# Repeated predictor variable analysis

8 messages
Open this post in threaded view
|

## Repeated predictor variable analysis

 good day all, I am trying to figure out the appropriate model following the problem.Say Participants are enrolled in a study and baseline covariates measured along with a predictor variable (but not the outcome variable). At Time 2, 3,4 (say),  The predictor variable is again measured. The outcome variable is then measured at future time 5 (say). Interest is to determine at which time point is best (in terms of predictive power) to measure the predictor.Below is an example of data. COVAR1, COVAR2, and Predt1 are measured at the same time while PredT2 is measured some times later - say after 5 weeks while Response is measured again sometimes after PredT2. Time difference between measurements is not fixed. Reference articles would also be appreciated. ThanksForchehid,COVAR1,COVAR2, PredT1,PredT2,Response1,22,1,90,36,752,28,1,49,49,903,25,1,62,16,914,23,0,22,66,695,32,1,66,87,856,24,0,77,18,657,31,1,48,76,868,31,0,70,80,649,26,1,17,59,7810,22,1,66,56,9211,30,0,4,44,6912,24,1,17,86,6213,30,0,55,59,5814,30,1,96,58,8415,26,1,22,33,9916,30,0,14,47,58 ===================== 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
Open this post in threaded view
|

## Re: Repeated predictor variable analysis

 Setting aside covariates, consider fitting a random coefficient model [using the MIXED procedure], with time incorporated as a fixed effect term and a random effect term, and the predictor variable treated as y:Level 1 Equation:y = B0J + B1J*time + eijLevel 2 Equations:B0J = Gamma00 + u0jB1J = Gamma10 + u1jFull Equation:y = Gamma00 +    + Gamma10*time    + ( u1j*time + u0j + eij )Next, regress your actual y on u0j and u1j, perhaps allowing u0j and u1j to covary [using AMOS]. This approach will allow you to evaluate the extent to which: (1) your predictor is related to y at baseline, u0j effect, (assuming you code time such that ‘0’ is BL);(2) change in your predictor (over time) is related to your y, u1j effect. Note: You could change the coding of time such that the effect of U0j is associated with a different time point. RyanSent from my iPhoneOn Aug 9, 2020, at 8:32 PM, Nkem Ntonghanwah <[hidden email]> wrote:﻿good day all, I am trying to figure out the appropriate model following the problem.Say Participants are enrolled in a study and baseline covariates measured along with a predictor variable (but not the outcome variable). At Time 2, 3,4 (say),  The predictor variable is again measured. The outcome variable is then measured at future time 5 (say). Interest is to determine at which time point is best (in terms of predictive power) to measure the predictor.Below is an example of data. COVAR1, COVAR2, and Predt1 are measured at the same time while PredT2 is measured some times later - say after 5 weeks while Response is measured again sometimes after PredT2. Time difference between measurements is not fixed. Reference articles would also be appreciated. ThanksForchehid,COVAR1,COVAR2, PredT1,PredT2,Response1,22,1,90,36,752,28,1,49,49,903,25,1,62,16,914,23,0,22,66,695,32,1,66,87,856,24,0,77,18,657,31,1,48,76,868,31,0,70,80,649,26,1,17,59,7810,22,1,66,56,9211,30,0,4,44,6912,24,1,17,86,6213,30,0,55,59,5814,30,1,96,58,8415,26,1,22,33,9916,30,0,14,47,58 ===================== 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
Open this post in threaded view
|

