7/9/10
Introduction Growth Curves Using Mplus Alan C. Acock University Distinguished Professor of Family Studies & Knudson Chair for Family Research & Policy Oregon State University College of Health and Human Sciences Summer Workshop Series July 2010
A brief history LISREL (Joreskog and Sorbom) was being
developed in the late 1960s and released commercially in the early 1970s Originally relied on entering 8 matrices specifying
all the parameters that were being estimated or fixed at a certain value Today has a graphic interface that generates the commands from a path diagram Extremely capable alternative to Mplus 1
Alan C. Acock, July, 2010 Introduction to MPlus
1
7/9/10
A brief history EQS (Bentler) was developed much later and
replaced the matrices with writing out a separate equation for each relationship It now has a nice “Diagrammer”
2
Alan C. Acock, July, 2010
A brief history AMOS (now an SPSS product) was developed
based on a graphic interface
It has the slowest introduction of new capabilities
3
Alan C. Acock, July, 2010
2
7/9/10
A brief history--Mplus Version 6 Version 5 Version 4 Version 3 Version 2 Version 1
4
April 2010 November 2007 February 2006 March 2004 February 2001 November 1998
Alan C. Acock, July, 2010
A brief history--MPlus Very rapid development Late development allowed a non graphic
interface to be highly efficient Destroys the idea that a picture is worth 1000 words Develop statistical applications, not drawing
5
Alan C. Acock, July, 2010
3
7/9/10
A brief history--MPlus MPlus was developed during the 1990s
when growth curves were being introduced to social and behavioral sciences Need separate drawing program, but this is best for publication quality Omni Graffle (Mac) for most figures here Office Visio or Open Office Draw (PC) 6
Alan C. Acock, July, 2010
A Brief History--Mplus Bengt Muthén is the statistician Linda Muthén is the language/interface/business Several people have contributed programming Economy of Scale idea is reversed Microsoft has 40,000 programmers so it takes a long time to
make a useful change Mplus has a couple programmers so it rapidly adds features Many new features are added between versions
7
Alan C. Acock, July, 2010
4
7/9/10
Buying Mplus Greatly reduced Student prices There are three modules (they apparently learned
this module idea from SPSS).You probably want all three There is an annual maintenance and this lets you Get “free” support Get “free” updates I started with Mplus 3.0 and have only paid for the annual maintenance fee ($175) since then 8
Alan C. Acock, July, 2010
Resources for Learning Barbara Byrne. (2010). Structural Equation
Modeling with MPlus: Basic Concepts, Applications, and Programming. (Was to be available July 1 www.statmodel.com Large, 752 page, User’s Manual as pdf file Short courses on video There are 8 of these, each is one day long Download handouts to follow videos 9
Alan C. Acock, July, 2010
5
7/9/10
Resources for Learning www.ats.ucla.edu/stat/seminars/ UCLA has several online examples and videos We will utilize files from the Mplus manual for many
of our examples. These typically involve simulated data. Sometimes we well assign hypothetical variable names to make these somewhat realistic Kline, Rex. (2010). Principles & practices of structural equation modeling (3rd ed.). N.Y. Guilford 10
Alan C. Acock, July, 2010
Introduction to the Concept Growth Curves are ideal for longitudinal studies Instead of predicting a person’s score on a variable (e.g.,
mean comparison among scores at different time points or relationships among variables at different time points), we predict their growth trajectory We will present a conceptual model, show how to apply the Mplus program, and interpret the results Once we can estimate growth trajectories, the more interesting issue becomes explaining individual differences in trajectories (why some people go up, down, or stay the same) 11
Alan C. Acock, July, 2010
6
7/9/10
Introduction to the Concept We will introduce growth curves for multiple groups such as
comparing women and men Time invariant and time variant covariates will be introduced Mediation will be introduced Develop your models incrementally Start with a simple grow curve Add complexities one at a time
12
Alan C. Acock, July, 2010
Sample data Data is from the National Longitudinal Survey of Youth that
started in 1997 We use the cohort that was 12 years old in 1997 and examined their trajectory for the BMI Some may not like using the BMI on this age group, but this is only to illustrate an application of growth curve modeling The following graph of 10 randomly selected kids was produced by Mplus
13
Alan C. Acock, July, 2010
