- Academic Research and Writing Tips / Tutorial

Effects of a Coach Education Program Aimed at Empowering the Coach-Created Motivational Climate (in Grassroots Football)



Sixty-eight (66 men and 2 women) grassroots football coaches (Mage = 37.6, SD = 9.9) of athletes aged 10 to 15 years old (Mteam age = 11.92, SD = 1.45) from the United Kingdom (n = 18), France (n = 17), Spain (n = 18), and Greece (n = 22) were recruited from the larger European-based PAPA project (Duda, 2013). The intervention condition consisted of 34 coaches and 34 coaches participated in the control condition. Participants gave informed consent to be filmed during training sessions at three time periods (T1–T3). Ethical approval for this study was granted by ethics boards at the University of Birmingham, United Kingdom, University of Valencia, Spain, the University of Thessaly, Greece, and the University Joseph Fourier, France. Guardians of the athletes were also informed of the filming and the purpose of the study and were given the opportunity to opt their child out of the study. Coach attrition occurred between the three time periods due to several reasons (e.g., team change, club change, profession change). In T2 data were obtained from XX coaches (United Kingdom, n = 18; France, n = 17; Spain, n = 16; Greece, n = 22). In T3 XX coaches were filmed (United Kingdom, n = 18; France, n = 17; Spain, n = 16; Greece, n = 22).

Athlets Education


Filmings were conducted three times across the course of an athletic season. The first filming took place during the first 4 weeks of the season, before coaches attended the Empowering Coaching™ workshops. The second filming occurred within a time window of 6 weeks, immediately after coaches attended the intervention workshops. The third filming took place approximately four to six weeks before the end of the season. The exact time of filming depended on the availability of coaches. Coaches in the control arm followed the same filming schedule, but did not receive a workshop.

Before the filming day, a researcher visited the training site to familiarize participants with the filming procedure (Van der Mars). During the day of the filming, the researcher arrived before the start time to set up video equipment and attach a wireless microphone to the coach. To minimize reactivity to the video cameras and prevent the cameras from interfering with regular sessions, cameras were placed in an unobtrusive section of the play area. Coaches were recorded during their training sessions with a digital camcorder, and audio was captured with a wireless microphone attached to the coach. After each filming, video files were analyzed for content and clarity to ensure the quality of audio and visual data was appropriate for the subsequent coding using the MMCOS.

For the present study, coding of video footage was conducted by two raters from each country. Rigorous rater training procedures were followed to ensure inter- and intra-rater agreement (Smith et al.). Reliability for all codings across all countries was established, exceeding the cutoff value (k > .70). Raters were postgraduate students in the discipline of sports psychology and had a strong knowledge of the conceptual and theoretical background of the study, as well as experience in teaching or coaching football (Smith et al.).

Since the duration of training sessions varied significantly, we adopted an approach for analyzing the footage that differed from methodologies utilized in previous systematic observations (e.g., van der Mars). This analytical approach involved splitting videos to four equal quarters to ensure that all recorded footage was coded, as even very brief interactions occurring in-between time blocks may have significantly affected the motivational atmosphere. Raters coded the footage according to a marking scheme that took into account the style and range of strategies employed by the coaches, as well as the impact (e.g., intensity, individual or group effects) of strategies on the climate. At the end of each quarter, raters coded the potency of each of the seven environmental dimensions on a scale of 0 (very low potency) to 3 (very strong potency). After raters coded the entire training session, they provided an overall rating of the degree to which the coaching atmosphere was empowering or disempowering using the same scale (Smith et al.).

Intervention Workshops

Coaches in the intervention condition attended the Empowering Coaching™ training program (Duda). Empowering Coaching™ draws from the two dominant motivational theory frameworks (AGT and SDT) and related research. The coach training program was implemented as part of the European-based PAPA project and aimed to provide a theory- and evidence-based intervention that fosters positive sport experiences for children and sustained sport participation (Duda). A detailed description of the conceptual and empirical foundations of the program as well as the key features of the intervention can be found elsewhere (see Duda).

Data Analysis

According to literature, although most coach training interventions require clustered data designs with multilevel modeling (Smith et al.; Smoll, Smith & Cumming) only a single intervention fulfilled this criterion. To examine our research question, we followed the suggestions of Bryk and Raudenbush and used multilevel linear modeling, as our repeated coach observations at first level were nested within coaches at the second level, which in turn, were nested within the intervention and control teams at third level. Nested data can potentially violate the independence assumption of ANOVA or the ordinary least-squares assumption in multiple regression (Hox). To address independence violations, multilevel modeling is suggested as the preferred analysis due to its ability to produce limited Type I errors and unbiased parameter estimates. Within the framework of multilevel modeling, repeated measurements in longitudinal design studies are also treated as nested data, where multiple observations are nested for individual participants (Peugh).

