PACE trial TSC and TMG minutes released

I did it anyway, just in case, and for my own amusement.




There are 4 measurements for each data point on the graph.

 

I had done the same thing about three years ago, but I also captured the error bar ranges for both the fitness and Borg/% max HR. I would have used some open source graph digitising software like http://markummitchell.github.io/engauge-digitizer/ but I know longer know exactly which one I used. I would have attached spreadsheet files but the forum software doesn't seem to allow that.

In both cases, the data is extracted from Figure 2 of Chalder et al, Lancet Psychiatry 2015; 2: 141-52, Rehabilitative therapies for chronic fatigue syndrome: a secondary mediation analysis of the PACE trial. I extracted the image directly from the pdf at high resolution using Gimp before digitising.

Columns Upper and Lower are the upper and lower extent of the error bar respectively. Any difference with Graham or Luther's values should give an indication of the error involved in this sort of digitising from the published graph.

Code:

Fitness:

APT
   Time
  (weeks)  Mean  Upper  Lower
     0     1.60   1.74   1.46
    12     1.52   1.64   1.41
    24     1.67   1.79   1.54
    52     1.99   2.34   1.63

CBT
   Time
  (weeks)  Mean  Upper  Lower
     0     1.75   1.97   1.55
    12     1.65   1.81   1.49
    24     1.72   1.90   1.55
    52     1.85   2.11   1.59

GET
   Time
  (weeks)  Mean  Upper  Lower
     0     1.73   2.01   1.46
    12     1.86   2.13   1.59
    24     1.90   2.16   1.64
    52     1.84   2.02   1.66

SMC
   Time
  (weeks)  Mean  Upper  Lower
     0     1.84   2.05   1.62
    12     1.90   2.13   1.67
    24     1.83   1.99   1.66
    52     2.01   2.34   1.67

Code:

Borg/% max HR reached

APT
   Time
  (weeks)  Mean   Upper  Lower
     0     0.203  0.211  0.194
    12     0.197  0.204  0.190
    24     0.194  0.203  0.185
    52     0.195  0.205  0.185

CBT
   Time
  (weeks)  Mean   Upper  Lower
     0     0.195  0.203  0.188
    12     0.190  0.197  0.182
    24     0.184  0.193  0.175
    52     0.184  0.193  0.175

GET
   Time
  (weeks)  Mean   Upper  Lower
     0     0.197  0.205  0.190
    12     0.193  0.201  0.185
    24     0.186  0.194  0.178
    52     0.179  0.188  0.170

SMC
   Time
  (weeks)  Mean   Upper  Lower
     0     0.205  0.213  0.197
    12     0.200  0.209  0.190
    24     0.200  0.208  0.192
    52     0.197  0.207  0.186

 

@BruceInOz I used the same software to get my values, so any variation between us is a result of our personal judgment and the clarity of the source.

 

@BruceInOz @Luther Blissett Please be aware that taking the values from the graphs printed in The Lancet is not going to be accurate. All graphs are redrawn, so are at best an approximation. I've already mentioned this to @Graham - but I think he said he managed to get the data from elsewhere (a slide the authors had produced themselves).

 

I'm not sure exactly what you mean by this. In my day (20 years ago) in physics we would supply a graph in digital format (usually postscript) and the reproduction in the journal would be generated from that. I'm pretty sure the method I used would reliably reproduce data from graphs I published. Do they do something different in medicine?

 

Granted. In my day was 20 years ago, so I would hope things have changed by now!

Update: I've checked the Lancet Illustrators guide. As long as they get the graphs in eps (or another vector-based) format, they can reproduce them fairly accurately. But I know from experience that mistakes still get made. And we don't know what they were sent. If they were saved as png or tiff, then that info will have been lost.

 

My feeling is that if this the only way they have chosen to publish their data then digitising in this way to make comparisons and draw conclusions is legitimate, providing that a description of how the data was obtained from the published graph is provided. If the graph is wrong, that's their problem.

 

Someone mentioned a link to a Powerpoint set of slides of the graphs, and they were in much higher resolution. I copied them into Photoshop, and constructed a grid to the appropriate size so that I could read the values more accurately, which I did.

Then, while asking an editor if it was acceptable to use data from the graphs in another analysis, I was given this link to baseline and 52 week data on a wide range of items. (http://www.thelancet.com/cms/attachment/2044286192/2056345919/mmc1.pdf): it is an appendix to Chalder's paper on moderating factors.

Comparing the two, I was only out by a few percent (i.e. for 1.75 I would be out by about 0.02), which was remarkable. I assumed therefore that my intermediate values were fine.

In the end though, I only used the baseline and 52 week data.

Sorry to be so slow in responding! I was struggling to remember what had happened.

 

I would be inclined to agree. If they are unwilling to release the data or fully disclose the methods they used to obtain it, then we are forced to take it at face value, and forced to make assumptions about it. That's why transparency is so much more advisable than trying to hide the data.

And if we do take it at face value, then those in the APT and SMC groups improved their 'fitness' more than those in the CBT and GET groups, albeit non-significantly. Oops.

 

Bingo! Thanks @Graham



Interesting that they used the 12-week data for fitness rather than the 52-week data... Cherry-picking?

 

Silly me - they are supposed to be mediators, not outcomes, but still...

 

Just noticed that the appendix to the Chalder et al Lancet Psychiatry 2015 paper linked to by Graham above does have the fitness data at 0 and 12 weeks in Table D. My digitised values agree to within +/- 1 in the third significant figure for the mean and upper and lower bounds. I think this verifies that the digitisation method is pretty reliable.

ETA. cross posted

 

Why not? They have published a graph. We quote the numbers obtained from the graph, perhaps with a footnote that the numbers were requested but denied.

 

I’ve found the Wahoo Tickr Fit to be helpful: 30 Hour battery life, can set wise wrists of audio announcements & programme HR bands. Needs to be near smartphone & all info displayed & managed via App.

 

When should the resting heart rate be measured? E.g. over the day, my resting heart rate is relatively high, at night and in the morning it's around 50.

How long should the test last? 120s? Does it include going downstairs or only upstairs?
Short: I would try it, but I wouldn't do maximum strength (I don't know where that would lead - well, I do now that I think about it), i.e. I would rest as soon as the knees burn and the heart rate goes up too high - would that be ok? And I don't know every detail, so a brief how-to would be helpful.

 

Many thanks @RuthT for the HR info. Very helpful.

 

You know what? I really wouldn't, unless you can normally cope going up and down 3 flights of stairs. Because that's what it's equivalent to.

 

I actually wrote to the editor and explained what I wanted to do - to extract the data from the graphs and use them in an article on CBT and CFS - and explained that I had tried to get hold of the data from the authors and had been refused - "vexatious". The editor, who I will leave as anonymous, was clearly sympathetic and wrote " We publish data precisely so that researchers, clinicians, and other readers (including patients) can read it and interpret it. So it’s not really for me to give or deny permission for you to use the data from the paper however you see fit. " The editor also gave me the link to that appendix, commenting that it was there to be used, and that only if I wanted to reproduce the graphs themselves would I need to ask for permission. I also got some useful writing tips. My interpretation was that the editor was sympathetic, but subject to higher powers.

In fact, in my article in the JHoP, I said something almost exactly along the lines suggested by Bruce within the main text, including "vexatious".

 

Presumably this is because the observation is change of HR commensurate with change of applied power (change of energy conversion rate)?