To solve the contradictive results from diversified methodology problems in long memory analysis on stock time series, we analysis step by step and the results show that the long memory which results from classical model decreases after the heteroscedasticity is controlled, and the evidence of long memory is diminished after controlled the short dependent and structural changes by Markov Switching VAR integrated model. Therefore, the long memory observed by many Chinese scholars in stock return series is mainly caused by changes in the structural of variance and short dependent. To Chinese Shenzhen and Shanghai stock markets, the structural change is more important than the volatility.