2324 Anal. Chem. 1909. 61, 2324-2327

is T ’A

c 250 40s mc boo

Figwe 1. Sections of the voltammograms (see text) for dopamine (DA) and ascorbate (AA) in PBS, pH 7.4: (top) DA, 100 pmol/L; AA, 200 pmol/L; at an unmodlfied carbon paste electrode; (middle) DA, 20 pmol/L; AA, 200 pmol/L; at a stearate-modlfied carbon paste elec- trode; (bottom) DA, 100 WmollL; AA, 500 Imol/L; at a tissue-treated stearate-modified carbon paste electrode.

The effect of brain tissue on the stearate-modified electrode is consistent with recent studies of similar action by brain tissue on unmodified carbon paste electrodes (24) and with reports for surfactant action on carbon paste electrodes (25). It has been proposed that in such cases the surfactant solu- bilizes the oil and other hydrophobic elements of the paste, leaving behind a “clean” graphite surface. We propose that a similar mechanism occurs in vivo. The stearate-modified electrode implanted in brain tissue is introduced to the hy- drophobic environment of lipids and proteins. These take the role played by the surfactants mentioned above and remove the oil and other hydrophobic /lipophilic moieties of the electrode surface, the result being a modification of the electrode surface and an increase in the rate of electron transfer as shown (Figure 1 and Table I). A reduction in sensitivity is found at the tissue-treated electrodes compared to the electrodes before treatment, and is most likely a result of partial blockage of the electrode surface due to the ad- sorption of lipids and proteins (26).

A linear sweep voltammetric wave, attributed to dopamine oxidation at the stearate-modified electrode in vivo, has been reported by Lane et al. (14). The wave, centered at +lo0 mV vs Ag/AgCl, has a peak height of the order of 1 nA. This peak occurs at a potential (vs SCE) corresponding to large ascorbate oxidation at the tissue-treated electrode. A current of 1 nA for dopamine in vitro would correspond to a concentration of approximately 5 wmol/L (10). Taking into account the restricted compartment environment of the electrode in the brain (27, 28), this concentration represents a gross under-

estimate of the concentration of dopamine needed to produce a current of this size. Considering the ratio of 1OooO:1 for ascorbate to dopamine concentrations in the striatum, it is likely that the current recorded in vivo is due almost entirely to ascorbate.

In conclusion, the results indicate that while stearate- modified electrodes have the desired properties for electro- chemical discrimination of ascorbate and dopamine before they are implanted in brain tissue, these properties are lost after implantation. Taken together, the literature data and the present findings also suggest that these electrodes are neither selective nor sensitive enough to detect dopamine levels in vivo.

LITERATURE CITED Marcenac, F.; Gonon, F. Anal. Chem. 1985, 57, 1778-1779. Crespi, F.; Martin, K. F.; Marsden, C. A. Neurosclence 1988, 27,

