Consider allowing JavaScript. Otherwise, you have to be proficient in reading since formulas will not be rendered. Furthermore, the table of contents in the left column for navigation will not be available and code-folding not supported. Sorry for the inconvenience.

The example in this article was generated with 4.0.5 by the package `truncnorm`

.^{1}

See also a collection of other articles.

- Click to show / hide R code.

Which trapezoidal rule should we use?

The linear trapezoidal rule goes back to the dark ages when people plotted concentrations on graph paper, connected points by means of a transparent ruler, cut the profile out, and weighed it on analytical scales.

“I suppose it is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail.

Apart from the few drugs which are subjected to capacity limited elimination like C_{2}H_{5}OH, distribution and/or elimination follows exponential decrease(s). Hence, for the vast majority of drugs the linear trapezoidal rule will lead to a positive bias for all sections of the profile where concentrations decrease.

“It is customary in biopharmaceutics to use the trapezoidal method to calculate areas under the concentration-time curve. The popularity of this method may be attributed to its simplicity both in concept and execution. However, in cases where […] there are long intervals between data points, large algorithm errors are known to occur.

That’s not what I call state of the art. I coded the linear-up/log-down trapezoidal rule for the TI-59 already in 1981…

For moonshine.

Data points are connected by straight lines and the *AUC* calculated as the sum of trapezoids.

We see that for \(\small{b=a}\) the trapezoid becomes a rectangle, where \(\small{A=h\times a}\). The anchor point is the *arithmetic* mean of \(\small{a}\) and \(\small{b}\).

For a fast bolus.

All concentrations are \(\small{\log_{e}}\)-transformed and the linear trapezoidal applied. At the end the *AUC* is back-transformed.

The anchor point is the *geometric* mean of \(\small{a}\) and \(\small{b}\). The log trapezoidal does not work if \(\small{a}\) or \(\small{b}\) are zero or if they are equal. In such a case software should fall back to the linear trapezoidal.

The concentration \(\small{C_0}\) has to be imputed. This is done either automatically by the software (extrapolated from the first two \(\small{\log_{e}}\)-transformed concentrations) or has to be provided by the user.

For an infusion or extravascular administration.

Straightforward. If subsequent concentrations increase or are equal (\(\small{C_{t_\textrm{i+1}}\geq C_{t_\textrm{i}}}\)), the linear trapezoidal is applied and the log trapezoidal otherwise (\(\small{C_{t_\textrm{i+1}}< C_{t_\textrm{i}}}\)).

Don’t import data blindly. Quite often some data are given with a nonnumeric code, *e.g.*. `BQL`

for values below the LLOQ, `NR`

for not reportable ones (say, due an interferring peak in chromatography), or simply `missing`

. This might lead to trouble, especially if you have to deal with lag times.

time | concentration |
---|---|

0.00 | BQL |

0.25 | BQL |

0.50 | BQL |

0.75 | BQL |

1.00 | 2.386209 |

1.25 | 3.971026 |

1.50 | 5.072666 |

2.00 | 6.461897 |

2.50 | 5.964697 |

3.00 | 6.154877 |

3.50 | 5.742396 |

4.00 | 4.670705 |

4.25 | 5.749097 |

4.50 | 6.376336 |

4.75 | 6.473850 |

5.00 | 6.518927 |

5.25 | 7.197600 |

5.50 | 7.171507 |

6.00 | 6.881594 |

6.50 | 7.345398 |

7.00 | 7.135752 |

7.75 | 5.797113 |

9.00 | 4.520387 |

12.00 | 3.011526 |

16.00 | 1.534004 |

24.00 | 0.369226 |

If you import this data and *press the button*,^{3} likely you will get an *AUC* of 77.26564 by the linear trapezoidal and 75.64961 by the linlog trapezoidal. But is this correct? Hey, the lag time shows up as zero! My trusty Phoenix/WinNonlin dropped all `BQL`

’s, imputed 0 at time zero and informs me:

` Note - the concentration at dose time `

` was added for extrapolation purposes. `

What does that mean? It calculated the first partial *AUC* as \(\small{(1-0)\times 2.386209/2=1.193105}\). That’s correct numerically but still false. Well, the software is right but its master, the wetware isn’t.

