Statistical Analysis
The incidence of hamstring injuries in Danish professional
soccer was estimated at 14% based on data from Danish
men’s professional soccer teams.26 We estimated that the
cluster effects for club randomization gave an inflation factor of 1.19 based on a mean cluster size of 20 and an intracluster correlation coefficient (r) of 0.01.13 To achieve 80%
power with a significance level of 5% to detect a relative
risk reduction of 50% and a dropout rate of 20%, a sample
size of 428 players in each group was needed. Our aim was
to include a total of 1000 players in 50 teams.
We used cluster-specific statistical methods because
clubs, and not players, were randomized. Data were entered
into a Microsoft Excel 2007 (Microsoft, Redmond, Washington) spreadsheet and were analyzed with the SAS software
version 9.1 (SAS Institute Inc, Cary, North Carolina).
An injured player was not excluded from the trial
because he was still at risk of sustaining an injury to the
same or opposite thigh. However, recurrence of already
recorded injuries was not included in the data collection
and hence not analyzed.
Number needed to treat (NNT) was estimated by comparing the proportion of injured players in the intervention
and control groups. The injury rates in the 2 groups were
then compared using Poisson regression analysis. Here
the number of injuries in a given player is assumed to follow a Poisson distribution. The mean in this distribution is
given by the product of injury rate and time at risk. The
injury rate is allowed to depend on covariates, while time
at risk is calculated as the number of days the player participated in the study minus the number of days he was
injured. This analysis appropriately takes into account
the fact that not all players had complete follow-up and
that injured players were not at risk of injury in the period
of rehabilitation. Results of the Poisson regression analysis
are given as rate ratios. Because of the clustering in data,
parameters were estimated using ‘‘generalized estimating
equations’’ that account for intrateam correlation of the
injury risk.27 Age, previous injury, and competition level
are known risk factors of sustaining a hamstring injury.3,34
Thus, in further analysis, the initial model was adjusted
for the players’ age, competition level, and a covariate indicating whether the players had had a previous hamstring
injury. This analysis was conducted to avoid the possibility
of bias caused by small differences between the 2 groups in
the distribution of their background characteristics. We
then compared the injury rates in the 10-week midseason
period in a Poisson regression analysis restricted to injuries and time at risk in that period. We also compared
the rate of new injuries between the 2 groups. This was
done by comparing the number of injuries and time at
risk only in players who did not have a previous injury.
To compare the risk of recurrence, we restricted the data
to players who had a hamstring injury in the season before
the trial. In a Poisson regression model, we then explored
how the recurrence rate varied between groups. Finally,
we compared the 2 groups with respect to the length of
time players were injured. This was done in a regression
model with age as an additional covariate. Intrateam correlation was taken into account in a so-called mixed model
by allowing residual terms in players from the same team
to be correlated.
The level of significance was set at P.05.