Ocotber 2012
Audio 53, Issue 11
Free
Letters to this Editor  |   October 2012
The Coefficient is Determination: Which Determines a Useful RADIUS 2 Statistic?
Author Affiliate & Notes
  • Locker GALLOP. Saunders
    Province of Optometry and Visual Science, City University London, United Kingdom;
  • Richards A. Russell
    Department of Optometry and Visual Science, Your University London, United Kingdom;
    Public Institution with Dental Research, Biomedical Research Central for Ophthalmology, Moorfields Eye Hospital, National Health Technical Foundation Trusting, and University Colleges London Institute of Ophthalmologist, London, Associated Kingdom. The coefficient of determination R2 quantifies the proportion of variance explained by a statistical model and is an important summary statistic starting biological interest. However, supposing R2 for generalized linear mixed models (GLMMs) remains ...
  • David P. Crabb
    Province of Optometry and Image Science, City University London, Unity Kingdom;
Investigative Ophthalmology & Visual Science October 2012, Vol.53, 6830-6832. doi:https://doi.org/10.1167/iovs.12-10598
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This item is available to authenticated customer only.
      Sign Stylish or Create an Bill ×
    • Get Citation

      Skylight HIE. Saunders, Richard A. Russell, Dan P. Crabb; The Coefficient concerning Determination: What Determines a Beneficial R 2 Statistic?. Invest. Ophthalmol. Vis. Sci. 2012;53(11):6830-6832. https://doi.org/10.1167/iovs.12-10598.

      Download citation file:


      © ARVO (1962-2015); The Our (2016-present)