Stamford, J. A. Anal. Chem. 1988, 58, 1033-1036. Kasser, R. J.; Renner, K. J.; Feng, J. X.; Brazell, M. P.; Adams, R. N. Brain Res. l98& 475, 333-344. Ewing, A. 0.; Wightman, R. M. J. Neurochem. 1984. 43. 570-577. Adams, R. N.; Marsden, C. A. I n Hendbook of Psychopharmacdogy; Plenum Press: New York, 1982; Vol. 15, pp 1-74. Voltammefry in the Neurosclences; Justice, J. B., Jr., Ed.; Humana Press: Cllfton, NJ, 1987. Measurement of Neurotransmitter Release In Vivo; Marsden, C. A,, Ed.; IBRO Handbook Series; J. Wiley and Sons: Chichester, 1984; Vol. 6. Marsden, C. A.; Joseph, M. H.; Kurk, 2. L.; MaMment, N. T.; 0”eill. R. D.; Schenk, J. 0.; Stamford, J. A. Neuroscience 1988, 25, 389-400. Blaha, C. D.; Lane, R. F. Brain Res. Bull. 1983, 10, 881-864. Gelbert, M. 8.; Curran, D. J. Anal. Chem. 1986, 58, 1028-1032. Lane, R. F.; Biaha, C. D.; Hari, S. P. Braln Res. Bull. 1987, 19, 19-27. Broderick, P. A. Life Sci. 1985, 36, 2269-2275. Lane, R. F.; Blaha, C. D.; Phillips, A. G. Brain Res. 1988. 397, 200-204. Glynn, G. E.; Yamamoto, B. K. Brain Res. 1989, 481, 235-241. Gonon. F. G.; Navarre, F.; Buda, M. J. Anal. Chem. 1984, 56. 573-575. Kelly, R. S.; Wightman, R. M. Brain Res. 1987, 423, 79-87. Church, W. H.; Justice, J. 6.. Jr. Anal. Chem. 1987, 59, 712-716. ONeill, R. D.; Fillenz, M.; Albery, W. J.; Goddard, N. J. Neurosclence 1983, 9, 87-93. OMham, K. 8. J. Electroanel. Chem. 1985, 184, 257-287. Sternson, A. W.; McCreery, R.; Feinberg, 6.; Adams, R. N. J . Nec- troanal. Chem. 1973, 46, 313-321. Dayton, M. A.; Ewing, A. G.; Wlghtman, R. M. Anal. Chsm. 1980, 52, 2392-2396. Kovach, P. M.; Ewing, A. 0.; Wilson, R. L.; Wightman, R. M. J. Neu- rosci. Mettwds 1984, 10. 215-227. Ormonde, D. E.; O‘Neill, R. D. J. Electroanal. Chem. 1989. 261,

Albahadily, F. N.; Mottob, H. A. Anal. Chem. 1987, 59, 958-962. Nelson, A.; Auffret, N. J. Electroanel. chem. 1988, 248, 167-180. Albery, W. J.; Goddard, N. J.; Beck, T. W.; Flllenz, M.; O’Neill. R. D. J. Electroanel. Chem. 1984, 161, 221-233. Cheng, H.-Y. J. Nectroanal. Chem. 1982, 135, 145-151.

885-896.

463-469.

Paul D. Lyne Robert D. O’Neill*

Chemistry Department University College Dublin Belfield, Dublin 4 Ireland

RECEIVED for review April 28, 1989. Accepted July 20,1989. We thank EOLAS for a grant to P.D.L. under the Basic Re- search Awards scheme.

Estimating Error Limits in Parametric Curve Fitting

Sir: The recent article by Phillips and Eyring in this journal (I) presented an interesting solution, based on the sequential simplex method, to the problem of the estimation of errors in nonlinear parametric fitting. The authors did not mention

two other simple, powerful, and reliable methods, the jackknife and the bootstrap (2-6).

The need for simple and robust procedures to assess con- fidence limits in estimated parameters is widely perceived in

0003-2700/89/0361-2324$01.50/0 0 1989 American Chemical Society

ANALYTICAL CHEMISTRY, VOL. 61, NO. 20, OCTOBER 15, 1989 2325

Table I. The Reference 'Experimental" Data Set"

t A t A

1.5 1.5 3.0 3.0 4.5 4.5 6.0 6.0 9.0

0.111 0.109 0.169 0.172 0.210 0.210 0.251 0.255 0.331

9.0 12.0 12.0 15.0 15.0 18.0 18.0 24.0 24.0

0.325 0.326 0.330 0.362 0.383 0.381 0.372 0.422 0.411

"These 'noisy" data (from ref 16) are simulated results of a first-order kinetics experiment. The absorbance (A) is a function of time ( t ) . The function has the form A = A& - exp(-kt)), where A,, the absorbance at t = a, and k, the rate constant, are the unknown parameters.

all physical sciences. The jackknife and the bootstrap al- gorithm can provide a particularly simple solution. Other methods (for example, likelihood and lack-of-fit, ref 7)) which will not be discussed here, can also be used.