7
7/9/10
Growth trajectory of 10 randomly selected kids
14
Alan C. Acock, July, 2010
Observations about these 10 kids A BMI value of 25 is considered overweight and a BMI of 30
is considered obese (I’m aware of problems with the BMI as a measure of obesity and with its limitations when used for adolescents) With just 10 observations it is hard to see much of a trend, but it looks like adolescents are getting a higher BMI score as they get older The X-axis value of 0 is when the adolescent was 12 years old; the 1 is when the adolescent was 13 years old, etc. We are using seven waves of data (labeled 0 to 6) from the panel study 15
Alan C. Acock, July, 2010
8
7/9/10
Observations about these 10 kids Clearly the kids have different intercepts. We need to
consider if there is significant random variation in the intercept It is less clear that the kids have different slopes. We can test to see if there is a significant random variation in the slope The intercept and the slope might be correlated. Those who start with a low BMI might not increase as much as those who start with a high BMI
16
Alan C. Acock, July, 2010
A figure for how MPlus conceptualizes a growth curve
17
Alan C. Acock, July, 2010
9
7/9/10
What’s in this figure? This figure is much simpler than it first appears. The key
variables are the two latent variables labeled the Intercept growth factor and the Slope growth factor The Intercept Growth Factor The intercept represents the initial level and is sometimes called the
initial level for this reason. It is the estimated initial level and its value may differ from the actual mean for BMI97 because in this case we are imposing a linear growth model. When covariates are added, especially when a zero value on
covariates is rare and covariates are not centered (household income) 18
Alan C. Acock, July, 2010
What’s in this figure? The Intercept Growth Factor A straight line may over or underestimate any one mean
including the initial mean The intercept is identified by the constant loadings of 1.0 going to each BMI score Some programs call the intercept the constant, representing the constant effect to which other effects are added or subtracted
19
Alan C. Acock, July, 2010
10
7/9/10
What’s in this figure? The Slope Growth Factor Is identified by fixing the values of the paths to each BMI
variable. In a publication you normally would not show the path to BMI97, since this is fixed at 0.0 We fix the other paths at 1.0, 2.0, 3.0, 4.0, 5.0, and 6.0. Where did we get these values? The first year is the base year or year zero The BMI was measured each subsequent year so these are
scored 1.0 through 6.0 Other values are possible. Suppose the survey was not done in 2000 or 2001 so that we had 5 time points rather than 7. We would use paths of 0.0, 1.0, 2.0, 5.0, and 6.0 for years 1997, 1998, 1997, 2002, and 2003, respectively 20
Alan C. Acock, July, 2010
What’s in this figure? The Slope Growth Factor It is also possible to fix the first couple years and then allow
the subsequent waves to be free This might make sense for a developmental process where the yearly intervals may not reflect the developmental rate Developmental time may be quite different than chronological time
21
Alan C. Acock, July, 2010
11
7/9/10
What’s in this figure? The Slope Growth Factor This has the effect of “stretching” or “shrinking” time to the
pattern of the data (Curran & Hussong, 2003) An advantage of this approach is that it uses fewer degrees of freedom than adding a quadratic slope and can fit better Compared to a quadratic for a curve, this approach doesn’t require a monotonic function
22
Alan C. Acock, July, 2010
What’s in this figure? The Slope Growth Factor You might use 0, 1.5, 2.0, 4.0, etc. to match the time
differences for the measurements Many national surveys take 6 months to administer so the baseline might be in January or it might be in June A person might be measured at month 6, month 13, & month 30 (last at first wave, first at second wave, last at third wave) Another person might be interviewed at month 1, 13, & 25
23
Alan C. Acock, July, 2010
12
7/9/10
What’s in this figure? The Slope Growth Factor Might use month rather than year for coding the wave First person would be coded 5, 12, & 29 Second person would be coded 0, 12, 24
Mplus has a feature that allows each participant to have a
different interval which is important when the time between waves varies TSCORE! This Mplus feature is underutilized
24
Alan C. Acock, July, 2010
What’s in this figure? Residual Variance & Random Effects The individual variation around the Intercept and Slope are
represented in Figure 1 by the RI and Rs.
