Understanding Very Scientific Academic Article

[kage65]

---

Members don't see this ad.

I'm struggling to understand a very ( to me ) high level academic article on Predicting of energy expenditure from heart rate monitoring.
Article is here http://www.braydenwm.com/cal_vs_hr_ref_paper.pdf

Specifically I'm trying to answer the question "what is the degree of accuracy of this method of measuring energy expenditure?".

The text which I'm having trouble with is in the abstract section, specifically:

"The correlation coefficient (r) between the measured and estimated
energy expenditure was 0.913. The model therefore accounted for 83.3%
([R.sup.2]) of the variance in energy expenditure in this sample"

Any help is appreciated. Thanks.

 

[starrsgirl]

Is this supposed to be a statistical answer or are you supposed to explain what's going on? I'm not a PT yet, but I do have a masters in exercise science and could help explain what's going on here. Unfortunately, I've only had basic pre req statistics so if you need more than that I'm useless.

 

[kage65]

Hi, basically trying to answer the question "what is the degree of accuracy of this method of measuring energy expenditure?" Meaning, can it be off by 10%, 20% , more? Tks

 

[AlanWattsBlues]

Hi, basically trying to answer the question "what is the degree of accuracy of this method of measuring energy expenditure?" Meaning, can it be off by 10%, 20% , more? Tks

It looks like Table VII in your article might give you some sense of things. It has what appears to be 95% confidence intervals for the errors, which can give you a sense of the spread that might reasonably be expected.

I guess it's not really a question of "Can it be off by 10%, 20%....?", it's a question of "What is the probability that the error exceeds 10%, 20%?". With regressions, I believe it should be the case that the residuals (the error between the predicted and actual value) are normally distributed with mean of zero. With a normal distribution, large deviations from the mean are unlikely, though always possible.

If you had the variance of the residuals, then you might be in a position to answer the question of "How likely is it that this estimate is off by greater than 10%?". To get the residual variance, you could look at the 95% confidence intervals and the mean biases in Table VII, then back into the standard deviations to answer the question of "What is the probability that the error exceeds 5,10,...?". Note that this is not measured in percentage terms, but is in absolute values.

So the 95% CI is MeanBias +/- 1.96*StandardError. Solve for the Standard Error then you can find the probability that the error is within a certain range using a standard normal table and the Mean Bias.

I'm no statistician, so I may be entirely off on some of these things. I know that things can very quickly get very complicated (Heteroscedasticity...I'll stop there). But this might be one way of thinking about your question.

 

[Azimuthal]

This is why no one should ever attempt to interpret literature via abstract.

 

[kage65]

So a statistician is the person who would be able to interpret this best?

 

[AlanWattsBlues]

So a statistician is the person who would be able to interpret this best?

If you have questions about this research, it might be worth directing them to L. R. Keytel, who is listed as the point-of-contact in the paper you linked to above (email address is on the last page). Who knows? You might get a reply from the source...

 

[Azimuthal]

So a statistician is the person who would be able to interpret this best?

Anyone who has taken a research methods course can interpret literature fairly well.

 

[goyo1010]

"The correlation coefficient (r) between the measured and estimated
energy expenditure was 0.913. The model therefore accounted for 83.3%
([R.sup.2]) of the variance in energy expenditure in this sample"

For your future reference:

Correlation coefficient (for regression), or R, is usually represented by a decimal or percentage. In this case, R = 0.913, or 91.3%. Meaning that the measured EE and estimated EE were highly correlated (in this case a very good thing) because the estimated EE (using their equation) was very similar to the values gained from the measured EE. The model they used, or the equation they managed to derive through a statistical process called multiple regressions (to identify predictors or variables that can explain a certain relationship), managed to explain 88.3% of the variance, or variability of the data. This is represented by the value of R^2.

R^2, when reported tells us the proportion of variance that is explained by the model. The more variance that is explained by the model, the more accurate the model is. This is because those predictors that were identified were found to be big players in determining that value of estimated EE. They found that the first model that included gender, age, weight, HR, and VO2 better explained the variance than when they used the second model, which did NOT include VO2 (R=0.857, R^2=73.4%). The first model more closely resembled actual measurements of EE.

Hope that wasn't too confusing. I didn't exactly go through if they met assumptions or not (a previous poster mentioned heteroscedasticity). So, yeah.

 

[kage65]

Thanks to all who replied.

 

[NewTestament]

Anyone who has taken a research methods course can interpret literature fairly well.

Fairly well, but an author could confuse any clinician with too many numbers without explanation. I myself have taken two levels of research and have read my share of articles and still have trouble understanding the data sometimes.

 

[Azimuthal]

Fairly well, but an author could confuse any clinician with too many numbers without explanation. I myself have taken two levels of research and have read my share of articles and still have trouble understanding the data sometimes.

If the data and methods used are difficult to interpret, is it clinically relevant to you? Maybe. Maybe not. That is why we review articles based on criteria such as the CEBM. If the discussion section require further explanation, you may contact the author. I know that this is sometimes difficult to do. Even more so when custom software is used for data collection and statistical analysis are used. However, if we're speaking of results using CI, p-value, r, r^2, etc., those statistical variables represent the same meaning. Articles that omit data or interpretation may speak to the quality of the article.