#6

Can you post the summary of that model? might be something useful in there.

#8

It would work better if you attach it as a CSV file. Also, something is a bit off because the PPT column has only two distinct values.

By summary of the model I meant
summary(anxiety_model)

#9

Hi Drew,

Sorry, this is probably really obvious but I do not know how to attach a .csv file as the only options I can see available are photo images or pdf's.

Please see summary of anxiety model below

> anxiety_model<- lmer(Anxiety2 ~ Intervention + Positive_Experience  +  Credibility
+                      + Expectancy + Attitude + Trait_Anxiety + Anxiety1 + (1|PPT) , DATA)
Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge with max|grad| = 0.0425793 (tol = 0.002, component 1)
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?

> summary(anxiety_model)
Linear mixed model fit by REML. t-tests use Satterthwaite's method ['lmerModLmerTest']
Formula: Anxiety2 ~ Intervention + Positive_Experience + Credibility +      Expectancy + Attitude + Trait_Anxiety + Anxiety1 + (1 | PPT)
   Data: DATA

REML criterion at convergence: 459.7

Scaled residuals: 
       Min         1Q     Median         3Q        Max 
-4.001e-06 -7.951e-07 -7.110e-08  8.578e-07  3.864e-06 

Random effects:
 Groups   Name        Variance  Std.Dev. 
 PPT      (Intercept) 1.636e+01 4.044e+00
 Residual             3.302e-11 5.746e-06
Number of obs: 98, groups:  PPT, 94

Fixed effects:
                      Estimate Std. Error         df t value Pr(>|t|)
(Intercept)         17.2989161  8.2550947  0.0054003   2.096    0.978
Intervention        -1.6009057  1.6031778  0.0458100  -0.999    0.903
Positive_Experience -0.0034879  0.0637182  0.0066724  -0.055    0.996
Credibility         -0.1336464  0.1208873  0.0560781  -1.106    0.883
Expectancy           0.0154761  0.0125173  0.0160547   1.236    0.953
Attitude             0.0782994  0.0727864  0.0005378   1.076    0.998
Trait_Anxiety        0.0219922  0.0974992  0.0847401   0.226    0.943
Anxiety1             0.5589693  0.0997523  0.0022963   5.604    0.988

Correlation of Fixed Effects:
            (Intr) Intrvn Pstv_E Crdblt Expctn Attitd Trt_An
Interventin -0.640                                          
Pstv_Exprnc -0.533  0.734                                   
Credibility -0.091  0.255 -0.090                            
Expectancy   0.332 -0.215 -0.045 -0.438                     
Attitude    -0.406  0.058 -0.077 -0.388 -0.224              
Trait_Anxty -0.493  0.070  0.061  0.011 -0.195  0.142       
Anxiety1    -0.353  0.072  0.161  0.039 -0.106 -0.123 -0.307
convergence code: 0
Model failed to converge with max|grad| = 0.0425793 (tol = 0.002, component 1)
Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?

P.S the variable PPT has been sorted, the two distinct values were an accident I created by playing around with the formating

#10

If you have a small anonymised data frame you wish to share. You can use the base function

dput()

This will cause a copy and pastable representation of your data frame that you can share as text

#11

Thank you, however my dataset is 121 obs of 13 variables so too large to copy and paste as text

#12

4 posts were split to a new topic: Can I share a dataset (csv, rdata, etc) on RStudio Community by renaming the filetype?

incomplete final line found by readTableHeader
#17

Hi @KarenMillin, 0.04 is quite off the targeted gradient of 0.002.
So indeed, the estimates may not be trustworthy.

Risking to state the obvious but it is not clear from your post: did you try to do what the warning says, i.e. rescaling the variables?

That may greatly help. You can use scale() for that, but then you must interpret each coefficient as the change triggered by an increase in one standard deviation of the focal covariate.

Hope this help.

