I am looking at land development using classified satellite imagery (0 = nondeveloped, 1 = developed, and multiple time periods of data are available as well as elevation data, etc). To model this, I'm considering a logistic regression (or some other model for binary response). The way I hypothesize the choice being made to develop land or not is as follows, people view collections of contiguous squares on the landscape and consider properties associated with the group of squares (composition of land cover, elevation, etc). They then decide whether or not they will transition any group of cells together from state 0 to 1 (develop them). The unit of observation is not the developer as many people may view a landscape and not develop it, this can?t be observed. Rather, I know which patches of cells become developed over time. So, someone must have looked at that aggregate of cells and decided they wanted to develop there. This way of looking at the system is perhaps overly complex, since the aggregates of cells considered are not observable and the many patches of cells that are considered overlap each other spatially and temporally. Also, either no patch a cell is a member of is developed or only one is, as a cell is assumed to only be developed once. Normally, researchers deal with each cell as an individual unit where a development decision is made. This has its obvious short-comings. Typically, these models perform poorly. Sometimes this approach is taken and steps are taken to account for spatial autocorrelation in error or the response values (whether or not developed) of cells neighboring a cell are included as an independent variable (autologistic, etc). Alternatively, the researcher groups the cells into square units (say 3x3 cells) and these are the observational unit (pretty arbitrary). I do not feel that these approaches adequately deal with the issue of what groups of cells the decision is made upon but is more out of convenience as these approaches are relatively simple and well known. I have been investigating nested models but from what I?ve seen these require that nests not overlap and are known (where as my nest overlap and aren?t known). I may be trying something that simply cannot be done. I would appreciate any ideas or specific areas to investigate. Thanks for reading.