Gait recognition is one of recently evolving techniques by which we can recognize individuals by one's gait. There are two major approaches; silhouette-based and model-based. In Japan, a method based on GEI (Gait Energy Image), which is one of the silhouette-based approaches, is used for forensic purposes. Sometimes, it is a problem of silhouettes' variabilities in one person due to different clothing that lessen recognition reliability under the GEI method. Here, we analyzed and evaluated the average error rates under clothing variation conditions using the method called Dynamic-features method, which we previously proposed. The Dynamic-features method was built inspired by previous studies of model-based gait recognition, which uses time-series of feature points and local shape features around the points automatically extracted from silhouette sequences. Before analysis, we roughly categorize whole data in the OU-ISIR gait database -treadmill dataset B-, which contains side-view data, into five clothing categories in order to deal with realistic off-line forensic situation, where we cannot strictly control the clothing conditions. As a result, the average increases of average error rate of GEI-based methods due to different clothing were ranged from approximately 8 to 11%, whereas that of the Dynamic-features method was approximately 3%. It was found that two representative dynamics of a feature point of one same person, where the point is influenced by different clothing conditions, showed different mean values but showed similar trends. Based on this fact, it is suggested that robustness of performances in Dynamic-features method under clothing variation conditions is obtained by effective utilization of dynamic properties of human's gait.