Instrument evaluation
C M Langton, C F Njeh in The Physical Measurement of Bone, 2016
The differences between methods (y axis) are likely to possess a normal distribution only if there is a random component to the error. The distribution of differences can be checked for normality by constructing a simple histogram (or by using a statistical package). A Bland-Altman plot should then be examined visually for any patterns that may exist in the differences of the two methods (figure 4.2). The possible relationship between the differences and the mean can be investigated formally by calculating the Spearman’s rank correlation coefficient (see section 4.4.3)
Principles of Recording and Analysis of Arterial Waveforms
Wilmer W Nichols, Michael F O'Rourke, Elazer R Edelman, Charalambos Vlachopoulos in McDonald's Blood Flow in Arteries, 2022
Basic scientists and clinicians often need to know whether a new method of measurement is equivalent to an established (or gold standard) one already in clinical use. A short review of the methodology used in method comparison studies is presented here. For a more in-depth evaluation of the methodology, see reviews by Bland and Altman (1999, 2003), Dewitte et al. (2002), Barnhart et al. (2007) and Smellie (2008). Currently available software eliminates the need for tedious statistical computation but does not reduce the burden of understanding the concepts underlying a method comparison study and accurate interpretation of the final results. Current guidelines for the combined graphical/statistical interpretation of method comparison studies (Brazdzionyte and Macas, 2007; Haber and Barnhart, 2008) include a scatter plot combined with correlation and regression analysis and/or a difference plot combined with calculation of the limits of the differences between the methods (the so-called 95 percent confidence limits of the bias or ±2 standard deviations) (Bland and Altman, 1986). The Bland–Altman plots (or Tukey mean-difference plots) are simple to construct and easy to interpret and are often used as a substitute for more rigorous statistical analysis. The general features of the Bland–Altman plot have been well described in the literature (Bland and Altman, 1986, 2003). The abscissa (x‑axis) shows the mean values of the two methods [(A + B)/2], whereas the ordinate (y-axis) represents the absolute difference between the values of the two methods (B − A). Some investigators report their results only in Bland–Altman plots (Scolletta et al., 2005; Compton et al., 2008b), while others report their results in linear regression and Bland–Altman plots (Giomarelli et al., 2004). A big mistake made by numerous investigators is comparing results to methods not considered gold standards (Compton et al., 2008b).
B
Filomena Pereira-Maxwell in Medical Statistics, 2018
In the context of method comparison studies, the Bland-Altman plot is a plot of the differences between measurements obtained by two different methods (on the y-axis), against the average of the two measurements (on the x-axis). The average of the two measurements gives the best estimate of the true value of what is being measured. The technique enables an assessment to be made of the extent of (dis)agreement between the methods in question. As shown in Figure B.5a, the bias or mean difference between the methods is represented on the plot as a solid horizontal line. When the plotted differences are randomly scattered around this line, their mean and standard deviation may be used to calculate limits of agreement between the methods. These are also marked on the plot, as dotted horizontal lines. If, on the other hand, there is greater scatter or variability of observed differences as the magnitude of the measurements increases, the calculations may be performed on the logsof the measurements, and the results back-transformed to the original scale. This approach gives a measure of relative or proportional agreement, rather than absolute differences, as above, so the limits of agreement now tell us that for any given individual, method A is expected to be within x% and y% of the measurement given by method B, as shown in Figure B.5b (SCHERPBIER-DE HAAN et al., 2011). A regression-based alternative approach is given by BLAND & ALTMAN (1999). The example in Figure B.5a is a Bland-Altman plot from the study by ADAM et al. (2002) that compares haemoglobin measurements in 108 pregnant women using a portable meter (the HemoCue®) and a lab- based automated haematology analyser. Haemoglobin measurements are used to screen for maternal anaemia, an important cause of maternal and perinatal morbidity and mortality. With regard to the extent of agreement between these two methods (using venous blood samples), the authors concluded that “The mean difference with limits of agreement between the two readings was 1.17 (-1.97, 4.31) g/dl. [...] According to the previously predefined clinical acceptable limits of ± 1 g/dl, the 2 methods could not be considered as interchangeable.” See also BLAND & ALTMAN (1986), ALTMAN (1991), BLAND (2015), and KIRKWOOD & STERNE (2003) for further discussion and illustrative examples. See also differencevs.average plots.
