Hi Dot, This specification would yield a single coefficient for the between-individual and within-individual effects. That is, you?re assuming the association is the same over time as it is across individuals at a single point in time. I wouldn?t expect this to be a safe assumption, and there?s a pretty straightforward fix: centre your time-varying predictors by their mean for each person. That will yield within effects equivalent to what you?d get from a fixed effects model. For more information about this, see: https://www.cambridge.org/core/services/aop-cambridge-core/content/view/S2049847014000077 and/or https://www.researchgate.net/publication/299604336_Fixed_and_Random_effects_models_making_an_informed_choice Hope that?s useful, Malcolm Malcolm Fairbrother Professor of Sociology Ume? University<http://www.umu.se/english> Sweden Date: Thu, 1 Feb 2018 21:07:10 +1030 From: Dot Dumuid <haveaballphysio at gmail.com<mailto:haveaballphysio at gmail.com>> To: r-sig-mixed-models at r-project.org<mailto:r-sig-mixed-models at r-project.org> Subject: [R-sig-ME] model specification for repeated measure Dear mixed model experts, We have a dataset of older adults. We measured their mental health (MH) 6 months before retirement and again 12 months post retirement. At both of these time points we also measured their physical activity (PA) (min/day), income (INC) and general health (GH). We would like to create a model that tells us if change in physical activity over the retirement threshold predicts change in mental health, and we'd like to use the model to predict how much mental health is predicted to change when physical activity is increased from perhaps 15 minutes to 60 minutes. We'd like to use a mixed model rather than just using change (difference) scores. And we'd like to control for things like change in general physical health and change in income. This is what the data look like *ID time MH PA GH INC* 01 pre 4 15 56 560 02 pre 5 30 30 1200 .. ..... .. .. .. ... 01 post 7 40 50 50 02 post 8 45 30 0 I'm not sure how best to build the model. Something like this? model <- lmer (MH ~ PA * time + GH + INC + (1|participant.ID) ) Thank you in advance. Dot
model specification for repeated measure
3 messages · Malcolm Fairbrother, Dot Dumuid
Thanks for the suggestions. I greatly appreciate you taking the time, and I look forward to trying out the ideas. Thanks, Dot On Fri, Feb 2, 2018 at 5:53 AM, Malcolm Fairbrother <
malcolm.fairbrother at umu.se> wrote:
Hi Dot, This specification would yield a single coefficient for the between-individual and within-individual effects. That is, you?re assuming the association is the same over time as it is across individuals at a single point in time. I wouldn?t expect this to be a safe assumption, and there?s a pretty straightforward fix: centre your time-varying predictors by their mean for each person. That will yield within effects equivalent to what you?d get from a fixed effects model. For more information about this, see: https://www.cambridge.org/core/services/aop-cambridge-core/content/view/ S2049847014000077 and/or https://www.researchgate.net/publication/299604336_Fixed_ and_Random_effects_models_making_an_informed_choice Hope that?s useful, Malcolm Malcolm Fairbrother Professor of Sociology Ume? University <http://www.umu.se/english> Sweden Date: Thu, 1 Feb 2018 21:07:10 +1030 From: Dot Dumuid <haveaballphysio at gmail.com> To: r-sig-mixed-models at r-project.org Subject: [R-sig-ME] model specification for repeated measure Dear mixed model experts, We have a dataset of older adults. We measured their mental health (MH) 6 months before retirement and again 12 months post retirement. At both of these time points we also measured their physical activity (PA) (min/day), income (INC) and general health (GH). We would like to create a model that tells us if change in physical activity over the retirement threshold predicts change in mental health, and we'd like to use the model to predict how much mental health is predicted to change when physical activity is increased from perhaps 15 minutes to 60 minutes. We'd like to use a mixed model rather than just using change (difference) scores. And we'd like to control for things like change in general physical health and change in income. This is what the data look like *ID time MH PA GH INC* 01 pre 4 15 56 560 02 pre 5 30 30 1200 .. ..... .. .. .. ... 01 post 7 40 50 50 02 post 8 45 30 0 I'm not sure how best to build the model. Something like this? model <- lmer (MH ~ PA * time + GH + INC + (1|participant.ID) ) Thank you in advance. Dot