#18

Hi Alexandre,

Thank you for your response.
I have scaled but I have possibly not done it correctly?
I used the scale function on the numeric values in the dataset (this excluded PPT and Intervention)

ind<-sapply(DATA, is.numeric)
DATA[ind] <- lapply(DATA[ind], scale)

and similar warnings come up with or without the scales variables

Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge with max|grad| = 0.050911 (tol = 0.002, component 1)
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?
3: In as_lmerModLT(model, devfun) :
  Model may not have converged with 1 eigenvalue close to zero: 5.7e-11

I have also used this coding to try scaling incase it was my coding which was the problem

NumDATA<-DATA[,c(3:13)]
scaled<-scale(NumDATA)
PPT_and_Intervention<-DATA[,c(1,2)]
DATASCALED<-cbind(PPT_and_Intervention,scaled)

and a similar warning comes up

Warning messages:
1: In optwrap(optimizer, devfun, getStart(start, rho$lower, rho$pp),  :
  convergence code -4 from nloptwrap
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  unable to evaluate scaled gradient
3: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge: degenerate  Hessian with 1 negative eigenvalues
4: In as_lmerModLT(model, devfun) :
  Model may not have converged with 1 eigenvalue close to zero: 4.8e-09
#19

What you have done for scaling the variables seem OK but you should take the habit to check.
Here the check would consist in testing that means are 0 and sd 1.
To do the whole thing, I would perhaps use {dplyr}, particularly if you are unsure of your *apply() calls.
Here is an example using the iris datset:

library(dplyr)

# scalling:
iris %>%
  mutate_if(is.numeric, scale) -> scaled_iris

# check if all means are 0
scaled_iris %>%
  summarise_if(is.numeric, ~ near(mean(.x), y = 0))
#>   Sepal.Length Sepal.Width Petal.Length Petal.Width
#> 1         TRUE        TRUE         TRUE        TRUE

# check if all sd are 1
scaled_iris %>%
  summarise_if(is.numeric, ~ near(sd(.x), y = 1))
#>   Sepal.Length Sepal.Width Petal.Length Petal.Width
#> 1         TRUE        TRUE         TRUE        TRUE

but indeed this does not seem to solve your problem...
Could you show the model summary of with the scaled variables?

#20

Ah this might be where things are going wrong...

> DATASCALED %>%
+   summarise_if(is.numeric, ~ near(mean(.x), y = 0))
  Intervention Positive_Experience Credibility Expectancy Attitude Trait_Anxiety Anxiety1 Positive_Affect_1
1        FALSE                  NA          NA         NA       NA          TRUE     TRUE              TRUE
  Negative_Affect_1 Anxiety2 Positive_Affect_2 Negative_Affect_2
1              TRUE       NA                NA                NA
> DATASCALED %>%
+   summarise_if(is.numeric, ~ near(sd(.x), y = 1))
  Intervention Positive_Experience Credibility Expectancy Attitude Trait_Anxiety Anxiety1 Positive_Affect_1
1        FALSE                  NA          NA         NA       NA          TRUE     TRUE              TRUE
  Negative_Affect_1 Anxiety2 Positive_Affect_2 Negative_Affect_2
1              TRUE       NA                NA                NA
#21

The missing means are likely due to missing values in the data. The mean function will return NA if at least one value is missing, unless the na.rm argument of mean is set to TRUE (it is FALSE by default). Looking at the summary of your data you posted earlier, the NA values above occurred for those columns with at least one missing value. Try changing mean(.x) to mean(.x, na.rm=TRUE) in your code.

Also, I'm not sure how Intervention is appearing at all in the output if it's coded as a factor. It looks like it must be numeric if it's passing the is.numeric test in summarise_if. I'm also not clear why Intervention isn't being scaled properly by the scale function. For example, if I do x=rep(1:2, c(68,53)) and then mean(x); sd(x) I get the expected result.

Regarding the convergence warnings, these troubleshooting tips might be helpful.

It would definitely help to have a sample of your real data. Could you paste into a reply the output of, say, dput(DATA[1:10, ]) and dput(DATASCALED[1:10, ]).

2 Likes
#22 
> DATASCALED %>%
+   summarise_if(is.numeric, ~ near(mean(.x, na.rm=T), y = 0))
  Positive_Experience Credibility Expectancy Attitude Trait_Anxiety Anxiety1 Positive_Affect_1
1                TRUE        TRUE       TRUE     TRUE          TRUE     TRUE              TRUE
  Negative_Affect_1 Anxiety2 Positive_Affect_2 Negative_Affect_2
1              TRUE     TRUE              TRUE              TRUE
> DATASCALED %>%
+   summarise_if(is.numeric, ~ near(sd(.x, na.rm = T), y = 1))
  Positive_Experience Credibility Expectancy Attitude Trait_Anxiety Anxiety1 Positive_Affect_1
1                TRUE        TRUE       TRUE     TRUE          TRUE     TRUE              TRUE
  Negative_Affect_1 Anxiety2 Positive_Affect_2 Negative_Affect_2
1              TRUE     TRUE              TRUE              TRUE

