“The probably distributions discussed inthis chapter are fundamental to the use of statis for judging the elaliy of cate and for testing various hypotheses. Te quincunx show nthe upper part of he photo isa mechanical device that forms anormal probability distribution, Every 10 minutes, 30,000 balls op ‘om the center op of the machine, which contains aregular pate of pesto randomly del the balls, Eachtime a alli @ peg, ithas a 5050 chance of fling tote right ort the ef Aer each ball passes ‘through the aay of pegs, it drops int one of the vical “bins” of the transparent case. The height f the alu of ballin each bin is proprtional othe probably of ball fling ito given bin. The smth cate show nthe bottom half ofthe phato traces out the probability distribution. Random, or indeterminate, errors can never be totally eliminated and are often the major source of uncertainty in a determination. Random errors are caused by the many uncontrollable variables that accompany every measurement, Us ally, most contributors to random error cannot be positively identified, Even if wwe can identify random error sources, itis often impossible to measure them because most afe s0 small that they cannot be detected individually. The aceu- mulated effect of the individual uncertainties, however, causes replicate results to fluctuate randomly around the mean of the set. For example, the seatcer of data in Figures 5-1 and 5-3 is a direct result of the accumulation of small random uncertainties. We have replotted the Kjeldahl nitrogen data from Figure 5-3 as a three-dimensional plor in to becter reveal the precision and accu- racy of each analyst, Note that the random etror in the results of analysts 2 and 4 is much larger than that seen in the results of analysts 1 and 3. The results of analyst 3 show outstanding precision but poor accuracy. The results of analyst 1 show excellent precision and good accuracy. CHAPTER 694 CHAPTERS Random Enrors in Chemical Analysis Figure 6-2 Ashree-dimensional plot showing absolute errr in Kjeldahl nitrogen determinations for fous dll cent analyses. Note thatthe results of analyst | are both precise and accurate “The eras of analy 3 ate precise, but the absolute error is large. The results ‘of analysts 2 and 4 are both imprecise and inaccurate [Naber of results [Numer of results 6A-1 Random Error Sources ‘We can get a qualitative idea of the way small undetectable uncertainties produce a detectable random error in the following way. Imagine a situation in which just four small random errors combine to give an overall error. We will assume that each error hhas an equal probability of occurring and that each can cause the final result to be high or low by a fixed amount + U. “Table 6-1 shows all the possible ways the four errors ean combine to give the indicated deviations from the mean value. Note that only one combination leads to ‘TABLE 6-1 Possible Combinations of Four Equal-Sized Uncertainties Ti6 = 0.0625 4n6 = 0250 6116 ~ 0375 46 = 0250 16 = 0.0625a deviation of +4 U, four combinations give a deviation of +2 U, and six give a deviation of © U. The negative errors have the same relationship. This ratio of 1:4:6:4:1 is a measure of the probability for a deviation of each magnitude. If we make a sufficiently large number of measurements, we can expect a frequency dist bution like that shown in Figure 6-2a, Note that the y-axis in the plot isthe relative frequency of occurrence of the five possible combinations Figure 6-2b shows the theoretical distribution for ten equal-sized uncertainties Again we see that the most frequent occurrence is zero deviation ftom the mean. AC the other extreme a maximum deviation of 10 UToecurs only about once in 500 results ‘When the same procedure is applied to a very large number of individual errors, bell-shaped curve like that shown in Figure 6-2c results. Such a plot is called a Gaussian curve or a normal error curve. 