Agreement between two methods

[Dennis Massue]

All:
I am new member to The STATA Forum.

I want to assess the agreement between methods (A-standard and B-new) used in assessing performance of bednets. I am using Bland and Altman analysis method. I summarize the results I got,

Mean of Method A- 68.4
Mean of Method B- 93.4

When I use batplot with trend, I get the follwowing results;
7/145 = 4.83% outside the limits of agreement
Mean Diff= -130.14 + 1.28*Average
Limits +/- 2.46* (36.93 + -0.35*Average)

When I use batplot with notrend, I get the following results;
Mean difference = -25.93678138995993
Limits of agreement = (-87.3091290668672, 35.43556628694733)
Averages lie between 22.917 and 100.000


Questions;
With such reults;
1) Which one should I use?
2) How to calculate percentage error?
3)How to interpret the results?

Thanks in advance,

[Nick Cox]

It wouldn't be responsible to advise without seeing the graph.

[Dennis Massue]

Thanks Nick,

From my earlier problem, the BA graphs are bellow; (i do not know if they can be seen clearly)

BA graph with trend.gph

BA graph with notrend.gph

[Nick Cox]

The graphs suggest that you need to compare measurements on some nonlinear scale. They seem a long way from two methods separated by different overall biases, which is the classic playground for this kind of analysis.

But please do see http://www.statalist.org/forums/help#stata 12.4 and 12.5 on posting .png not .gph.

What is unclear about

Home page:

Please do read the Statalist FAQ for crucial advice before you try to post a message to Statalist.

New topic prompt:

First read Advice on Posting.

?

Will someone please tell me?

[Dennis Massue]

Hi Nick,

Attached herewith are the correct format (in .png) of BA graphs from my earlier examples

BA graph with notrend

BA graph with trend

[Nick Cox]

Thanks for posting graphs that are more easily readable. I trust you now see why we advise that format.

I don't know what A and B are or what "bednets" are. At a guess, each measure has a bound of 100 (%) and this measurement scale affects what you can do. It's not obviously appropriate for Bland-Altman type analysis as for example any relationship between A and B must respect the corner (100, 100) and a linear relationship of any kind is most unlikely.

You could try posting the data, but I'd suggest that you need support from your advisor or supervisor on this.

I have never used batplot (where does it come from? you are asked to explain that), as I tend to use concord (SJ) instead. Why is one data point so much bigger?

[Dennis Massue]

Dear Nick.
Thanks for the reply and Yes, I now understand how to insert graphs. Just to clarify further, A and B are methods that were used to assess performance of bednets. Outcomes compared is proportion dead (mortality- percent dead). What I did was to compare the agreement between the two methods on how they measure bednet perfomance on killing mosquitoes. This is done for the aim of comparing the method B (new method) with the method A (standard method). Therefore the graphs I posted were from Bland-Altman analysis (Ref: Bland and Altman, 1986).
Thanks for the suggestion of using concord (concordance correlation coefficient). Thank you for this. Just a small qustion on this, can I use this command for comparing the agreement between two methods (as I explained above)? Is the command available in STATA 13?

[Nick Cox]

concord can be used in Stata 13. You just need to install it.

search concord

and use the most recent update. Alternatively, type

net describe st0015_6, from(http://www.stata-journal.com/software/sj10-4)

and follow the install instructions.

Your dataset doesn't look very large, although I don't know how many points lie in that big blob. If you were able and willing to post the raw data, I could have a look. I think you need to analyse your data on a transformed scale.

It doesn't seem that you have read the FAQ all the way to the end! http://www.statalist.org/forums/help#spelling