RI is the variance in the intercept around the mean slope RS is the variance in the slope around the mean slope These are what statisticians call the random effects
The value of the Intercept Growth Factor & the Slope Growth
Factor is really what statisticians call a fixed effect—what would happen if every body was the same We expect substantial variance in both of these as some individuals have a higher or lower starting BMI (intercept) and some individuals will increase (or decrease) their BMI at a different rate (slope) than the average growth rate 25
Alan C. Acock, July, 2010
13
7/9/10
What’s in this figure? Residual Variance & Random Effects In our sample of 10 individuals shown above, notice One adolescent starts with a BMI around 12 and three adolescents
start with a BMI around 30. Clearly there is a random effect for the intercept growth factor Some children have a BMI that increases and others do not. Perhaps there is a random effect for the slope growth factor The variances, RI and R2 are critical if we are going to explore
more complex models with covariates (e.g., an intervention, gender, psychological problems, race, household income, physical activity, fidelity of implementation) These covariates might explain why some individuals have a steeper or less steep initial level and growth rate than the average 26
Alan C. Acock, July, 2010
What’s in this figure? Residual Variance & Random Effects What’s this about a random intercept model and a random
coefficients (slope) model A random intercept model would fix the variance of RS at 0.0 and let the variance of RI be free A random coefficients model would let both RS & RI be free We can estimate the model with both free Estimate the model with just RI free Estimate the model with neither free
27
Alan C. Acock, July, 2010
14
7/9/10
What’s in a figure? Random Errors The ei are random error terms We are usually fitting some functional form of a growth rate
such as linear of quadratic Some years may move above or below the growth trajectory A lot of error is eliminated when we have a random coefficient model since some people may be systematically going up/down quicker than the fixed effect
28
Alan C. Acock, July, 2010
The Mplus program for a linear growth curve
29
Alan C. Acock, July, 2010
15
7/9/10
What’s new compared to day 1? Usevariables are: subcommand to only include the
BMI variables since we are doing a growth curve for these variables We drop the Analysis: section if we had a single processor because we are doing basic growth curve and can use the default options. With multiple processors, this is included to tell Mplus how many processors to utilize We have a Model: section because we need to describe the model. Mplus was designed after growth curves were well understood. There is a single line to describe our model:
! 30
i s | bmi97@0 bmi98@1 bmi99@2! !bmi00@3 bmi01@4 bmi02@5 bmi03@6;!
Alan C. Acock, July, 2010
What’s new compared to day 1? i s | bmi97@0 bmi98@1 bmi99@2! ! !bmi00@3 bmi01@4 bmi02@5 bmi03@6;! The i represents the intercept. We could enter intercept, start,
initial, level, whatever The s represents the slope. We could enter slope, growth, increase, change, whatever A few words make a picture because of defaults Linda assume the constant of 1.0 for each year to represent the intercept The slope is defined by fixing the paths from the slope to bmi97 at 0.0, the path from the slope at bmi98 at 1.0, etc. 31
Alan C. Acock, July, 2010
16
7/9/10
What’s new compared to day 1? A few words make a picture because of defaults Mplus assumes there is a random effect and makes RI & RS random
32
effects Mplus assumes the intercept and slope covary, i.e., cov(RI,RS) is free to be estimated The intercepts for BMI97 – BMI 2003 are all assumed to be zero. There is an implicit command [
[email protected]] where the [] refer to the intercepts Mplus assumes the means for the intercept and slope are free (fixed effects) The errors, e97 – e03, are assumed free and uncorrelated. To correlate e97 with e98 we would add a command e97 with e98 !
Alan C. Acock, July, 2010
What’s new compared to day 1? A big change from day 1 is to ask for a plot of the growth
trajectory. This is initiated by Plot: The type of plot we want for a growth trajectory is Plot3, Type is plot3;! Mplus must know what to plot so we enter the variable used for our waves to define the series The asterisk at the end of the series (*) is used to tell Mplus we are using the same values for the waves that we listed in the Model command, 0 – 5 33
Alan C. Acock, July, 2010
17
7/9/10
Selected Growth Curve Output—Missing
34
Alan C. Acock, July, 2010
Selected Growth Curve Output—Means
You can see a steady increase in the BMI between 1997 and 2003 The rate of increase seems to be leveling off Increases by 1.3 units between 1997 and 1998, by about 0.8 units between 1998 and 1999, but this gradually decreases to under 0.6 units between 2002 and 2003. A quadratic might work with a decreasing rate of increase over time Biggest problem for any definite functional form is the big increase between 1997 and 1998 35
Alan C. Acock, July, 2010
18
7/9/10
Selected Growth Curve Output— Correlations
Correlations get weaker as waves get farther apart. BMI97 & BMI03 a bit high. BMI01 & BMI02 a bit high.