64-2 Distribution of Experimental Results From experience with many determinations, we find that the distribution of re cof the licate data from most quantitative analytical experiments approaches Gaussian curve shown in Figure 6-2. As an example, consider the data in the spreadsheet in Table 6-2 for the calibration of a 10-ml. pipet.! In chis experiment 2 small flask and stopper were weighed. Ten milliliters of water were transferred to the flask with che pipet, and the flask was stoppered. The flask, the stoppet, and the water were then weighed again. The temperature of the w: to determine its density: The mass of the water was then calculated by raking the difference between the wo masses, The mass of water divided by its density is the volume delivered by che pipet. The experiment was repeated 50 times. The data in Table 6-2 are typical of those obtained by an experienced worker ‘weighing to the nearest milligram (which corresponds o 0.001 ml.) on a rop-loading balance and being careful to avoid systematic error. Even so, the results vary from a low of 9.969 ml to a high of 9.994 mL. This 0.025-mL spread of data results directly from an accumulation of all random uncertainties in the ex ‘The information in Table 6-2 is easier to visualize when the data are rearranged into frequency diseribution groups, as in Table 6-3. In this instance, we count and tabulate the number of data points falling inco a series of adjacent 0.003-mL ranges and caleulate the percentage of measurements in each range. Note that 26% of the results occur in the volume range from 9.981 to 9.983 mL. This isthe group cor taining the mean and median value of 9.982 mL. Note also chat more than half the results are within 0.004 ml. of this mean. The frequency distribution data in Table 6-3 are plotted as a bar graph, of histogram in Figure 6-3, labeled A. We ean imagine that as the number of was also mea measurements increases, the histogram approaches the shape of che continuous ‘curve shown as plot B in Figure 6-3. This plot shows a Gaussian curve, oF normal error curve, which applies to an infinitely large set of data. The Gaussian curve has the same mean (9.982 mL), the same precision, and the same area under the ccurve as the histogram. \ gee Section SAA for an expeient on ealibrton of pipet 6A The Nature of Random Errors 95 Coeeraines ere ne sae f\ al 5-4 200120 aU 46 Deviation from mean 04 LAN, -U-4U-aU 0 sav aU Relative frequeney © fos . Zot ~~ 0 ++ Devistion rom mean © Figure 6-2 Requenyditibuon sasurements containing (a) four random sncertaities,(b) tea random uncertainties, and (¢) a very large number of random uncertainties, ‘The spread in a set of replicate mea surements is the diference between the highest and lowest revul. A histogram is bar graph such as thar shown by plor Ain Figure 6-396 CHAPTERS Random Enrors in Chemical Analysis TABLE 6-2" 1 998 18 2) gars 9 3) 9a Zz a 9990 2 S| 9975 2 6 aa 2B 7) 9986 ™ a ga 3 9 ose % 0 9990 z nN 9980 x 2 9909 zB 3B 9978 a a7 3 16) 9982 2 6 9a 2 7 $958 a "Data listed ithe order obtained ‘Mean 9962 Maxirum 2| Meson 9982 Mewnum 9975, 9580 390 son 9.934 9901 9.987 9978 9983 9982 3331 S981 9959 9.995 9377 9975 3553 saat 9959, 002s, SSBSHRPOA* Sees eR LAT ete Te Fs 7 a] TRephcate Oata forthe Catbraton of a 1O-mi Pipet™ [2 Trial Volome, mt Trial Volume, mi. Tiiat Volume, mt 9976 3990 988 aa) 9995 9978 9.986 9982 a7? as? 9986 oa78 9983 9380 94 9379 ‘or Fact clelations af the stitial quam Listed at he bottom of Table 6-2 se 8. R Grouch and FJ. Holley Applcasions of Mics scl im Analy Chemie, 2ad ed, Belmont, CA: Brooks/Cole, 2014, ch. 2 AGaussian, or normal error curve, is Variations in replicate measurements, such as those in Table 6-2, result from sconeshatsbowsteaymmetol dis pgmerous small and individwally undetectable random errors that are caused tribution of data stound the mean of an infinite set of data such as the one in Figure 6-2. by unconcrollable variables in the experiment, Such small errors usually tend to cancel one another and thus have a minimal effect on the mean value. Ocea- sionally, however, they occur in the same direction and produce a large positive for negative net error. Sources of random uncertainties in the calibration of a pipet include: (1) visual judgments, such as the level of the water with respect to the marking on the pipet and the mercury evel in the thermometer; (2) variations in the drainage TABLE Frequency Distribution of Data from Table 6-2a Py 2 Percentage of measurements 9969 9.972 9975 9978 9.981 994 9.947 9.990 9.993 9971 9974 9977 9940 9.943 9.945 9.949 9.952 9.995 Range of meavured valves, mi. time and in the angle of the pipet as it drains; (3) temperature fluctuations, which allect the volume of the pipet, the viscosity of the liquid, and the performance of the balance: and (4) vibrations and drafts that cause small variations in the bal- ance readings. Undoubtedly, there are many other sources of random uncertainty in this calibration process that we have not listed. Even the simple process of calibrating a pipet is allected by many small and uncontrollable variables. The cumulative influence of these variables is responsible for the observed scatter of results around the mean ‘The normal distribution of daca that results ftom a large