discontinuous growth models paul d. bliese walter reed army institute of research
DESCRIPTION
Discontinuous Growth Models Paul D. Bliese Walter Reed Army Institute of Research. Purpose. Introduce discontinuous growth models Illustrate where these models might be used in organizational research Provide details on how to set up the level-1 time variables - PowerPoint PPT PresentationTRANSCRIPT
Discontinuous Growth Models
Paul D. BlieseWalter Reed Army Institute of Research
Purpose
• Introduce discontinuous growth models• Illustrate where these models might be used in
organizational research• Provide details on how to set up the level-1 time
variables• List resources for researchers interested in applying
the models
Outline
• Review of basic growth models• Discontinuities in longitudinal data• Sleep changes over 27 days• Task-change paradigm to study adaptability• Specifying time in the models
– Dummy code for transition– Dummy code and post-transition slope– Adding quadratic terms
• Practical advice• Other applications of model• References
Review of Basic Growth Model
• Growth modeling is widely applied to the analyses of longitudinal data
• Useful when trends are expected and where there are no anticipated discontinuities between measurement intervals– New employees’ skill acquisition– Individuals’ performance when learning a new task– Sales growth in start-up organizations– Changes in childrens’ height with age
• Growth models– Test for individual differences in the outcome– Variability in the rate of change over time
Review: Heights of Boys in Oxford
Age
Hei
ght
130
140
150
160
170
-1.0 -0.5 0.0 0.5 1.0
10 9
-1.0 -0.5 0.0 0.5 1.0
2 6
-1.0 -0.5 0.0 0.5 1.0
7
17 16 15 8
130
140
150
160
170
20
130
140
150
160
170
1 18 5 11 3
12
-1.0 -0.5 0.0 0.5 1.0
13 14
-1.0 -0.5 0.0 0.5 1.0
19
130
140
150
160
170
4
Review: Heights of Boys in Oxford
• Average height reliably varies across boys– ICC(1) = 0.74
• Slopes randomly vary
-1.0 -0.5 0.0 0.5 1.0
13
01
40
15
01
60
17
0
Age
heig
ht
-1.0 -0.5 0.0 0.5 1.0
13
01
40
15
01
60
17
0
Age
heig
ht
Review: Heights of Boys in Oxford
• Growth model can explain differences in overall height and differences in slopes as a function of level-2 characteristics such as:– Genetics– Nutrition– Average height of parents
• Typical growth models work in these cases because– There is an underlying trend to the data– There is no explicit transition point between measurement
intervals where change takes on a distinct non-linear pattern
Discontinuities in Longitudinal Data
• In many situations, events representing distinct transition points occur during longitudinal data collection
• Events can be planned– A new HR initiative targeting turnover rates– An intended but unexpected change in the nature of the task
in research on individual adaptability
• And unplanned– The unexpected passage of an economic stimulus package
during a study of individual consumer spending
Discontinuities in Longitudinal Data
• In longitudinal data, these transition points may:– Be the topic of research interest– Mask sub-trends in the overall growth pattern
• Even longitudinal data without any apparent growth may reveal important information when transition points are examined– Over a 10 year period, sales in established markets may be
flat, but nonetheless contain important information about a variety of distinct events
• New marketing initiatives
• Changes in management
– Over a 27-day period, adult sleep patterns may not change, but may nonetheless mask information about transitions
Sleep Changes over Time (Bliese et al., 2007)
• Minutes of sleep over 27 days (NS linear trend)• On surface, not a good candidate for growth model
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25Day
Min
utes
of S
leep
All Participants
Sleep Changes over Time
• Model with a term for the transition from sleeping in barracks to sleeping in a field exercise setting
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25Day
Min
utes
of S
leep
All Participants
Sleep Changes over Time
• The transition point masked underlying patterns in the longitudinal data.– Significant pre-change slope (p<.05)– Significant transition (p<.10)– Significant post-change slope (p<.05)
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25
Day
Min
utes
of S
leep
All Participants
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25
Day
Min
utes
of S
leep
All Participants
Sleep Changes over Time
• The form of the relationship is partially captured in a quadratic trend– Quadratic term approximates the nature of the data– Misses the distinct transition phase
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25
Day
Min
utes
of S
leep
All Participants
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25
Day
Min
utes
of S
leep
All Participants
Individual Differences in Sleep Patterns
• Importantly, individuals can differ in each term of the discontinuous growth model
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25Day
Min
utes
of S
leep
All Participants Participant A Participant B
Individual Differences in Sleep Patterns
• Individual differences can be predicted using level-2 variables (such as participant age).