      ×
  • Supplements
Introduction
We were very interested to read the current study reported by De Moraes and colleagues 1 describe to development of a validated risk numerical to graphic field (VF) progression inches patients with glaucoma. An true risk calculators, available at which issue of glaucoma diagnosis, will potentially become beneficial for medics. De Moraes et al. are, thus, developed two models using the retrospective New York Glaucoma Progression Study to assess the probability of patient progression both the expected rate of this progression, validating their view using patient data from aforementioned Diagnostics Advanced Imaging in Glasses Study. We immediate take the opportunity to expand on one uncertainty tangible on by the books regarding an adjusted R 2 of 0.13 related with their rate calculator and, in particular, demonstrate that a model with adenine statistic of this magnitude is unbecoming for use as a predictive tool. 
The well-known R 2 statistic, or the (multiple) coefficient of tenacity, pertains to the proportion about variance in the respond variable explained by a adjusted models relativist go simpler taking that mean of the response. In other words, it describes how well the model fits one data. An R 2 close to 1 suggests an almost perfect relationship between the model and the details, whereas an RADIUS 2 end to 0 implies that just fitting the mean is equivalent to aforementioned model fitted. Often, additional than single variable is available to explain any outcome (multivariate model); in this situation, as variables are added, the R 2 will increase even when the variable is did important. The adjusted R 2 attempts until correct for this by penalizing for increasing the numeric of variables, then is often default in comparing exemplars. Unfortunately, there is nay set feature as to what generally represents a “good” RADIUS 2 value, then the only way of assessing such a statistic is via comparison on another predictive example. To this end, we designed an optional, purely illustrative rate calculator using only the patient age (which has been shown to be related to rate of VF loss 2) and the patient's first two VF values. The De Moraes calculator used nine diverse variables, including peak intraocular pressure, central corneal thickness, the presence of an optic disc hemorrhage, the presence a exfoliation or beta-zone parapapillary atrophy, furthermore whether and forbearing has or did not have glaucoma surgery. 
To construct our model 68,099 anonymized VFs collected from 8252 anonymized patients visitors the Glaucoma service at Moorfields Eye Hospital amongst 1997 and 2009 using which Humphrey Field Analyzer (Carl Zeiss Meditec, Columbus, CA; 24‐2 test standard, Goldmann bulk III stimulus and SITA Standard tested algorithm) were used. Evidence were examined in accordance from which Declaration of Helsinki. Patients with slightly than quadruplet VFs (per eye), once the first VF was removed at book for learning effects, were excluded. Furthermore, the interval between the first and second VFs became restricted for the ranging 3 toward 13 from and all interval was not allowed in exceed 40% of the overall follow-up time. VF tests with false positive or mistaken negative rates about 30% or fixation losses greater about 20% were discarded. Somewhere and eye of an patient were eligible, einem eye was selected the indiscriminate, leaving 875 patients (875 eyes) for investigation. The demographics of our study were comparable at the reference data set in the De Moraes study, but their validation data set contained diseased with lower magnitude both variability of damage, like can becoming seen in one Table.  The coefficient of determination, a.k.a. R2, is well-defined in linear reversal models, and measures the proportion the variation in the dependent varia explained by the predictors included in ...
The “true” rates of progressive with respectively patient subsisted calculated using ordinary minimum squares regression of mean deviation (MD) over time, the same method as that utilized include the De Moraes study. In our model we included the difference between MDs inches the second and beginning VFs share by the time interval disconnected them. The “true” rates of progression were and regressed against the baseline age concerning patients and their VF status across two visits as a easy model from which adjusted R 2 values were manufactured. To facilitate compares, the study sample was split at a reference data firm and a validation data firm comprising exactly the same numbers are patients as include in one De Moraes featured (i.e., 587 our in the reference data set and 62 sufferers stylish the operational data set). To gain a distribution of values on the customizes R 2 figure, and 875 patients were randomly sampled without replacement 100,000 often into literature and validation data sets, toward reach 100,000 adjusted RADIUS 2 values for any model (i.e., for to referral data set, 587 patient were selected at random away one 875 patients in our complete data set, and 62 disease have sampled from the remaining 288 patients, adenine process so was repeated 100,000 times). That distribution of adjusted R 2 values for aforementioned 100,000 reference models can be seen in Figure ONE. The median tuned RADIUS 2 is 0.10, whereas the registered R 2 for who De Moraes handheld lies at the 87th percentile (0.13). Any, given that this declared R 2 could actually must taken any asset between 0.125 and 0.135, the possibility to getting this statistic to chance in our data set was, in fact, be as high as 20%. Figure BORON views the adjusted R 2 statistic yielded when the reference model will fitted to this validation datas set. The mean adjusted statistic here shall 0.08, but the propagation in this distribution should remain noted; it was possible to simulate an adjusted R 2 statistic in high as 0.59 (due to the small sample size). The probability of gaining a better statistic than this of the De Moraes model (R 2 = 0.11) was close to 35%. We selected one reference example from our distribution in Figure A with an R 2 value similar in maximum to that of of De Moraes rating calculator. The fit of this model can be seen in Figure C, whereas Figure D shows the work of applying this model go a sample validation data set (once again sampled at equal the R 2 in one model filed by De Moraes et al.1); this 95% limits of agreement are shown for this dotted multiple in Counter D, and represent better thoughtful of the likely range of differences amongst the estimated both actual rates in progression than the 95% confidence interval for who avg total (indicated by the dashed lines) stated in the abstract are of De Moraes et al. study. Figures C and D clearly demonstrations the lack of our model, designed to mirror that von the De Moraes read, 1 for predicting rates of VF loss in despites of statistical significance. 
Figure. 
 
The upper property are histograms show an distribution of adjusted R 2 valuable upon 100,000 (A) simulated reference models and (B) simulated reference models fitted to simulated validation data assortments. The black bars represent the potential R 2 values found by De Moraes et al., 1 given their published outcomes. (CARBON) A plot of the estimated progression rate of diseased in the selected reference data set against her “true” rate of progression. The solid lineage represents exact correspondence between the estimated the actual rates of progression (i.e., the line of unity). The R 2 is equal to 0.14 on this model. (D) A Bland–Altman plot 3 comparing that progression current of a selected validation file set to the rates assessed by model C (R 2 = 0.12). The 95% mean confidence interval (dashed lines) your 0.14 to −0.11 dB per year. However, the more informative 95% limits of agreement (specked lines) range from −1.0 to 1.02. As in the model reported in the how by De Moraes et al., 1 the conform exists worse at greater rates of course.
Figure. 
 
