multi-level imputation error

[bryorsnef]

Hi there,

I'm trying to set up a multi-level imputation, but am hitting an error that looks like a bug. The model involves imputation of subject and item level variables (i.e. 2 sets of level 2 variables), and therefore two random intercepts - one for the subject factor, and one for the item factor. When I try to set things up with impute.2l.lmer, I get this error:

Error in str2lang(x) : :1:105: unexpected ')'
1: dv._lmer~1+subid+itemid+cwwt+cagtb+hub_connections+tdiff+rdiff+freq.diff+phonsimdiff+rt+(1+subid+itemid|)

Where the formula seems mangled - the subid and itemid should go on the other side of the |. As well, it seems like subid and itemid are also being included as fixed effects, even though they are coded as 2 for the imputed variable in the predictor matrix.

[stefvanbuuren]

Sorry I can't do magic. Please provide a reproducible example in a really simplified case.

[bryorsnef]

Appologies for not posting right away with an example - I was hoping this was a known issue or if I broke something obvious. I'm waiting for confirmation from a collaborator that I can share the original data, because I'm having issues reproducing the issue on my own.

[gerkovink]

There is a pipe symbol | in your code. I believe the language interpreter reads this as a conditional and expects some input, but instead finds a closing parenthesis ). I believe your error will disappear if you adjust the following part in your code:

(1+subid+itemid|)

All the best,

Gerko

[gerkovink]

I can't even reprex your error because the above code breaks the reprex()

> reprex()
Rendering reprex...
Error: callr subprocess failed: <text>:24:43: unexpected ')'
23: imp <- mice(boys, m = 3, maxit = 2) # proof of concept
24: with(imp, lmer(age ~ hgt + (1 + bmi + tv |)
                                              ^
Type .Last.error.trace to see where the error occured

But in mice the following code also breaks, which proves my point.

> imp <- mice(boys, m = 3, maxit = 2) # proof of concept

 iter imp variable
  1   1  hgt  wgt  bmi  hc  gen  phb  tv  reg
  1   2  hgt  wgt  bmi  hc  gen  phb  tv  reg
  1   3  hgt  wgt  bmi  hc  gen  phb  tv  reg
  2   1  hgt  wgt  bmi  hc  gen  phb  tv  reg
  2   2  hgt  wgt  bmi  hc  gen  phb  tv  reg
  2   3  hgt  wgt  bmi  hc  gen  phb  tv  reg
> with(imp, lmer(age ~ hgt + (1 + bmi + tv|)))
Error: unexpected ')' in "with(imp, lmer(age ~ hgt + (1 + bmi + tv|)"

[bryorsnef]

Thanks for the help. The error isn't happening in the analysis step though, it's happening when I call mice() to do the imputation step. I don't explicitly state an lmer formula anywhere. I've been searching around for any resources on multi-level imputation in mice or other packages or functions to use. I see the mice.impute.ml.lmer in miceadds, but I'm having a hell of a time getting that set up. I'd like to stick to an FCS approach because the analyses I want to do eventually involve random slopes, so joint modelling seems less appealing. But finding clear guidance with good documentation hasn't been easy. Any suggestions on good tutorials or resources for multi-level imputation would be helpful.

[stefvanbuuren]

@gerkovink Thanks, R expects a variable name after |. It does not make sense to leave it empty. @bryorsnef did you already try specify your class variable?

[gerkovink]

finding clear guidance with good documentation hasn't been easy. Any suggestions on good tutorials or resources for multi-level imputation would be helpful.

Have a look at the Multilevel Imputation Vignette

Can't help you any further without a proper reprex since the error seems to be dependent on your imputation specification.

All the best,

Gerko

[gerkovink]

Closing as the issue is not reproducible.

[atanasj]

I am having the same error. I have not included data, however below is the output. clus is the cluster variable. I have tried to follow the online vignettes as closely as possible. What seems to be happening is that it is dropping the cluster variable in the hierarchical regression that I assume is being used in the backend to impute the data. Am I doing something wrong? Is there a different way to specify the class variable?

