Digital Visibility, Algorithmic Misclassification, and Welfare Exclusion Among Informal Workers: A Mixed-Methods Study in Indonesia

Abstract

Algorithmic governance is reshaping how states allocate social protection, yet the distributive consequences of automated welfare targeting in the Global South remain poorly understood. Grounded in street-level bureaucracy and procedural-justice theory, this study examined how digital transaction visibility and employment informality predict false-negative welfare exclusion, whether perceived algorithmic misclassification mediates these effects, and whether algorithmic procedural justice moderates them. An explanatory-sequential mixed-methods design combined an audit of 2,500 automated eligibility decisions and a cross-sectional survey of 640 household heads served by a public organization in Palembang, South Sumatera, Indonesia (response rate 84.2%), with phenomenological interviews of false-negative cases. All scales were reliable (Cronbach's alpha .79-.88). The audit showed aggregate accuracy of 72.3% but a 23.4% false-negative rate, rising to 55.1% among undocumented households. Hierarchical regression revealed that employment informality (beta=0.28), algorithmic misclassification (beta=0.31), and digital transaction visibility (beta=0.19) positively predicted welfare exclusion, while algorithmic procedural justice was protective (beta=-0.22); the model explained 52% of variance (F(9,630)=75.84, p<.001, Cohen's f2=1.08). Misclassification partially mediated the informality-exclusion link (indirect effect=0.17, 95% CI [0.12, 0.23]), and procedural justice buffered it (interaction beta=-0.14, p=0.002). Logistic regression confirmed that non-digital informal workers faced 4.27 times higher odds of exclusion (95% CI [2.98, 6.12], p<.001). Qualitatively, gross e-wallet throughput was misread as income, and frontline discretion collapsed into a computer-says-no bureaucracy. Algorithmic fairness in Indonesia is better understood as parameterization bias against the informal economy; restoring conditional human-in-the-loop discretion is recommended.