Table 9.

The details of implementing bias mitigation for the neural network model using stochastic gradient descent after matching.

Sociodemographic conditions		Optimal threshold	Recall	Specificity	Accuracy	Differencea
Sex		52.10				1.88 b
	Male		50.00	97.85	97.65
	Female		51.85	97.88	97.49
Marital status		5.80				0.07
	Never been married		100.00	50.26	50.51
	Other groups of marital status		100.00	50.33	50.69
Working condition		12.50				5.12
	Working ≥35 hours		88.89	89.27	89.27
	Other groups of working condition		87.50	85.54	85.56
Race		8.10				0.03
	White		100.00	73.39	73.56
	Black		100.00	73.42	73.55
Income		52.60				3.07
	An income of <US $20,000		50.00	96.96	96.74
	Other groups of income		51.61	98.42	98.09

^aDifference between recall and specificity after bias mitigation.

^bItalicized values indicate an improvement compared with the initial values (50% threshold).