Weird parameter estimates....

Amanda Pollitt

unread,

Oct 29, 2013, 5:29:23 PM10/29/13

to lav...@googlegroups.com

So I'm attempting to replicate a longitudinal panel model that was also done in MPlus as our project is switching to R. (yay!) However, not only am I getting different numbers than the MPlus model, I'm getting out of bounds parameter estimates. Note that I can't correlate the residuals of w3access and w4ladder as they are not the same constructs.

Here is my code:

aim2.model <- '

# measurement model.

w3fin =~ NA*w3income + w3strain + w3access

w4fin =~ w4income + w4strain + w4ladder

w3alc =~ NA*w3alc1 + w3alc2 + w3ex1 + w3ex2

w4alc =~ w4alc1 + w4alc2 + w4ex2 + w4ex2

# regressions

w4alc ~ w3alc + w3fin

w4fin ~ w3alc + w3fin

# residual correlations

w3alc1 ~~ w4alc1

w3alc2 ~~ w4alc2

w3ex1 ~~ w4ex1

w3ex2 ~~ w4ex2

w3income ~~ w4income

w3strain ~~ w4strain

w3fin ~~ w3alc

w3alc ~~ w4alc

#constraining latent variance

w3alc ~~ 1*w3alc

w3fin ~~ 1*w3fin

'

fit3 <- sem(aim2.model, data = aim2model, missing = "ML")

summary(fit3, standardized = TRUE)

Here is the covariance matrix:

w3incm w3strn w3ccss w4incm w4strn w4lddr w3alc1 w3alc2 w3ex1 w3ex2

w3inco 2.188

w3stra 0.000 0.035

w3acc 0.000 0.000 0.032

w4inco 0.347 0.000 0.000 2.771

w4stra 0.000 0.008 0.000 0.000 0.032

w4ladd 0.000 0.000 0.000 0.000 0.000 2.155

w3alc1 0.000 0.000 0.000 0.000 0.000 0.000 0.918

w3alc2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 7.244

w3ex1 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.499

w3ex2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.481

w4alc1 0.000 0.000 0.000 0.000 0.000 0.000 0.302 0.000 0.000 0.000 1.102

w4alc2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.135 0.000 0.000 0.000 5.015

w4ex2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.083 0.000 0.000 0.489

w4ex1 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.105 0.000 0.000 0.000 0.000 2.320

I thought that maybe this was an identification issue, but when I let it run with lavaan defaults it gives me even weirder estimates. Any clue where I might be making a mistake?

Thanks!

Amanda

Edward Rigdon

unread,

Oct 29, 2013, 9:15:16 PM10/29/13

to lav...@googlegroups.com

Amanda--

Are these two lines correct?

w3alc =~ NA*w3alc1 + w3alc2 + w3ex1 + w3ex2

w4alc =~ w4alc1 + w4alc2 + w4ex2 + w4ex2

In the second line, it looks like the same observed variable, w4ex2, appears twice. Could that be part of your problem?

--Ed Rigdon

--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To post to this group, send email to lav...@googlegroups.com.
Visit this group at http://groups.google.com/group/lavaan.
For more options, visit https://groups.google.com/groups/opt_out.

Amanda Pollitt

unread,

Oct 30, 2013, 12:04:20 PM10/30/13

to lav...@googlegroups.com

Good lord it's always something like that haha. I fixed the typo but I'm still getting an out of bounds parameter, here is the new matrix:

w3incm w3strn w3ccss w4incm w4strn w4lddr w3alc1 w3alc2 w3ex1 w3ex2

w3inco 2.190

w3stra 0.000 0.035

w3acc 0.000 0.000 0.032

w4inco 0.352 0.000 0.000 2.800

w4stra 0.000 0.008 0.000 0.000 0.032

w4ladd 0.000 0.000 0.000 0.000 0.000 2.147

w3alc1 0.000 0.000 0.000 0.000 0.000 0.000 0.927

w3alc2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 7.252

w3ex1 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.485

w3ex2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.483

w4alc1 0.000 0.000 0.000 0.000 0.000 0.000 0.311 0.000 0.000 0.000 1.222

w4alc2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.107 0.000 0.000 0.000 5.168

w4ex1 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.024 0.000 0.000 0.000 0.407

w4ex2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.072 0.000 0.000 0.000 0.394

Thanks so much for the help!

Amanda

Edward Rigdon

unread,

Oct 30, 2013, 12:14:43 PM10/30/13

to lav...@googlegroups.com

How about these two lines, farther down:

w3fin ~~ w3alc

w3alc ~~ w4alc

The first of these seems to break the pattern that otherwise characterizes your syntax.

