Page 153 - 71_01

VOL. 71 (1), 153-173, 2005 CALIBRACIÓN, COMPARACIÓN DE MÉTODOS Y...

isolated non interfering measurements. Thus, assumption normality, nevertheless,
is a plausible assumption as an error term is made up of the combination of a large
number of small chance effects arising from several sources. Such a combination
tends to produce a normal distribution, regardless of the distribution of the sepa-
rate errors (the Central Limit Theorem) if its variance is finite.

In the context of most calibration problems the assumption relative to the abs-
cissa variable is reasonable because the analyte concentrations (x values) are pre-
cise enough. Particular attention must been given to equation in which one va-
riable is involved on both sides. Then an error in this quantity appears in both
coordinates mutually correlated in both conditions, i.e., the independent varia-
ble x is not an exact quantity and the independence of errors is not fulfilled.

When the abscissa range, e.g. concentration, span several orders of magnitude,
the precision of the y values vary greatly over the range of the x values. There two
main solutions to the problem of non constant variance: Transform the data, or
perform a weighted least squares regression analysis as several authors have po-
inted out is a better solution.

It is obvious that estimates of error variances independent both of the assumed
model and the method of fitting can only be obtained from replicates at each point.
It is important to understate that repeated runs must be genuine repeats and not
just repetitions of the same reading. The question of how many replicate measu-
rements to take must include consideration of the magnitude of variability, avai-
lability of the test material and reagent, the time required, the cost of each mea-
surement, and the variability required in the final result. Even within the concepts
that are based on the construction of a calibration curve, there is no consensus
about the choice of calibration samples and the number of replicates. As the num-
ber of replicate increases, however, the central limit theorem states that the fre-
quency distribution for the mean value approaches normality (very rapidly indeed,
especially if the parent distribution is symmetric). This fortunate circumstance
provides a very important, but little recognized basis for replication of analyses.

INTRODUCCIÓN

Las técnicas estadísticas se basan en suposiciones, y la validez de
los resultados obtenidos en la práctica depende de que las condicio-
nes supuestas se satisfagan, al menos con un grado suficiente de
aproximación. Exner ha suministrado recientemente ejemplos va-
riados en los que los datos experimentales se procesan de manera
incorrecta desde el punto de vista estadístico. Vamos a tratar la pro-
blemática del ajuste de una línea recta a un conjunto de datos biva-
riantes, cuando se llevan a cabo observaciones replicadas, sin perder
de vista el ámbito de las medidas físicas y químicas, de tanta apli-
cación en Química y Farmacia. La «Conferencia Internacional sobre

155