topic Re: Log transformation of negative values in SAS/IML Software and Matrix Computations

Log transformation of negative values

iaatoloye — Wed, 20 Oct 2021 12:54:24 GMT

Hello,

I have some data on net nitrogen mineralization and the values range from negative to positive. I measured nitrogen mineralization over several days and obtained the net nitrogen mineralization. I tried the approach below but could not obtain an output to merge to my data set. @Rick_SAS proc iml outputoriginal data

Thanks,

proc iml;
use work.idnet2;
read all var {NetNit };

LNetNit = log10(NetNit + 1 - min(NetNit )); /* translate, then transform */

LogNetNit = j(nrow(NetNit ),1,.); /* allocate missing */
idx = loc(NetNit > 0); /* find indices where Y > 0 */
if ncol(idx) > 0 then
LogNetNit [idx] = log10(NetNit [idx]);

print NetNit LNetNit LogNetNit ;

How do I obtain my output for further analysis? How do I obtain back-transformed means and standard errors??

With the LNetNit transformation, all values were positive? Is this acceptable?

With the LogNeNit, I had a lot of missing data and negative values.

Transforming and interpreting net values with positive and negative values

iaatoloye — Wed, 20 Oct 2021 12:19:13 GMT

Hello,

Thanks,

proc iml;
use work.idnet2;
read all var {NetNit };

LNetNit = log10(NetNit + 1 - min(NetNit )); /* translate, then transform */

How do I obtain my output for further analysis? How do I obtain back-transformed means and standard errors??

With the LNetNit transformation, all values were positive? Is this acceptable?

With the LogNeNit, I had a lot of missing data and negative values.proc iml outputData

Re: Transforming and interpreting net values with positive and negative values

Ksharp — Wed, 20 Oct 2021 12:31:33 GMT

Plz post it at IML forum since it is about SAS/IML .

And calling @Rick_SAS

Re: Log transformation of negative values

Rick_SAS — Wed, 20 Oct 2021 12:59:56 GMT

If I understand your question, you don't want to check whether NetNit is positive, you want to check whether the translated data is positive:

proc iml;
use work.idnet2;
read all var {NetNit };
LNetNit = log10(NetNit + 1 - min(NetNit )); /* translate, then transform */

LogNetNit = j(nrow(NetNit ),1,.); /* allocate missing */
Y = NetNit + 1 - min(NetNit );   /* <== use the translated data */
idx = loc(Y > 0);      /* find indices where Y > 0 */
if ncol(idx) > 0 then
   LogNetNit[idx] = log10(Y[idx]);

print NetNit LNetNit LogNetNit ;

Re: Log transformation of negative values

iaatoloye — Wed, 20 Oct 2021 13:18:45 GMT

Yes, NetNit contains both negative and positive values and when I take the log of NetNit in Proc Glimmix, I end up having missing values.

However, with the code below, I have the data as positive values. I am not sure if this is a valid transformation approach. If it is, how do I use the output from proc iml in proc glimmix and back-transform the means?

LNetNit = log10(NetNit + 1 - min(NetNit )); /* translate, then transform */

I was hoping to get positive values with the code below but I ended up having missing and negative values.

LogNetNit = j(nrow(NetNit ),1,.); /* allocate missing */

Y = NetNit + 1 - min(NetNit ); /* <== use the translated data */

idx = loc(Y > 0); /* find indices where Y > 0 */

if ncol(idx) > 0 then

LogNetNit[idx] = log10(Y[idx]);

Thank you,

Re: Log transformation of negative values

Rick_SAS — Wed, 20 Oct 2021 13:49:55 GMT

> I was hoping to get positive values with the code below but I ended up having missing and negative values.

The code will give you only nonnegative values for Y. If NetNit is missing, then Y will be missing (as it should). The code translates the data by some data-dependent value into the interval [1, infinity).

The interpretation is that you are measuring the data using some different baseline reference. For example, if you measure elevation from sea level, the elevations of Death Valley and the Mariana Trench are negative. But if you measure from one meter below the bottom of the Mariana Trench, then all elevations on earth are positive. In the language of distributions, you are translating the data distribution so that it has a new mean, but preserving other properties of the distribution.

I don't know your model, so I can't really comment on PROC GLMMIX. But I encourage you to consider whether you should log-transform negative values in the first place.