Thank you so far! This is the data scaled correctly using na.rm=T

#23

Hi Joel,
Here is a sample, I have excluded PPT for confidentiality

> dput(DATA[1:10,-1 ])
structure(list(Intervention = structure(c(2L, 2L, 1L, 1L, 2L, 
1L, 1L, 2L, 1L, 2L), .Label = c("1", "2"), class = "factor"), 
    Positive_Experience = c(28L, 25L, 48L, 44L, 24L, 56L, 44L, 
    27L, 54L, 24L), Credibility = c(14L, 15L, 28L, 18L, 19L, 
    26L, 27L, 17L, 19L, 10L), Expectancy = c(67L, 130L, 148L, 
    62L, 41L, 148L, 126L, 92L, 116L, 47L), Attitude = c(58L, 
    52L, 59L, 57L, 62L, 69L, NA, 68L, NA, NA), Trait_Anxiety = c(50L, 
    49L, 48L, 55L, 47L, 51L, 34L, 38L, 43L, 37L), Anxiety1 = c(46L, 
    45L, 40L, 43L, 39L, 41L, 28L, 45L, 39L, 40L), Positive_Affect_1 = c(20L, 
    27L, 24L, 29L, 15L, 36L, 13L, 26L, 29L, 15L), Negative_Affect_1 = c(13L, 
    14L, 23L, 20L, 30L, 13L, 15L, 23L, 10L, 12L), Anxiety2 = c(40L, 
    46L, 36L, 51L, 40L, 42L, 33L, 47L, 40L, 37L), Positive_Affect_2 = c(19L, 
    30L, 26L, 33L, 14L, 42L, 20L, 28L, 30L, 19L), Negative_Affect_2 = c(11L, 
    11L, 12L, 18L, 28L, 10L, 10L, 17L, 10L, 10L)), row.names = c(NA, 
10L), class = "data.frame")
> dput(DATASCALED[1:10,-1 ])
structure(list(Intervention = structure(c(2L, 2L, 1L, 1L, 2L, 
1L, 1L, 2L, 1L, 2L), .Label = c("1", "2"), class = "factor"), 
    Positive_Experience = c(-0.701604116517139, -0.939112124832279, 
    0.881782605583798, 0.56510526116361, -1.01828146093733, 1.51513729442417, 
    0.56510526116361, -0.780773452622186, 1.35679862221408, -1.01828146093733
    ), Credibility = c(-1.16909004675983, -1.00167858421666, 
    1.17467042884461, -0.499444196587137, -0.332032734043962, 
    0.839847503758257, 1.00725896630143, -0.666855659130311, 
    -0.332032734043962, -1.83873589693253), Expectancy = c(-0.926956916586304, 
    0.424855253435389, 0.811087302013016, -1.03424359674676, 
    -1.48484765342065, 0.811087302013016, 0.339025909307028, 
    -0.390523515784045, 0.124452548986124, -1.35610363722811), 
    Attitude = c(-0.601886922263556, -1.32415122897982, -0.481509537810845, 
    -0.722264306716268, -0.120377384452711, 0.722264306716268, 
    NA, 0.601886922263556, NA, NA), Trait_Anxiety = c(0.40727588473911, 
    0.217004139359062, 0.0267323939790145, 1.35863461163935, 
    -0.163539351401033, 0.597547630119158, -2.63707204134166, 
    -1.87598505982146, -0.924626332921225, -2.06625680520151), 
    Anxiety1 = c(0.360164280390108, 0.148611474918249, -0.909152552441049, 
    -0.27449413602547, -1.12070535791291, -0.697599746969189, 
    -3.44778621810336, 0.148611474918249, -1.12070535791291, 
    -0.909152552441049), Positive_Affect_1 = c(-1.23988545925122, 
    -0.22226628872236, -0.658388790377586, 0.0684820457144573, 
    -1.96675629534327, 1.08610121624332, -2.25750462978008, -0.367640455940769, 
    0.0684820457144573, -1.96675629534327), Negative_Affect_1 = c(-0.69514055812505, 
    -0.525559897775995, 1.00066604536549, 0.491924064318332, 
    2.18773066780888, -0.69514055812505, -0.355979237426941, 
    1.00066604536549, -1.20388253917221, -0.864721218474104), 
    Anxiety2 = c(-0.896585529996111, 0.32415015315244, -1.71040931876181, 
    1.34142988910957, -0.896585529996111, -0.48967363561326, 
    -2.32077716033609, 0.527606100343866, -0.896585529996111, 
    -1.50695337157039), Positive_Affect_2 = c(-1.4696357765789, 
    -0.0873261838257026, -0.589984217554138, 0.289667341470624, 
    -2.09795831873944, 1.4206479173596, -1.34397126814679, -0.33865520068992, 
    -0.0873261838257026, -1.4696357765789), Negative_Affect_2 = c(-0.575431702367056, 
    -0.575431702367056, -0.387860042479452, 0.737569916846171, 
    2.61328651572221, -0.763003362254659, -0.763003362254659, 
    0.549998256958567, -0.763003362254659, -0.763003362254659
    )), row.names = c(NA, 10L), class = "data.frame")
>
#24

