Observation is one of the most commonly used approaches in pedagogical research, but the collected data have often been analyzed inappropriately. The purpose of this study was, using the compositional data analysis, to investigate student teachers' behaviors in teaching physical education (PE) classes. Forty-four senior students from a Midwestern urban university were involved the study and they were all teaching PE in elementary or middle schools for their teaching practice. Using a teaching behavior scale developed by Weigand and Hawkins (1989), the student teachers' teaching behaviors were observed in a 15-second interval, and were recorded into the following five categories: General observation, specific observation, management, instruction, and modeling/instruction. The observations were completed by an experienced pedagogical researcher, and each student teacher was observed from four to eight times during their practicum experience. For the convenience of the data analysis, five behavior categories were collapsed into three categories: Observation (including both general and specific observation), management, and instruction (including modeling/instruction), and only first two observations were compared. Each student teacher's total teaching time was computed by summing the time spent in each category together, and the percentage time of each category was calculated by diving a particular category time by the total time. The assumption of the logistic normal distribution was examined using Anderson-Darling, Cramer-won Mises, and Watson tests. The ternary diagrams were used to illustrate the interrelated patterns among three behavior categories. The partial least square and maximum likelihood estimation tests were conducted to examine the pattern changes of the teaching behavior. The results of the Anderson-Darling, Cramer-won Mises, and Watson tests showed that the data met the assumption of the logistic normal distribution. The maximum likelihood tests indicated that, in general, no statistically significant changes happened in student teachers' teaching behavior pattern. However, the examination of the ternary diagrams indicated that, at the beginning of their teaching, student teachers spent relative more time on observation (39%; 11% in specific observation) than management (30%) and instruction (30%; 10% in modeling/ instruction). As they became more familiar with teaching, there was a trend that many of them spent more time in instruction (2nd observation): Observation (36%; 14% in specific observation), management (29%) and instruction (33%; 13% in modeling/instruction). In conclusion, as student PE teacher spent more time in their teaching practice, they became better managers of their teaching. As a result, they were able to spend more time in instruction.