Of course, that’s nonsense. All ‘hidden’ imputed concentrations are above the LLOQ, the one at 0.75 h is already more than 5times the LLOQ.

However, help is on the way. Apply a set of ‘BQL Rules’ (see the User’s Guide or the online help for details). A screenshot of mine:

Then we get what we expect. The lag time is reported with 0.75 h, the *AUC* with 76.37082 by the linear trapezoidal and with 74.75478 by the linlog trapezoidal. All is good because the first partial *AUC* is now \(\small{(1-0.75)\times 2.386209/2=0.2982761}\).

I picked a difficult one. A biphasic release formulation, one compartment model, PK-parameters: \(\small{D=100}\), dose fractions of the IR and DR components \(\small{\left\{0.6,0.4\right\}}\), fraction absorbed \(\small{f=0.70}\), volume of distribution \(\small{V=5}\), absorption rate constants of the components \(\small{k_{\,01.\textrm{IR}}=1.4\,\textrm{h}^{-1}}\), \(\small{k_{\,01.\textrm{DR}}=0.7\,\textrm{h}^{-1}}\), lag times of the components \(\small{tlag_{\,01.\textrm{IR}}=25\,\textrm{min}}\), \(\small{tlag_{\,01.\textrm{DR}}=4\,\textrm{h}}\), elimination rate constant \(\small{k_{\,10}=0.18\,\textrm{h}^{-1}}\).

The sampling schedule was 0, 0.25, 0.5, 0.75, 1, 1.25, 1.5, 2, 2.5, 3, 3.5, 4, 4.25, 4.5, 4.75, 5, 5.25, 5.5, 6, 6.5, 7, 7.75, 9, 12, 16, and 24 hours in order to get reliable estimates of both peaks.

Error distributions were uniform for \(\small{f}\) (0.5–1), log-normal for \(\small{V=5}\) (*CV* 50%), \(\small{k_{\,01}}\) (*CV* 25%), \(\small{k_{\,10}}\) (*CV* 40%), and truncated normal for \(\small{tlag_{\,01.\textrm{IR}}}\) (limits 5–45 min), \(\small{tlag_{\,02.\textrm{DR}}}\) (lower limit 3 h). Analytical error was log-normal with a *CV* of 5% of the simulated concentration.

1,000 simulated profiles, the LLOQ was set to 5% of the theoretical *C*_{max}, *i.e.*, simulated concentrations below the LLOQ were forced to `NA`

. The *AUC* was calculated by both methods and compared to the true one (without analytical error).

Cave: 210 LOC.