## Re: Repeated predictor variable analysis

 Why not just use all the predictor time point values in ordinary regression?  There aren't that many of them.On Sun, Aug 9, 2020 at 7:36 PM Ryan Black <[hidden email]> wrote:Setting aside covariates, consider fitting a random coefficient model [using the MIXED procedure], with time incorporated as a fixed effect term and a random effect term, and the predictor variable treated as y:Level 1 Equation:y = B0J + B1J*time + eijLevel 2 Equations:B0J = Gamma00 + u0jB1J = Gamma10 + u1jFull Equation:y = Gamma00 +    + Gamma10*time    + ( u1j*time + u0j + eij )Next, regress your actual y on u0j and u1j, perhaps allowing u0j and u1j to covary [using AMOS]. This approach will allow you to evaluate the extent to which: (1) your predictor is related to y at baseline, u0j effect, (assuming you code time such that ‘0’ is BL);(2) change in your predictor (over time) is related to your y, u1j effect. Note: You could change the coding of time such that the effect of U0j is associated with a different time point. RyanSent from my iPhoneOn Aug 9, 2020, at 8:32 PM, Nkem Ntonghanwah <[hidden email]> wrote:﻿good day all, I am trying to figure out the appropriate model following the problem.Say Participants are enrolled in a study and baseline covariates measured along with a predictor variable (but not the outcome variable). At Time 2, 3,4 (say),  The predictor variable is again measured. The outcome variable is then measured at future time 5 (say). Interest is to determine at which time point is best (in terms of predictive power) to measure the predictor.Below is an example of data. COVAR1, COVAR2, and Predt1 are measured at the same time while PredT2 is measured some times later - say after 5 weeks while Response is measured again sometimes after PredT2. Time difference between measurements is not fixed. Reference articles would also be appreciated. ThanksForchehid,COVAR1,COVAR2, PredT1,PredT2,Response1,22,1,90,36,752,28,1,49,49,903,25,1,62,16,914,23,0,22,66,695,32,1,66,87,856,24,0,77,18,657,31,1,48,76,868,31,0,70,80,649,26,1,17,59,7810,22,1,66,56,9211,30,0,4,44,6912,24,1,17,86,6213,30,0,55,59,5814,30,1,96,58,8415,26,1,22,33,9916,30,0,14,47,58 ===================== 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 -- Jon K Peck[hidden email] ===================== 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
Open this post in threaded view
|

## Re: Repeated predictor variable analysis

 In reply to this post by Nkem Ntonghanwah ﻿I assumed there were more than three time points based on “(say)”. If there are only three time points, what Jon suggests seems reasonable to me. OTOH, if there are many time points, I’d consider an alternative approach such as the one in my initial reply.Sent from my iPhoneOn Aug 9, 2020, at 9:50 PM, Jon Peck <[hidden email]> wrote:﻿Why not just use all the predictor time point values in ordinary regression?  There aren't that many of them.On Sun, Aug 9, 2020 at 7:36 PM Ryan Black <[hidden email]> wrote:Setting aside covariates, consider fitting a random coefficient model [using the MIXED procedure], with time incorporated as a fixed effect term and a random effect term, and the predictor variable treated as y:Level 1 Equation:y = B0J + B1J*time + eijLevel 2 Equations:B0J = Gamma00 + u0jB1J = Gamma10 + u1jFull Equation:y = Gamma00 +    + Gamma10*time    + ( u1j*time + u0j + eij )Next, regress your actual y on u0j and u1j, perhaps allowing u0j and u1j to covary [using AMOS]. This approach will allow you to evaluate the extent to which: (1) your predictor is related to y at baseline, u0j effect, (assuming you code time such that ‘0’ is BL);(2) change in your predictor (over time) is related to your y, u1j effect. Note: You could change the coding of time such that the effect of U0j is associated with a different time point. RyanSent from my iPhoneOn Aug 9, 2020, at 8:32 PM, Nkem Ntonghanwah <[hidden email]> wrote:﻿good day all, I am trying to figure out the appropriate model following the problem.Say Participants are enrolled in a study and baseline covariates measured along with a predictor variable (but not the outcome variable). At Time 2, 3,4 (say),  The predictor variable is again measured. The outcome variable is then measured at future time 5 (say). Interest is to determine at which time point is best (in terms of predictive power) to measure the predictor.Below is an example of data. COVAR1, COVAR2, and Predt1 are measured at the same time while PredT2 is measured some times later - say after 5 weeks while Response is measured again sometimes after PredT2. Time difference between measurements is not fixed. Reference articles would also be appreciated. ThanksForchehid,COVAR1,COVAR2, PredT1,PredT2,Response1,22,1,90,36,752,28,1,49,49,903,25,1,62,16,914,23,0,22,66,695,32,1,66,87,856,24,0,77,18,657,31,1,48,76,868,31,0,70,80,649,26,1,17,59,7810,22,1,66,56,9211,30,0,4,44,6912,24,1,17,86,6213,30,0,55,59,5814,30,1,96,58,8415,26,1,22,33,9916,30,0,14,47,58 ===================== 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 -- Jon K Peck[hidden email] ===================== 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
Open this post in threaded view
|

## Re: Repeated predictor variable analysis

Open this post in threaded view
|