Previously, we have discussed the evolution of the acute:chronic workload ratio (ACWR) here: http://sportsdiscovery.net/journal/2016/11/23/the-debate-on-the-acutechronic-workload-ratio/
Prior to that we also discussed a case study of using the ACWR in rehabilitation here:
Part One of our ACWR debate series outlined some of the contemporary research that include the metric in investigations of training load and injury risk. As the attention on the ACWR increased, the spotlight turned to the calculation itself. For example, a number of correspondences debated whether it should be determined using rolling or exponentially weighted moving averages.
Since that post in late 2016, the debate surrounding this calculation has continued to grow. There has been a wealth of research that has incorporated this dimension from which to analyse training load. Simultaneously, researchers have continued to critically assess, and attempt to refine the calculation itself. In this post, we will carry on where we left off and share some of the publications that have since deliberated over the computation of this topical metric.
The Role of Injury Definition
Much like these posts, Billy Hulin published an editorial in BJSM in early 2017 discussing the debate on the ACWR, entitled the “never-ending search for the perfect acute:chronic workload ratio: what role injury definition?”. He raises a point regarding the definition of injury, which could clearly influence the findings of workload-injury examinations. Hulin demonstrates the difference in likelihood of injury depending on the definition of injury used, specifically; medical attention injury, time-loss injury and missed-match injury. Clearly there is a need for clarity.
Rolling Averages vs Exponentially Weighted Moving Averages
Exponentially weighted moving averages were proposed as an alternative method to rolling averages, due to the limitations discussed in part one. The first study to investigate the differences between the rolling average (RA) and exponentially weighted moving average (EWMA) calculations of ACWR and the subsequent injury risk was by Nicholas Murray and colleagues here.
Both approaches demonstrated a high ACWR was associated with an increase in injury risk, in both preseason and in-season phases. The authors found the EWMA method was more sensitive to detect changes in injury risk than the RA method. Specifically, the EWMA had a greater sensitivity in identifying increases in injury likelihood at higher ranges of the ACWR. In fact, rolling averages may underestimate the risk of injury at higher acute:chronic workload ratios.
Lazarus and colleagues used “smoothed” and “differential” measures of training load in this free access Frontiers paper. The smoothed load is also an exponentially weighted rolling average but these authors employed a different time constant to that used in Sean Williams’s BJSM correspondence.
Mathematical Coupling and Spurious Correlations
This open access BJSM editorial by the guys at Teesside University, Manchester United and Liverpool John Moores University attempts to dissect the mathematical rubrics behind the ACWR. They demonstrate that because the “acute load” also makes up part of the “chronic load”, these two variables will be mathematically coupled. This results in a spurious correlation; defined as “one that exists between two variables irrespective of any true biological or physiological association between the variables.” (Pearson, 1896)
Lorenzo Lolli and colleagues generated data to simulate the data reported in Murray et al. (2016) for a hypothetical Australian football club with 1000 players. The data are discussed in the editorial alongside their hypothesis. The authors conclude that such spurious correlations lead to bias inferences.
Indeed, it has never sat well with me that the acute timeframe is incorporated into the chronic load. On one hand these two should be treated as two separate entities; assessing the acute workload in light of the chronic timeframe it is superseding. On the other hand, they are of course not separate entities. Any training stimulus surely becomes a part of your chronic workload immediately. The body does not draw a line in the sand after a week of training. More on that in a moment…
Greg Atkinson has eluded that this is the first of a number of papers exploring the mathematical concepts of the calculation, one of which is covered later on…
Now published: Stage 1 of our full analysis of the Acute:Chron ratio. Stages 2 (scaling properties) and 3 (prognostic utility) in prep. More superb work from @Lorenzo_Lolli90 with @spswgreg @Alan_Batterham @StrudwickTony @robbyt05 Rich Hawkins & Dave Kelly https://t.co/xWJPieEtuT
— Greg Atkinson (@Greg_at_TeesUni) November 4, 2017
Mathematical Coupling and Spurious Correlations – The Response
In May 2018, Windt and Gabbett published a BJSM editorial to further explore the concept of mathematical coupling with the ACWR. They walk through a comparison between using the traditional coupled formula and an uncoupled version. The uncoupled version is defined as “the ratio of the most recent week of work with the average of the three preceding weeks”. Appreciating how both of these methods work is important when interpreting the outcomes of them. At this stage, neither one is superior to the other but it is essential that researchers and practitioners alike comprehend and outline how they are calculating the ACWR, including whether the formula is coupled or uncoupled.