36
Alan C. Acock, July, 2010
Selected Growth Curve Output—Fit
Correlations get weaker as waves get farther apart. BMI97 & BMI03 a bit high. BMI01 & BMI02 a bit high.
37
Alan C. Acock, July, 2010
19
7/9/10
Selected Output—Estimates
Growth curve is: BMI’ = 21.035 + 0.701×Year! 38
Alan C. Acock, July, 2010
Selected Output—Res. Var & Errors
RI = 15.051, the SD = 3.88. This is significant. About 95% of the kids have an intercept between 15.051 ± 2×3.88 or 7.29 and 22.81. This is fitting the intercept and not the actual mean at 1997 because we are using a linear growth curve. The initial mean was much higher. RS = 0.255, the SD = 0.50. About 95% have an increase in their BMI of 0.70 ± 2×0.50 or between 0.20 BMI units per year and 1.20 BMI units. Both seem like subsantial random effects
39
Alan C. Acock, July, 2010
20
7/9/10
Selected Output—Modification Indices
Rarely one to free the intercept loadings of 1.1 even though that has a huge M.I. The M.I. = 37.34 for BMI03 with BMI97 suggests we are not fitting at both ends. A quadratic might be needed. We do NOT automatically free parameters that have a large M.I. Could improve fit with covariates 40
Alan C. Acock, July, 2010
Graphic results Click on Graphs Observed individual values Set seed, random order, number of curves (20)
41
Alan C. Acock, July, 2010
21
7/9/10
Graphic results Most who started high show an increase (one dropped)
42
Alan C. Acock, July, 2010
Graphic results Growth curve versus actual means GraphsView graphsSample & estimated means
43
Alan C. Acock, July, 2010
22
7/9/10
Quadratic Growth Curve
44
Alan C. Acock, July, 2010
Mplus program for quadratic curve
45
Alan C. Acock, July, 2010
23
7/9/10
Tests for Model Fit—quadratic vs. linear
46
Alan C. Acock, July, 2010
Quadratic—Selected Output
47
Alan C. Acock, July, 2010
24
7/9/10
Quadratic—Plot
There is a good fit. The sample means and the estimated means are very close 48
Alan C. Acock, July, 2010
Quadratic—Plot
There is still significance variance (random effect) for both the intercept and the slope among individual adolescents that needs to be explained 49
Alan C. Acock, July, 2010
25
7/9/10
They say I need 3 waves, why have more: Information available with 3 waves 3 waves of data: H1: model is called unrestricted 3 means 3 variances 3 covariances
My1, My2, My3 Var(Y1),Var(Y2),Var(Y3) Cov(Y1,Y2), Cov(Y1,Y3), Cov(Y2,Y3)
9 known statistics
50
Alan C. Acock, July, 2010
Counting parameters--Linear
51
Alan C. Acock, July, 2010
26
7/9/10
3 waves linear growth curve needs 8 3 waves of data: H0: model of simple growth curve A B C D E F,G,H 8
variance of intercept variance of slope covariance of intercept & slope Mean of intercept growth factor Mean of slope growth factor 3 error variances Total number of parameters you are estimating
With 3 waves we have 9 – 8 = 1 degree of freedom
52
Alan C. Acock, July, 2010
4 waves gives huge increase in information 3 waves of data: H1: model is called unrestricted 4 means 4 variances 6 covariances
My1, My2, My3, My4 Var(Y1),Var(Y2),Var(Y3),Var(Y4) Cov(Y1,Y2), Cov(Y1,Y3), Cov(Y1,Y4) Cov(Y2,Y3), Cov(Y2,Y4), Cov(Y3,Y4)
14 known statistics With 4 waves we are still estimating 8 parameters so we have 14
– 8 = 6 degree of freedom
53
Alan C. Acock, July, 2010
27
7/9/10
Alternative to Quadratic Instead of using 0,1,2,3,4,5,6 for linear component
waves and 0,1,4,9,16,25,36 for quadratic component waves Could use 0, 1, *, *, *, *! Could use 0, *, *, *, *, 1 If using months, e.g., 12, 24, 36 and want quadratic, need to rescale, 36 square is 1,296. May divide by 10, e.g., 1.2, 2.4, 3.6 to keep scale of quadratic reasonable These are not nested under quadratic so can’t test, can us BIC
54
Alan C. Acock, July, 2010
Missing values Missing completely at random (MCAR) Missing at random (MAR) John is measured at waves 1, 2, 5 Sue (a part of refresh sample) is measured at waves 2, 3, 4, 5 Antoine is measured a just wave 1 Juan is measured at waves 1,2,3,4 Use all available information to estimate mean/variance for
each wave (Antoine gives information about wave 1) Use all available information to estimate covariances (John & Sue do this for some, but not all covariances 55
Alan C. Acock, July, 2010
28
7/9/10
Missing values—Auxiliary Variables These variables are not part of your model but still serve two