Relatively few explicit references to the jackknife and the bootstrap are found in the chemical literature. Some appli- cations of the bootstrap have been determination of the confidence limits of correlation coefficients between elemental concentrations in meteorites (8)) analysis of the correlation between blood lead content and blood pressure in policemen (9)) determination of confidence bounds of means, errors, and correlation coefficients in air quality data (IO), determination of confidence intervals for the timing of the DNA molecular clock in human filogenesys (11)) and analysis and interlabo- ratory comparison of exponential fits in specific-heat mea- surements (12). The jackknife has found application in the estimation of confidence intervals for parameters associated with quantitative structure-activity relationships (13)) in uncertainty analysis in reactor risk estimation (14), and in outlier detection and error estimation in geothermometer calibration (15).

Concise operational descriptions are presented here and are applied to a simple test problem in two parameters. Extension to problems with more parameters and/or variables is straightforward. Formal descriptions and proofs can be found in ref 3 and 6 and references therein.

The jackknife is a finite algorithm (it requires a finite and a priori computable number of calculations) while the boot- strap is not: the number of computations needed is propor- tional to the precision asked of the results. The bootstrap can provide a better representation of the geometry of the confidence regions, at the cost of a significantly larger com- putational load. Both can use any curve-fitting program.

Let us consider the now classic data set (from ref 16) re- produced in Table I.

The 18 "experimental" pairs (A, absorbance; t , time) sim- ulate the results of a first-order kinetics experiment. They are to be fitted by the function

A = A,(1 - exp(-kt))

where A,, the absorbance at t = m, and k, the rate constant, are the unknown parameters to be determined.

Assuming that all points have identical statistical weight, that the error is normal (random with Gaussian distribution), and that it affects only the dependent variable A, use of any nonlinear least-square parametric curve fitting program gives

A, = 0.4043

& = 0.1698 as extensively published. The fitting program used in this work was SIMP, a public domain implementation of the

Table 11. Parameters Fitting the Ful l Data Set (Table I) and the 18 Jackknife Subsets of 17 Points Each"

A , k SSR

0.404 275 02

0.405 939 18 0.405 777 97 0.405 177 94 0.405 540 86 0.403 581 31 0.403 581 31 0.403 652 97 0.403 988 70 0.404 293 19 0.404 278 38 0.406 488 95 0.406 13861 0.406 020 11 0.402 573 00 0.405 275 02 0.407 392 05 0.396 140 77 0.399 689 70

0.404 196 11

0.002 592 61

Full Set

Diminished Sets 0.169 830 49 0.003 642 02

0.166 835 86 0.003 217 55 0.167 121 95 0.003 29786 0.168 074 16 0.003 576 97 0.167 383 43 0.003 515 64 0.171 362 37 0.003 600 41 0.171 362 37 0.003 600 41 0.171 502 97 0.003 580 29 0.170 592 54 0.003 629 21 0.168 054 53 0.003 414 61 0.168 799 02 0.003 564 66 0.170 112 28 0.002 932 80 0.170 070 34 0.003 137 11 0.168 898 68 0.003 516 72 0.170 752 77 0.003 522 67 0.169 083 22 0.003 620 91 0.167 523 33 0.003 437 54 0.177 452 69 0.002 866 60 0.174 060 41 0.003 404 72

0.169 946 83

0.002 594 33

Mean

Standard Deviation

(point 1 deleted) (point 2 deleted) (point 3 deleted) (point 4 deleted) (point 5 deleted) (point 6 deleted) (point 7 deleted) (point 8 deleted) (point 9 deleted) (point 10 deleted) (point 11 deleted) (point 12 deleted) (point 13 deleted) (point 14 deleted) (point 15 deleted) (point 16 deleted) (point 17 deleted) (point 18 deleted)

"The full data set and the one-minus subsets were fitted by a PC-DOS Pascal port of program SIMP (ref 17), a public domain implementation of the simplex algorithm.

