standardized values in multilevel
Locked
standardized values in multilevel
 
|
I know this is a controversial request because I have seen Cam's (and
other's) citations on both Multilevel and Semnet on this topic and I don't want to have that discussion. I'd like a response to the technical question of whether (and how) standardized values of fixed effects can be computed given spss (I may have missed it but I don't think mixed can output those values). If there is a computational citation, that would be fine. Concretely, given Mixed f with b c/fixed b c/print solution/random intercept | subject(xx) covtype(id). I get unstandardized regression coefficients for b and c. How do I standardize them? I know this is possible in mplus. I'd like to do it in spss. Thanks, Gene Maguin ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Re: standardized values in multilevel
Administrator
|
Does running the analyses using Z scores on the dep and pred variables suffice?
Or maybe I am missing something as I have not followed the 'controversy'. input program. loop XX=1 to 50. compute b=normal(1). Compute c=normal(1). compute f=b*1.5 + c*3 +normal(.5). end case. end loop. end file. end input program. Mixed f with b c/fixed b c/print solution /random intercept | subject(xx) covtype(id). DESCRIPTIVES VARIABLES=f b c /SAVE. Mixed zf with zb zc/fixed zb zc/print solution /random intercept | subject(xx) covtype(id). regression dep f / method enter b c. ** USING RAW ** Estimates of Fixed Effects Parameter Estimate Std. Error Intercept -.2177999 .0632562 B 1.3827831 .0602746 C 3.0975385 .0602287 a Dependent Variable: F. ** USING Z-SCORES **. Estimates of Fixed Effects Parameter Estimate Std. Error Intercept -6.1015506E-18 .0185465 ZB .4393847 .0191525 ZC .9850044 .0191525 a Dependent Variable: ZF Zscore(F). **USING REGRESSION**. Coefficients Model Unstandardized Coefficients Standardized Coefficients B Std. Error Beta 1 (Constant) -.217800 .063256 B 1.382783 .060275 .439385 C 3.097539 .060229 .985004 a Dependent Variable: F
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
Re: standardized values in multilevel
|
|
I have not followed the controversy either, but can say that federal education evaluations use just this all the time, and it's considered not just acceptable but often necessary or preferable to other approaches, including using the native scale values.
With SPSS, I only know to do this by converting the values to standardized scores (Z-scores) beforehand, and working from there. In many cases, this is preferable to letting the program do it, as it allows you to choose how the Z-score is calculated. For instance, when I'm working with State standardized test scores, I don't want to use the Mean and SD of the sample, I want it for the population, and I have it for the population, but SPSS (nor any other program I know of) could do that without me converting it as such. Matthew J Poes Research Data Specialist Center for Prevention Research and Development University of Illinois -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of David Marso Sent: Tuesday, January 03, 2012 12:13 PM To: [hidden email] Subject: Re: standardized values in multilevel Does running the analyses using Z scores on the dep and pred variables suffice? Or maybe I am missing something as I have not followed the 'controversy'. input program. loop XX=1 to 50. compute b=normal(1). Compute c=normal(1). compute f=b*1.5 + c*3 +normal(.5). end case. end loop. end file. end input program. Mixed f with b c/fixed b c/print solution /random intercept | subject(xx) covtype(id). DESCRIPTIVES VARIABLES=f b c /SAVE. Mixed zf with zb zc/fixed zb zc/print solution /random intercept | subject(xx) covtype(id). regression dep f / method enter b c. ** USING RAW ** Estimates of Fixed Effects Parameter Estimate Std. Error Intercept -.2177999 .0632562 B 1.3827831 .0602746 C 3.0975385 .0602287 a Dependent Variable: F. ** USING Z-SCORES **. Estimates of Fixed Effects Parameter Estimate Std. Error Intercept -6.1015506E-18 .0185465 ZB .4393847 .0191525 ZC .9850044 .0191525 a Dependent Variable: ZF Zscore(F). **USING REGRESSION**. Coefficients Model Unstandardized Coefficients Standardized Coefficients B Std. Error Beta 1 (Constant) -.217800 .063256 B 1.382783 .060275 .439385 C 3.097539 .060229 .985004 a Dependent Variable: F Gene Maguin wrote > > I know this is a controversial request because I have seen Cam's (and > other's) citations on both Multilevel and Semnet on this topic and I > don't want to have that discussion. I'd like a response to the > technical question of whether (and how) standardized values of fixed > effects can be computed given spss (I may have missed it but I don't > think mixed can output those values). If there is a computational > citation, that would be fine. > Concretely, given > Mixed f with b c/fixed b c/print solution/random intercept | > subject(xx) covtype(id). > I get unstandardized regression coefficients for b and c. How do I > standardize them? I know this is possible in mplus. I'd like to do it > in spss. > > Thanks, Gene Maguin > > ===================== > To manage your subscription to SPSSX-L, send a message to > LISTSERV@.UGA (not to SPSSX-L), with no body text except the command. > To leave the list, send the command SIGNOFF SPSSX-L For a list of > commands to manage subscriptions, send the command INFO REFCARD > -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/standardized-values-in-multilevel-tp5117328p5117590.html Sent from the SPSSX Discussion mailing list archive at Nabble.com. ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