Neonate auditory brainstem response repeatability with controlled force gauge bone-conducted stimulus delivery
Published in International Journal of Audiology, 2018
Andrew Stuart, Hannah M. Dorothy
Bland–Altman plots (Bland and Altman 1986) were constructed to examine absolute repeatability between tests for both wave V latencies and amplitudes (Figure 4). Each Bland–Altman plot is a bivariate scatterplot of the difference of two test measurements on the Y-axis and the average of the two test measurements on the X-axis. Three horizontal reference lines are superimposed on each plot: the average difference between the two test measurements (i.e. termed the bias) and the 95% limits of agreement (i.e. the mean difference ± 1.96 SD). If test measures were repeatable, Bland and Altman (1986) state, “the mean test difference should be zero.… [and] we expect 95% of the differences to be less than two standard deviations” (p. 7). This was evident in each plot. Mean intratester differences were approximately ±0.40 ms and ±0.25 μV for wave V latency and amplitude, respectively. Mean intertester differences were approximately ±0.65 ms and ±0.25 μV for wave V latency and amplitude, respectively. There should also be no systematic variation with the mean differences of the two measurements. Proportional/systematic bias was examined in two ways. First, test differences were examined in each plot with separate t-tests for paired samples. There were no statistically significant differences between any tests (p > 0.05). Second, linear trends between the differences of two test latency and amplitude measurements were examined with linear regressions. There were no linear predictive relationships between averaged and difference measures in any plot.
Machine learning models using non-linear techniques improve the prediction of resting energy expenditure in individuals receiving hemodialysis
Published in Annals of Medicine, 2023
Alainn Bailey, Mohamed Eltawil, Suril Gohel, Laura Byham-Gray
We used a modified Bland–Altman plot to measure the levels of agreement between mREE and eREE from each model [38]. The original Bland-Altman plot graphically assesses agreement between two methods of measurement by examining one method on the Y-axis by comparison with either the true measure on the X-axis or the mean of both measures if the criterion is not known [38]. In this case, we used residual values calculated via percentage on the Y-axis and mREE (the criterion measure) on the X-axis. A full description of the method was previously published by this group [27]. Limits of agreement for predictive equations have been established at ± 10% from zero difference from mREE in the nutrition literature [38]. Those limits have been used for validation by Byham-Gray et al. [21,23], Morrow et al. [25] and Bailey et al. [27] when assessing equations for people receiving dialysis [21,23,25,27]. This graphical analysis was applied to each of the best models (and the MHDE) across the complete validation sample for which REE was generated. The analysis was subsequently repeated with the validation set divided into subgroups of BMI. Individuals with a BMI less than 24.9 kg/m2, 25–29.9 kg/m2, or ≥ 30 kg/m2 were categorized as underweight/normal weight, overweight, or obese.
Validation of Two Point of Care Devices for Hemoglobin Estimation in Blood Donors
Published in Hemoglobin, 2020
Deepika Chenna, Shamee Shastry, Kalyana-Chakravarthy Pentapati
Of the 100 samples, 95.0% were obtained from male donors. The mean ± SD of Hb measured using the reference method, CompoLab TM and True Hb (g/dL) were 14.79 ± 1.24, 15.15 ± 1.46 and 14.3 ± 1.76, respectively. The Bland–Altman plot, a graphical method to plot the difference scores of two measurements against the mean for each sample was plotted. Ninety-five percent of the data points were within ±1.96 SD of the mean difference, limits of agreement for a new device recording the measurements when compared to the standard. On average, CompoLab TM measured 0.4 units more than the reference method and True Hb measured 0.4 units less than the reference method (Figures 1 and 2). The value of two standard deviations of the difference of Hb measurement was >2.0 g/dL for both CompoLab TM and True HB.
Related Knowledge Centers
- Biomedicine
- Ma Plot
- Analytical Chemistry
- Assay
- Medical Statistics
- Gold Standard
- Medcalc
- Statsdirect
- Standard Deviation
- Inter-Rater Reliability