Table 111. Error Analysis of Data in Table I

error A,, abs a t t = m k, rate constant

method value, 0.404 value, 0.170 ref

jackknife 0.0107 0.0107 a bootstrap 0.0102 0.0103 a sequential simplex 0.012 0.013 1 Marquardt 0.009 0.010 1 Newton 0.006 0.007 1

" Present work.

simplex algorithm (17) ported to Borland Pascal in the PC-DOS environment. The questions to be answered now are what is the error on the estimates of the parameters A, and 12 and what are the limits (e.g. at the 0.5,0.9,0.98 confidence level) on the possible range of these estimates?

THE JACKKNIFE To estimate error by the jackknife method 1. Delete the first data point from the triginal data set. 2. Fit the "jackknifed" set to compute Am(1), k(l). 3. Repeat steps 1 and 2 n times (where n is the number

of data poinJs) deleting now point 2, 3, ..., n, to obtain a set of n (A,('), kc')) parameter sets.

4. Compute the jackknife estimate of the standard error of parameter p (here, p = A,, k) as

where

f i ~ ) = l/n?fi(i) i = l

Application to the example data set produced the jackknife replicates in Table 11, from which the standard errors reported in Table I11 were obtained, in good agreement with results

2326 ANALYTICAL CHEMISTRY, VOL. 61, NO. 20, OCTOBER 15, 1989

previously obtained by other methods. The jackknife can be applied to small data sets, such as the

one studied here, with no need to develop computer code. All that is needed is to run the fitting program n times, where n is the number of data points in the set. The fitting program can utilize, of course, any algorithm.

The jackknife can be programmed in the following way: Instruct or modify the fitting program (called FITTER) to

append the parameters it computes to a given result file. Write a program (called KNIFE) to produce diminished

data sets (input the data set and the ordinal number of the point to be deleted).

Write a program (possibly a batch/script file, called JACK) to call n times KNIFE and FITTER.

JACK will produce a file containing the n parameter sets obtained by fitting the n diminished sets of (n - 1) data points. The variance of these parameters can be computed by any statistical package (or most pocket calculators). Multiplying the square root of the variance by (n - 1)1/2 gives the standard error of each parameter.

Outliers can be easily detected by examining the sum of the squares of the residuals (SSR) of the jackknifed sets: if deleting a data point decreases significantly the SSR of the resulting reduced set, that point is probably in error. The decision to reject an outlier from the experimental set can then be made according to definite rules.

If the error on a parameter is known to obey a specific distribution (for example normal or Student’s law), then tabulated values of the appropriate function can be used to translate standard error to confidence intervals. The vari- ance-covariance matrix of the jackknifed results can be used to construct elliptical joint confidence regions, but this technique implicitly assumes Gaussian distribution and has been found unreliable (18), at least for a highly nonlinear function and a data set affected by large and probably in- homogeneus errors (19).

THE BOOTSTRAP The bootstrap method studies the statistical properties of

a set of parameter sets obtained by fitting a large number of simulated data sets. These simulated data sets (or bootstrap replicates) are simply obtained by random sampling (with replacement) of the “experimental” data set.

Operationally: 1. Copy a point at random from the original data set to

2. Repeat step 1 n times, the number of samples in the

3. Fit the bootstrap replicate, computing Am(1), R ( I ) .

4. R_epeat steps 1, 2, and 3 m times to obtain a set of m

5. Compute the bootstrap estimate of the standard error

a simulated data set.

original set, to produce a bootstrap replicate.-

( i L ( i ) , k(i)) sets.

on parameter p (here, p = A,, k ) as

where m

The bootstrap can be programmed in the following way: Instruct or modify the fitting program (called FITTER) to

append the parameters it computes to a given result file. Write a program (called STRAP) to produce a bootstrap

replicate by copying n times a point at random from the original set to the replicate set.

Write a program (it can be a batch/script fiie, called BOOT) to call m times STRAP and FITTER.