All results aggreed with Certara’s Phoenix/WinNonlin.

```
library(truncnorm)
setwd("E:/Public/Documents/BEBAC/homepage/articles")
function(x, y, first.x, last.x) {
trap.lin <-# linear trapezoidal rule
0
pAUC.lin <-for (i in seq(which(x == first.x), which(x == last.x))) {
if (x[i] == first.x) { # start triangle
c(pAUC.lin, 0.5*(x[i]-x[i-1])*y[i])
pAUC.lin <-else {
} c(pAUC.lin, 0.5*(x[i]-x[i-1])*(y[i]+y[i-1]))
pAUC.lin <-
}
} cumsum(pAUC.lin)
AUC.lin <-return(tail(AUC.lin, 1))
} function(x, y, first.x, last.x) {
trap.linlog <-# lin-up/log-down trapezoidal rule
0
pAUC.linlog <-for (i in seq(which(x == first.x), which(x == last.x))) {
if (x[i] == first.x) { # start triangle
c(pAUC.linlog, 0.5*(x[i]-x[i-1])*y[i])
pAUC.linlog <-else {
} if (y[i] >= y[i-1]) { # linear up
c(pAUC.linlog, 0.5*(x[i]-x[i-1])*(y[i]+y[i-1]))
pAUC.linlog <-else { # log down
} c(pAUC.linlog,
pAUC.linlog <--x[i-1])*(y[i]-y[i-1])/
(x[i] (log(y[i]/y[i-1])))
}
}
} cumsum(pAUC.linlog)
AUC.linlog <-return(tail(AUC.linlog, 1))
}set.seed(123456)
1000 # no of profiles
n <- c(0,0.25,0.5,0.75,1,1.25,1.5,2,2.5,3,3.5,
t <-4,4.25,4.5,4.75,5,5.25,5.5,6,6.5,7,7.75,9,12,16,24)
100
D <- c(0.6, 0.4) # dose fractions of components
D.f <-# Index ".d" denotes default values, ".c" CV
0.7 # fraction absorbed (BA)
F.d <- 0.5 # lower limit
F.l <- 1 # upper limit
F.u <- 5 # volume of distribution
V.d <- 0.50 # CV 50%, lognormal
V.c <-1.d <- 1.4 # absorption rate constant (1st component)
k01.2.d <- 0.7 # absorption rate constant (2nd component)
k01. 0.25 # CV 25%, lognormal
k01.c <- 0.18 # elimination rate constant
k10.d <- 0.40 # CV 40%, lognormal
k10.c <- 0.25 # 1st lag time
tlag1.d <- 5/60 # lower truncation 1
tlag1.l <- 0.75 # upper truncation 1
tlag1.u <- 4 # 2nd lag time
tlag2.d <- 3 # lower truncation 2
tlag2.l <- 0.5 # CV 50%, truncated normal
tlag.c <- 0.05 # analytical error CV 5%, lognormal
AErr <- 0.05 # fraction of theoretical Cmax
LLOQ.f <-
function(x, D = D, D.f = D.f, F = F.d, V = V.d,
Ct <-k01.1 = k01.1.d, k01.2 = k01.2.d, k10 = k10.d,
tlag1 = tlag1.d, tlag2 = tlag2.d) {
F.d*D*D.f[1]/V.d*k01.1.d/(k01.1.d-k10.d)*
C1 <- (exp(-k10.d*(x-tlag1))-exp(-k01.1.d*(x-tlag1)))
F.d*D*D.f[2]/V.d*k01.2.d/(k01.2.d-k10.d)*
C2 <- (exp(-k10.d*(x-tlag2))-exp(-k01.2.d*(x-tlag2)))
< 0] <- 0
C1[C1 < 0] <- 0
C2[C2 return(C1 + C2)
} data.frame(subject = rep(1:n, each = length(t)), t = t, C = NA)
