Error-in-variables (EIV) models are useful tools to account for measurement error in the independent variable. For studies of agreement, this is particularly useful where there are paired measurements of the paired measurements (X & Y) of the same underlying value (e.g., two assays of the same analyte).
Deming regression is one of the simplest forms of of EIV models promoted by W. Edwards Deming1. The first to detail the method were Adcock (1878) followed by Kummell (1879) and Koopmans (1936). The name comes from the popularity of Deming’s book (Deming 1943), and within the field of clinical chemistry, the procedure was simply referred to as “Deming regression” (e.g., Linnet (1990)).
We can start by creating some fake data to work with.
library(SimplyAgree)
dat = data.frame(
x = c(7, 8.3, 10.5, 9, 5.1, 8.2, 10.2, 10.3, 7.1, 5.9),
y = c(7.9, 8.2, 9.6, 9, 6.5, 7.3, 10.2, 10.6, 6.3, 5.2)
)Also, we will assume, based on historical data, that the measurement error ratio is equal to 2 (i.e., the error variance of x is twice the error variance of y).
The data can be run through the dem_reg function and the
results printed.
dem1 = dem_reg(y ~ x,
data = dat,
error.ratio = 2,
weighted = FALSE)
dem1
#> Deming Regression with 95% C.I.
#>
#> Call:
#> dem_reg(formula = y ~ x, data = dat, weighted = FALSE, error.ratio = 2,
#> conf.level = 0.95)
#>
#> Coefficients:
#> (Intercept) x
#> 0.1285 0.9745The resulting regression line can then be plotted.
The assumptions of the Deming regression model, primarily normality
and homogeneity of variance, can then be checked with the
check method for Deming regression results. Both plots
appear to be fine with regards to the assumptions.
After fitting a Deming regression model, you can extract the
estimated true values and residuals. The fitted() function
returns the estimated true Y values, accounting for measurement error.
The residuals() function returns the optimized residuals by
default, which correspond to the perpendicular distances from each point
to the regression line.
For this example, I will rely upon the “ferritin” data from the
deming R package.
library(deming)
data('ferritin')
head(ferritin)
#> id period old.lot new.lot
#> 1 1 1 1 1
#> 2 2 1 3 3
#> 3 3 1 10 9
#> 4 4 1 13 11
#> 5 5 1 13 12
#> 6 6 1 15 13Let me demonstrate the problem with using simple Deming regression when the weights are helpful. When we look at the two plots below, we can see there is severe problem with using the “un-weighted” model.
dem2 = dem_reg(
old.lot ~ new.lot,
data = ferritin,
weighted = FALSE
)
summary(dem2)
#> Deming Regression with 95% C.I.
#>
#> Call:
#> dem_reg(formula = old.lot ~ new.lot, data = ferritin, weighted = FALSE,
#> conf.level = 0.95)
#>
#> Coefficients:
#> coef bias se df lower.ci upper.ci t p.value
#> Intercept 5.2157 -0.235818 2.18603 160 0.8985 9.533 2.386 0.01821
#> Slope 0.9637 0.002597 0.02505 160 0.9143 1.013 -1.448 0.14949
#>
#> 160 degrees of freedom
#> Error variance ratio (lambda): 1.0000
plot(dem2)Now, let us see what happens when weighted is set to
TRUE.
dem3 = dem_reg(
old.lot ~ new.lot,
data = ferritin,
weighted = TRUE
)
summary(dem3)
#> Weighted Deming Regression with 95% C.I.
#>
#> Call:
#> dem_reg(formula = old.lot ~ new.lot, data = ferritin, weighted = TRUE,
#> conf.level = 0.95)
#>
#> Coefficients:
#> coef bias se df lower.ci upper.ci t p.value
#> Intercept -0.02616 0.0065148 0.033219 160 -0.09176 0.03945 -0.7874 4.322e-01
#> Slope 1.03052 -0.0001929 0.006262 160 1.01815 1.04288 4.8729 2.626e-06
#>
#> 160 degrees of freedom
#> Error variance ratio (lambda): 1.0000
plot(dem3)The weighted model provides a better fit.
