Author(s) This pre-registration is currently anonymous to enable blind peer-review. It has one author.
Pre-registered on 2021/09/01 00:27 (PT)
1) Have any data been collected for this study already? No, no data have been collected for this study yet.
2) What's the main question being asked or hypothesis being tested in this study? How does algorithmic fairness impact algorithmic permissibility? Which personality variables play a role?
H1: The unbiased algorithm is perceived as fairer than the two biased algorithms.
H2: People will find it less permissible for a biased (vs. unbiased) algorithm to make loan decisions.
H3: There is a positive relationship between perceived (group) fairness and permissibility.
RQ1: Is there a difference in perceived (group) fairness and permissibility between the two biased algorithms (favoring males vs. favoring married people)?
H4: There is an interaction between gender and algorithm condition; females perceive the algorithm favoring males as less fair and less permissible than males. No gender difference is expected in the unbiased condition.
RQ2: Is there also a gender effect in the favoring married people-condition?
RQ3: Do (subscales of the) social dominance orientation, modern sexism, political orientation, or marital status moderate the effects of condition on perceived fairness and permissibility?
3) Describe the key dependent variable(s) specifying how they will be measured. There are four fairness items – one general (this algorithm is fair), two focusing more on procedural fairness and one explicitly addressing group fairness. We will run an exploratory factor analysis and calculate Cronbachs alpha to test whether participants differentiate between different aspects or not. Depending on the results, we will either use the group fairness item (most direct test) or a scale containing this item as fairness measure.
Permissibility will be measured with 3 items from Bigman and Gray (2018).
4) How many and which conditions will participants be assigned to? Three conditions: unbiased algorithm, biased: favoring males, biased: favoring married people; operationalized via false positive rates for different subgroups (8.4% vs. 8.3% in the unbiased condition; 6.2% vs. 10.8% (females vs. males; singles vs. married people, respectively) in the biased conditions
Gender of the participant will be used as quasi-experimental factors
Modern sexism will be measured with 4-items adapted from Swim et al. (1995) and Kam & Archer (2021).
Social dominance orientation will be measured with six items from Ho et al. (2015): one from the portrait and contrait dominance subscale, both from the portrait and constraint anti-egaliarianism subscale
Political orientation is measured on a 9-point scale from left – right.
5) Specify exactly which analyses you will conduct to examine the main question/hypothesis. To test H1, H2, H4 and explore RQ1 and RQ2, we will conduct a 3 (condition) x 2 (gender) MANOVA with perceived fairness and permissibility as dependent variables. H3 will be tested with a correlation analysis. RQ3 will be tested with several regression models, testing different subscales and interaction terms with bias-condition.
6) Describe exactly how outliers will be defined and handled, and your precise rule(s) for excluding observations. no exclusions planned
7) How many observations will be collected or what will determine sample size? No need to justify decision, but be precise about exactly how the number will be determined. N = 600. For the analyses involving gender, this should result in roughly 100 participants per cell as in prior work on algorithm aversion (e.g. Bigman & Gray, 2018).
8) Anything else you would like to pre-register? (e.g., secondary analyses, variables collected for exploratory purposes, unusual analyses planned?) This research is largely exploratory to derive hypotheses for a grant proposal. It is, therefore, run after an unrelated study on the effects of networking and video conference use. The networking data will be used to see whether a prior cluster analysis replicates during COVID-19. We also test a new scale for ambient awareness, include the news-find-me scale with the new algorithm subscale (Zúñiga & Cheng, 2021) and some measures on social media and news use. We will explore relationships between these variables.