Don't forget to do this please:

#25

Sorry Alexandre, here it is now

> summary(anxiety_model)
Linear mixed model fit by REML. t-tests use Satterthwaite's method ['lmerModLmerTest']
Formula: Anxiety2 ~ Intervention + Positive_Experience + Credibility +  
    Expectancy + Attitude + Trait_Anxiety + Anxiety1 + (1 | PPT)
   Data: DATASCALED

REML criterion at convergence: 131.9

Scaled residuals: 
       Min         1Q     Median         3Q        Max 
-6.345e-07 -1.271e-07 -7.250e-09  1.407e-07  6.334e-07 

Random effects:
 Groups   Name        Variance  Std.Dev. 
 PPT      (Intercept) 6.772e-01 8.229e-01
 Residual             3.664e-14 1.914e-07
Number of obs: 98, groups:  PPT, 94

Fixed effects:
                     Estimate Std. Error        df t value Pr(>|t|)
(Intercept)          0.141053   0.154859  0.101408   0.911    0.839
Intervention2       -0.279668   0.310413  0.101408  -0.901    0.839
Positive_Experience  0.004366   0.159561  0.101408   0.027    0.992
Credibility         -0.157852   0.146184  0.101408  -1.080    0.825
Expectancy           0.135684   0.117723  0.101408   1.153    0.820
Attitude             0.147302   0.121889  0.101408   1.208    0.816
Trait_Anxiety        0.026961   0.098053  0.101408   0.275    0.927
Anxiety1             0.527898   0.094465  0.101408   5.588    0.700

Correlation of Fixed Effects:
            (Intr) Intrv2 Pstv_E Crdblt Expctn Attitd Trt_An
Interventn2 -0.840                                          
Pstv_Exprnc -0.618  0.717                                   
Credibility -0.213  0.257 -0.094                            
Expectancy   0.156 -0.183 -0.010 -0.435                     
Attitude     0.014  0.007 -0.133 -0.398 -0.209              
Trait_Anxty -0.059  0.009  0.003  0.010 -0.161  0.102       
Anxiety1    -0.050  0.067  0.163  0.058 -0.085 -0.184 -0.362
convergence code: 0
Model failed to converge with max|grad| = 0.0247718 (tol = 0.002, component 1)
Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?
#26

Do you have a reference to that? I thought it would quietly drop those rows of the data with missing values in fields in the formula. Which is handling missing data, but maybe not in the way you think.

Regards,
Ron.

#27

You have 68 participants (PPT) in one intervention, and 53 in the other, giving a total of 121. I think elsewhere you describe your data as 121 rows with 13 columns. So do you only have one observation per level of PPT? If so, this should be the source of your problem - residual effects being completely confounded with the PPT effect. In this case I think you want an lm rather than an lmer.

Or are participants reused either within or across interventions in some way?

Ron.

#28

Hi Ron,

Thank you for your response.
I have a total of 121 participants in a between-subjects design.
Yes, I only have one observation of each variable per level of PPT

The lmer function is one I have previously been advised to use, do you think this is not the case?

#29

It is not possible to fit the model you are trying to fit, ie with random, independent PPT, with only one observation per PPT. In this case there is no difference between the PPT and the random, residual error terms, hence the failure to converge and residual variance of zero (3.302e-11).

#30

Hi Ron,

Thanks again for your response
I understand what you are saying and I wonder is there any way to use this model but fit it in another way? The only reason I ask is because I have been assured this is model I should use.

Thanks, Karen