Grid Search Method - Details

New for Version 4.5.0.0

For version 4.5.0.0, we have made some changes to the layout and values for Grid Search. But don’t worry, everything still works much as you remember from older versions. The old layout was a little confusing with parameters being shared between methods when they really shouldn’t be. The new interface and values are more precise and specific to each method. Below is what has changed and why:

• The “Number of Observations” controls have been moved under Grid Search. In the old interface, the number of observations controls were used for both Grid Search and Hudson’s. But this resulted in some ambiguity for the parameter values. So we’ve moved the controls under Grid Search to make it clear that these are specific only to the Grid Search Method. Hudson’s will have its own parameter controls (in time).
• The description for the “Number of Observations” controls has changed. The old descriptions used for these controls were not accurate or clear for all possible cases. We’ve reworded the descriptions for these controls so that they are accurate in all situations.
• The default values for “Number of Observations” have changed. You may notice that the default values for the number of observations are different than in previous versions. That is correct, but the new values mean the same thing. Due to the change in the descriptions for each control, the old default values were no longer correct. The new default values are accurate and obtain the same results as the old default values. If you load a session from an older version of Joinpoint, these parameters will automatically be updated to the new values. In this way, if you re-run an old session, you’ll get the exact same results.
• The “Number of Joinpoints” controls have been moved after the method selection. The controls are now in the correct order. The maximum number of joinpoints is actually dependent on the settings for “Number of Observations”, so it makes sense for the “Number of Joinpoints” controls to come after the “Number of Observations” controls.

How Grid Search Works

Using the default settings, the Grid Search Method only allows the joinpoints to occur exactly at observations. This does not, however, find the best fit. A better fit can be achieved by using a finer grid - by changing the setting for "Number of points to place between adjacent observed x values in the grid search" to something larger than the default of zero. So, the Grid Search Method creates a "grid" of all possible locations for joinpoints specified by the settings, and calculates the SSE at each one to find the best possible fit. With lower values for “Number of points to place between...”, this method is computationally more efficient.

Min number of observations from a joinpoint to either end of the data (excluding the first or last joinpoint if it falls on an observation)

Use this parameter to keep joinpoints from being placed too close to the end points. This value must be at least one (1) and no more than nine (9), but the default is set to two (2). So, for example if this value is set to 2, there is annual data from 1990-2014, and joinpoints are restricted to fall on the data points, then the first possible Joinpoint is 1992.

Min number of observations between two joinpoints (excluding any joinpoint if it falls on an observation)

Use this parameter to keep joinpoints from getting too close together. This value can be set as low as zero (0) and as high as nine (9), but the default is set to two (2). For example, if this value is set to 2, there is annual data from 1990-2014, joinpoints are restricted to fall on the data points and there was a joinpoint at 2000, then the closest possible other joinpoints would be at either 1997 or 2003.

The permutation tests are valid for any of the allowable choices for the above two settings (the minimum number of observations from a joinpoint to either end of the data, and the minimum number of observations between two joinpoints); however, some statistics (the standard error of the slope parameters and the associated p-values) cannot always be calculated when there are not at least three observations on a line segment (excluding observations at the joinpoints). See the statistical notes about Estimated Regression Coefficients (Beta). Similarly, some statistics cannot be calculated if a line segment in the joinpoint model is an exact fit.

Number of points to place between adjacent observed x values in the grid search

To allow joinpoints to occur between observations, specify the number of points to place between adjacent observed x-values in the grid search as 1-9. So, for example, if this value is set to 3, and the data is annual from 1990-2014, then the joinpoints could occur at values like 1995.25, 2000.5, or 2010.75, as well as on the observations. The minimum allowed value is 0 and the maximum is 9.  Note that a fine grid search with this value set at >5 may take a long time to process. An alternative to this would be to use Hudson's Method. 

Minimum APC Difference Worth Detecting (MADWD)

In some situations Joinpoint may determine that there are one or more pairs of segments where the difference in the Annual Percent Change (APC) between some consecutive segments is quite small. Users can enter a minimum APC difference worth detecting as a minimum percentage point difference (Min. PPD). Joinpoint will start with the final selected model that was chosen using the specified analysis criteria and lower the number of Joinpoints until all consecutive segments pass the Min. PPD criteria.