In reply to this post by Maguin, Eugene
Gene, std coeff = [(unstd coeff) * (sd of x)] / (sd of y)
You could enter "standardized" variables into the linear mixed model, but keep in mind that the variance components will likely change. The formula above, along with a detailed discussion, can be found in "Multilevel Analysis: techniques and applications" by J.J. Hox. HTH, Ryan
On Tue, Jan 3, 2012 at 11:27 AM, Gene Maguin <[hidden email]> wrote: I know this is a controversial request because I have seen Cam's (and |
Re: standardized values in multilevel
|
|
Hi Gene, Hox outlines this strategy often in proximity to a discussion about centering, which can influence coefficients and the interpretation thereof. Not sure what you’re after, but Enders & Tofighi (2007) have an excellent and thorough article on centering, the implications, and how to interpret results under various strategies. From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of R B Gene, A simple approach would be to fit the linear mixed model using the original variables and then to apply the following formula: std coeff = [(unstd coeff) * (sd of x)] / (sd of y) You could enter "standardized" variables into the linear mixed model, but keep in mind that the variance components will likely change.
HTH, Ryan On Tue, Jan 3, 2012 at 11:27 AM, Gene Maguin <[hidden email]> wrote: I know this is a controversial request because I have seen Cam's (and |
Re: standardized values in multilevel
|
|
In reply to this post by Ryan
Thanks, for all the replies to my request (and the citation). I’ll do the standardized analysis if I need to but what I was hoping for was what Ryan posted. However, here is the point I’m not sure about. In the model were x is the level 2 predictor of variation of the y intercept Mixed y with x/fixed x/. . . . I can get sd of x from the univariate stats. However, for sd of y do I want the univariate stats value of y, as would be true in an ordinary one level regression, or do I need the sd of the level 1 intercept? And, if I need the sd of the level 1 intercept, how can I get that from mixed? Thanks, Gene Maguin From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of R B Gene, A simple approach would be to fit the linear mixed model using the original variables and then to apply the following formula: std coeff = [(unstd coeff) * (sd of x)] / (sd of y) You could enter "standardized" variables into the linear mixed model, but keep in mind that the variance components will likely change.
HTH, Ryan On Tue, Jan 3, 2012 at 11:27 AM, Gene Maguin <[hidden email]> wrote: I know this is a controversial request because I have seen Cam's (and |
|
|
Gene, For the record, I am generally not in favor of standardizing regression coefficients in order to assess relative importance for reasons you've probably heard a number of times.
Ryan
On Wed, Jan 4, 2012 at 9:51 AM, Gene Maguin <[hidden email]> wrote:
|
Re: standardized values in multilevel
|
|
In reply to this post by Maguin, Eugene
Gene, Raudenbush standardizes the level 2 predictor variables on the variance between level 2 units in the y variable, reasoning that level 2 variables predict the between-unit variance, rather than the within-unit variance. There is a pretty extensive “worked example” of this in Raudenbush et al (1992) On the job improvements in teacher competence, which can be accessed on Google Books. For your purposes, you would need the sd of the level 1 intercept. Best, Steve www.StatisticsDoc.com From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Gene Maguin Thanks, for all the replies to my request (and the citation). I’ll do the standardized analysis if I need to but what I was hoping for was what Ryan posted. However, here is the point I’m not sure about. In the model were x is the level 2 predictor of variation of the y intercept Mixed y with x/fixed x/. . . . I can get sd of x from the univariate stats. However, for sd of y do I want the univariate stats value of y, as would be true in an ordinary one level regression, or do I need the sd of the level 1 intercept? And, if I need the sd of the level 1 intercept, how can I get that from mixed? Thanks, Gene Maguin From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of R B Gene, A simple approach would be to fit the linear mixed model using the original variables and then to apply the following formula: std coeff = [(unstd coeff) * (sd of x)] / (sd of y) You could enter "standardized" variables into the linear mixed model, but keep in mind that the variance components will likely change.
HTH, Ryan On Tue, Jan 3, 2012 at 11:27 AM, Gene Maguin <[hidden email]> wrote: I know this is a controversial request because I have seen Cam's (and |
| Free forum by Nabble | Edit this page |