number of experiments is illustrated in Feature 6-1 Flipping Coins: A Student Activity to Illustrate a Normal Distribution If you flip a coin 10 times, how many heads will you get? Try it, and record your results, Repeat the experiment, Are your results the same? Ask friends or members of your class co perform the same experiment and tabulate the results, The table below Contains the results obtained by several classes of analytical chemistry students over an, 18-year period, ia (continued) 6K The Nature of Random Exors 97 Figure 6-3. Abistogram (4) showing distribution of che 50 results in Table {6-3 and a Gaussian curve (8) for data having the same mean and standard deviation as the data inthe histogram.98 CHAPTERS Statistical analysis only reveals inomatonthaisaieady present in a data set. Nonew information is created by statistical treatments, Statistical methods, do allow us to categorize and characterize ddatain different ways and to make objective and intelligent decisions about data quality and interpretation. A population is che collection ofall ‘measurements of interest co the ‘experimenter, while a sample isa subset of measurements selected from the population, Random Ezrors in Chemical Analysis 4 6 [Number of heads Figure 6F-1 Results of a coin-Bipping experiment by 395 students over an 18-year period, [Gad STATISTICAL TREATMENT OF RANDOM ERRORS We can use statistical methods co evaluate the random errors discussed inthe preceding, section. Generally, we base statistical analyses on the assumption that random extors in analytial results follow a Gaussian, or normal, distribution, such as that illustrated in Figure 6-2c, by curve B of Figure 6-3, or by the smooth curve in Figure 6F-1. Ana- Iycical daca can follow distributions other than the Gaussian distribution, For exam- ple, experiments in which there is either a successful outcome ota failure produce data that follow the binomial distribution. Radioactive or photon-counting experiments produce results that follow the Poisson distribution. However, we often use a Gaussian distibucion to approximate these distributions. The approximation becomes better in the limit ofa large number of experiments, As a rule of thumb, if we have more than. 30 results and the data are not heavily skewed, we can safely use a Gaussian distribu- tion. Thus, we base this discussion entirely on normally distributed random errors. 6B-1 Samples and Populations ‘Typically ina scientific study, we infer information about a population or universe from observations made on a subset or sample. The population isthe collection of all measurements of interest and must be carefully defined by the experimenter, In some eases, the population is finite and real, while in others, he population is hypo- ‘hetical or conceptual in nature. ‘As an example of a real population, consider a production run of multivitamin tablets that produces hundreds of thousands of tabless. Although the population is finite, we usually would not have the time or resources to test all the tablets for qual- ity control purposes. Hence, we select a sample of tablets for analysis according to statistical sampling principles (see Section 8B). We then infer the characteristics of the population from those of the sample. In many of the cases encountered in analytical chemistry, the population is conceptual. Consider, for example, the determination of calcium in a community ‘water supply to determine water hardness. In this example, the population is the very large, neatly infinite, number of measurements that could be made if we6B Statistical Treatment of Random Errors 99 analyzed the entire water supply. Similarly, in determining glucose in the blood of a patient, we could k ments if we used the ‘ochetically make an extremely large number of measure tire blood supply. The subset of che population analyzed in both these cases is the sample. Again, we infer charact 1m those obtained with the sample. Hence, itis w populasion being characterized Statistical laws have been derived for populations, but chey can be used for samples after suitabl ristice ofthe population tant to define the mp Do not confuse the statistical ‘sample with the analytical sample. Consider: sample: taken from the came codification. Such modifications are needed for small samples because a few data points may not represent the entire population. In the discussion thar follows, we first describe the Gat sian statsties of popula tions. Then we show how these relationships can be modified and applied to small samples of data supply and analyzed in the for caleiuin, The