data <-# individual PK parameters
runif(n = n, min = F.l, max = F.u)
F <- rlnorm(n = n, meanlog = log(V.d)-0.5*log(V.c^2+1),
V <-sdlog = sqrt(log(V.c^2+1)))
.1 <- rlnorm(n = n, meanlog = log(k01.1.d)-0.5*log(k01.c^2+1),
k01sdlog = sqrt(log(k01.c^2+1)))
.2 <- rlnorm(n = n, meanlog = log(k01.2.d)-0.5*log(k01.c^2+1),
k01sdlog = sqrt(log(k01.c^2+1)))
rlnorm(n = n, meanlog = log(k10.d)-0.5*log(k10.c^2+1),
k10 <-sdlog = sqrt(log(k10.c^2+1)))
rtruncnorm(n = n, a = tlag1.l, b = tlag1.u,
tlag1 <-mean = tlag1.d, sd = tlag.c)
rtruncnorm(n = n, a = tlag2.l, b = Inf,
tlag2 <-mean = tlag2.d, sd = tlag.c)
seq(min(t), max(t), length.out=2500)
x <-# theoretical profile
Ct(x = x, D = D, D.f = D.f, F = F.d, V = V.d,
C.th <-k01.1 = k01.1.d, k01.2 = k01.2.d, k10 = k10.d,
tlag1 = tlag1.d, tlag2 = tlag2.d)
LLOQ.f * max(C.th)
LLOQ <-# individual profiles
for (i in 1:n) {
Ct(x = t, D = D, D.f = D.f, F = F[i], V = V[i],
C <-k01.1 = k01.1[i], k01.2 = k01.2[i], k10 = k10[i],
tlag1 = tlag1[i], tlag2 = tlag2[i])
for (j in 1:length(t)) { # analytical error (multiplicative)
if (j == 1) {
rnorm(n = 1, mean = 0,
AErr1 <-sd = abs(C[j]*AErr))
else {
} c(AErr1, rnorm(n = 1, mean = 0,
AErr1 <-sd = abs(C[j]*AErr)))
}
} C + AErr1 # add analytical error
C <-< LLOQ] <- NA # assign NAs to Cs below LLOQ
C[C $C[data$subject == i] <- C
data
} data.frame(subject = 1:n, tlag = NA, tlast = NA, Clast = NA,
res <-th = NA, lin = NA, linlog = NA)
# NCA
for (i in 1:n) {
data$t[data$subject == i]
tmp.t <- data$C[data$subject == i]
tmp.C <- t[!is.na(tmp.C)][1]
tfirst <- rev(tmp.t[!is.na(tmp.C)])[1]
tlast <-$tlag[i] <- tmp.t[which(tmp.t == tfirst)+1]
res$tlast[i] <- tlast
res$Clast[i] <- tmp.C[tmp.t == tlast]
res$lin[i] <- trap.lin(tmp.t, tmp.C, tfirst, tlast)
res$linlog[i] <- trap.linlog(tmp.t, tmp.C, tfirst, tlast)
res x[x >= tfirst & x <= tlast]
compl.t <- Ct(x = compl.t, D = D, D.f = D.f, F = F[i], V = V[i],
C <-k01.1 = k01.1[i], k01.2 = k01.2[i], k10 = k10[i],
tlag1 = tlag1[i], tlag2 = tlag2[i])
$th[i] <- 0.5*sum(diff(compl.t)*(C[-1]+C[-length(C)]))
res
}$RE.lin <- 100 * (res$lin - res$th) / res$th
res$RE.linlog <- 100 * (res$linlog - res$th) / res$th
res# print(res[, c(1:4, 6:7)], row.names = FALSE) # many lines
summary(res[, 8:9])
wilcox.test(res$RE.lin, res$RE.linlog, conf.int = TRUE)
wt <-
# spaghetti plot
Ct(x = x, D = D, D.f = D.f, F = F.d, V = V.d,
C <-k01.1 = k01.1.d, k01.2 = k01.2.d, k10 = k10.d,