L55 - t

t LJ5 - 360 380 400 420 ~ <40

A - X E o

Flgure 1. Contour plot of the density of distribution of the results of the fit of 2000 bootstrap replicates of the data set in Table I, in the absorbance-rate constant plane. The plot was obtained by mapping the values of the 2000 pairs of fltted parameters in a 64 X 64 grid. The grid was smoothed (Blackman window of period 4) in the two dimensions. The correlation between the two parameters and the skewdness of their distribution are visually evidenced.

The output of BOOT will be a file containing the m pa- rameter sets obtained by fitting m bootstrap replicates con- taining the same number of data points (n) as the original set. The best estimate of the error on each parameter is the square root of the sample variance (the sample standard deviation) of each parameter.

Application of the bootstrap (m = 2000) to the example data set gave error estimates in agreement with jackknife results (Table 111). Jackknife error estimates can be proven to be larger than bootstrap estimates by a factor [n/(n - 1)l1l2 (4) .

It is perhaps not sufficiently appreciated that function parameters are often strongly correlated. This can be evi- denced, in this example, by the contour plot in the absor- bance-reaction rate plane of the density (probability) dis- tribution of the results of fitting 2000 bootstrap replicates (Figure 1). The knowledge, with higher certainty, of the value of one of the parameters will allow a more correct estimate of the value of the other. For example, if other measurements were to indicate that the value of A, is exactly 0.380 (and not 0.403 f 0.010 as determined), the best estimate of 12 would become 0.192 (and not 0.170 f 0.010). This can be seen in Figyre 1-or obtained from the least-squares li2ear regression of k vs A , computed from the m bootstrap (A,(’), kc’)) pairs

k = 0.5347 - 0.9012A- Confidence intervals around best parameter estimates are

usually assumed to be normal (symmetrical and shaped as Gaussians); the joint confidence curves are then elliptical in parameter space. While this is true in linear models, at least for large number of points and/or normal error distribution on these points, confidence intervals in nonlinear regression are in general nonnormal, even if the errors on the data are themselves normally distributed their nonnormality is a consequence of the nonlinearity of the fitting function.

So, the histograms illustrating the bootstrap distribution of the probable values of the parameters A , and k (Figures 2 and 3) are asymmetric, reflecting the asymmetry apparent in Figure 1 along the regression line of the contour plot.

Caution should then be used in applying standard tools of statistical analysis, such as the normal probability or Student’s t distribution functions, to obtain confidence limits from standard error estimates. Rather, the whole set of bootstrap results should be used to observe the ranges withiin which the fitted parameters have a definite probability of being found.

From the results of 2000 bootstrap replicates it was possible to compute the asymmetric confidence intervals reported as fractiles in Table IV. The median (not the average) of the distributions is in excellent agreement with the results ob-

ANALYTICAL CHEMISTRY, VOL. 61. NO. 20. OCTOBER 15, I989 0 2927

Table IV. Confidence Intervals of the Computed Parameters (from Bootstrap Analysis, m = 2OOO)"

fractile 0.50

parameter 0.01 0.05 0.25 (median) 0.75 0.95 0.99

A. 0.3758 0.3847 0.3974 0.4042 0.4103 0.4180 0.4218 k 0.1498 0.1565 0.1641 0.1699 0.1773 0.1899 0.1996

'The fractiles were computed by program Aayst (hyst Software Technologies, Inc) on the fits of 2WO bootstrap replicates. Note that the median of the replicates is in agreement with the fit of the original data set. Since the distributions are skewed (Figures 2 and 3). the average values of the parameters in the replicates are slightly but significantly different from the median: A. = 0.4033, k = 0.1112. The table can be wed to indicate, for example, that there is a 90% probability that the true value of k lies between 0.1565 and 0.1899, but only a 1% nrobabilitv that it lies above 0.1996.

-. Y OZl. ff s I,,. I

A."O -re 2. Histogram representing the disblbution probability of A,, obtained from 2000 bwtstrap replicates 01 the data set in Table I. Note that the distribution is skewed.