Passing-Bablok regression is a robust, nonparametric alternative to Deming regression for method comparison studies (Passing and Bablok 1983). Unlike Deming regression, it makes no assumptions about the distribution of the measurement errors. The method estimates the slope as the shifted median of all pairwise slopes between data points, which makes it resistant to outliers.
Passing-Bablok regression is particularly useful when:
The pb_reg() function implements three variants of
Passing-Bablok regression:
# Create example data
pb_data <- data.frame(
method1 = c(69.3, 27.1, 61.3, 50.8, 34.4, 92.3, 57.5, 45.5, 33.3, 60.9,
56.3, 49.9, 89.7, 28.9, 96.3, 76.6, 83.2, 79.4, 51.7, 32.5,
99.1, 14.2, 84.1, 48.8, 61.5, 84.9, 93.2, 73.8, 62.1, 98.6),
method2 = c(69.1, 26.7, 61.4, 51.2, 34.7, 88.5, 57.9, 45.1, 33.4, 60.8,
66.5, 48.2, 88.3, 29.3, 96.4, 77.1, 82.7, 78.9, 51.6, 28.8,
98.4, 12.7, 83.6, 47.3, 61.2, 84.6, 92.1, 73.4, 61.9, 98.6)
)
# Fit Passing-Bablok regression
pb1 <- pb_reg(method2 ~ method1, data = pb_data)
#> Warning in pb_reg(method2 ~ method1, data = pb_data): Bootstrap confidence
#> intervals are recommended for 'invariant' and 'scissors' methods. Consider
#> setting replicates > 0.
pb1
#> Passing-Bablok (scissors) with 95% C.I.
#>
#> Call:
#> pb_reg(formula = method2 ~ method1, data = pb_data)
#>
#> Coefficients:
#> (Intercept) method1
#> 0.09622 0.99440
#>
#> Kendall's Tau (H0: tau <= 0):
#> tau: 0.9632
#> p-value: 0.0000
#>
#> CUSUM Linearity Test:
#> Test stat: 1.0000
#> p-value: 0.1844The summary provides details about the regression coefficients and diagnostic tests:
summary(pb1)
#> Passing-Bablok (scissors) with 95% C.I.
#>
#> Call:
#> pb_reg(formula = method2 ~ method1, data = pb_data)
#>
#> Coefficients:
#> term coef se lower.ci upper.ci df null_value reject_h0
#> 1 Intercept 0.09622 0.390603 -0.7206 0.8105 28 0 FALSE
#> 2 method1 0.99440 0.005918 0.9840 1.0072 28 1 FALSE
#>
#> 28 degrees of freedomThe plot() method displays the regression line with
data:
The check() method provides diagnostic plots specific to
Passing-Bablok regression, including the CUSUM linearity test and
Kendall’s tau correlation:
For more robust inference, especially with the “invariant” or “scissors” methods, bootstrap confidence intervals are recommended:
pb2 <- pb_reg(method2 ~ method1,
data = pb_data,
replicates = 999)
summary(pb2)
#> Passing-Bablok (scissors) with 95% C.I.
#>
#> Call:
#> pb_reg(formula = method2 ~ method1, data = pb_data, replicates = 999)
#>
#> Coefficients:
#> term coef se lower.ci upper.ci df null_value reject_h0
#> 1 Intercept 0.09622 0.486542 -0.9375 0.8516 28 0 FALSE
#> 2 method1 0.99440 0.006857 0.9838 1.0090 28 1 FALSE
#>
#> 28 degrees of freedom
#>
#>
#> Pairs bootstrap CIs based on 999 resamplesAn experimental enhancement to the dem_reg and
pb_reg functions is the addition of joint
confidence regions for the slope and intercept parameters.
