Hi Mike,

I am currently trying to fit a multivariate metaSEM model. I am getting a non-positive definite matrix when fitting a random effects model, and this is likely expected due to a lot missing data in the attached datafile. Would you agree that this is the reason behind the non-positive definite error? Please also note that I tried running the model with fewer studies that provided more data (the number of studies went from 108 to 44) but the errors were the same.

Additionally, I can’t seem to fit the fixed effects model. Given that I am new to R, I was hoping you could look at my code below and let me know how you have handled “NA” or missing values in metaSEM. Below is all the code I have used, as well as the errors.

Thanks in advance,

Yusra

#importing the dataset with a matrix for each sample

my.full <- readFullMat("T:/matrix.dat")

#creating an object with the sample sizes

n <- ("100 100 100 100 100 100 241 45 70 122 119 119 447 56 41 38 103 57 192 105 62 48 56 123 56 123 121 103 300 128 106 20 17 75 10 19 21 492 153 107 71 123 23 15 14 22 131 727 30 297 60 125 37 140 88 182 735 45 300 300 77 136 162 65 166 60 75 87 59 57 97 129 242 83 54 78 75 30 60 105 65 64 92 101 120 88 144 120 464 104 296 120 121 95 76 60 256 251 317 52 622 80 297 171 98 354 690 88")

#random effects model

> random.full <- tssem1(my.full, n, method = "REM", RE.type="Symm", RE.startvalues=0.1, RE.lbound=1e-10, I2="I2q", model.name=NULL,suppressWarnings=TRUE)

Error in function (x, n, cor.analysis = TRUE, dropNA = FALSE, as.matrix = TRUE, :

x is not positive definite!

#fixed effects model

> fixed1 <- tssem1(my.full$data, n$n, method = "FEM", cor.analysis = TRUE, cluster = NULL, RE.type= "Symm", suppressWarnings=FALSE)

Error in !all.equal(my.range[1], my.range[2]) : invalid argument type

In addition: Warning messages:

1: In min(x, na.rm = na.rm) :

no non-missing arguments to min; returning Inf

2: In max(x, na.rm = na.rm) :

no non-missing arguments to max; returning -Inf

Attachment | Size |
---|---|

matrix.dat | 62.18 KB |

Hi, Yusra.

Many of the matrices are non-positive definite. You may check the matrices with "FALSE".

> is.pd(my.full)

1 2 3 4 5 6 7 8 9 10 11 12 13

NA NA NA NA NA NA FALSE TRUE NA FALSE FALSE NA TRUE

14 15 16 17 18 19 20 21 22 23 24 25 26

TRUE TRUE NA NA NA TRUE NA NA NA NA NA TRUE TRUE

27 28 29 30 31 32 33 34 35 36 37 38 39

TRUE FALSE FALSE TRUE FALSE FALSE FALSE TRUE FALSE TRUE TRUE NA FALSE

40 41 42 43 44 45 46 47 48 49 50 51 52

NA TRUE TRUE NA TRUE TRUE NA NA TRUE NA NA TRUE FALSE

53 54 55 56 57 58 59 60 61 62 63 64 65

FALSE NA TRUE TRUE FALSE TRUE TRUE NA NA FALSE NA TRUE TRUE

66 67 68 69 70 71 72 73 74 75 76 77 78

FALSE NA NA NA NA TRUE NA FALSE FALSE TRUE TRUE TRUE TRUE

79 80 81 82 83 84 85 86 87 88 89 90 91

NA NA FALSE FALSE TRUE TRUE NA NA NA TRUE NA TRUE TRUE

92 93 94 95 96 97 98 99 100 101 102 103 104

NA TRUE FALSE FALSE FALSE FALSE NA NA NA NA TRUE FALSE FALSE

105 106 107 108

TRUE TRUE TRUE TRUE

For example, the correlation between x5 and x12 in my.full[7] is 1.47.