Association vs Prediction
Correlation does not imply causation; association does not imply prediction. These are not interchangeable terms. Pepe and colleagues (2004) discuss in some depth in this open access article that even strong associations, for example measured via odds ratios, are not suitable for classifying prediction.
Maurizio Fanchini and colleagues explore this discussion in relation to training load and injury, including the application of the ACWR (available here on ResearchGate). They found strong associations between ACWR calculations and non-contact injury risk in Serie A footballers. However, these ACWR measures showed poor predictive ability and this also stood true for very high ACWR (>85th percentile).
A shrewd part of the methodology in this study was to include different chronic timeframes (two, three and four weeks). This is a similar approach to that which David Carey and colleagues published, and we discussed previously here. Fanchini and colleagues found similar associations using all three chronic timeframes. They therefore acknowledge that further investigation is required to identify the optimal combination of timeframes for acute and chronic duration.
Billy Hulin and Tim Gabbett recently published another BJSM editorial about the never-ending search for the perfect ACWR, this time with the title “Indeed association does not equal prediction”. In response to the Fanchini paper, they reflect upon their regret about using the terms “predict” and “predictive” in their early work on the ratio (Hulin et al, 2015; 2016). They aim to clarify the practical application of the ACWR and expand in three important areas:
- An ACWR of 1.5 is not the magical boundary where all training should cease and desist
- Acute and chronic workloads can be used for the team and the individual
- Translating research to practice will always be about considering multiple variables and outcomes
An Unnecessary Normalisation Process?
In their next BJSM editorial, Lorenzo Lolli and his colleagues at Teesside, Man Utd and Liverpool John Moores further explore the mathematical foundation of the ACWR. Their critique centres around the scaling properties of the ratio; entitled “an inaccurate scaling index for an unnecessary normalisation process?”.
According to Curran-Everett (2013), a ratio of this sort should have a true and proportional association between the numerator and denominator and the ratio should normalise the denominator for all individuals in a consistent manner across the measurement range. The ACWR may not meet these fundamentals. The slideshare below from Professor Greg Atkinson, originally from his presentation at the Isokinetic Conference in Barcelona, may be useful in trying to understand some of these concepts:
The group analysed the data from Thorpe et al’s (2016) study tracking morning fatigue status across the in-season with Manchester United players. They suggest that acute load by itself may be a useful absolute marker. Normalising via a ratio, as per the ACWR, may just be adding noise to the analysis. This paper certainly takes the stats to another level. There is a need for such research to analyse the minutia of the ACWR approach. The challenge is finding a balance between a practical and accessible approach to analysing training load, and a mathematically sound methodology. As is always the case, the applied environment struggles to wait for academia to publish its scrutiny.
My Final Thoughts (so far… again)
The acute:chronic workload ratio has come a long way in a relatively short space of time. What started as a training stress balance percentage has now been critically examined and its flaws debated at length. There is a seemingly endless possibility of factors to take into account, and while we cannot possibly calculate all of them on a regular basis, it is important to be aware of these considerations. Also, it is important to be critical and deliberate in the method you utilise.
This debate is not going away and rightly so. It is our responsibility to continue to attempt to refine the calculation and discover the most valid, sensitive and practical method to employ in the applied environment. What we can be sure of is 7:28 day rolling average is not the only way to calculate the acute:chronic workload ratio!
Until part three (probably)…