A major advantage to employing multilevel modeling techniques with longitudinal data in a pretest–posttest design—compared to ANCOVAs—is that it adds more intermediate observations. In turn, these additional observations significantly increase the statistical power of tests between the experimental and control groups (Maxwell; Willett). More importantly, hierarchical linear modeling allows incomplete data to be included in the analysis without losing statistical power (Hox).

Hypothesized Model (pattern of change over time)

A two- level hierarchical model assessed the effects of the PAPA coach training program on the overt coaching behaviors and thus on the objective coach-initiated motivational climate dimensions as assessed by MMCOS. It was expected that the Empowering Coaching™ training program would have a positive effect (i.e., increasing) on the empowering dimensions of the climate (i.e., autonomy, task, relatedness supportive, and structured) as well as on the overall Empowering climate. On the other hand it was expected that the Empowering Coaching™ training program would have a negative effect (i.e., decreasing) (AS is changing over time/ is not changing over time/ changes in a lower pace/degree) compared to on the disempowering dimensions of the climate (i.e., controlling, ego, and relatedness thwarting) as well as on the overall Disempowering climate across three measurements during a period) . First- level units were coaching behavior observations of the coaches that participated in the PAPA project in France, UK, Spain and Greece, as rated by independent raters, resulting in a total of 68 coaches corresponding to 193 training sessions for analysis.

Second- level units were the participant coaches which were allocated to experimental and control groups. The experimental group consisted of 34 coaches while 34 coaches comprised the control group. Multilevel modeling was implemented through SPSS MIXED MODELS, Version 22. Hierarchical models are those in which data collected at different levels of analysis (e.g., observations, groups) may be studied without violating assumptions of independence in linear multiple regression. Multilevel modeling takes account of these dependencies by estimating variance associated with group (e.g., experimental or control) differences in average response (intercepts) and group differences in associations (slopes) between predictors and the DV (e.g., group differences in the relationship between time and group). This is accomplished by declaring intercepts and/or slopes to be random effects. In the hypothesized model, individuals and households are declared random effects to assess variability among individuals within households as well as variability among households. Also, one of the predictors, noise level, was declared a random effect, reflecting the hypothesis that there would be individual differences in the association between noise level and annoyance.

Multilevel Modeling

Thus, we followed an exploratory model building process (Hox) to estimate the best fitting model to our data. Prior to the analyses, the dataset was restructured (variables to cases) in order to meet the needs for HML analysis (Peugh & Enders). We used the Maximum Likelihood (ML) parameter estimation method to compute coefficients for all models that we tested except for null model. First, a null or unconditional model (level 1) was estimated by inserting dependent variables (autonomy support, task, ego, controlling, relatedness supportive, relatedness thwarting, structured, as well as higher order factors, empowering and disempowering) and subjects (i.e., coaches) but no predictors were imported at this stage. Level 1 model calculates the intraclass correlation coefficient (ICC) which estimates the degree of variation that exists in level 2 units and evaluates the necessity for multilevel modeling. It should also be noted that the procedure was repeated for each dependent variable separately. We then proceeded in building gradually more complex models to check for improvement in the model fit. In model 2 we included the predictor “time”. “Time” variable consisted of three measurement time points (one pre-intervention and two post-intervention measurements of coach behavior). In addition, the intervention group modality was fixed at 0. The third model included coaches grouping into intervention and control groups, as level 2 variable intended to explain variations in level 1 intercepts across coaches. In the last step, a final model examined the “group” X “time” interaction effects. Following Rasbash et al.’s recommendations, to appreciate the efficacy of model 1 compared to the empty model, and model 2 compared to model 1, we assessed the significant improvement in the fit statistic. We calculated the deviance statistic of the model (i.e. Δ−2 log L/df), which follows a χ2 distribution at k degree of freedom, k representing the number of adding parameters to estimate. When χ2 is significant at p { .05, this indicates a greater improvement in the fit statistic compared to the previous model.


The results were not entirely conclusive as could have been expected with a study that depends on quantizing qualitative factors in such a crude manner. However, some general patterns were observed, thanks to the combination of regression and variance analysis. It was clearly noticeable that the ego-involving categorization was most effective in creating patterns between the different models, with a variance of intercepts between 0.46 and 0.73 from the empty model to the level-2 interaction model. This suggests a strong and direct effect between ego-involving environmental dimensions and positive result in coaching.