260
280
300
320
340
360
380
400
0 2 4 6 8 10 12 14 16 18 20 22 24Day
Min
utes
of S
leep
22 Year Old (Average Age)20 Year Old25 Year Old
Task-Change Paradigm
• Research on adaptability frequently uses the task-change paradigm
• Task-change paradigm– Uses complex tasks– Nature of task unexpectedly changes– Tasks are not learned to asymptotic performance prior to
change• Produces large individual differences
• Individual response to change is used to make inferences about adaptability
Task-Change Paradigm (Lang, 2007)
• Performance and the task change paradigm– Task unexpectedly changes after 6th trial
-10
-8
-6
-4
-2
0
2
4
6
8
10
0 1 2 3 4 5 6 7 8 9 10 11Trial
Per
form
ance
Overall Trend
Task-Change Paradigm (Lang, 2007)
• Data contain strong individual differences in transition parameter
-15
-10
-5
0
5
10
15
0 1 2 3 4 5 6 7 8 9 10 11Trial
Per
form
ance
Individual "01122105"Overall TrendIndividual "s10r8"
Task-Change Paradigm
• Theoretically, Lang & Bliese (in review) conceptualize adaptability as:– Response to transition– Re-acquisition slope
• These two elements of adaptability are separate from– Basal performance (overall performance)– Initial skill acquisition
• Discontinuous growth model provides a framework to accurately mirror what is occurring in the experiment
Specifying the Level-1 Time Model
• Details of the discontinuous growth model are provided in Chapter 5 of Singer & Willett (2003)– Singer, J. D., & Willett, J. B. (2003). Applied Longitudinal
Data Analysis: Modeling Change and Event Occurrence. Oxford University Press.
• The remainder of this lecture provides the basic foundation for understanding how models can be specified.
• Focus on specifying the level-1 time component of the models
Coding Time in Growth Model
• Growth models are one specific form of a class of mixed-effects models for longitudinal data.
• The standard way of coding time in growth models is by using a vector from 0 to n observations.
Subject time height1 0 140.51 1 143.41 2 144.81 3 147.11 4 147.71 5 150.21 6 151.71 7 153.31 8 155.8
Age
Hei
ght
130
140
150
160
170
-1.0 -0.5 0.0 0.5 1.0
10 9
-1.0 -0.5 0.0 0.5 1.0
2 6
-1.0 -0.5 0.0 0.5 1.0
7
17 16 15 8
130
140
150
160
170
20
130
140
150
160
170
1 18 5 11 3
12
-1.0 -0.5 0.0 0.5 1.0
13 14
-1.0 -0.5 0.0 0.5 1.0
19
130
140
150
160
170
4
Coding Time in Mixed-Effects Models
• Coding time as a vector with a linear trend line is an efficient way (in terms of Degrees of Freedom) to account for time effects.– 1 DF for fixed effects, 3 for random effects
• The other extreme is model time as a set of n-1 dummy codes– 8 DF for fixed effects– Lots and lots of random effects
Coding Time for Only Discontinuity
• Coding for discontinuity involves adding additional vectors to the level-1 predictor matrix
• The simplest way to model discontinuity is to add a dummy code that is 0 before the change and 1 after the change
• Coding of Lang & Bliese task-change paradigm data
ID TIME TRANS1 0 01 1 01 2 01 3 01 4 01 5 01 6 11 7 11 8 11 9 11 10 11 11 1
Coding for Only Discontinuity
• The fixed-effects table indicates whether the transition point represents a significant change
• Task-change paradigm data (Lang, 2007) shows– A significant linear increase in performance– A significant decline when the task was unexpectedly
changed
Value Std.Error DF t-value p-value
(Intercept) -2.16 0.59 2022 -3.65 0.00
TIME 1.20 0.09 2022 13.42 0.00
TRANS -4.37 0.62 2022 -7.06 0.00
Coding for Only Discontinuity
• Potential limitation is that this model specification restricts pre-slope and post-slopes to be equal– Visual representation of Lang & Bliese adaptability data and
Bliese et al., 2007 sleep data
• Random effects for sleep data shown (significant)
-10
-8
-6
-4
-2
0
2
4
6
8
10
0 1 2 3 4 5 6 7 8 9 10 11Trial
Per
form
ance
All Participants
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25Day
Min
utes
of S
leep
Participant A All Participants Participant B
Coding for Discontinuity and Slope Differences
• Adding one more vector to the level-1 model provides a way to determine if the post-transition slope varies from the pre-transition slope
• The vector that was previous labeled “TIME” now represents the pre-transition slope
ID PRE TRANS POST1 0 0 01 1 0 01 2 0 01 3 0 01 4 0 01 5 0 01 6 1 01 7 1 11 8 1 21 9 1 31 10 1 41 11 1 5
Coding for Discontinuity and Slope Differences
• Task-change paradigm data model reveals:– A significant linear increase in pre-change performance– A significant decline when the task was unexpectedly
changed– A post-change slope that is significantly smaller than the
pre-change slope
Value Std.Error DF t-value p-value
(Intercept) -3.69 0.63 2021 -5.84 0.00
PRE SLOPE 1.81 0.13 2021 14.46 0.00
TRANS -4.98 0.62 2021 -8.05 0.00
POST SLOPE -1.22 0.18 2021 -6.88 0.00
Coding for Discontinuity and Slope Differences
