Jet Lag Does Not Impact Football Performance: Using Randomization Inference to Handle Complexity

Abstract

Introduction It is commonly accepted that traveling across time zones affects sport performance (i.e. via jet lag). This belief is based on poor quality evidence for team sports and simplistic analyses, such as t-tests and linear regression, to explore complex phenomena. For instance, Roy & Forest used such analyses to examine win percentages for the NFL, NBA, and NHL, concluding that East Coast teams were disadvantaged. Similarly, Smith et al. primarily used t-tests to show that West Coast NFL teams were more likely than East Coast teams to beat the Vegas spread in evening games (non-coastal teams were omitted). Neither analysis considered time zone change or game time as continuous constructs nor did they account for important contextual information. We used modern causal inference methods and a decade of Collegiate Football games to determine if jet lag and kickoff time have any causal effect on beating the Vegas spread. This required fitting nonlinear splines for both data re-weighting and analysis; however, using weights in a generalized additive model (GAM) presents challenges for standard frequentist inferences. Thus, non-parametric simulations were developed to obtain valid causal inferences via randomization inference (RI). Methods Pro Football Focus data from college football seasons 2013-2022 were paired with time zone data from Google Maps, weather data from gridMET, and Vegas spread data from collegefootballdata.com. GAM-based propensity scores were calculated from turf type, stadium type, precipitation, humidity, temperature, and wind speed. These propensity scores orthogonalized these variables relationship to the treatments (i.e., game time and hours gained due to time zone change) consistent with the Potential Outcomes framework. The propensity scores were used to weight the observations in a GAM logistic regression, which modeled beating the Vegas spread as a function of a splined interaction for game time and hours gained in travel. Since valid standard errors cannot be calculated from GAMs with weights, we used RI to compare the interaction effect to random chance. We simulated 5,000 datasets of random treatments under the positivity assumption. Each RI dataset was analyzed with the same GAM used for the observed data to obtain a distribution of noise F-statistics. The real data F-statistic was contrasted to the RI distribution for inferences. Results The real data were highly compatible with the null hypothesis of no effect for hours lost/gained in travel and game time (p = 0.471). Conclusion We need to rigorously interrogate assumptions regarding what affects performance in team sports. There is no clear indication that jet lag and game time affect team performance when appropriate analyses are performed in a causal inference framework. Similarly rigorous analysis should be undertaken to confirm or refute other assumptions in sport science, such as workload management, sleep practices, and dietary/supplementation regimens.