tlag1 = tlag1.d, tlag2 = tlag2.d)
== 0] <- NA
C[C FALSE
log.plot <-if (log.plot) {
"y"
log.y <- c(LLOQ*0.9, 1.2*max(data$C, na.rm = TRUE))
ylim <-else {
} ""
log.y <- c(0, 1.04*max(data$C, na.rm = TRUE))
ylim <-
}dev.new(width = 4.5, height = 4.5, record = TRUE)
par(no.readonly = TRUE)
op <-par(mar = c(4, 2.9, 0.2, 0.1), cex.axis = 0.9,
xaxs = "r", yaxs = "i")
plot(x, C, type="n", axes = FALSE, ylim = ylim, log = log.y,
xlab = "time", ylab = "")
mtext("concentration", 2, line = 2)
axis(1, at = c(seq(0, 6, 2), c(9, 12, 16, 24)))
axis(2, las = 1)
abline(h = LLOQ, col = "red")
rug(t, ticksize = -0.015)
for (i in 1:n) {
data$C[data$subject == i]
tmp.C <- t[!is.na(tmp.C)][1] # 1st C >= LLOQ
tfirst <- rev(t[!is.na(tmp.C)])[1] # last C >= LLOQ
tlast <-for (j in 1:(length(t)-1)) {
if (!is.na(tmp.C[j]) & !is.na(tmp.C[j+1])) {
if (tmp.C[j+1] >= tmp.C[j]) { # up (linear)
lines(x = c(t[j], t[j+1]),
y = c(tmp.C[j], tmp.C[j+1]), col = "#0000FF80")
else { # down (linlog)
} NULL
tk <- NULL
Ck <-for (k in seq(t[j], t[j+1], length.out = 100)) {
c(tk, k)
tk <- c(Ck, exp(log(tmp.C[j])+abs((k-t[j])/(t[j+1]-t[j]))*
Ck <- (log(tmp.C[j+1])-log(tmp.C[j]))))
}lines(tk, Ck, col = "#FF00FF80")
}
}
}
}lines(x, C, type = "l", lwd = 3, col = "#00000080")
c(expression(C[t[italic(i)]]>=C[t[italic(i)-1]]*" (linear up)"),
leg <-expression(C[t[italic(i)]]>=C[t[italic(i)-1]]*" (log down)"),
"theoretical profile", "LLOQ")
legend("topright", box.lty = 0, y.intersp = 1.2, lwd = c(1, 1, 3, 1),
col = c("#0000FF80", "#FF00FF80", "#00000080", "red"),
legend = leg, seg.len = 2.5)
box()
# box plots
1.04*c(-1, +1)*max(abs(range(res[, 8:9])))
ylim <- par(no.readonly = TRUE)
op <-par(mar = c(0, 2.9, 0, 0), cex.axis = 0.9)
plot(c(0.5, 2.5), c(0, 0), type = "n", axes = FALSE,
ylim = ylim, xlab = "", ylab = "")
abline(h=0, lty=3)
boxplot(res$RE.lin, res$RE.linlog, boxwex = 0.35, notch = TRUE,
notch.frac = 0.6, ylim = ylim, names = "", ylab = "",
main = "", axes = FALSE, col = "bisque", outcol = "red",
add = TRUE)
axis(2, las = 1)
mtext("linear", at = 1, 1, line = -1.1)
mtext("lin-up / log-down", at = 2, 1, line = -1.1)
mtext("relative error (%)", 2, line = 2)
paste0("\u2206 ", sprintf("%+.3f", wt$estimate),
wilc.diff <-"\n95% CI ",
sprintf("(%+.3f", wt$conf.int[1]), ", ",
sprintf("%+.3f)", wt$conf.int[2]))
text(x = 1.5, y = min(res[, 8:9]), labels = wilc.diff, cex = 0.9)
par(op)
```

