Missing data cannot be directly handled by an estimation routine. You must "fix" your data appropriately in advance. Some things to try:
1. See MISDATA where you can fit in the missing values if you have some knowledge of the underlying process that generated the data. See: http://www.mathworks.in/help/toolbox/ident/ref/misdata.html
2. If the missing samples are found in small isolated clusters, you can split the data into independent NaN-free segments. Represent each segment as an iddata object. Then MERGE the resulting data objects into one "multi-experiment" data object. See: http://www.mathworks.in/help/toolbox/ident/ref/mergeiddata.html
The merged multi-experiment data object can then be used for estimation.
