Power analysis

An a priori power analysis was conducted to determine the sample size needed to detect a small effect size. Data was simulated in R [37] by creating condition columns (NA, PA, and control—dummy coded), an error term (Gaussian distribution with a mean of 0 and standard deviation of 1), and a choice column designating the four possible outcomes (1=alcoholic beverage + alone, 2=non-alcoholic beverage + alone, 3=alcoholic beverage + with other, and 4=non-alcoholic beverage + with other). The choice column was based on probabilities for cells 1 and 3 equating to small odds ratios (~1.4); for the NA condition, this was equal to a 59% chance of selecting choice 1 (0.59/(1-0.59)=1.44) and a 24% chance of selecting choice 2 (0.24/((1-0.59)-0.24)=1.41). This was reversed for the PA condition (59% for choice 3 and 24% for choice 1). The remaining probabilities (1-0.83=0.17) were split between choices 2 and 4. Because we didn’t have specific hypotheses for the control condition, all 4 choice outcomes had a 25% chance of being selected. Using these probabilities, a Monte Carlo simulation was conducted in which 10,000 multinomial logistic regressions were run to see how many significant p-values emerged per comparison (i.e., 1 compared to 2, 1 compared to 3, and 1 compared to 4) depending on different sample sizes. Using an alpha of 0.05, power of 0.80, and a small effect size (odds ratio=1.4), 150 participants were required (50 participants per condition). Due to the COVID-19 pandemic and closure of the Carnegie Mellon University campus, only 128 participants were run. Two participants were excluded before data analyses were conducted because they did not meet eligibility requirements during the experimental session, despite meeting requirements during the phone screen (i.e., they reported past solitary drinking during the phone screen, but denied past solitary drinking during the experimental session). Thus, the final sample included 126 individuals.