full2010.pdf - page 414

376
Ášµ„n ´ 0.5 ×¥ªœÄ®‡µn n n ¦o°¥¨³ °Š‹» —Š‹³¤¸‡µ¨—¨Š n n ¨³š¸É‡µ°ºÉ œÇ n ×¥ªœÄ®‡µn n n ¦o°¥¨³ °Š‹» —Š‹³¤¸‡µ n n
¨—¨Š‹œ™¹Š¦³—´‡ªµ¤´¤¡´œ›r¦³®ªµŠn ˜´ªÂž¦°· ¦³Ášµ„ n ´ 0.67 ¨³‹³Á¡·¤ ¹œÁ¨È„œo°¥®¨´Š‹µ„œ´œ É Ê Ê
„¦–¸š¸É‹Îµœªœ˜´ªÂž¦°· ¦³, ´—ªœ ° n Š‡ªµ¤¨o¤Á®¨ª °Š¨´ „¬–³š¸É œÄ‹«¹„¬µÂ¨³¦³—´‡ªµ¤´ ¤¡´œ›r
¦³®ªµŠ˜´ªÂž¦°· ¦³‡Šš¸É n ¡ªµn Á¤ºÉ ° œµ—˜´ª°¥µŠÁ¡·¤ ¹œ n É Ê š¸É ´—ªœ °Š‡ªµ¤¨o¤Á®¨ª °Š¨´„¬–³š¸É œÄ‹«¹„¬µ¤¸ ‡µ n n
Ášµ„n ´ 0.1 ×¥ªœÄ®‡µn n n ¦o°¥¨³ °Š‹»—Š¤¸‡µ¨—¨Š n n , š¸É 0.5 ‡µn ¦o°¥¨³ °Š‹»—ŠÅ¤Â˜„˜µŠ„œ¤µ„œ´n n n ´ „ ×¥‡µ °Š n
‹»—Š¤¸‡µž¦³¤µ– n n 0.4623 ¨³š¸É 0.9 ×¥ªœÄ®‡µn n n ¦o°¥¨³ °Š‹»—Š¤¸ ‡µ¼Š ¹œ n n Ê
„¦–¸š¸Éœµ—˜´ª°¥µŠn , ¦³—´‡ªµ¤´ ¤¡´œ›r¦³®ªµŠ˜´ªÂž¦°· ¦³Â¨³ n ´—ªœ °Š‡ªµ¤¨o¤Á®¨ª °Š¨´„¬–³š¸É n
œÄ‹«¹„¬µ‡Šš¸É ¡ªµÁ¤ºÉ°‹Îµœªœ˜´ªÂž¦°· ¦³Á¡·¤ ¹ œ n É Ê Ã—¥ªœÄ®‡µn n n ¦o°¥¨³ °Š‹»—Š¤¸n ‡µ¨—¨ n Š
Ÿ¨„µ¦ª·Á‡¦µ³®r „µ¦™—™°¥Á·Š¡®» š¸É¤¸Ÿ¨°´œ˜¦„¦· ¥µ¡ªµ‡µ °Š´¤ž¦³· š›·„µ¦˜´—· œÄ‹ · n n Í (R2) ¤¸ ‡µ¼Š ‡º n °
Ášµ„n ´ 0.907 —´Šœ´œ¤„µ¦œ¸‹¹Š Ê Ê µ¤µ¦™œÎµÅžÄož¦³¤µ–®µ‡µ‹»—Šš¸ÉÁ®¤µ³¤š¸É »—Äœ™µœ„µ¦–r°ºÉ œ Ç Å—o n n
‡ÎµÎ µ‡´ : „µ¦™—™°¥Ã¨‹· ˜·„ 2 ž¦³Á£š °´˜¦µ‡ªµ¤Ÿ·—¡¨µ—Äœ„µ¦‹ÎµÂœ„„¨»¤n ‹»—Š n
Abstract
The objective of this study is to find out the optimal cut-off point for predictive classification of ungrouped
data using binary logistic regression model. The factors interested are number of independent variables (p), sample size
(n), proportion of failure (a), and degree of multicollinearity among independent variable (M). The data are generated
using Monte Carlo technique through R-program. The independent data are simulated having uniform distribution, the
number of independent variables are generated having 3 levels; low level (p=1, 2), medium level (p= 3, 4), and high
level (p=5, 6), the sample size are also generated having 3 levels; low level (n=20, 40) medium level (n=60, 80), and
high level (n=100, 120), the proportion of failure are generated using a equal to 0.1, 0.5 and 0.9, and the degree of
multicollinearity are generated with M equal to 0, 0.33, 0.67 and 0.99. Each situation is repeatedly 500 times. The
cut-off point is captured using Hadjicostas P. theory. The average of cut-off points is computed from all runs in each
situation as the optimal cut-off point for the situation. The optimal cut-off point when provides the minimum
classification error rate. The results can be summarized as follow:
As the sample size, the number of independent variables, and the degree of multicollinearity are kept
constant, with the low level of the number of independent variables (p=1, 2), the optimal cut-off point will decrease
when the proportion of failure increases. It is also found that when the number of independent variables are medium
and high level (p=3, 4, 5, and 6), the optimal cut-off point will be the maximum cut-off point at the proportion of
failure equal to 0.5, this cut-off point is in the neighborhood of 0.4542.