This is confirmed by the Tax Authorities / Allowances in answers to questions from RTL Nieuws and Trouw. Those with a low income ran a greater chance of being singled out by the so-called risk classification model of Benefits.
This risk classification model (RCM) was taken off the air last July, after a critical report by consultancy firm KPMG. The Tax and Customs Administration then tried to improve the system and get it up and running again.
–
–
Selection model was ‘stigmatizing’
An internal follow-up investigation in March this year found that the model may have been ‘illegitimate’ and ‘discriminatory’. Citizens could be ‘stigmatised’, according to their own analysis provided to RTL Nieuws and Trouw. After that, it was decided not to use the RCM for good. (The core of this report you read here).
The risk classification model was used from April 2013 to determine which applications for benefits would receive additional scrutiny. This is a self-learning algorithm that itself determines which indicators indicate that applications for benefits have an increased risk of errors or fraud. All applications received a risk score. Applications with the highest risk score were then subject to additional checks.
–
–
The tax authorities acknowledge that ‘at least since March 2016’ the income of people ‘with variable limit values’ has been used. In practice, this means that low incomes received a higher risk score, while higher incomes received a lower risk score. According to the Tax Authorities, there appeared to be ‘a statistical relationship between the level of income and the chance of an (in)correct application’. The service would not say which limit values were used.
‘Not a big factor’
The tax authorities state that income had only a ‘small weight’ in the risk selection for the childcare allowance, because the amount of the allowance mainly depends on the number of children who go to childcare. “Income was not a big factor and was only included in part of the selections.”
According to the Tax Authorities, the intention was to prevent high recoveries of benefits, which mainly had an impact on people with low incomes.
–
–
Alarm bells about outcome
In practice, the model mainly came down to lower incomes, according to documents released earlier. Of the 1000 highest risk scores, no less than 82.3 percent had a household income of less than 20,000 euros per year. That is more than 11 times the average: of all applications, only 7.3 percent had a household income of less than 20,000 euros.
Also striking is the very high proportion of single parents selected for control: almost 90 percent.
–
–
Experts in the field of risk selection and machine learning algorithms reacted with dismay at the outcome. “How exactly the model worked cannot be ascertained on the basis of public information. But if the outcome zooms in on a specific group, in this case low incomes, then alarm bells should go off among officials who work with those models,” responds Cynthia Liem, associate professor of Artificial Intelligence at Delft University of Technology.
Zoomed in on low incomes
According to Liem, the overrepresentation of low incomes points to ‘selection bias’. In short: if the model suspected that there is a greater chance of fraud with lower incomes, it is mainly those applications that are checked. As a result, applications from low incomes in particular were corrected. Because the model is again being trained with those results, the focus has been on low incomes.
–
–
“Anyone who knows anything about risk selection knows that you have to be aware of selection bias,” says Liem. For example, by also checking applications at random. But that did not happen with Allowances, according to the research into the model from March this year, which the Tax Authorities made available for inspection.
“We only examine the high scores and there is no supervisory process that also examines other characteristics,” the service writes in response to the risk of selection bias.
Test when applying for housing benefit
Earlier research by Trouw and RTL Nieuws showed that the Tax and Customs Administration also used an indicator to select whether someone had Dutch nationality. After critical reports from the Dutch Data Protection Authority and Amnesty International, Member of Parliament Renske Leijten (SP) asked at the end of October for the RCM model to be made public.
State Secretary Alexandra van Huffelen (Supplies) wants to do this if ‘a careful substantive response’ has been prepared. that would end of this week go to the House of Representatives, she said.
Supervision of the shovel
The supervision of applications for rent, care and childcare allowance has now been overhauled. The old risk classification model RCM is no longer used. Allowances started a trial in September with checking applications for housing benefit. From now on, concerned citizens will be informed about the reason for extra checks (this did not happen in the past).
–
–
The role of the civil servant is also changing: he is also informed about the reason for the ‘ejection’ by the new model, so that it can be specifically examined whether or not there is a problem with the application for housing benefit.
Extensive hearing on behalf of State Secretary Van Huffelen meet you here and answers for follow-up questions here.
–
–