56
purposes First they help predict the score on a missing value, e.g., education may not be relevant to your model, but it would help predict income Second, they help meet the MAR assumption by explaining who does and who does not have a missing value. Gender is an example as men are more likely to skip items. Observations have missing values on a random basis (MAR)— AFTER you control for all variables in your model and all auxiliary variables Mplus has a simple way to add auxiliary variables
Alan C. Acock, July, 2010
Full Information Maximum Likelihood with Auxiliary variables
57
Alan C. Acock, July, 2010
29
7/9/10
Full Information Maximum Likelihood with Multiple Imputation No values are imputed for missing values under the full
information maximum likelihood approach, just means, variances, and covariances are analyzed Although we will not show it here, Mplus can use multiple imputation It can use multiple imputation dataset from other packages
(Stata, SAS, etc.) It can do the multiple imputation itself starting in version 6 There is much debate about the best solution to missing
values but it usually doesn’t matter and they are asymptotically equivalent
58
Alan C. Acock, July, 2010
Multiple Cohort Extension Many national surveys will have multiple cohorts. For example,
the National Longitudinal Study of Youth, 1997 began in 1997 with kids who were 12 to 18. In year two these kids were 13 to 19, in year three they were 14 to 21, and in year four they were 15 to 22 With just four waves of data we can represent a growth curve for people from 12 to 21 Perhaps a study was discontinued at this point, but 3-years later they did it again The follow-up survey would have the people 18 to 25 Using five waves this way we would have data from 12 to 25—
transition from pre-adolescence to adulthood
59
Alan C. Acock, July, 2010
30
7/9/10
Multiple Cohort Extension Mplus uses the command, Data cohort to rearrange
your data We will not illustrate how this is done here but will show this in Stata on day 3
60
Alan C. Acock, July, 2010
Multiple Cohort Extension Instead of organizing the data by age, it is organized by birth
cohort There is massive missing values. In their example there are 90 possible observations, but 50 of these are missing This is MCAR unless there is a cohort effect such as people born in 1963 have a different set of historical circumstances than people born in 1964 or 1965 Muthén did this with a few waves of data to describe growth curves for drinking behavior from adolescence to mid 30s
61
Alan C. Acock, July, 2010
31
7/9/10
Multiple group comparisons Suppose we are interested in differences between Males and 62
Females in their growth on BMI We could treat these as subgroups and simultaneously estimate the model in both groups—Many comparisons: Do they have the same intercept growth factor Same slope growth factor Same variance for intercept growth factor Same variance for slope growth factor Same covariance of intercept and slope residuals Same residual errors Same covariance of residual errors
Alan C. Acock, July, 2010
Multiple group comparisons At the least we can test whether they have the same intercept
growth factor and same slope growth factor (fixed effects comparison) The problem with this approach to multiple group comparison is that it quickly becomes overwhelming We will check the intercept and slope invariance here First, we estimate the model without any constraints (as we’ve done before, however We will estimate it simultaneously for two groups
63
Alan C. Acock, July, 2010
32
7/9/10
Two group solution, same form only
64
Alan C. Acock, July, 2010
Two group solution, same form only We add our categorical variable to the Usevariables
command. This is binary, but could have more categories We tell Mplus how the data is grouped and add labels for each set of outcome grouping is male (0 = female 1 = male);! This simultaneously estimates the model separately for males and females and reports two sets of results with no invariance constraints You do NOT need to sort data to do this 65
Alan C. Acock, July, 2010
33
7/9/10
Multiple Groups—Form only
66
Alan C. Acock, July, 2010
Multiple Groups—Form only
67
Alan C. Acock, July, 2010
34
7/9/10
Multiple groups: Different Intercept & Slope?
