Show summary
Background: Sport injuries burden professional and recreational athletes. In 2021, Norwegian hospitals operated 1 462 anterior cruciate ligaments, and 62% of these happened during sports activity. To prevent injuries, it may be possible to change the training load. Unfortunately, how training load can be altered to achieve desired outcomes is unknown, because the relationship between training load and injury risk has proven difficult to study. The ability of currently used statistical methods to capture this complex relationship is either limited, or unknown. Consequently, studies have employed a plethora of statistical approaches. Systematic reviews have reported inconsistent and even conflicting findings both within and between studies, and declared the studies too variable to compare in analyses. Experts have questioned the evidence supporting training load as an injury prevention tool, and called for improved statistical methodology. Despite this, few studies have recommended alternatives, and those who have, have not tested the methods’ accuracy or precision. The validity of recommended methods is therefore unknown. To improve research on injury prevention programs, knowledge is needed on how to statistically determine the relationship of training load and injury risk. Aims: To identify statistical methods suitable for assessing the relationship between training load and injury risk. Specifically, to find methods for dealing with 1) missing data, 2) non-linearity, 3) time-dependent effects, and 4) the effects of relative training load. Main Methods: We analyzed three football datasets and one handball dataset: Norwegian Premier League men’s football (42 players, 38 injuries), Norwegian U-19 football (81 players, 81 injuries), Norwegian elite youth handball (205 players, 471 injuries), and Qatar Stars League (QSL) football (1 465 players, 1 977 injuries). In all Norwegian cohorts, training load was defined as the number of minutes in training/match activity multiplied by the athlete’s rating of perceived exertion on a scale from 1 to 10 (sRPE). The Norwegian Premier League data additionally had measures of distance and speed registered by Global Positioning Systems (GPS) devices in football. In the QSL cohort, training load was defined as the number of minutes in football training/activity. The Norwegian Premier League football and Norwegian elite youth football were the basis for three simulation studies (Paper I–III). We simulated a relationship between training load and probability of injury under different scenarios of missing data, non-linearity, and time-dependent effects. With the aid of accuracy and uncertainty measures, we compared the ability of various statistical methods to model the simulated relationships in the respective scenarios. Regression analyses were used to check whether there were any signs of non-linearity between sRPE and injury risk in the three Norwegian cohorts (Paper II), and also signs of time-dependent effects between training load and injury risk in the handball and QSL cohorts (Paper III–IV). In addition, we applied a novel approach of estimating the effect of recent training load relative to past training load on injury risk (relative training load) on the Norwegian elite U-19 and QSL data (Paper IV). Main Results: In each of the simulations, the performance of a few methods stood out from the rest. Firstly, for handling missing data, multiple imputation using predicted mean matching had, generally, the lowest percentage bias of all compared methods, and had acceptable bias (< |5%|) up to 50% missing data in sRPE and up to 90% missing data in the total distance GPS measure. Secondly, when we modelled parabolic non-linear relationships, fractional polynomials, quadratic regression and restricted cubic splines had the lowest root-mean-squared error, and highest coverage of 95% prediction intervals. Lastly, in the simulation of time-dependent effects, the distributed lag non-linear model was the only method that accurately modelled more than one scenario. It had the lowest root-mean-squared error and the narrowest 95% confidence intervals, by far, compared with the other methods. The handball model presented a parabolic J-shaped relationship between sRPE and injury risk (p < 0.001). The QSL model displayed time-dependent effects, where effect estimates of past training load decreased exponentially for each day in the past. The QSL model also showed highest injury risk at low levels of past training load, lowest risk at medium levels, and intermediate risk at high levels of past training load, for each level of recent training load. This demonstrated that relative training load can be modelled with this novel approach. Conclusion: Missing data in training load should be imputed with multiple imputation using predicted mean matching. Researchers in training load and injury risk should consider the potential for non-linearity and time-dependent effects, and explore such effects by specifying fractional polynomials or restricted cubic splines in distributed lag non-linear models. Modelling recent and past training load separately can be used to study the effects of relative training load on injury risk.