f jes recul is the population. They ated 6B-2 Properties of Gaussian Curves single satistical sample. This Figure 6-4a shows two Gaussian curves in which we plot the relative frequency y of various deviations from the mean versus the deviation from the mean. As showa in the margin, curves such as these can be described by an equation that contains just ‘wo paramecers, the population mean 1 and the population standard deviation or ‘The term parameter refers to quantities such as and o that define a population eos ‘or distribution. Data values such as x ate variables. The cerm statisti refers an Wa ‘estimate of parameter that is made from a sample of data as discussed below. The sample mean and the sample standard deviation are examples of statistics that esti- mate paramecers 1 and respectively quation for a normalized Gaussian curve has the form over The Population Mean sand the Sample Mean ® Scientists find ic useful co differentiate between the sample mean and the population mean. The sample mean ¥ is the arithmetic average of a limited sam- ple drawn from a population of data. The sample mean is defined as the sum of the measurement values divided by the number of measurements as given by Equation 5-1, page 84. In that equation, NV represents the number of measurements os ao Eon on o -=—_ 0 —+. = oo 2 Deviation from mean, =p re o » 3 Figure 6-4 Normal eroe curves. The standard deviation for curve Bis twice that for curve A, that is, ry = 2ory In (a) che abscsta isthe deviation from the mean (x ~ pin the unit of measurement In (b) the abscisa is the deviation from the mean in units of” For this pl, che «wo curves A and B are identical100 CHAPTERS Random Ezzors in Chemical Analysis The sample mean Fisfound from ) in she sample set. The population mean jt, in contrat, ie the true mean forthe x population. It is also defined by Equation 5-1 with the added provision chac N Ds represents the coral number of measurements in the population. In the absence of z systematic error the population mean is aso the true value for the measured quantity. v To emphasize the difference between the two means, the sample mean is symbol- where Nis the number of ied by ¥ and the population mean by 41. More often than not, particularly when measurements in the sample set Nis small, ¥ differs from yx because a small sample of data may not exactly rep- ‘The ame equation ie used to The sine gouston sede resent its population, In most cases we do not know st and must infer its value from ¥. The probable difference between ¥ and pt decreases rapidly as the num- ber of measurements making up the sample increases; usually by the time N reaches 20 to 30, this diflerence is negligible. Noce thac the sample mean ¥ is a statistic that estimates the population parameter J where Nis now the total number of The Population Standard Deviation & roessarement inthe popalation The population standard deviation @, which is a measure of the precision ofthe poptlaton, is given by che equation Inthe absence of systematic tnonthepopuation means Sine taeeduesfemeasied quantiy. 6) ‘The quantity (x,— 4) in Equation #1 isthe devaionof where NY isthe number of datapoints making up che population Spores komibe meant ‘The two curves in Figure 6-4a are for ewo populations of data chat differ only Equation 64 whieh is fora in their standard deviations. The standard deviation for the daca set yielding the sample of data_ broader but lower curve B is ewice that for the measurements yielding curve A. The breadth ofthese curves i messure ofthe precision ofthe two sets of data, Thus the precision of the data set leading to curve A is twice as good as that of the data set represented by curve B. Figure 6-4b shows another type of notmal error curve in which the x axis is now a new variable z, which is defined ae ew) F 69 The quanti 2 represent che Tontoncfcsenitionds > forelaton mean lative [Note chat xis the relative deviation of a data point from the mean, that is, the devia~ tion relative co the standard deviation. Hence, when x — je = 07, zis equal to one: the standard deviation Itis when x — jt = 20, zis equal to wor and so forth, Since z is the deviation from the commonly given asa variable ean relative to the standard deviation, a plot of relative frequency versus z yields in statistical tables since iti a a single Gaussian curve that describes all populations of daca regardless of standard. dimensionless quantity. deviation. Thus, Figute G-4b is the normal error curve for both sets of data used co ploccurves A and B in Figure 6-4a, ‘The equation for the Gaussian error curve is ed 9 