68
Alan C. Acock, July, 2010
Multiple groups: Different Intercept & Slope? There are two model statements Fist is an overall and describes the group with the lowest score on
the grouping variable (females were 0; males were 1). Second models the next higher score on the grouping variable, males. Don’t list things that stay the same Only list things that are different or That you constrain to be equal
Mplus puts the same number whenever it wants to force
coefficients to be equal. When Mplus puts [] around a latent variable, it is referring to the mean. When Mplus puts [] around an observed variable it is referring to the intercept The intercept and slope are forced to be the same in the two groups 69
Alan C. Acock, July, 2010
35
7/9/10
Multiple groups—equal I and s
70
Alan C. Acock, July, 2010
Multiple groups—equal I and s
71
Alan C. Acock, July, 2010
36
7/9/10
Multiple groups—equal I and s
72
Alan C. Acock, July, 2010
Multiple groups—equal I and s Although we can say there is a highly significant difference
between the level and trend for girls and boys, we need to be cautious because this difference of chi-square has the same problem with a large sample size that the original chi-squares have In fact, the measures of fit are hardly changed whether we constrain the intercept and slope to be equal or not. Moreover, the visual difference in the graph is not dramatic We could also put other constraints on the two solutions such as equal variances and covariances, and even equal residual error variances, but we will not 73
Alan C. Acock, July, 2010
37
7/9/10
Alternative to Multiple Group Analysis
74
Alan C. Acock, July, 2010
Alternative to Multiple Group Analysis Enter the grouping variable as a predictor Males have a higher intercept and a steeper slope Positive slope from male to intercept growth factor means
males have a higher intercept Positive slope from male to slope growth factor means males have a steeper slope Limitations This imposes restrictions that everything else is equal for
females and males Same error terms, same residual variances, same covariance of residual variances If those differences are important need the two-group approach 75
Alan C. Acock, July, 2010
38
7/9/10
Alternative to Multiple Group Analysis
76
Alan C. Acock, July, 2010
Alternative to Multiple Group Analysis
77
Alan C. Acock, July, 2010
39
7/9/10
Alternative to Multiple Group Analysis
78
Alan C. Acock, July, 2010
Alternative to Multiple Group Analysis
79
Alan C. Acock, July, 2010
40
7/9/10
Alternative to Multiple Group Analysis— Graphic representation We see that the intercept is 20.385 and the slope is .625. How is
gender related to this? For girls the equation is: Est. BMI = 20.911 + .656(Time) + .242(Male) + .086(Male)(Time)
= 20.911 + .656(Time) + .242(0) + .086(0)(Time) = 20.911 + .656(Time)
80
Alan C. Acock, July, 2010
Alternative to Multiple Group Analysis— Graphic representation We see that the intercept is 20.385 and the slope is .625. How is
gender related to this? For boys the equation is: Est. BMI = 20.911 + .656(Time) + .242(Male) + .086(Male)(Time)
= 20.911 + .656(Time) + .242(1) + .086(1)(Time) = (20.911 + .242) + (.625 + .086)(Time) = 21.153 + .711(Time)
81
Alan C. Acock, July, 2010
41
7/9/10
Alternative to Multiple Group Analysis— Graphic representation Using these we estimate the BMI for girls is initially 20.911. By the
seventh year when she is 18(Time = 6) her estimated BMI will be 20.385 + .656(6) or 24.847 Using these results, we estimate the BMI for boys is initially 21.153. By the seventh year it will be 21.153 + .711(6) or 25.419 Since a BMI of 25 is considered overweight, by the age of 18 we estimate the average boy will be classified as overweight and the average girl is not far behind! We could use the plots provided by Mplus, but if we wanted a nicer looking plot we could use another program. I used Stata getting this graph 82
Alan C. Acock, July, 2010
Alternative to Multiple Group Analysis— Graphic representation using Stata
Note, When we treat the grouping variable as a predictor, as in this example, we only test whether the intercept and slope are different for the two groups. This is an interaction between gender and growth trajectory. We do not allow the other parameters to be different and do not test whether this is reasonable or not. 83
Alan C. Acock, July, 2010
42
7/9/10
Alternative to Multiple Group Analysis— Graphic representation using Stata