The jackknife and-to an even larger extent--the bootstrap can be extremely demanding of computer resources. Pro- ducing and fitting the Moo replica& of this example required more than 20 min on a 16-MHz PC-DOS machine based on 80386 and 80387 processors. On the other hand, the growing availability of inexpensive computer power, the high cost of software development, and the power of these methods make the jackknife and the bootstrap a practical solution and a valid alternative to possibly faster, but less powerful and far more complex methods.

In conclusion, the computation of error estimates and of confidence intervals is posaihle and relatively simple in non- linear m e fitting, as i t is in linear problems. The jackknife and the bootstrap require no exceptional code development effort and can applied together with any curve fitting program. It is hoped that the methods illustrated here will be of use in those cases-more and more common given the present proliferation of computers and of computer-based instrumentation-in which data are processed by nonlinear curve-fitting methods.

LITERATURE CITED (1) Phillip. G. R.: Eyrlng. E. M. Anal. them. 1988, BO, 73.9-741. (2) Efron. E.: Gong. 0. Am. statbtkL?n 1982. 37(1). 36-48. (3) Mlller. R. 0. BiomeMka 1974, 61. 1-16. (4) Efron. E. Can. J . Star$rks 1981, 9(2). 139-172. (5) Efran. E. Ann. statist. 1979. 7. 1-26. (6) BBbU. 0. 1.: b e . A. Statistics Frobabury Len. 1989. 7. 151-160. (7) Donamaon. J. R.: Scimabel. R. B. ~echromemcs 1987, 28. 67-82. (8) Florenskii, P. V.: Pevenko, A. S. MteorilpiB 1987. 46, 172-177. (9) Weiss. S. T.: Munoz, A.: Stein, A,: Sparrow. D.: Spekeer. F. E. E W .

Environ. HBslth Perwect. 1988, 76. 53-56. (10) Hanna. S. R. JAPCA 1988. 38(4). 406-412. (11) Hasegawa. M.: Kishlno. H.: Yano. T. J . M. EuN. 1987. 26(1-2).

,lr)-ld, , . . . (12) Crary. S. E.: Fahey. D. A. Fiys. Rev. 8 : CMdenS. ManSr 1087,

(13) Dietrich. S. W.: Cxepr. N. 0.: Hansch. C.; Bentby. 0. L. J. Med.

(15) Powell. R. J . Mtamorph. oeol. 1085. S(3). 231-243. (16) Deming. S. N.: Morgan. S. L. Anel. Own. 1972. 45. 27SA-283A. (17) Caceci, M. S.: Cacheris. W. P. Syi.9 Mag. 1984. 96). 340-362.

FYun 9. Histogram representing the distribution probability of k . obmM t" 2000 boostrap repilcates of h data set in Table I. Note

2. resumng in "spikes" in me distribution. of no significance.

tained from the fit of the original data set.

35(4). 2102-2104.

Chem. 1980, 23(11). 1201-1205. that Me ~istribution 1% skewed. A finer grM was used than in Figure (14) M.; T. w. Ew, -, l,w, 85(31, 39H99.

Since the bootstrap method depends on the existence of random numbers, its results are not, in a strict sense, repro- ducible. There is a finite probability that the STRAP routine will produce "degenerate" replicatea. So, in the example given, there is about a 4.57 X lWB probability that one replicate set will contain 18 times the same data point; such a data set cannot be fitted. It is often convenient then to trim the bwtstrap results set and to perform statistics, for example, only using values within a given range about the mean.

11;; ~ ~ ~ ~ " . ~ ~ ~ ~ 9 ~ ~ ~ ~ g i S ~ ~ 6 1 2 3 - 1 2 9 . Marco s. Caceci

CEA-DRDD/SESD B.P. F-92265 FontenaY-aux-RmeS Cedex France

R E C ~ for review May 5,1989. Accepted July 17,1989.