How to choose alpha values in CHAID? In CHAID control parameters we have to specify the alpha value for Merging Threshold and Splitting Threshold. Typically this alpha (p-value) is set at 0.05. How do we select the alpha2 and alpha4 parameters? What if we choose one of them as 0.05 and another as 0.01?
 A: The original paper by Kass, 'An Exploratory Technique for Investigating Large Quantities of Categorical Data' (there is a copy here http://www4.stat.ncsu.edu/~dickey/Analytics/Datamine/Reference%20Papers/kass80.pdf) doesn't appear to address what the alpha values ought to be. However, there is an example, where the splitting criterion is set to 0.049 and the merge criterion is set to 0.05, although no explanation or justification is offered for this particular level. Most likely the use of the 0.05 level in the original paper has at least contributed to its status as the 'default' alpha.
The paper 'CHAID and earlier supervised tree methods' by Ritschard,2001 (http://www.unige.ch/ses/metri/cahiers/2010_02.pdf) , says that thresholds are set to 'usual critical values considered for statistical significance, 1%, 5% and 10%' hence it seems that the level set may be determined to be the level the analyst believes corresponds to statistical significance, in which case, if you have a preference, you can change it.
Ritschard also notes that there are four stopping rules applied in CHAID - the alpha threshold, maximum depth, minimum parent node size, minimum child node size - so that the alpha itself is not the sole factor in determining splits, and may not be invoked under some settings (as it was not in the example he discusses).
