Bland Altman plot: Log-transformation and back transformation of LOA

[morten1]

Hi,

I want to compare two methods with a Bland Altman plot. The difference between measurement 1 and measurement 2 are non-normally distributed data, therefore, I want to log-transform the data. I get the bias, SD and LOA on the log-transformed data, and can make the BA-plot.

My problem is when I am transforming back the LOA to original scale, which I want to insert on the original BA-plot, for easier interpretation. How do I calculate the slopes of the LOA and how do I know the intersection with the y-axis?

I also want to make 95 % CI of the mean difference of this data, how can I do this on log transformed data when I want the 95 % CI to be on the original scale? (for comparison to other normally-distributed data). Or should I just calculate 95 % CI of the mean difference on non-normally distributed data?

Thanks!!

[Oligolas]

Not sure I understood but I will try to answer.

My problem is when I am transforming back the LOA to original scale, which I want to insert on the original BA-plot, for easier interpretation.

When I plot with a log scale i am not doing any transformation, I specify "TYPE=LOG" in my GTL or SGPLOT code, SAS does the job.

Generally speaking avoid any back-transforming of your data, maybe create a macro that does the plot depending of the type of analysis you want (linear or logarithmic).

"How do I calculate the slopes of the LOA"... If you mean the regression line use PROC REG:

   ods select none;
   proc reg data=inputDS;
      model yvar = xvar ;
      ods output ParameterEstimates=outputDS;
   run;quit;
   data _null_;
      set outputDS;
      if variable eq 'Intercept' then call symput('Int', put(estimate, BEST.));    
      else            call symput('Slope', put(estimate, BEST.));  
   run;
   ods select all;

how do I know the intersection with the y-axis?

y = slope*x +intercept  intersection with y-axis is your intercept.

I also want to make 95 % CI of the mean difference of this data, how can I do this on log transformed data when I want the 95 % CI to be on the original scale? (for comparison to other normally-distributed data). Or should I just calculate 95 % CI of the mean difference on non-normally distributed data?

 

As I said, avoid back-transforming, calculate the 95%CI on original scale data.

Cheers