Table 4 Causes of discrimination in data analytics

1. Algorithmic causes

 1.1. Definition of the target variable

Barocas and Selbst 2016 [8], d'Alessandro et al. 2017 [25]

 1.2. Data issues

Training data (Historically biased data sets)

Kamiran and Calders 2012 [42], Barocas and Selbst 2016 [8], Brayne 2017 [14], d'Alessandro et al. 2017 [25]

 1.3. Data issues

Training data (manual assignment of class labels)

Barocas and Selbst 2016 [8], d'Alessandro et al. 2017 [25]

 1.4. Data issues

Data collection (Overrepresentation and underrepresentation)

Barocas and Selbst 2016 [8], d'Alessandro et al. 2017 [25]

 1.5. Proxies

Schermer 2011 [73], Kamiran and Calders 2012 [42], Barocas and Selbst 2016 [8], Zliobaite and Custers 2016 [95], d'Alessandro et al. 2017 [25]

 1.6. Feedback loop

Mantelero 2016 [54], Brayne 2017 [14], d'Alessandro et al. 2017 [25]

 1.7. Overfitting

Kamiran and Calders 2012 [42], Mantelero 2016 [54]

 1.8. Feature selection

Barocas and Selbst 2016  [8]

 1.9. Cost function

Error by omission

d'Alessandro et al. 2017 [25]

 1.10 Masking

Proxies

Peppet 2014 [ 61], Zarsky 2014 [93], Barocas and Selbst 2016 [8], Zliobaite and Custers 2016 [95], Kroll et al. 2017 [45]

2. Digital divide

 2.1. Skills

Boyd and Crawford 2012 [12], Casanas i Comabella and Wanat 2015[18]

 2.2. Resources

Barocas and Selbst 2016 [8], Pak et al. 2017 [60]

 2.3. Geographical location

Casanas i Comabella and Wanat 2015 [18], Barocas and Selbst 2016 [8], Pak et al. 2017 [60]

 2.4. Age

Casanas i Comabella and Wanat 2015 [18]

 2.5. Income

Barocas and Selbst 2016 [8], Pak et al. 2017 [60]

 2.6 Gender

Boyd and Crawford 2012 [12]

 2.7. Education

Boyd and Crawford 2012 [12]

 2.8 Race

Bakken and Reame 2016 [6], Sharon 2016 [74]

3. Data linkage

Susewind 2015 [76], Cato et al. 2016 [19], Zarate et al. 2016 [91], Ploug and Holm 2017 [64]