordering and delivering dishes, without social contact.

All multi-item constructs with existing scales were adapted from the public health, marketing, and tourism literature. Validity and reliability were ensured by back-translating the measures. Before our formal survey, we invited three professors and three Ph.D. students to examine our items. Based on their advice, we revised the items and kept the language of the items clear, specific, and simple. We also conducted a pretest and collected 53 surveys. Factor analysis was used to test the reliability and validity of the measurements to ensure the effectiveness of the follow-up survey further.

In order to guarantee the confidentiality and quality of data, we invited our respondents randomly who received service from service robots in the hospitality industry. Every responder spent about 4 min answering this survey. All respondents received RMB 4 as the payment for participating in our survey.

We invited the respondents randomly to participate in our survey through Wenjuanxing (www.wjx.cn) (accessed on 1 September 2020), the biggest survey platform in China. A total of 647 customers were invited from September to October 2020, when the outbreak of COVID-19 was largely controlled in China. In total, 36 respondents were removed because of failing to pass the attention tests or taking an unreasonably short time (i.e., less than two minutes), and 22 respondents were discarded because of incomplete data (>25% of answers omitted). In total, 589 valid respondents were used for our data analyses. The demographic profile of the sample is shown in Table 1. Approximately 51.4% of respondents were female, whereas 48.6% were male. The majority of respondents were aged 18 to 39 (97.1%) and had a bachelor's degree (68.8%). In addition, a plurality (38.5%) of respondents had yearly income between 5000 and 100,000 RMB. The second most common was an income between 10,000 and 20,000 RMB (24.7%), and third most common was less than 5000 RMB (20%).


**Table 1.** Demographic profile of the sample (*n* = 589).

### *3.2. Measures*

We measured all multi-item constructs with existing scales drawn from the tourism, marketing, and healthcare literature (Table 2), using a seven-point Likert format (1 = strongly disagree/not at all; 7 = strongly agree/extremely) for all measures except attitude towards risk. Specifically, the perceived risk of COVID-19 was evaluated by two items adopted from Kim and Lee [69] and Gidengil et al. [68]. Customer–robot engagement was measured in terms of attention (four items), enthusiasm (four items), and interaction (four items), and this methodology was adopted from So et al. [40]. Social distancing was assessed by two items adopted from Aron [70]. Health consciousness was evaluated by four items adopted from Gineikiene et al. [71].


#### **Table 2.** Measured items and CFA results.

Notes. α, Cronbach's α; CR, composite reliability; AVE, average variance extracted.

We also measured risk attitude, which was assessed by five items adopted from Forlani and Mullins [56], i.e., please answer the following 5 items by circling the alternative ("a" or "b") you would feel most comfortable with. 1. (a) an 80% chance of winning \$400, or (b) receiving \$320 for sure; 2. (a) receiving \$300 for sure, or (b) a 20% chance of winning \$1500; 3. (a) a 90% chance of winning \$200, or (b) receiving \$180 for sure; 4. (a) receiving \$160 for sure, or (b) a 10% chance of winning \$1600; 5. (a) a 50% chance of winning \$500, or (b) receiving \$250 for sure.

Finally, the technology adoption model (TAM) literature deems that the customer behavior related to new technology is influenced by customer-level factors regarding the perception of the technology, such as perceived usefulness and perceived ease of use [72–74]. Therefore, we controlled for these variables to minimize omitted variable bias and account for factors that explained significant variance in customer–robot engagement. We measured perceived usefulness (four items) and perceived ease of use (four items) with scales adapted from Davis [72] and Agarwal and Karahanna [75].

#### *3.3. Data Analysis*

The marker-variable technique [76] was employed to statistically identify the threat of common method variance (CMV). Confirmatory factor analysis (CFA) was performed to evaluate the reliability and validity, and structural equation modeling (SEM) was used to examine the direct hypotheses. The bootstrapping approach based on PROCESS macro [77] was used for the mediation analysis and moderation analysis. These data analyses were conducted using SPSS 24.0 (IBM, New York, NY, USA) and Amos 24.0 (IBM, New York, NY, USA).

#### **4. Results**

#### *4.1. Reliability and Validity*

Table 2 shows the results of the CFA. The CFA resulted in good fit to the data (χ2/df = 2.71, GFI = 0.904, NFI = 0.980, CFI = 0.987, RMSEA = 0.054). The composite reliability was satisfactory as well because the scores for all constructs ranged from 0.87 to 0.97, exceeding the threshold of 0.70 [78]. Our instrument demonstrated convergent validity, as all factor loadings were between 0.70 and 0.97, greater than the recommended minimum value of 0.50; the average variance extracted (AVE) for each construct ranged from 0.62 to 0.94, greater than the threshold of 0.50 [79].

The results in Table 3 indicated strong discriminant validity, as the square roots of the AVEs were greater than the corresponding correlation coefficients between the factors [80].

**Table 3.** Descriptive statistics and correlation matrix of variables.


Note. The values in the lower diagonal of the table present the correlations between the constructs, while the values in the diagonal of the table present the square roots of the AVEs of the construct. We take education level as a marker variable 3. *n* = 589; \*\* *p* < 0.01. Bold: the square roots of the AVE for each construct.

### *4.2. Common Method Biases*

In addition to program control, statistical controls were employed to assess the common method biases [81]. We adopted the marker-variable technique [76] to evaluate the common method biases and took education level as a marker variable. As shown in Table 3, the correlation coefficients between education level and other variables were small and not significant (*p* > 0.05). Thus, the common method biases of the current study were not serious.

Consistent with Schwepker's study [82], we used the CFA technique to analyze potential common method biases using three steps. First, all items point to the latent variables measured by them, and carry out an eight-factor model CFA, which is called model C1. Second, all items point to the common method biases variable and carry out a one-factor model CFA, which is called model C2. Third, we compared the changes of model fit indexes of model C1 and model C2 to see if a significant difference emerged. As shown in Table 4, the model fit of model C2 was poor, and the model fix of model C1 improved fit significantly (Δχ<sup>2</sup> = 3735.12, Δdf = 28, *p* < 0.001), which means that the common method biases were not serious.