--Ed Rigdon

Terrence Jorgensen

unread,

Oct 30, 2013, 12:22:45 PM10/30/13

to lav...@googlegroups.com

I fixed the typo but I'm still getting an out of bounds parameter, here is the new matrix:

w3incm w3strn w3ccss w4incm w4strn w4lddr w3alc1 w3alc2 w3ex1 w3ex2
w3inco 2.190
w3stra 0.000 0.035
w3acc 0.000 0.000 0.032
w4inco 0.352 0.000 0.000 2.800
w4stra 0.000 0.008 0.000 0.000 0.032
w4ladd 0.000 0.000 0.000 0.000 0.000 2.147
w3alc1 0.000 0.000 0.000 0.000 0.000 0.000 0.927
w3alc2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 7.252
w3ex1 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.485
w3ex2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.483
w4alc1 0.000 0.000 0.000 0.000 0.000 0.000 0.311 0.000 0.000 0.000 1.222
w4alc2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.107 0.000 0.000 0.000 5.168
w4ex1 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.024 0.000 0.000 0.000 0.407
w4ex2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.072 0.000 0.000 0.000 0.394

What out of bounds estimate are you talking about? If this is the model-implied covariance matrix, you can't tell whether there are out of bounds estimates very easily -- try cov2cor() to turn it into a correlation matrix, and see if there are any values that exceed [1, -1].

If this is the sample covariance matrix, you can do the same cov2cor() check, but you might have out of bounds estimates because you have missing data, IF this sample matrix is estimated using pairwise deletion (instead of extracting the ML estimate from lavaan).

If the correlations among your variables are as low as they seem from the covariance matrix, I wouldn't trust results from any package, unless you have a very large sample size and use more items to identify your constructs, in order to compensate for the low communalities (see attached article for advice). But it's hard to tell whether this is a problem from your (unstandardized) covariance matrix. It's also worth noting that you have some wildly different variances among the indicators within each construct (e.g., 2.8 vs. 0.03, or 7.25 vs. 0.48). That can cause problems, so you might consider rescaling your variables in some way to make their variances more similar. I don't have a specific guideline, but certainly the variables with more than 10 times, and some close to 100 times, greater variance than other indicators would be problematic (although I'm not sure it's related to why you get different numbers in the two packages).

It's also possible you still aren't specifying the exact same model in both packages. Have you checked your degrees of freedom and number of parameters being estimated?

MacCallumetal1999.pdf

Terrence Jorgensen

unread,

Oct 30, 2013, 12:25:42 PM10/30/13

to lav...@googlegroups.com

How about these two lines, farther down:
w3fin ~~ w3alc
w3alc ~~ w4alc

The first of these seems to break the pattern that otherwise characterizes your syntax.
--Ed Rigdon

Good eye, Ed, although I think it's the second of those that is inconsistent/problematic. The "w4" variables are both regressed on both of the "w3" variables. So the exogenous "w3" variables should be correlated, and the endogenous "w4" variables should be correlated. But the line "w3alc ~~ w4alc" is specifying a covariance path where there is already a regression path: "w4alc ~ w3alc", so it's not identified.

Message has been deleted

Amanda Pollitt

unread,

Oct 31, 2013, 2:54:15 PM10/31/13

to lav...@googlegroups.com

Thanks for everyone's input. I've been staring at this code and feeling like I'm going crazy. I made the edit that Terrence suggested, so I figured I'd post the code and the output so you guys can see what's happening.

aim2.model <- '

# measurement model

w3fin =~ w3strain + w3access + w3income

w4fin =~ w4strain + w4ladder + w4income

w3alc =~ w3alc1 + w3alc2 + w3ex1 + w3ex2

w4alc =~ w4alc1 + w4alc2 + w4ex1 + w4ex2

# regressions

w4alc ~ w3alc + w3fin

w4fin ~ w3alc + w3fin

# residual correlations

w3alc1 ~~ w4alc1

w3alc2 ~~ w4alc2

w3ex1 ~~ w4ex1

w3ex2 ~~ w4ex2

w3income ~~ w4income

w3strain ~~ w4strain

w3alc ~~ w3fin

w4fin ~~ w4alc

'

fit3 <- sem(aim2.model, data = niaaa, missing = "ML")

summary(fit3, standardized = TRUE)

Here you can see where I made the changes to the correlations.

