Data extraction sheet and included-studies list for: Software Engineering Decision-Making in Low-Resource Environments — A Scoping Review
Description
Searched six databases (ACM Digital Library, Google Scholar, Scopus, IEEE Xplore, ScienceDirect, EBSCO) in November 2025 using the same three-block Boolean query, adjusted per database syntax: ("low-resource" OR "resource-constrained" OR "limited resource") AND ("software development" OR "application design" OR "software developer") AND ("software architecture" OR "design decision" OR "software design" OR "design choice") Retrieved 1,677 records (ACM: 1,000; Google Scholar: 410; Scopus: 240; IEEE Xplore: 16; ScienceDirect: 6; EBSCO: 5). Removed 135 duplicates, leaving 1,542 records for title-and-abstract screening. Two reviewers independently screened all 1,542 titles/abstracts, excluding non-English records, records whose primary contribution was ML/AI model performance without SE-decision discussion, and records using "low-resource" only in the NLP data-scarcity sense, alongside a broad check for low-resource setting, a concrete SE decision, and constraint-driven motivation. Reviewers calibrated on a pilot of 30 records (Cohen's kappa = 0.75), then resolved disagreements by discussion. Retained 144 records for full-text assessment. Two reviewers independently assessed all 144 full texts against the complete criteria, resolving disagreements by consensus. Included studies had to be conducted in or explicitly designed for a low-resource environment; document at least one concrete SE decision demonstrably motivated by a specific low-resource constraint; present primary evidence (a design artefact, architecture, framework, or empirical study); and be peer-reviewed or a substantive grey literature report with identifiable authorship. Excluded studies addressed low-resource contexts only through infrastructure, policy, or socioeconomic analysis; lacked sufficient design detail for extraction; were themselves a review or meta-analysis; or had an unobtainable full text. This excluded 7 unretrievable records and 105 on criteria (mainly insufficient design detail or no established constraint-decision link), leaving 32 included studies. Developed a structured extraction form, piloted on 5 studies, then applied to all 32 included studies. For each study, recorded: authors/year, application domain, a verbatim quotation of the low-resource constraint(s) with page reference, a verbatim quotation of the documented SE decision with page reference, and any reported outcomes/evaluation. Coded the extracted material inductively: two coders independently assigned a decision category, constraint-to-decision rationale, principle code, pattern code, and trade-off code to each study, resolving disagreements through discussion against the original text. Each code remains traceable to its verbatim evidence quote.
Files
Steps to reproduce
Searched six databases (ACM Digital Library, Google Scholar, Scopus, IEEE Xplore, ScienceDirect, EBSCO) in November 2025 using the same three-block Boolean query across all databases, adjusted only for each database's query syntax: ("low-resource" OR "resource-constrained" OR "limited resource") AND ("software development" OR "application design" OR "software developer") AND ("software architecture" OR "design decision" OR "software design" OR "design choice") Retrieved 1,677 records (ACM: 1,000; Google Scholar: 410; Scopus: 240; IEEE Xplore: 16; ScienceDirect: 6; EBSCO: 5). Removed 135 duplicates, leaving 1,542 records for title-and-abstract screening. Two reviewers independently screened all 1,542 titles/abstracts against a subset of the inclusion/exclusion criteria (language, ML/AI focus, NLP-specific "low-resource" usage, and a broad relevance check), first calibrating on a pilot sample of 30 records (Cohen's κ = 0.75). Compared decisions and resolved disagreements by discussion. Retained 144 records for full-text assessment. Two reviewers independently assessed all 144 full texts against the complete inclusion/exclusion criteria set (IC1–IC5, EC1–EC7), resolving disagreements by consensus. Excluded 7 unretrievable records and 105 on criteria (primarily insufficient design documentation or no established constraint–decision link), leaving 32 included studies. Developed a structured data extraction form, piloted and refined on 5 studies, then applied to all 32 included studies. For each study, recorded: authors/year, application domain, a verbatim quotation of the low-resource constraint(s) with page reference, a verbatim quotation of the documented SE decision with page reference, and any reported outcomes/evaluation. Coded the extracted material inductively: two coders independently assigned a decision category, constraint-to-decision rationale, principle code, pattern code, and trade-off code to each study, resolving disagreements through discussion with reference to the original text. Each code remains traceable to its verbatim evidence quote.
Institutions
- Makerere UniversityCentral Region, Kampala