**3. Methodology**

In this paper, we carried out a systematic literature review to provide a complete, exhaustive summary of relevant literature addressing the role of agritourism in supporting the sustainable development of rural areas. Following the principles and the process of a Systematic Literature Review proposed by Denyer and Tranfield [47], our research methodology was organized in three phases: papers location and selection, papers analysis and classification, the definition of themes.

#### *3.1. Papers Location and Selection*

We selected Scopus as the scientific database to perform our search. Scopus delivers a comprehensive overview of the world's research output in our domain of reference and it can handle advanced queries. Elsevier Scopus is a citation database containing more than 50 million records from around 5000 publishers, for publications in peer-reviewed journals, omitting books, book chapters, discussion papers, and non-refereed publications.

Based on the prior experience of the review team and previous literature on Agritourism Studies, an initial set of keywords was defined. First, we have considered synonyms of "Agritourism" as search items. We initialized a List A of search keywords with English terms related to agritourism based activities (including "agritourism", "agrotourism", "agri-tourism", "agro-tourism", "Farm based tourism", "Farm tourism", "Rural tourism" [21]. We also initialized a List S of sustainability-related terms (including the terms "sustainability", "sustainable", "development", and related synonymous).

The keywords were constructed into search strings, in order to administer the search to the Scopus scientific database. The following search string was structured: *The search must contain at least one keyword of the Agritourism Domain (A) and one keyword from the Sustainability Domain (S).* Through this procedure, we identified an initial sample of 212 papers. We manually analyzed metadata (authors, title, source, and year) in order to detect new keywords to add to the lists *A* and *S* respectively. We iteratively performed this phase until no newer keywords or new papers were found. Through this procedure, we identified a list of 405 scientific works.

After, the objective of the process was to select papers with high scientific quality. As a consequence, we have kept only those articles in the sample that were published in academic journals, removing conference proceedings as source type. A total amount of 325 entries is indexed as journal papers. In order to assess the quality of scientific publications, we selected only journals with impact factors indexed in the Thompson Reuters Journal Citation Reports. At the end of this cycle, we obtained the final set P of 192 papers (published in 66 journals) to be analyzed. In Figure 1, a graphical representation of sampled papers in P clustered by publication year is shown.

#### *3.2. Papers Analysis and Classification*

The set P was analyzed through quantitative techniques with the aim to identify relevant topics in the investigated knowledge domain and to group them in macro themes. In particular, we applied a text mining solution based on the Latent Dirichlet Allocation (LDA) technique [48]. This allowed us to build a Document–Term Matrix, that is, a matrix describing the relative presence of keywords in a

corpus of documents. The LDA technique leverages Bayesian Estimation Techniques to infer a vector representing the degree of membership (topic proportion) of each document to each topic. The LDA technique takes as input the documents to be analyzed (192 papers) and the number of topics *k* to be extracted. As suggested by Chang et al. [49] and Blei [48], we selected *k* using a reasonable practice of evaluation among alternative values in such a way that the interpretation of the machine-generated model results becomes as easy as possible from the point of view of a human reader. We have evaluated multiple outputs of the LDA with *k* ranging from 2 to 30 and have consensually agreed that the most meaningful set of topics is reached with *k* = 10.

**Figure 1.** Papers in P by publication year.

The LDA procedure gave as output (Table 2) a group of significant keywords associated with each topic and the document–term matrix.


**Table 2.** Keywords grouped by topics (LDA output).

#### *3.3. Definition of Topics*

In order to deduce meaningful descriptions of each topic, we implemented a human-based review of a restricted, representative, and relevant subset Q ⊆ P of high-quality papers. Q consisted of those articles in P that match ALL the following criteria [50]:

Were published in academic journals ranked at a "C" level or higher of the German Academic Association for Business Ranking or equivalent values of ISI Impact Factor (IF >= 0.7) or ABS Academic Journal Quality Guide (higher than 2◦ category).

Have a topic proportion (TP) value of 0.25 or higher.

Papers included in the subset Q are 34 and are listed in Appendix A.

The 10 topics detected with the LDA procedure are named and discussed according to the 34 selected papers.