lavaan (0.5-14) converged normally after 136 iterations

Used Total

Number of observations 17836 20747

Number of missing patterns 101

Estimator ML

Minimum Function Test Statistic 2209.101

Degrees of freedom 65

P-value (Chi-square) 0.000

Parameter estimates:

Information Observed

Standard Errors Standard

Estimate Std.err Z-value P(>|z|) Std.lv Std.all

Latent variables:

w3fin =~

w3strain 1.000 0.055 0.282

w3access 3.671 0.185 19.840 0.000 0.201 0.749

w3income 10.664 0.493 21.610 0.000 0.583 0.367

w4fin =~

w4strain 1.000 0.110 0.522

w4ladder 8.458 0.210 40.201 0.000 0.931 0.536

w4income 13.009 0.342 38.055 0.000 1.432 0.650

w3alc =~

w3alc1 1.000 1.465 0.836

w3alc2 1.640 0.018 90.501 0.000 2.402 0.666

w3ex1 0.947 0.007 133.918 0.000 1.387 0.894

w3ex2 0.843 0.006 131.447 0.000 1.234 0.871

w4alc =~

w4alc1 1.000 1.434 0.792

w4alc2 1.381 0.016 86.804 0.000 1.980 0.657

w4ex1 0.965 0.008 127.552 0.000 1.383 0.908

w4ex2 0.801 0.006 123.974 0.000 1.148 0.878

Regressions:

w4alc ~

w3alc 0.527 0.009 59.388 0.000 0.538 0.538

w3fin 0.045 0.273 0.166 0.868 0.002 0.002

w4fin ~

w3alc -0.006 0.001 -6.573 0.000 -0.074 -0.074

w3fin 1.417 0.055 25.871 0.000 0.704 0.704

Covariances:

w3alc1 ~~

w4alc1 0.311 0.012 26.443 0.000 0.311 0.292

w3alc2 ~~

w4alc2 1.107 0.058 18.959 0.000 1.107 0.181

w3ex1 ~~

w4ex1 0.024 0.007 3.598 0.000 0.024 0.054

w3ex2 ~~

w4ex2 0.072 0.006 13.039 0.000 0.072 0.166

w3income ~~

w4income 0.352 0.031 11.399 0.000 0.352 0.142

w3strain ~~

w4strain 0.008 0.000 24.269 0.000 0.008 0.249

w3fin ~~

w3alc -0.010 0.001 -10.272 0.000 -0.127 -0.127

w4fin ~~

w4alc -0.006 0.001 -4.139 0.000 -0.065 -0.065

Intercepts:

w3strain 0.099 0.002 63.219 0.000 0.099 0.510

w3access 0.445 0.002 206.830 0.000 0.445 1.660

w3income 6.622 0.013 500.272 0.000 6.622 4.163

w4strain 0.124 0.002 74.183 0.000 0.124 0.587

w4ladder 5.978 0.014 433.207 0.000 5.978 3.444

w4income 4.606 0.018 259.152 0.000 4.606 2.092

w3alc1 2.131 0.014 152.157 0.000 2.131 1.216

w3alc2 3.203 0.029 110.230 0.000 3.203 0.888

w3ex1 1.158 0.012 93.352 0.000 1.158 0.746

w3ex2 1.093 0.011 96.487 0.000 1.093 0.771

w4alc1 2.215 0.014 155.152 0.000 2.215 1.224

w4alc2 2.662 0.024 111.229 0.000 2.662 0.883

w4ex1 1.118 0.012 93.104 0.000 1.118 0.734

w4ex2 0.958 0.010 92.820 0.000 0.958 0.732

w3fin 0.000 0.000 0.000

w4fin 0.000 0.000 0.000

w3alc 0.000 0.000 0.000

w4alc 0.000 0.000 0.000

Variances:

w3strain 0.035 0.000 0.035 0.920

w3access 0.032 0.002 0.032 0.440

w3income 2.190 0.029 2.190 0.865

w4strain 0.032 0.000 0.032 0.728

w4ladder 2.147 0.032 2.147 0.712

w4income 2.800 0.057 2.800 0.577

w3alc1 0.927 0.014 0.927 0.302

w3alc2 7.252 0.091 7.252 0.557

w3ex1 0.485 0.009 0.485 0.202

w3ex2 0.483 0.008 0.483 0.240

w4alc1 1.222 0.017 1.222 0.373

w4alc2 5.168 0.063 5.168 0.569

w4ex1 0.407 0.009 0.407 0.175

w4ex2 0.394 0.007 0.394 0.230

w3fin 0.003 0.000 1.000 1.000

w4fin 0.006 0.000 0.485 0.485

w3alc 2.146 0.035 1.000 1.000

w4alc 1.460 0.026 0.711 0.711

The regressions are where I'm concerned. I'm getting a value over 1 there with standardized output. You can also see that my sample is very large and I'm using ML for missing data. Thank you so much for the help.

