# Why doesn't my Garmin VO2 max estimate agree with the Cooper test estimate?

I've been performing Cooper tests each month to track my fitness. Basically you run as far as you can in 12 minutes and estimate your VO2 max using the formula,

where d12 is the distance covered in 12 minutes, measured in metres.

Yesterday I ran 2730m, which gives a VO2 max of 49.75 mL/min/kg, which is substantially higher than the estimate from my Garmin watch of 45.99 mL/min/kg. So what are the likely causes of this and what can I learn about my running from the difference?

I have worked out that the Cooper test estimate is not very precise (see my answer below) but I suspect another part of it is that the Cooper test is maximal effort, whereas the Garmin estimate is an average over a long period of time and is based on mostly sub-maximal efforts. If the algorithm for extrapolating from sub-maximal efforts is badly calibrated for me, then it might under-estimate my true VO2 max. I am wondering whether this means my running or cardiovascular efficiency is less than average when I am running slowly than it is for maximal efforts (and whether I can do anything to improve that).

Not an answer per se, but here's, to my knowledge, the whitepaper Garmin used for their VO2max calculations : https://assets.firstbeat.com/firstbeat/uploads/2017/06/white_paper_VO2max_30.6.2017.pdf

Therein, they state their error is about 5% (but highly dependent on an accurate value for MaxHR). From your extracted data, you could easily calculate the average error in Cooper's method to compare!

• Unfortunately IIRC the white paper doesn't explain exactly how the 5% figure is calculated - it seems a bit optimistic, especially given the number of YouTube videos comparing Garmin estimates with lab based one - many of which show much larger differences. I did calculate the uncertainty in Cooper's method in my answer - the 95% prediction interval is roughly +/- 6.5 mL/min/kg around the predicted value - very imprecise. Commented May 10 at 8:20
• Rereading the paper, it does say that the model was validated using 79 runners who were all preparing for a marathon, which means they are most certainly not a representative sample of recreational runners, and AIUI marathon running is not really about VO2 max. I suspect the errors would be higher for a wider range of runners. They do say a "vast majority" of them has errors below 3.5mL/kg.min, which suggests it is more accurate for marathon runners than the Cooper test is for a less specific set of runners. Thanks for introducing the white paper into the discussion (+1) Commented May 10 at 8:58

Just to add some (hopefully) useful information - one issue is that the Cooper test estimate is highly uncertain. I got Cooper's paper from 1968

Cooper, Kenneth H. (15 January 1968). "A Means of Assessing Maximal Oxygen Intake: Correlation Between Field and Treadmill Testing". JAMA. 203 (3): 203. doi:10.1001/jama.1968.03140030033008. ISSN 0098-7484.

I then digitised the data in Figure 1 by hand, so that I could re-analyse it. I could only identify 104 distinct points and apparently there are 115 subjects, so I suspect that I have missed some co-located crosses. The reanalysis therefore is not quite the same, but close enough for the point I want to make. Here is my reproduction of figure 1, with the addition of lines indicating my most recent test:

My reconstruction gives a VO2 Max of 49.26 mL/min/kg instead of 49.75 mL/min/kg, but close enough.

I think Cooper made a mistake in regressing distance as a function of VO2 max, since the aim is to estimate VO2 max from a measured distance, I think it makes more sense to perform the regression the other way round:

This shows the 95% prediction interval for my VO2 max based on my most recent performance. It gives an expected VO2 max of 48.92 mL/min/kg, but the prediction interval is very wide, covering 42.42 - 55.43 mL/min/kg, so my Garmin estimate is evidently consistent with the Cooper test estimate, providing you take into account the uncertainty in the estimate.

Perhaps on-line Cooper test calculators should give the prediction interval rather than just a single most likely value?

The Garmin algorithm is also likely to have a prediction interval - it would be very useful to know that as well. Being based on a larger amount of data from previous activities seems likely to make it narrower, but that may be offset by the uncertainty in the correction for sub-maximal effort.

• @downvoter why the downvote? Answering your own questions is permitted on SE, and this is only a partial answer that I thought others might find useful. If you have a technical criticism of what I have written, it would be more productive if you could state it clearly in the comments so I can address it. Commented May 9 at 14:31