Traditional confidence intervals treat each parameter separately, but in
regression analysis, these parameters are correlated.
Joint confidence regions account for this correlation with a region in the (intercept, slope) parameter space. This approach, promoted by Sadler (2010), provides several advantages:
The power advantage is most pronounced when the ratio of maximum to minimum X values is small (< 10:1), which is common in clinical method comparisons.
Two regions are available through the method
argument:
"sadler" (default for Deming
regression): the sum-of-squares region of Sadler
(2010). It uses the Deming sum of
squares directly, not a variance-covariance matrix, and is not exactly
an ellipse. In simulations it rejected a true identity line 4.3-5.9% of
the time at a nominal 5%, from n = 10 upward."wald" (the only option for
Passing-Bablok regression): an ellipse from the variance-covariance
matrix of the estimates. For Deming regression that matrix is a
delete-one jackknife, and the test is liberal in small samples (about
8.5% at n = 10 and 7.4% at n = 20 at a nominal 5%). For Passing-Bablok
regression, use bootstrap or delete-d jackknife resampling (both held
about 5% in simulations).The plot_joint() function allows visualization of the
confidence region in parameter space:
This plot shows:
method = "wald" draws the ellipse)Notice how the region is narrower than the rectangle, especially in the directions that matter for detecting bias.
For Passing-Bablok regression, bootstrap (or delete-d jackknife) resampling must be used to obtain the variance-covariance matrix needed for the joint confidence region:
The joint_test() function provides a formal hypothesis
test for whether the identity line (slope = 1, intercept = 0) falls
within the joint confidence region:
joint_test(dem1)
#>
#> Joint Confidence Region Test [Sadler sum-of-squares, F-based] (H0:
#> intercept = 0, slope = 1)
#>
#> data: y ~ x
#> F = 0.059522, df1 = 2, df2 = 8, p-value = 0.9426
#> alternative hypothesis: true intercept and slope are not equal to the null values
#> null values:
#> intercept slope
#> 0 1
#> sample estimates:
#> intercept slope
#> 0.1284751 0.9744516This returns an htest object with Sadler’s F statistic
and p-value. The Wald version (Mahalanobis distance) is available with
joint_test(dem1, method = "wald"). The test can also be
applied to Passing-Bablok models fitted with bootstrap, where it is
always the Wald test:
joint_test(pb2)
#>
#> Joint Confidence Region Test [Wald, F-based] (H0: intercept = 0, slope
#> = 1)
#>
#> data: method2 ~ method1
#> X-squared = 7.5672, df1 = 2, df2 = 28, p-value = 0.03512
#> alternative hypothesis: true intercept and slope are not equal to the null values
#> null values:
#> intercept slope
#> 0 1
#> sample estimates:
#> intercept slope
#> 0.09621849 0.99439776Deming regression assumes paired measures (\(x_i, \space y_i\)) are each measured with error.
\[ x_i = X_i + \epsilon_i \]
\[ y_i = Y_i + \delta_i \]
We can then measure the relationship between the two variables with the following model.
\[ \hat Y_i = \beta_0 + \beta_1 \cdot \hat X_i \]
Traditionally there are 2 null hypotheses.
First, the intercept is equal to zero:
\[ H_0: \beta_0 = 0 \space vs. \space H_1: \beta_0 \ne 0 \]
Second, that the slope is equal to one:
\[ H_0: \beta_1 = 1 \space vs. \space H_1: \beta_1 \ne 1 \]
Deming regression with error ratio \(\lambda\) and weights \(w_i\) minimises
\[ S(a, b) = \sum_{i=1}^n \frac{w_i (y_i - a - b x_i)^2}{1 + \lambda b^2} \]
Following Sadler (2010) (after Draper and Smith), with \(\hat S\) the minimum, the joint \((1-\alpha) \times 100\%\) confidence region for \((\beta_0, \beta_1)\) is the set of \((a, b)\) with
\[ F = \frac{(S(a, b) - \hat S)/2}{\hat S / (n - 2)} \leq F_{2, n-2, 1-\alpha} \]
and the joint test of a hypothesized \((a_0, b_0)\) (e.g., \((0, 1)\) for identity) compares \(F\) at that point with the \(F(2, n - 2)\) distribution. No variance-covariance matrix is needed, and only the error ratio (not the size of the errors) matters. For a fixed slope the condition is quadratic in the intercept, so the region is drawn exactly.