Furthermore, the time it took to complete the examination varied between the control and experimental groups visibly, as evident with the level-2 interaction model. This is mostly due to the random factors such as noise level, which were expected to create artifacts of variance. However, they did not seem to damage the data set in the slightest as most environmental dimensions were relatively up to expectation. The controlling environmental dimensions, in particular, the random ones, showed a within group variance of 0.22 throughout all groups expect for level-2 where it was 0.21, which is nearly identical, meaning the random effects were contributory, but did not vary between the different groups, further affirming the accuracy of the results.

As for the linear relationships between the results, there is a clear proportional pattern between age of coach and that coach’s performance, specifically when measured in time. Most coaches with more than 5 years of experience received a good time result, with the expected outlier case. Furthermore, it is evident that there are patterns of method between different age groups and combinations of methodology that developed over time with the coaches, due to their personal experience.

It would appear from the results that the coaches all start with a mostly empowering approach with very little ego involving influence and they are for the most part not very controlling. This might be related to their inexperience and inability to dominate over the people they coach. Furthermore, it is important to note the relationship between positive and negative reinforcement, as sometimes it does seem to stem from an early stage of coaching. In those cases, it continues to grow over time and it more or less reduces the presence of positive reinforcement down the line. One can conclude this from the following relationships:

The coaches with much experience, who were of older age and had, disempowering tendencies, were also quite ego involved and controlling. Furthermore, they did not support autonomy. These patterns could also be noted in a small number of younger coaches with less experience and while they might be too young to express a high disempowering value, one may come to the conclusion that they will grow into these patterns at a later time. Within these values and analyses, there were many patterns discovered, such as the above-mentioned disempowering trend. These values seem to be interconnected in such a way that one causes another over a given period of time, as long as the age is right. One might also consider the possibility of zeitgeist influence, where the older coaches might have been brought up in a different time, with a slightly different social opinion and training, where empowering versus disempowering trends were defined differently as they are today. Perhaps twenty or thirty years ago, disempowering as a coaching mechanism was simply preferred.

Another visible pattern that directly stems from this one was the relationship between empowering and disempowering together, within single individuals, when compared against their age. Both younger and more inexperienced coaches seemed to cling harder to a single choice between the two, whereas older coaches that are more experienced often employed both methods simultaneously. This might speak more towards the spirit of the times during which the coaches initially trained, although it is a far more sensible conclusion to assume that with age comes the ability to appreciate different methods of approach in training people.

Conclusively, the ego involving property alone seems to vary quite a lot after a certain age point. Namely, coaches over the age of 30 seem to vary in ego involving factors almost randomly, whereas those younger than 30 had mostly low values of ego involving factors. This suggests that ego, on its own, develops randomly, or that it is unaffected by the given factors, but can be used as a variable that other variables depend on. A good example for this is controlling factor, which seems to increase proportionally from ego involving factors.

Another interesting pattern that was found with the analysis of covariance was the relationship between empowering, controlling and autonomy support. Namely, on their own, the linear relationships between empowering and controlling versus autonomy support seem somewhat weak, or rather, as if a linear relationship does not exist. However, a multivariate analysis revealed otherwise, as the combination of empowering approach and controlling seem to contribute directly towards autonomy support in opposite effect. In other words, coaches that were highly empowering and not at all controlling always have a high autonomy support. This multivariate analysis was necessary because analyzing them directly and linearly led to the conclusion that controlling behavior had no visible effect on autonomy support, nor did empowering factor.


The results of this analysis were very conclusive simply because of the use of multivariate analyses. The quantification of real world qualitative factors seldom functions when entered directly into a linear analysis, but in this case, a wise combination of the various variables seemed to yield a satisfactory and credible result. The analysis of significance also served a great role in ensuring that no variables were given too much solitary credence.

As for the real world implications of the study, I would personally conclude that while this study found many relationships between various factors, the most important real world implication found is the result amalgamation of all these factors towards high performance. I believe that this study would be very beneficial to any coaching institution or organization towards teaching the coaches of the various statistical results of their actions and how those decisions will affect them in the long term. Furthermore, I believe that it was quite revolutionary to analyze both empowering and disempowering approaches fairly and side by side, as this is seldom done due to the natural negative stigma behind negative reinforcement. I believed however, that every variable needed to be factored in, regardless of stigma, and it quite paid off in the long run.