> my.full[7]

$`7`

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13

x1 1.00 NA NA NA 0.99 0.95 NA NA NA NA NA 0.93 0.44

x2 NA NA NA NA NA NA NA NA NA NA NA NA NA

x3 NA NA NA NA NA NA NA NA NA NA NA NA NA

x4 NA NA NA NA NA NA NA NA NA NA NA NA NA

x5 0.99 NA NA NA 1.00 1.00 NA NA NA NA NA 1.47 0.25

x6 0.95 NA NA NA 1.00 1.00 NA NA NA NA NA 0.78 0.20

x7 NA NA NA NA NA NA NA NA NA NA NA NA NA

x8 NA NA NA NA NA NA NA NA NA NA NA NA NA

x9 NA NA NA NA NA NA NA NA NA NA NA NA NA

x10 NA NA NA NA NA NA NA NA NA NA NA NA NA

x11 NA NA NA NA NA NA NA NA NA NA NA NA NA

x12 0.93 NA NA NA 1.47 0.78 NA NA NA NA NA 1.00 0.26

x13 0.44 NA NA NA 0.25 0.20 NA NA NA NA NA 0.26 1.00

Mike

Hi Mike,

Thanks for pointing it out! The correlations were in Fisher's z metric but once I changed the correlations to the r metric, I still get the same errors ("not positive definite" for the random effects model, and "In min(x, na.rm = na.rm) :

no non-missing arguments to min; returning Inf error" for the fixed effects model). I'm also getting the following error when checking if the matrices are positive definite:

is.pd(my.full3)

Error in eigen(x, only.values = TRUE) : infinite or missing values in 'x'

I've attached the dataset once again. Is the problem related to how missing values are being handled? Thanks again for all your help!

Yusra

Hi, Yusra.

There are still problems in your data.

> my.full3 <- readFullMat("rmatrix.dat")

> is.pd(my.full3)

1 2 3 4 5 6 7 8 9 10 11 12 13

NA NA NA NA NA NA FALSE TRUE NA FALSE FALSE NA TRUE

14 15 16 17 18 19 20 21 22 23 24 25 26

TRUE TRUE NA NA NA TRUE NA NA NA NA NA TRUE TRUE

27 28 29 30 31 32 33 34 35 36 37 38 39

TRUE TRUE FALSE TRUE TRUE TRUE FALSE TRUE FALSE TRUE TRUE NA TRUE

40 41 42 43 44 45 46 47 48 49 50 51 52

NA TRUE TRUE NA TRUE TRUE NA NA TRUE NA NA TRUE TRUE

53 54 55 56 57 58 59 60 61 62 63 64 65

TRUE NA TRUE TRUE TRUE TRUE TRUE NA NA TRUE NA TRUE TRUE

66 67 68 69 70 71 72 73 74 75 76 77 78

FALSE NA NA NA NA TRUE NA FALSE TRUE TRUE TRUE TRUE TRUE

79 80 81 82 83 84 85 86 87 88 89 90 91

NA NA FALSE TRUE TRUE TRUE NA NA NA TRUE NA TRUE TRUE

92 93 94 95 96 97 98 99 100 101 102 103 104

NA TRUE TRUE TRUE TRUE TRUE NA NA NA NA TRUE TRUE TRUE

105 106 107 108

TRUE TRUE TRUE TRUE

For example, the 7th matrix is not positive definite. The correlation between x5 and x6 is 1.

> index.na <- is.na(diag(my.full3[[7]]))

> my.full3[[7]][!index.na, !index.na]

x1 x5 x6 x12 x13

x1 1.00 0.76 0.74 0.73 0.41

x5 0.76 1.00 1.00 0.90 0.24

x6 0.74 1.00 1.00 0.65 0.20

x12 0.73 0.90 0.65 1.00 0.25

x13 0.41 0.24 0.20 0.25 1.00

Regards,

Mike

Thanks! I will look into this.