This was my original solution for this joint testing problem. I would, for Deming regression, stick with Sadler’s region.
With method = "wald" the region is
\[ (\hat{\beta} - \beta_0)^T V^{-1} (\hat{\beta} - \beta_0) \leq 2 F_{2, n-2, 1-\alpha} \]
(or \(\chi^2_{2, 1-\alpha}\) with
test_method = "asymptotic"), where:
This forms an ellipse in parameter space that accounts for the correlation between slope and intercept.
A Deming regression model also assumes the measurement error (\(\sigma^2\)) ratio is constant. In
SimplyAgree, the error ratio is always the error variance
of x divided by the error variance of y.
\[ \lambda = \frac{\sigma^2_\epsilon}{\sigma^2_\delta} \]
In SimplyAgree, the error ratio can be set with the
error.ratio argument. It defaults to 1, but can be changed
by the user. If replicate measures are taken, then the user can use the
id argument to indicate which measures belong to which
subject/participant. The measurement error, and the error ratio, will
then be estimated from the data itself.
If the data was not measured in replicate then the error ratio (\(\lambda\)) can be estimated from the coefficient of variation (if that data is available) and the mean of x and y (\(\bar x, \space \bar y\)).
\[ \lambda = \frac{(CV_x \cdot \bar x)^2}{(CV_y \cdot \bar y)^2} \]
In some cases the variance of X and Y may increase proportional to
the true value of the measure. In these cases, it may be prudent to use
“weighted” Deming regression models. The weights used in
SimplyAgree are the same as those suggested by Linnet (1993).
\[ \hat w_i = \frac{1}{ [ \frac{x_i + \lambda \cdot y_i}{1 + \lambda}]^2} \]
Weights can also be provided through the weights
argument. If weighted Deming regression is not selected
(weighted = FALSE), the weights for each observation is
equal to 1.
The estimated mean of X and Y are then estimated as the following.
\[ \bar x_w = \frac{\Sigma^{N}_{i=1} \hat w_i \cdot x_i}{\Sigma^{N}_{i=1} \hat w_i} \]
\[ \bar y_w = \frac{\Sigma^{N}_{i=1} \hat w_i \cdot y_i}{\Sigma^{N}_{i=1} \hat w_i} \]
First, there are 3 components (\(v_x, \space v_y, \space cov_{xy}\))
\[ v_x = \Sigma_{i=1}^N \space \hat w_i \cdot (x_i- \bar x_w)^2 \] \[ v_y = \Sigma_{i=1}^N \space \hat w_i \cdot (y_i- \bar y_w)^2 \] \[ cov_{xy} = \Sigma_{i=1}^N \space \hat w_i \cdot (x_i- \bar x_w) \cdot (y_i- \bar y_w) \]
The slope (\(b_1\)) can then be estimated with the following equation.
\[ b_1 = \frac{(\lambda \cdot v_y - v_x) + \sqrt{(v_x-\lambda \cdot v_y)^2 + 4 \cdot \lambda \cdot cov_{xy}^2}}{2 \cdot \lambda \cdot cov_{xy}} \]
The intercept (\(b_0\)) can then be estimated with the following equation.
\[ b_0 = \bar y_w - b_1 \cdot \bar x_w \]
The standard errors of b1 and b0 are both estimated using a jackknife method (detailed by Linnet (1990)).
Deming was a titan of the fields of statistics and engineering and I would highly recommend reading some of his academic work and books↩︎