The existing spectrum handoff techniques for cognitive radio perform handoff based solely on channel availability. This can result in a frequent number of channel handoffs and cannot ensure the quality requirement of applications such as delay bounds. This paper proposes the use of cumulative probability based on past backlog measurements to determine whether and how to perform handoffs. In order to prevent unnecessary handoff operation, secondary user should keep the same channel as long as the cumulative probability estimation does not violate some bound. However, the estimation from past observations may not correspond to the the real behavior in the immediate future even if efficient prediction models are employed. Thus, we propose the use of backup channels for short time periods to alleviate this problem. We analyze our strategies through extensive simulations and we compare them to random and classic approaches. Results show that our proposed strategies significantly reduce the number of channel handoffs up to 76% while still supporting the delay bound requirement.