6 days later
I'm sorry to come back to this question (see below for original question)... but I have one more thing to ask. Because I am *not* interested in whether mental health *changes* between the two time points, but I am only interested in whether physical activity predicts mental health (when controlling for other things like general health and income), is it true that I don't put "time" in the model as a fixed effect? By using a random intercept for ID, the model already caters for the fact that observations are matched by ID. So my model could just be: model <- lmer (MH ~ PA + GH + INC + (1|participant.ID) ) (Assuming, off course, no assumptions are violated,...and I am ignoring random slopes at the moment) Then I can predict MH for various min/day of PA (e.g., mean baseline PA, mean baseline PA+10, mean baseline PA+20... etc), keeping GH and INC constant. Subtracting the new predicted MHs from the mean predicted MH should tell me how MH is predicted to change when PA changes across the retirement transition. Does this sound OK? Thank you, Dot (This is the original question....) *Dear mixed model experts,* *We have a dataset of older adults. We measured their mental health (MH) 6 months before retirement and again 12 months post retirement.* *At both of these time points we also measured their physical activity (PA) (min/day), income (INC) and general health (GH).* *We would like to create a model that tells us if change in physical activity over the retirement threshold predicts change in mental health, and we'd like to use the model to predict how much mental health is predicted to change when physical activity is increased from perhaps 15 minutes to 60 minutes. We'd like to use a mixed model rather than just using change (difference) scores. And we'd like to control for things like change in general physical health and change in income.* *This is what the data look like* *ID time MH PA GH INC* *01 pre 4 15 56 560* *02 pre 5 30 30 1200* *.. ..... .. .. .. ...* *01 post 7 40 50 50* *02 post 8 45 30 0 * *I'm not sure how best to build the model. Something like this?* *model <- lmer (MH ~ PA * time + GH + INC + (1|participant.ID) )* *Thank you in advance.* On Fri, Feb 2, 2018 at 6:22 AM, Dot Dumuid <haveaballphysio at gmail.com> wrote:
Thanks for the suggestions. I greatly appreciate you taking the time, and I look forward to trying out the ideas. Thanks, Dot On Fri, Feb 2, 2018 at 5:53 AM, Malcolm Fairbrother < malcolm.fairbrother at umu.se> wrote:
Hi Dot, This specification would yield a single coefficient for the between-individual and within-individual effects. That is, you?re assuming the association is the same over time as it is across individuals at a single point in time. I wouldn?t expect this to be a safe assumption, and there?s a pretty straightforward fix: centre your time-varying predictors by their mean for each person. That will yield within effects equivalent to what you?d get from a fixed effects model. For more information about this, see: https://www.cambridge.org/core/services/aop-cambridge-core/ content/view/S2049847014000077 and/or https://www.researchgate.net/publication/299604336_Fixed_and _Random_effects_models_making_an_informed_choice Hope that?s useful, Malcolm Malcolm Fairbrother Professor of Sociology Ume? University <http://www.umu.se/english> Sweden Date: Thu, 1 Feb 2018 21:07:10 +1030 From: Dot Dumuid <haveaballphysio at gmail.com> To: r-sig-mixed-models at r-project.org Subject: [R-sig-ME] model specification for repeated measure Dear mixed model experts, We have a dataset of older adults. We measured their mental health (MH) 6 months before retirement and again 12 months post retirement. At both of these time points we also measured their physical activity (PA) (min/day), income (INC) and general health (GH). We would like to create a model that tells us if change in physical activity over the retirement threshold predicts change in mental health, and we'd like to use the model to predict how much mental health is predicted to change when physical activity is increased from perhaps 15 minutes to 60 minutes. We'd like to use a mixed model rather than just using change (difference) scores. And we'd like to control for things like change in general physical health and change in income. This is what the data look like *ID time MH PA GH INC* 01 pre 4 15 56 560 02 pre 5 30 30 1200 .. ..... .. .. .. ... 01 post 7 40 50 50 02 post 8 45 30 0 I'm not sure how best to build the model. Something like this? model <- lmer (MH ~ PA * time + GH + INC + (1|participant.ID) ) Thank you in advance. Dot