Terrence Jorgensen

unread,

Oct 31, 2013, 4:39:47 PM10/31/13

to lav...@googlegroups.com

Estimate Std.err Z-value P(>|z|) Std.lv Std.all

Regressions:
w4alc ~
w3alc 0.527 0.009 59.388 0.000 0.538 0.538
w3fin 0.045 0.273 0.166 0.868 0.002 0.002
w4fin ~
w3alc -0.006 0.001 -6.573 0.000 -0.074 -0.074
w3fin 1.417 0.055 25.871 0.000 0.704 0.704

The regressions are where I'm concerned. I'm getting a value over 1 there with standardized output.

I'm not seeing any out-of-bounds estimates, anywhere in your output. When you request standardized results using "summary(fit3, standardized = TRUE)", you will get unstandardized results under the first column (labeled "Estimate"), estimates where only the latent variables are standardized in the 5th column (labeled "Std.lv"), and estimates where both the latent and observed variables are standardized (labeled "Std.all"). For comparison, see the results for excluding the "standardized" argument: summary(fit3).

So the regression parameter that is 1.417 is not out of bounds because it is not standardized. The standardized solution for this estimate, and for all estimates in your model, are within bounds (see the far right column).

Terry

Edward Rigdon

unread,

Oct 31, 2013, 6:46:38 PM10/31/13

to lav...@googlegroups.com

Collinearity can drive standardized regression coefficients beyond +/- 1.0, as can marginal identification problems (which are somewhat the same thing) and misspecification. How severe is Collinearity among predictor sets?

--Ed Rigdon

Sent from my iPad

On Oct 31, 2013, at 2:30 PM, Amanda Pollitt <ampoll...@gmail.com> wrote:

No -- just kidding. I followed what Terrence says so here is the new code:

model2 <- '
# measurement model
w1alc =~ NA*h1to15 + h1to16 + h1to17 + h1to18
w2alc =~ h2to19 + h2to20 + h2to21 + h2to22
w3alc =~ w3alc1 + w3alc2 + w3ex1 + w3ex2
w4alc =~ w4alc1 + w4alc2 + w4ex1 + w4ex2
w3fin =~ NA*w3strain + w3access + w3income

w4fin =~ w4strain + w4ladder + w4income

# regressions
w2alc ~ w1alc
w3alc ~ w2alc

w4alc ~ w3alc + w3fin
w4fin ~ w3alc + w3fin
# residual correlations

w3strain ~~ w4strain
w3income ~~ w4income
h1to15 ~~ h2to19
w3alc1 ~~ w4alc1
h1to16 ~~ h2to20
w3alc2 ~~ w4alc2
h1to17 ~~ h2to21
w3ex1 ~~ w4ex1
h1to18 ~~ h2to22
w3ex2 ~~ w4ex2
w3alc ~~ w3fin
w4alc ~~ w4fin
w1alc ~~ 1*w1alc
w3fin ~~ 1*w3fin
'
And here is my regression output:

Regressions:
w4alc ~
w3alc 0.768 2.499 0.308 0.758 0.536 0.536
w3fin -0.021 19.665 -0.001 0.999 -0.015 -0.015
w4fin ~
w3alc -0.123 NA -0.086 -0.086
w3fin 1.006 NA 0.703 0.703

Ugh.
Amanda

On Thursday, October 31, 2013 11:22:52 AM UTC-7, Amanda Pollitt wrote:
Ed -- Great! So that was the problem. Thank you for catching my typos. It should read w3fin ~~ w4fin. Fixed that, and things look like they should.

Terrence -- Thank you for your advice. We do have a sample of 20,747 but I'll bring the scaling information to the project group as I think that's a valid concern. When I meant out of bounds parameter estimates, I meant regression coefficients -- I was getting values over 1 in my standardized output.

So although this fixed MY problem, I'm still getting differences from MPlus. I went back and checked and it looks like MPlus doesn't constrain one of the factor loadings in the standardized output like lavaan doesp; it constrains them only in unstandardized output. I don't expect anyone here to know MPlus since this is a lavaan board, but it's certainly weird. I think that's where the differences lie.

But either way, thank you for solving my problem!!

Amanda

On Wednesday, October 30, 2013 9:25:42 AM UTC-7, Terrence Jorgensen wrote:

Reply all

Reply to author

Forward