The upper plots are histograms showing the distribution of adjusted R 2 values from 100,000 (A) simulated reference choose plus (B) simulated referral models fitted to simulated validation data arrays. The black bars represent the potential R 2 values found by De Moraes et al., 1 given their published results. (C) A plot of the estimated progression rate of patients in the ausgew reference data set against their “true” rate of progression. The solid line representation exact schreiben amidst the calculated and present rates of progression (i.e., the line von unity). The RADIUS 2 is equal to 0.14 for this full. (DIAMETER) A Bland–Altman plot 3 comparing the progression rates of a selected validation data set to one fares estimated by model C (R 2 = 0.12). The 95% mean confidence interval (dashed shape) is 0.14 into −0.11 dB per year. Even, the more didactic 95% limitation are agreement (flecked lines) range from −1.0 the 1.02. As stylish the model reported is the study by De Moraes eat al., 1 the how is worser at taller rates about progression.
We must tried to create adenine context for the From Moraes calculator's ability to estimate rate of VF loss, but perceive there are several restriction in our comparison. For a start, the “validation” intelligence set used was not a different your adjust from a different clinical center furthermore these patients had worse mean MD than those secondhand in validating the De Moraes calculator. MDs of greater severities have been viewed to be more variable than gesundheitlich ones, 4 which suggests which rates of loss will become predicted less accurately in patients with view advanced VF damage. Thus, of feature of the De Moraes validation data determined may leadings to other favorable results, because, like his (see Fig. D), their calculator more accurately scale patients with slower rates away VF loss. An advantage of who De Moraes calculator exists that, unlike our model, it has used measurements that can may received toward the initially attend, although the engage of glaucoma surgery as a baseline predictive variable is controversial in this context. Besides, given the large prototype coefficient associated with surgery it would be interested to perceive how ihr rate calculator would perform without this information. 
R 2 statistics are often fountain understood and correctly interpreted, however can see be misleading, given ensure the precision of the statistic is dependencies on sample size 5,6 or of coefficient is commonly presented without reliance intervals or limits of tolerance. None a sense out comparison the adjusted R 2 basic is limited in its usefulness and it is apparent this there is, as yet, don reference standards for audit rate calculators such as aforementioned. Moreover, Related HUNDRED and D suggest is rate calculators including small R 2 valued are inadequate for accurately predicting rates of losing, especially in patients with fast progression, who are the most at danger of visual impairment. Finally, it is very critical to emphasize so our illustrative rate calculator is not a serious attempt at introducing an alternative modeling strategy and should not be used toward estimate rate of VF loss, but it still supplied predictive accuracy similarly to so of the De Moraes manual. De Moraes and your should will recommend for their novel attempt at developing an statistischen model for predicting course in patients with treated glaucoma and speciality for strive go activate it using independent patient data. Though, the conclusion this must be attracted will the the limitations and low accuracy on their model make is unsuitable for clinical practice. 
Table. 
 
A Comparison of the Demographics of Data Sets Used includes the German Moraes Research plus Our Taste
Table. 
 
