Missing observations are a common feature of many time series. We show that the pattern of the missing values is important for selecting an estimation method that results in small variance. If missing values are in consecutive time points, then using the ‘true’ data generating process produces estimates with much lower MSE. On the other hand, the performance of a ‘wrong’ model, such as the cubic spline, is acceptable in the case of random non-consecutive missing values. The differences in the performance of the cubic spline are validated through simulations for the case where the ‘true’ model is an AR(1).