In the last week or so I have refined existing unit tests and added new ones to extend coverage to all of the user classes (Gravity, Production, Attraction, and Doubly) as well as to the BaseGravity class. Instead of testing every user class for every possible parameterization, the baseGravity class is tested over different parameterizations such that the tests for the user classes primarily focus on testing the building of dummy variables for the respective model formulations. In contrast, the BaseGravity class tests different cost functions (power or exponential) and will also be used for different variations that occur across the user classes. Unit tests were also added for the CountModel class, which serves as a dispatcher between the gravity models and more specific linear models/estimation routines. Finally, unit tests were added for the GLM class which is currently being used for all estimation on all existing spatial interaction mdoels within SpInt. This will be expected to change when gradient-based optimization is used for estimation of zero-inflated models in a MLE framework instead of the IWLS estimation currently used in he GLM framework.

In addition to unit tests code was also completed for handling overdispersion. First, several tests were added for testing the hypothesis of overdispersion within a Poisson model. Second, the QuasiPoisson family was added to the GLM framework. This is essentially the same as the Poisson family, a scale parameter, phi, (also known as the dispersion parameter) is estimated using a chi-squared statistic approximation and used to correct the covariance, standard errors, etc. to be less conservative in the face of overdisperison. QuasiPoisson capabilities were then added to the gravity model user classes as a boolean flag that defaults to false so one can easily adopt a quasi-MLE poisson modleing approach if they use the dispersion tests and conclude there is significant overdispersion. It was decided to push the development of the zero-inflated poisson model until the end of the summer of code schedule, which is also where graident-based optimization now resides. This makes sense, since these go hand-in-hand.

Next up on the agenda is an explortory data analysis technique to test for spatial autocorrelation in vectors and a helper function to calibrate origin/destinaton specific models so that the results can be mapped to explore possible non-stationarities.