*3.2. Experimental Design*

A choice experiment (CE) approach was used to evaluate urban consumers' WTP for the attributes of enhanced mandatory labelling of the GM soybean oil. The CE model relies on random utility theory and factor value theory, and they indicate that the utility is from the attributes possessed by the item rather than item itself [36]. As for the enhanced mandatory GM food labelling, the combination of the labelling attributes and choice scenarios are formulated in the CE. Specifically, the consumer can obtain the utility *vk* from the *k-th* labelling attribute, and the utility *V*, obtained from enhanced mandatory GM food labelling, equals to the sum of the utility *vk*(k = 1, 2, . . . , K).

$$V = \lambda\_1 \upsilon\_1 + \lambda\_2 \upsilon\_2 + \dots + \lambda\_k \upsilon\_k \tag{1}$$

where *λk* is the unknown parameter, referring to individual's preference for utility *vk*. Consumer *i* must evaluate the utility *Uimn* from the enhanced mandatory GM food labelling associated with the alternative *m* = 1, 2, ... , *M* in the *n-th* choice set. Within a given group of alternatives relating to a choice set, the consumer selects the utility-maximizing alternative. *Uimn* is a random variable that can be expressed as:

$$M\_{imn} = V\_{imn}a + \mu\_{mn} \tag{2}$$

where *αmn* refers to the estimated parameter vector. *μmn* is the random disturbance term. The vector *Vimn* means sum of the utility obtained from the mandatory labelling attribute and paymen<sup>t</sup> vehicle associated with the alternative *m* = 1, 2, ... , *M* in the *n-th* choice set [36].

This study adopts a choice experiment model on the GM soybean oil sales and labelling in China Including Regulations on Administration of Agricultural Genetically Modified Organisms Safety and Administrative Measures for Agricultural GMOs Labeling issued by the Ministry of Agriculture and Rural Affairs of P. R. China. This study follows

the relevant literature [26,27,35,37–40], and the representative GM organism safety management policies. The CE model contained three labelling attributes and the paymen<sup>t</sup> vehicle (Table 1), that is, allergen presence labelling, nutrient and compositional change labelling, traceability codes, and price.



a: Seven texts drafted by the Codex Alimentarius Commission (CAC) on Food Identification Subcommittee in the past decades. b: Food and Drugs Act, Food and Drug Regulations, and Consumer Packaging and Labeling Act by Canada; c: Draft Guidance for Industry: Voluntary Labeling Indicating Whether Foods Have or Have Not Been Developed Using Bioengineering by US; d: Traceability and Labelling Regulation and Regulation (EC) on Novel Foods and Novel Food Ingredients issued by EU; e: Genetically Modified Organisms (Traceability and Labelling) (England) Regulation 2004.312; RMB 6.80 = USD 1.

> Each of the first three attributes include two levels (disclosure or nondisclosure), and the price includes three levels. Thus, there were 24 possible combinations in total. We can constitute 276 CE scenarios by pairing those combinations. After eliminating both the overlapping and theoretically contradictory CE scenarios, we conduct the screening experiment, and obtain twelve CE scenarios. These scenarios are randomly divided into two groups, with each contain six CE scenarios. A sample CE scenario is shown in (Figure 1).


**Figure 1.** Example choice scenario.

In the experiment, GM soybean oil is selected as the analysis unit for four reasons. Firstly, the studies showed that Chinese consumers preferred to accept foods derived from bioengineered food rather than directly edible GM foods like GM soybean oil [41]. Therefore, it can be inferred that consumer demand for labelling information for the directly edible GM foods is the most urgent. Secondly, according to the GM organism safety certificates for both commercial planting and GM organisms imported as raw materials approved by the Ministry of Agriculture and Rural Affairs of China, currently there are

only three kinds of directly edible GM foods on the Chinese market; that is, locally grown GM papaya, GM soybean oil and GM canola oil made from imported GM soybeans or rapeseed. The GM oil is labeled "The processing material is GM soybeans or rapeseeds." Thirdly, soybean oil is not only the most daily consumption edible oil in the majority of Chinese cities, but also the most popular with food processing enterprises and the catering industry. Fourthly, there are a wide range of alternatives to GM soybean oil and non-GM soybean oil available on the Chinese market, such as non-GM peanut oil, non-GM corn oil, non-GM sunflower oil, non-GM canola oil, non-GM rapeseed oil, and many kinds of oil blends. There are a variety of brands of edible oil in the Chinese market. Some only sell non-GM soybean oil (such as the Northeast soybean oil, Xinheshun and Qiansuihao, etc.), while some only sell GM soybean oil (such as Fortune, Jinlongyu, YuanBao, Fivelakes, etc.). Most enterprises produce only one or a few of the edible oils (such as only producing soybean oil, peanut oil, corn oil, sunflower oil, olive oil, rapeseed oil, or blended oil). In order to ensure the GM soybean oil, non-GM soybean oil and oil blends are identical in the brand, capacity and other aspects, this section uses "X" brand edible oil as the experimental unit, which is one of the top ten well-known brands of edible oil in China. Except for peanut oil, all other soybean oil substitutes are supplied in a 5-liter jug.

The prices of all kinds of "X" brand 5-liter edible oil are shown in (Table S1), which was presented to the respondents in the experiment. In this study, the price of 5L GM soybean oil (RMB 45.8, 1 USD=6.80 RMB) is set as the lower limit, and the price of 5L non-GM soybean oil (RMB 66.8) is set as the upper limit. According to the principle of isometric and rounding, the price is set at three levels: RMB 46, RMB 53, and RMB 60.

Additionally, the CE model follows a "randomized design" developed by Sawtooth Software, Inc. [42]. Compared to the fixed design, the randomized design can eliminate order and psychological context effects [43]. Additionally, the randomized designs are more efficient in asymmetric choice experiments when not all attributes have equal levels [44].