• Graphs below contrast task-change paradigm data and sleep data using:– The model with only a transition variable– The model with both a transition and post-transition slope
variable
-10
-8
-6
-4
-2
0
2
4
6
8
10
0 1 2 3 4 5 6 7 8 9 10 11Trial
Per
form
ance
Slope and TransitionTransition Only
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25Day
Min
utes
of S
leep
Slope and TransitionTransition Only
Individual Differences in Sleep Patterns
• Individual differences can exist for the:– Pre-transition slope– Transition– Post-transition slope
• Differences can be modeled with level-2 variables
260
280
300
320
340
360
380
400
1 3 5 7 9 11 13 15 17 19 21 23 25Day
Min
utes
of S
leep
All Participants Participant A Participant B
260
280
300
320
340
360
380
400
0 2 4 6 8 10 12 14 16 18 20 22 24Day
Min
utes
of S
leep
22 Year Old (Average Age)20 Year Old25 Year Old
Flexibility of Coding: Quadratic Effects
• The approach has considerable flexibility and can incorporate curvilinear effects
• Lang & Bliese data with quadratic terms
ID TIME TRANS POST QUADPRE QUADPOST1 0 0 0 0 01 1 0 0 1 01 2 0 0 4 01 3 0 0 9 01 4 0 0 16 01 5 0 0 25 01 6 1 0 25 01 7 1 1 25 11 8 1 2 25 41 9 1 3 25 91 10 1 4 25 161 11 1 5 25 25
Flexibility of Coding: Quadratic Effects
• Using this coding, the fixed-effects model identifies a significant quadratic component to the pre-transition slope
• Post-transition slope has no significant quadratic form
Value Std.Error DF t-value p-value
(Intercept) -4.73 0.69 2019 -6.83 0.00
PRE SLOPE 3.37 0.45 2019 7.57 0.00
TRANS -5.53 0.69 2019 -7.96 0.00
POST SLOPE -2.74 0.63 2019 -4.35 0.00
PRE QUADRATIC -0.31 0.09 2019 -3.65 0.00
POST QUADRATIC -0.01 0.09 2019 -0.09 0.93
Flexibility of Coding: Quadratic Effects
• Graphs shows typical learning curve prior to task change
-10
-8
-6
-4
-2
0
2
4
6
8
10
0 1 2 3 4 5 6 7 8 9 10 11
Trial
Per
form
ance
Quadratic Model
Practical Advice on Estimating Models
• With flexibility comes complexity!• The goal of these models is typically to explain
individual-level variability in the model parameters• As level-1 growth parameters increase, the number
of random effects also increases– With three growth parameters (pre-linear, transition, and
post-linear), the methods must estimate 4 variances and 6 covariances
– Practically speaking, models often fail to converge when numerous random terms are included
• Choice of how to specify the random components of the model must be guided by theory and a systematic approach to examining the model
Practical Advice on Estimating Models
• Recommend following a model specification strategy such as that outlined by Bliese & Ployhart (2002).– Estimate the ICC for the outcome– Identify the significant fixed effects for time– Identify which effects for time randomly vary across
individuals– Determine whether other adjustments are needed to level-1
error structure (e.g., autocorrelation)– Include level-2 predictors of randomly varying level-1 effects.
Do not rely only on empirical results of step 3. Also use theory as a guide.
Some Other Applications
• Lang & Kersting (2007) used a discontinuous model with 4 data points to examine teachers’ effectiveness ratings after implementing feedback
2.10
2.15
2.20
2.25
2.30
2.35
2.40
2.45
2.50
Time 1 Time 2 Time 3 Time 4
Teacher Effectiveness
Rat
ing
Teacher TRANSPOST
TRANS1 0 01 1 01 1 11 1 22 0 02 1 02 1 12 1 2
Some Other Applications
• Bliese, Wesensten & Balkin (2006) used the approach to mirror design elements of a sleep study
• Residual individual variance was highly significant – Indicates strong individual differences in ability to perform
during sleep restriction
240250260270280290300310320330340350360
BL E1 E2 E3 E4 E5 E6 E7 R1 R2 R3
Study Phase
Rea
ctio
n T
ime
(in
mse
c)
2.96 Hrs TST4.80 Hrs TST6.78 Hrs TST
260
270
280
290
300
310
320
330
340
350
360
BL E1 E2 E3 E4 E5 E6 E7
Day of Phase
Rea
ctio
n T
ime
(in
mse
c)
Age 29Age 39 (Average)Age 48
Selected References
Bliese, P. D., McGurk, D., Thomas, J. L., Balkin, T. J., & Wesensten, N. (2007). Discontinuous growth modeling of adaptation to sleep setting changes: Individual differences and age. Aviation, Space, and Environmental Medicine, 78, 485-492.
Bliese, P. D., Wesensten, N., & Balkin, T. J. (2006). Age and individual variability in performance during sleep restriction. Journal of Sleep Research, 15, 376-385.
Lang, J. W. B. (2007). General Mental Ability and Two Types of Adaptation to Unforeseen Change. Dissertation. Rheinisch-Westfälische Technische Hochschule, Aachen, Germany.
Lang, J. W. B. & Bliese, P. D. (in revision). General Mental Ability and Two Types of Adaptation to Unforeseen Change: Applying Discontinuous Growth Models to the Task-Change Paradigm.
Lang, J. W. B. & Kersting, M. (2007). Regular feedback from student ratings of instruction: Do college teachers improve their ratings in the long run? Instructional Science, 35, 187-205.