Lag times^{4} were in 230 profiles 0.5 h, in 407 0.75 h, in 337 1 h, and in 26 1.25 h. In 493 profiles *t*_{last} was at 16 hours and in 507 at 24 hours.

After 16 hours the profiles may give a false impression. Any kind of average plot would be positively biased. This will be elaborated in another article.

The linear and lin-up/log-down trapezoidal rules (compared to the theoretical *AUC* of profiles without analytical error) clearly show the positive bias of the former and that it performs significantly worse than the latter.

Better methods to estimate the absorption phase(s) were proposed^{5} but are currently not implemented in any software.

There is no good reason to stick to the linear trapezoidal rule any more. The lin-up/log-down trapezoidal rule is implemented for decades (‼) in commonly used software. In Phoenix/WinNonlin it’s just one click in the NCA Options (see the online help for details). Don’t tell me that it’s complicated.

In the R-package `bear`

^{6} it is the default. THX to Yung-jin!

`method = 'lin up/log down'`

is also the default of the function `pk.calc.auxc()`

in the R-package `PKNCA`

.^{7} ^{8}

The bias of the linear trapezoidal rule gets especially nasty if there a missing values. Here I removed the 16 h value of the first simulated profile.

```
# you have to run the simulation script before
data$t[data$subject == 1]
t.full <- data$C[data$subject == 1]
C.full <- C.full[-which(t.full == 16)]
C.miss <- t.full[-which(t.full == 16)]
t.miss <- t[!is.na(C.full)][1]
tfirst <- rev(t.full[!is.na(C.full)])[1]
tlast <- data.frame(dataset = c("full", "16 h missing"),
comp <-lin = NA, linlog = NA)
$lin[1] <- trap.lin(t.full, C.full, tfirst, tlast)
comp$linlog[1] <- trap.linlog(t.full, C.full, tfirst, tlast)
comp$lin[2] <- trap.lin(t.miss, C.miss, tfirst, tlast)
comp$linlog[2] <- trap.linlog(t.miss, C.miss, tfirst, tlast)
compprint(comp, row.names = FALSE)
dev.new(width = 4.5, height = 4.5)
par(no.readonly = TRUE)
op <-par(mar = c(4, 2.9, 0.2, 0.1), cex.axis = 0.9,
xaxs = "r", yaxs = "i")
plot(t.miss, C.miss, type="n", axes = FALSE,
ylim = c(0, 1.04*max(data$C, na.rm = TRUE)),
xlab = "time", ylab = "")
mtext("concentration", 2, line = 2)
axis(1, at = c(seq(0, 6, 2), c(9, 12, 16, 24)))
axis(2, las = 1)
abline(h = LLOQ, col = "red")
rug(t, ticksize = -0.015)
points(16, C.full[which(t.full == 16)], pch = 19,
cex = 1.25, col = "red") # the ‘missing’ one
# linear
lines(t.miss, C.miss, col = "#FF00FF80", lwd = 2)
# lin-up / log-down
for (i in 1:(length(t.miss)-1)) {
if (!is.na(C.miss[i]) & !is.na(C.miss[i+1])) {
if (C.miss[i+1] >= C.miss[i]) { # up (linear)
lines(x = c(t.miss[i], t.miss[i+1]),
y = c(C.miss[i], C.miss[i+1]),
col = "#0000FF80", lwd = 2)
else { # down (linlog)
} NULL
tk <- NULL
Ck <-for (k in seq(t[i], t.miss[i+1], length.out = 200)) {
c(tk, k)
tk <- c(Ck, exp(log(C.miss[i])+abs((k-t.miss[i])/
Ck <- (t.miss[i+1]-t.miss[i]))*
(log(C.miss[i+1])-log(C.miss[i]))))
}lines(tk, Ck, col = "#0000FF80", lwd = 2)
}
}
}legend("topright", box.lty = 0, y.intersp = 1.2, lwd = c(2, 2, 1),
col = c("#FF00FF80", "#0000FF80", "red"), seg.len = 2.5,
legend = c("linear", "lin-up / log-down", "LLOQ"))
box()
par(op)
```

```
R> dataset lin linlog
R> full 77.42499 75.71357
R> 16 h missing 82.97551 76.56601
```

Connecting the 12 and 24 hours data by a straight line in linear trapezoidal fires back (*AUC*_{0–t} +7.2%). The lin-up/log-down is much less affected (*AUC*_{0–t} +1.1%).

Hence, if the linear trapezoidal is used, in a comparative BA study missing values bias the T/R-ratio.

However, as seen in Fig. 1 and Fig. 2, the more similar two concentrations (in the Figures \(\small{a}\) and \(\small{b}\)), the more similar the areas obtained by the linear and log trapezoidal. Or the other way ’round: The narrower the sampling interval, the smaller the difference obtained by the two methods. Hence, if the linear trapzoidal is used, missings in a part of the profile with rich sampling hurt less than in a part with wide intervals.

Always state in the protocol / report which trapezoidal rule you will use / used.^{9} It supports assessors^{10} to verify your results.

Nowadys the linear trapezoidal rule is of historical interest only and should be thrown into the pharmacometric waste bin.

If your SOPs still call for the linear trapezoidal rule, it’s high time for an update.

**Abandon calculating the AUC by the linear trapezoidal rule. Your results will inevitably be biased and lead into trouble if values are missing.**

It’s okay to talk about the linear trapezoidal rule at the beginning of the pharmacology curriculum but only in a *historical context*.^{11} Heck, Yeh and Kwan wrote about the bias of the linear trapezoidal already 43 years ago.