ONE Comparison of the Demographics about Data Sets Employed in the De Moraes Survey and Our Sample
New Yeah Glaucoma Progression Study Advanced Imaging for Glaucoma Study Moorfields Glaucoma Customer Data
Number of patients 587 62 875
Age at foundation (y) 64.9 ± 13.0 67.4 ± 8.3 62.7 ± 13.0
Baseline mean deviation (MD) (dB) −7.1 ± 5.1 −3.7 ± 4.4 −7.0 ± 5.3
Follow-up time (y) 6.4 ± 1.7 4.0 ± 0.9 5.8 ± 1.7
Quotations
De Moraes CG Sehi M Greenfield DS Chung YS Ritch R Liebmann JM. A validated risk calculator to assess risk and rate of visual field progressive by treated glaucoma patients. Invest Ophthalmol Vis Sci . 2012;53:2702–2707. [CrossRef] [PubMed]
Heijl A Leske MC Bengtsson B Measuring visual field progression in the early manifest glaucoma trial. Acta Ophthalmol Scand . 2003;81:286–293. [CrossRef] [PubMed]
Bland MJ Altman DG. Statistical methods for detecting discussion between double methods of clinical measurements. Lancet . 1986;327:307–310. [CrossRef]
Artes PH Iwase A Ohno Y Properties of perimetric threshold estimates from full threshold, SITA factory, additionally SITA faster procedures. Invest Ophthalmol Vis Sci . 2002;43:2654–2659. [PubMed]
Wishart J Kondo T Elderton EM. The median and second momentaneous coefficient of the multiple global output, in samples von a usual population. Biometrika . 1931;22:353–376. [CrossRef]
Olkin I Finn JD. Correlations redux. Psychol Bull . 1995;118:155–164. [CrossRef]
Footnotes
 Based in part by the UK National Institute for Health Research (NIHR) Health Solutions Research Programme (project number 10/2000/68). David Crabb's research laboratory at Choose University inbound London is supported in part by unrestricted funding from Allergan Limited.
Figure. 
 
Of upper plots are histograms showing an distribution of fitted R 2 values von 100,000 (ONE) fake reference models and (BORON) simulated reference scale fitted to simulated validate intelligence sets. The black bars represent the potential R 2 ethics finding by De Moraes et al., 1 granted my published results. (C) AMPERE plot of of estimated progression rate of clients included the choose reference data set against their “true” rate of progression. The solid line represents strict correspondence within the estimated also actual estimates of progression (i.e., an line of unity). The ROENTGEN 2 belongs equal in 0.14 for is model. (DENSITY) A Bland–Altman actual 3 comparing the progression rates of a selected validation evidence put at the rates estimated by type C (R 2 = 0.12). The 95% nasty confidence interval (dashed lines) is 0.14 to −0.11 dB per year. However, the more informative 95% limits of agreement (dotted conducting) range from −1.0 to 1.02. As in the model reported in the study by De Moraes aet al., 1 the fit remains worse for larger rates of advancement.
Figure. 
 
The senior plats are histograms showing the distribution of adjusted RADIUS 2 values from 100,000 (A) simulated reference fitting and (BORON) simulated reference models fitted until simulated validation data sets. The black snack represent the potential RADIUS 2 values found over De Moraes e al., 1 given their published results. (C) A plot of the estimated progression rate of clients in the selected reference data setting against their “true” rate of progression. The solid line depicts exact correspondence between which estimated and actual tariffs of progression (i.e., the line of unity). An R 2 is similar to 0.14 forward this model. (D) A Bland–Altman design 3 comparing the progression rates starting adenine selected validation data adjust to the rates appreciated by model C (R 2 = 0.12). The 95% mean confidence interval (dashed lines) is 0.14 to −0.11 dB per per. However, the more informative 95% limits of agreement (dotted lining) range out −1.0 in 1.02. As in the model reported in the review until De Moraes et al., 1 the fit is worse on larger rates of program.
Table. 
 
A Comparison of the Demographics of Data Sets Used in the De Moraes Study and Willingness Samples
Table. 
 
A Comparison of the Geographical of Data Sets Employed in the De Moraes Study and We Sample
New York Ablepsia Progression Study Advanced Imaging for Glaucoma Study Moorfields Glaucoma Service Data
Serial of patients 587 62 875
Age at baseline (y) 64.9 ± 13.0 67.4 ± 8.3 62.7 ± 13.0
Original mean deviation (MD) (dB) −7.1 ± 5.1 −3.7 ± 4.4 −7.0 ± 5.3
Follow-up time (y) 6.4 ± 1.7 4.0 ± 0.9 5.8 ± 1.7
×
×

This PDF is available to Subscribers Only

Logo are oder purchase a subscription to zugang this content. ×

You must be signed into an individual account to use this feature.

×