NOTE: Data available upon request.

> 
+ pred_ <-
+   readxl::read_xlsx(
+     here("data/tmp/mi_working.xlsx"),
+     sheet = "pred_matrix",
+     range = "A1:L12"
+   ) %>%
+   .mice_pred_matrix()
+ pred_
+ 
>      clus ID pb ac bhs tb age sex dep anx si
clus    0  0  0  0   0  0   0   0   0   0  0
ID      0  0  0  0   0  0   0   0   0   0  0
pb     -2  0  0  0   0  0   0   0   0   0  0
ac     -2  0  1  1   1  1   1   1   1   1  1
bhs    -2  0  1  1   1  1   1   1   1   1  1
tb     -2  0  1  1   1  1   1   1   1   1  1
age     0  0  0  0   0  0   0   0   0   0  0
sex     0  0  1  1   1  1   1   1   1   1  1
dep    -2  0  1  1   1  1   1   1   1   1  1
anx    -2  0  1  1   1  1   1   1   1   1  1
si     -2  0  1  1   1  1   1   1   1   1  1
> 
+ imp <-
+   mice(
+     lpa_cov2,
+     pred = pred_,
+     meth = "2l.pmm",
+     maxit = 0,
+     printFlag = FALSE
+   )
+ 
> 
+ imp.5 <-
+   mice.mids(
+     imp,
+     maxit = 5,
+     print = FALSE
+   )
+ 
+ boundary (singular) fit: see ?isSingular
boundary (singular) fit: see ?isSingular
Error in str2lang(x) : <text>:1:43: unexpected ')'
1: dv._lmer~1+pb+ac+bhs+tb+age+dep+anx+si+(1|)
                                              ^
> 

[stefvanbuuren]

The imp_t0$loggedEvents is NULL implicate that mice does not remove the cluster (or any other) variable.

Multilevel imputation is tricky and considerably less robust than flat file imputation because there are many more parameters to fit. So, if you can, evade multilevel imputation!

If you must use multilevel, carefully build up the model, just as shown in my chapter 7. I interpret these messages that your imputation model is too complex for your data. Try simplifying to just two or three variables, and build it up from there. Only add cluster means carefully for the variables of interest (not everything at once). This is breakable technology, so proceed slowly.

[atanasj]

The funny thing is, I don't always get these errors about the convergence warning (the end), but always get the warning about singular fit. Should I be worried about one or both of these warnings? Also, what is a good rule of thumb regarding convergence when inspecting the plots? Regarding the above example, I just included the clusters and not the cluster means, which seemed to simplify things. But if on some runs I do not get this warning, and on runs preceding it I do, does that mean that all subsequent runs (as I increase the maxit values) are all unusable? Specifically:

• First run with maxit = 10 gives a warning.
• Next run with maxit = 10 (adding to the one above) does not provide a warning. Because the run above did, this run is unusable too?

[stefvanbuuren]

The singular warning is serious, and suggests that somewhere in your data there is duplicate information. You know your data best, so try to find the (combination of) variables that are causing the warning, and remove the superfluous ones.

[atanasj]

Thanks @stefvanbuuren for your advice. Can you please provide an example of what you mean:

somewhere in your data there is duplicate information.

Are there any references or examples that I can learn from to help me identify the potential cause of these problems? Also, when you say "remove the superfluous ones", do you mean remove them as predictors of the imputation or remove them from imputation all together?

Thanks again for your guidance.

[atanasj]

@stefvanbuuren, as an update, the only time I can get this to work is to remove the random effect of the cluster variable. This was based on the answer to this question. Do you have an opinion about this or any other suggestions?

[isaactpetersen]

I received this error (#> Error in str2lang(x): <text>:1:24: unexpected ')' 1: dv._lmer ~ 1+extrav+(1|)) when using multilevel imputation with miceadds with a cluster variable that is either a character or factor variable. I also receive an error (albeit a different error) when using a multilevel imputation method from the mice package instead of from the miceadds package. By changing the cluster variable to an integer, it resolved the issue. I describe the bug here: #657