“[…] the log trapezoidal method can be be used advantageously in combination with a scond method, such as the linear trapezoidal rule, to yield optimal estimates.

For a good while we are living in the 21^{st} century. 😉 Nowadays, the lin-up/log-down trapezoidal rule should be elaborated.^{13} ^{14}

Defining in NCA the lag time as the time point before the time of the first measurable concentration is a conventient consensus. It’s quite obvious that the true one lies somewhere in between. A more realistic approach was suggested^{15} but is not implemented in any software. Before you think about a home brewn solution: It might be judged by regulatory assessors as borderline modeling – which it isn’t.

I used up to eight significant digits in the example above only to allow accurate recalculation. Our bioanalytical methods are never that precise (more than up to four digits belong to the realm of science fiction). In the past I experienced questions from assessors when data in full precision (15 digits) were used in calculations but were not reproducible based on the ‘nice’ rounded numbers given in the analytical and statistical report. This issue will be elaborated in another article.

top of section ↩︎ previous section ↩︎

License

Helmut Schütz 2021

1^{st} version April 03, 2021.

Rendered 2021-04-13 17:38:37 CEST by rmarkdown in 0.25 seconds.

Footnotes and References

Mersman O, Trautmann H, Steuer D, Bornkamp B.

*truncnorm: Truncated Normal Distribution.*2018-02-27. CRAN.↩︎Yeh KC, Kwan KC.

*A Comparison of Numerical Integrating Algorithms by Trapezoidal, Lagrange, and Spline Approximation.*J Pharmacokin Biopharm. 1978; 6(1): 79–98. doi:10.1007/BF01066064.↩︎*»You press the button, we do the rest.«*was an advertising slogan coined by George Eastman, the founder of Kodak, in 1888.↩︎In NCA defined as the time point before the time of the first measurable concentration.↩︎

Purves RD.

*Optimum Numerical Integration Methods for Estimation of Area-Under-the-Curve (AUC) and Area-under-the-Moment-Curve (AUMC).*J Pharmacokin Biopharm. 1992; 20(3): 211-26. doi:10.1007/bf01062525.↩︎Lee, H-y, Lee Y-j.

*bear: the data analysis tool for average bioequivalence (ABE) and bioavailability (BA).*v2.8.9. 2020-09-21. bear↩︎Denney W, Duvvuri S, Buckeridge C.

*Simple, Automatic Noncompartmental Analysis: The PKNCA R Package.*J Pharmacokin Pharmacodyn. 2015; 42(1): 11–107, S65. doi:10.1007/s10928-015-9432-2.↩︎Denney B, Buckeridge C, Duvvuri S.

*PKNCA. Perform Pharmacokinetic Non-Compartmental Analysis.*version 0.9.4. 2020-06-01. CRAN.↩︎WHO. Technical Report Series, No. 992. Annex 6.

*Multisource (generic) pharmaceutical products: guidelines on registration requirements to establish interchangeability.*Section 7.4.7*Parameters to be assessed.*2017. download.↩︎… and consultants. For me it is regularly a kind of scavenger hunt.↩︎

Dost FH.

*Der Blutspiegel. Konzentrationsabläufe in der Kreislaufflüssigkeit.*Leipzig: VEB Thieme; 1953. German.↩︎Gibaldi M, Perrier D.

*Pharmacokinetics. Appendix D. Estimation of Areas.*New York: Marcel Dekker; 1982.↩︎Papst G.

*Area under the concentration-time curve.*In: Cawello W. (ed)*Parameters for Compartment-free Pharmacokinetics.*Aachen: Shaker Verlag; 2003. p. 65–80.↩︎Gabrielsson J, Weiner D.

*Pharmacokinetic and Pharmacodynamic Data Analysis.*Stockholm: Apotekarsocieteten; 5^{th}ed. 2016. p. 141–58.↩︎Csizmadia F, Endrényi L.

*Model-Independent Estimation of Lag Times with First-Order Absorption and Disposition.*J Pharm Sci. 1998; 87(5): 608–12. doi:10.1021/js9703333.↩︎