Table 8 — Gate 4 westside equations and classification results 



A. Equations 



Logistic regression Discriminant analysis 



Variable 



Coefficient 



(Std Err) < 



Coefficient 



(Std Coeff) 



Ln(STUMPMILL) 



-5.308 



(0.716) 



-3.307 



(0.830) 



(SPLT)^ 



.00002 



(.0000) 







Ln(SPLT) 







.824 



(.194) 



Ln(HAULRAT) 



.648 



(.248) 







HAULRAT 







.744 



(.173) 



DEAD 



-.036 



(.012) 



-.023 



(.270) 



Ln(DENSE) 







.183 



(.148) 



(VPA)"2 







.113 



(.150) 



Ln(ACRES) 







-.114 



(-.137) 



(%CABLE)2 



-.00018 



(.00004) 



-.0001 



(.487) 



Ln(UNCUT, 3) 



-9.340 



(3.632) 







(UNCUT,_3) 







-.0006 



(.149) 



COMPMILL 



.864 



(.288) 



.636 



(.318) 



PMETH 



.924 



(.406) 



.593 



(.218) 



LMBRPROD 



.026 



(.010) 







EXCH^ 



-5.964 



(3.124) 



-6.856 



(.419) 



Constant 



110.470 



(28.442) 



23.213 





B. Classification results 





Logistic regression 



Discriminant analysis 



Actual Total 



Correct 



Percent 



Correct 



Percent 



group sales 



predict 



correct 



predict 



correct 



Unsold 145 



100 



69.0 



115 



79.3 (78.6)' 



Sold 204 



170 



83.3 



157 



77.0 (75.0)' 



All sales 349 



270 



77.4 



272 



77.9 (76.5)' 



'Indicates percent correctly classified using the holdout method. 



of higher valued species, located near several competitive 

 milling centers, sale contract allows the winning bid price 

 to fluctuate with the lumber market, lumber markets in 

 an upswing, high volumes per acre harvested, does not 

 require undue movement of harvesting equipment and 

 labor. 



As they rise in value, or are present in the timber sale 

 the following variables negatively influence salability: 

 stump-to-mill costs (STUMPMILL), percentage volume 

 dead white pine or dead lodgepole (DEAD), percentage 

 volume cable yarded (%CABLE), U.S./Canadian exchange 

 rates (EXCH^ g), and the uncut volume under contract 

 (UNCUT^ 3) for the westside National Forests of Region 1. 

 Once again these variables indicate sale quality in terms 

 of costs and the type of volume harvested. The most im- 

 portant characteristic in determining salability is 

 STUMPMILL, with %CABLE second. 



Overall correct classification is approximately 78 per- 

 cent (table 8b). The logistic regression correctly classified 

 69 percent of the unsold sales and 83.3 percent of the sold 

 sales. The discriminant function correctly classified 79.3 

 percent of the unsold sales and 77.0 percent of the sold 

 sales. The holdout method indicates that the discriminant 

 function's classification results are stable. 



Adding gate 4 and market information allowed the 

 overall correct classification to increase approximately 

 12 percent. In addition to the overall improvement in 

 correct classification, the gate 4 equations improved 

 individual class results. Therefore we not only gain 

 overall classification accuracy, but also accuracy within 

 individual groups. 



Statistical Evaluation 



Up to this point I have discussed the classification re- 

 sults achieved when modeling sold and unsold timber 

 sales at different gates, by different classification tech- 

 niques, and by geographical zones. But the question still 

 remains as to the statistical significance of the classifica- 

 tion results achieved by the gates, classification tech- 

 nique, and geographical zone. The practical significance 

 of the results are left to the reader to determine. 



In general, the categorical analysis of variance (see 

 table 9) indicates that a statistically significant difference 

 exists in classification results when considering the geo- 

 graphical zone (eastside vs. westside), and when moving 

 fi-om gate 1 to gate 4. But there was no statistically sig- 

 nificant difference in classification success with respect 

 to the classification method (logistic regression vs. 

 discriminant analysis). Also, there were no statistically 

 significant interaction terms in the analysis of variance 

 equation. 



A statistically significant improvement in classification 

 results was observed when moving fi-om gate 1 through 

 gate 4. As better information is used in the equation 

 development process, the classification results improve 

 significantly. The sale can be described in terms of aver- 

 age elevation, average slope, total volume, and sale acres 

 at gate 1. These measures indicate the type of yarding 

 systems that will be needed and the general size of the 

 sale, but the general nature of the information will not 

 allow an accurate model to be developed. The process 

 needs specific information with regard to the percentage 

 of volume cable yarded, the miles of road construction, 

 the number of pieces moved, and so forth. Also, the equa- 

 tion is improved as market information is added to the 

 process. 



The results show that a statistically significant higher 

 percentage of sales were correctly classified in the east- 

 side National Forests. In general, the eastside equations 

 had fewer variables than the westside equations. 



Table 9 — Categorical analysis of variance results^ 



Source 



df 



Chi-square 



Probability 



Intercept 

 Region^ 

 Gate" 

 Method^ 



1 

 1 

 2 

 1 



2803.82 

 10.16 

 27.51 

 .08 



0.0000 

 3.0014 

 ^0000 

 .7724 



'Interaction terms were not significant. 



Variable that represents the eastside and westside geographical 

 zone. 



^Indicates a statistically significant factor. 

 ^Variable that represents gates 1 , 2-3, and 4. 

 'Variable that represents the statistical methods. 



8 



