> This message is in MIME format. Since your mail reader does not understand this format, some or all of this message may not be legible.
--B_3126969183_162382 Content-type: text/plain; charset="US-ASCII" Content-transfer-encoding: 7bit On 2/1/03 10:12 AM, in article 001701c2c9b1$a0e8a230$9650bc3f@DHJC4K01, "Karl L. Wuensch" <[EMAIL PROTECTED]> wrote: > My colleague's query was, however, "if you are going to dichotomize a > continuous subject variable and compare the resulting two groups on a second > continuous variable, even though that is not generally a good idea, is it more > useful to use a median split (upper half vs lower half) or to compare the > tails (such as upper third versus lower third)?" I suggested to my colleague > that this would depend, in part, on the form of the relationship between the > two continuous variables (not necessarily strictly linear), and reminded him > that throwing out the middle of the distribution would reduce N and thus might > reduce power too. My colleague confessed that he was fishing for a citation > to justify having done something that I told him earlier was not a good thing > to do. ;-) >> >> If the relationship is linear and the distribution of the variable is approximately linear, then my prior post definitely answers this question. Median split will reduce the expected r^2 to 64% of what it would have been, the extreme thirds will reduce the expected r^2 to 79% of what it would have been. If you have enough degrees of freedom to burn, then the extreme thirds is a bad idea, but not as bad an idea as median splits. If the relationship is nonlinear, then dividing into two groups, whether the extreme third tails or median splits, precludes any possibility of detecting the nonlinearity. Furthermore, Maxwell & Delaney (Psych Bulletin, 1993, 113, 181-190) demonstrate that obscuring nonlinearity in that way can produce a spurious interaction. Why anyone continues to split data after that article is beyond me, but subsequent articles like the recent MacCollum et al. article in Psych Methods (indeed a gem) remain necessary. Irwin & McClelland (Journal of Marketing Research, forthcoming) squashes another false belief that perhaps median splits are a good idea when the predictor variables are very skewed, non-normal distributions. Even in those situations, splitting the data remains a bad idea. Gary [EMAIL PROTECTED] > --B_3126969183_162382 Content-type: text/html; charset="US-ASCII" Content-transfer-encoding: quoted-printable <HTML> <HEAD> <TITLE>Re: Dichotomization</TITLE> </HEAD> <BODY> <FONT FACE=3D"Verdana">On 2/1/03 10:12 AM, in article 001701c2c9b1$a0e8a230$9= 650bc3f@DHJC4K01, "Karl L. Wuensch" <[EMAIL PROTECTED]> = wrote:<BR> <BR> </FONT><BLOCKQUOTE><FONT FACE=3D"Arial"> My col= league's query was, however, "if you are going to dichotomize a continu= ous subject variable and compare the resulting two groups on a second contin= uous variable, even though that is not generally a good idea, is it more use= ful to use a median split (upper half vs lower half) or to compare the tails= (such as upper third versus lower third)?" I suggested to my col= league that this would depend, in part, on the form of the relationship betw= een the two continuous variables (not necessarily strictly linear), and remi= nded him that throwing out the middle of the distribution would reduce N and= thus might reduce power too. My colleague confessed that he was fishi= ng for a citation to justify having done something that I told him earlier w= as not a good thing to do. ;-)<BR> </FONT><BLOCKQUOTE><FONT FACE=3D"Verdana"><BR> <BR> </FONT></BLOCKQUOTE></BLOCKQUOTE><FONT FACE=3D"Verdana">If the relationship i= s linear and the distribution of the variable is approximately linear, then = my prior post definitely answers this question. Median split will redu= ce the expected r^2 to 64% of what it would have been, the extreme thirds wi= ll reduce the expected r^2 to 79% of what it would have been. If you h= ave enough degrees of freedom to burn, then the extreme thirds is a bad idea= , but not as bad an idea as median splits.<BR> <BR> If the relationship is nonlinear, then dividing into two groups, whether th= e extreme third tails or median splits, precludes any possibility of detecti= ng the nonlinearity. Furthermore, Maxwell & Delaney (Psych Bulleti= n, 1993, 113, 181-190) demonstrate that obscuring nonlinearity in that way c= an produce a spurious interaction. Why anyone continues to split data = after that article is beyond me, but subsequent articles like the recent Mac= Collum et al. article in Psych Methods (indeed a gem) remain necessary. &nbs= p;Irwin & McClelland (Journal of Marketing Research, forthcoming) squash= es another false belief that perhaps median splits are a good idea when the = predictor variables are very skewed, non-normal distributions. Even in= those situations, splitting the data remains a bad idea.<BR> <BR> Gary<BR> [EMAIL PROTECTED]<BR> </FONT><BLOCKQUOTE><FONT FACE=3D"Verdana"><BR> </FONT></BLOCKQUOTE><FONT FACE=3D"Verdana"><BR> </FONT> </BODY> </HTML> --B_3126969183_162382-- . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