84
Alan C. Acock, July, 2010
Growth Curves with Time Invariant Covariates A covariate is time invariant if it is constant for the duration
of the growth curve (gender, race, condition, etc.) Our alternative approach using gender illustrates this It has been called conditional latent trajectory modeling (Curran & Hussong, 2003)—the intercept growth factor and slope growth factor are conditioned on other variables You could call this moderated growth trajectories with gender serving as the moderator You could call this an interaction between gender and the growth trajectory Here is a slightly more complex example 85
Alan C. Acock, July, 2010
43
7/9/10
Adding Time Invariant Covariates
86
Alan C. Acock, July, 2010
Adding Time Invariant Covariates Here we add two covariates White is coded 1 for white and 0 for nonwhites Emotional problems is a latent variable that was measured in
1997 (not any other year)
There is a youth self-report of emotional problems There is a parent report of the youth’s emotional problems
Both covariates influence the intercept, linear slope and
quadratic slope The time invariant covariates moderate (interact with) the linear slope growth factor and the quadratic slope growth factor 87
Alan C. Acock, July, 2010
44
7/9/10
Program for time invariant covariates
88
Alan C. Acock, July, 2010
Program for time invariant covariates The Useobservations command restricts the sample
to males who are not Asian nor Other on race/ethnicity Emotional problems is a latent variable
89
Alan C. Acock, July, 2010
45
7/9/10
Program for time invariant covariates
90
Alan C. Acock, July, 2010
Mediation & Moderation with Time Invariant Covariates We are predicting the growth in drinking problems among
adolescents The amount of parental drinking may contribute to this Parental monitoring may contribute Peer influence may contribute These variables MODERATE (interact with) the trajectory. The rate of
change varies depending on these three variables
Parental drinking may be mediated by monitoring and peers If parents monitor their adolescent and there are positive peers
then the effect of parental drinking may be greatly diminished
91
Alan C. Acock, July, 2010
46
7/9/10
Program for time invariant covariates— Mediation & Moderation
Note, the mediation modeled using the Model Indirect: subcommand 92
Alan C. Acock, July, 2010
Time Varying Covariates Time varying covariates can be very important We might have a randomized trial with a treatment and
control group The trial might involve a four year intervention The treatment group should have a more positive trajectory than the control group The fidelity of the program may vary from year to year
93
Alan C. Acock, July, 2010
47
7/9/10
Time Varying Covariates If the treatment is having a positive effect (positive
trajectory), there might be annual variations around this trajectory, depending on the fidelity with which the program was implemented each year Years of high fidelity might push the outcome variable above the overall trajectory; Years of low fidelity might do the opposite These annual perturbations may be explainable by corresponding variations in program fidelity
94
Alan C. Acock, July, 2010
Time Varying Covariates Performance in Treatment Group
0
1
2
3
4
5
Baseline (0) was normal level of implementation,Year 2 (1) was a high level of implementation, but year 3 (2) and year 4 (3) there was a staff problem that resulted in a lower level of implementation. This was corrected in year 4 and year 5 as okay. These variations in fidelity of implementation pushed the score each year above or below the overall trajectory. 95
Alan C. Acock, July, 2010
48
7/9/10
Time Varying Covariates The error terms are correlated for adjacent years The W represents a vector of time invariant covariates that
moderate the intercept and the slope Intervention vs. control Gender
The Ai variables represent the time varying covariates Fidelity of implementation Peer group influence
The time varying covariates directly influence the Yi scores
96
Alan C. Acock, July, 2010
Time Varying Covariates Might study decline in drinking behavior from age 22 to age
27 Slope is negative Time varying covariates might include: Drinking Behavior of Peers Work status (student to employed full time) Number of children Percent of friends who have children
97
Alan C. Acock, July, 2010
49
7/9/10
Time Varying Covariates
98
Alan C. Acock, July, 2010
50