Bias can creep in at every stage of the data collection and analysis. In the early stages, the process by which data are generated may reflect existing inequalities. The collection and assumptions made in the processing of data may be biased. The training datasets that are developed for machine learning, which by definition reflect a sample of the population, can be a biased subset. The labelling of those datasets, which is often done by human classification, can reflect conscious or unconscious human bias. Finally the decisions that are made also be biased. (1)

The measurement of employment and unemployment in household surveys in the Current Population Survey (CPS) is a good example. Only a small number of households are sampled, and so it's important to pay attention to non-response bias, particularly since COVID forced data collection to rely more heavily on phone interviews. The small sample size also means that a number of processing adjustments are made to create state and local estimates; the assumptions may induce processing bias. Measurement bias is can occur when the measures that are reported do not fully reflect the underlying concept - either due to issues in the measurement process itself or because the concept is complex. The measurement of employment in the CPS is primarily defined as worked at least one hour a week in paid work in the survey week(4)), which is very different from the "quality" employment of interest in reports like the MSPSR, or the Infrastructure and Jobs Act.

Biases don't just occur with employment measures, of course. Bias due to lack of representativeness can occur if the data that is collected is systematically different from the underlying population; a non employment example is 311 calls about crime are likely to underrepresent poor neighborhoods. Survivorship bias, a subset of selection bias, only considers the “survivors” of a process, overlooking those who didn’t make it through; another non employment example is gender differences in publications(5). Algorithmic bias has come more to the fore with increased use of Artificial Intelligence tools, and can result when the algorithms used for data processing or analysis are biased, often due to biased training data or assumptions in the algorithm’s design. For example, the training data for voice recognition used by businesses can be biased because they use standard American English. Disclosure bias can result from attempts to protect privacy, like differential privacy, and have disparate impact on certain racial and ethnic groups. And sometimes the survey population can simply be different from what people think it is: for example, employment data are collected from a survey of the civilian, non-institutional population, so it doesn't capture information on the full population - the military and incarcerated individuals are excluded by design.

While it is almost impossible to remove all bias from data, analysts should try to understand where biases might exist and try to be ethical in their approach. One good definition of data ethics is “a focus on data practices with impact on people and society. It recommends responsible and sustainable data collection, processes, and storage practices. Additionally, it ensures ethical use of data”(6). Ethical issues include considering issues of data ownership, confidentiality, consent, and transparency, particularly with respect to different minority groups.

Ethics issues should be considered in every aspect of policy analysis. Machine learning provides a very clear example of what issues can arise. Machine learning techniques typically predict classification outcomes as a score between zero and one. In the public policy case, a high threshold will typically result in fewer people being identified as recipients of a particular policy, with lower tax payer costs; a low threshold will be the reverse. So the analyst should always ask - how is the threshold decision made, and by whom?

Similar ethical issues arise with data collection. Better data is necessary to produce more targeted policies, but that means a greater privacy risk. The analyst should ask - how is the privacy/utility tradeoff decision be made, and by whom? Data on at risk populations is important to better serve those populations – and reduce the possibility of discrimination – but that also introduces risks. Again, who makes those decisions? Informed consent is not a useful framework with much data: Helen Nissenbaum has pointed out that notifications are often either comprehensible or comprehensive, but rarely both (2, 3). Privacy is also rarely an absolute, because the context differs, as shown in work by Kreuter and coauthors (4). In sum, the answers to ethical questions are not clear-cut, but an open and transparent discussion of the decision-making process, and understanding the sources of any bias, are essential to an ethical analysis

References:

1. Bias and Fairness, Chapter 11 https://textbook.coleridgeinitiative.org/chap-bias.html (Sections 11.1 - 11.2)

2. Nissenbaum H. Contextual integrity up and down the data food chain. Theoretical Inquiries in Law. 2019;20(1):221-56.

3. Nissenbaum H. A contextual approach to privacy online. Daedalus. 2011;140(4):32-48.

4. Kreuter F, Haas G-C, Keusch F, Bähr S, Trappmann M. Collecting survey and smartphone sensor data with an app: Opportunities and challenges around privacy and informed consent. Social Science Computer Review. 2020;38(5):533-49.