Because it appeats in the Gaussian error curve expression, the square of the standard deviation 0? is also important. This quantity is called the variance (see Section 6B-5)6B Statistical Treatment of Random Esrors 101 A normal error curve has several general properties: (a) The mean occurs at the 20, (0.0054 and 0.0058 mL) The rapid improvement in the reliability of s with increases in N makes it feasi- ble to obtain a good approximation of @ when the method of measurement is not excessively time consuming and when an adequate supply of sample i available For example, if the pH of numerous solutions is to be measured in the course of an investigation, it is useful to evaluate s in a series of preliminary experiments. ‘This measurement is simple, requiring only that a p trodes be immersed in the test solution and the pH read fom a scale or a display. To derermine 5, 20 to 30 portions of a buffer solution of fixed pH can be mea- sured with all steps of the procedure being followed exactly. Normally, it is safe to assume that the random error in this test is the same az that in subsequent mea- surements. The value of s calculated from Equation 6-4 is a good estimator of the population value, «7 of rinsed and dried elec- Spreadsheet Summary In Chapter 2 of Applications of Microsoft® Excel in Analytical Chemistry 2nd ed, we introduce the use of Excel's Analy- sis ToolPak co compute che mean, standard deviation, and other quantities ‘The Descriptive Statistics option finds the standard error of the mean, the median, the range, che maximum and minimum values, and parameters that reflec che symmetry ofthe data sec Pooling Data to Improve the Reliability of s Ie'we have several subsets of data, a better estimate of the population standard deviation can be obtained by pooling (combining) the data instead of using only. cone daca set, Again, we must assume the same sources of random error in all the measurements, This assumption is usually valid if the samples have similar com- positions and have been analyzed in exactly the same way. We must also assume that the samples are randomly drawn from the same population and thus have a common value of6B Statistical Treatment of Random Errors The pooled estimate of s, which we call s,s i8 a Weighted average of the indi- vidual estimates. To calculate 5,.ca: deviations from the mean for each subset are squared; the squares of the deviations of all subsets are then summed and divided by the appropriate number of degrees of freedom. The pooled s is obtained by tak- ing the square root of the resulting number. One degree of freedom is lost for each subset, Thus, the number of degrees of freedom for the pooled ris equal co the total number of measurements minus the number of subsets. Equation 6-7 in Feature 6-4 gives the fll equation for obtaining ja for # data sets. Example 6-2 illustrates che application of this type of computation. Equation for Calculating the Pooled Standard Deviation ‘The equation for computing a pooled standard deviation from several ses of data takes the form Dei H + Bes H+ Bile Bt N +N AN +N, en where Nj is the number of results in set 1, Ny is the number in set 2, and so forth. ‘The term Nis the total number of daca secs pooled. EXAMPLE 6-2 Glucose levels are routinely monitored in patients suffering from diabetes, The glucose concentrations in a patient with mildly elevated glucose levels were determined in different months by a spectrophotometric analytical method, ‘The patient was placed on a low-sugar diet to reduce the glucose levels, The following results were obtained during a study to determine the effective ness of the diet, Calculate a pooled estimate of the standard deviation for the method. 1108, 1122, 1075, 1099, 1115, 168743 1083, 1100 992, 975, 1022, 1001, 991 1182.80 1788, 85,779, $22, 800 1086.80 799, 745,750, 774, 777, 800, 758 2950.86 “Toul number of measurements = 24 “Total sum of squares = 6907.89 Solution For the first month, the sum of the squares in che next to last column was calculated as follows: Sum of squares = (1108 ~ 1100.3)? + (1122 — 1100.3)" + (1075 — 1100.3)? + (1099 — 1100.3)? + (1115 ~ 1100.3)" +085 — 1100.37 + (100 — 11003)= 168743 107108 CHAPTERS Random Ezrors in Chemical Analysis ‘The other sums of squares were obtained similaly. The pooled standard deviation is then 8.58 ~ 19 mg/L Note this pooled value is a better estimate of than any of the individual s values in che last column. Note also that one degree of freedom is lost for each of the four sets, Because 20 degrees of freedom remain, however, the calculated value of scan be considered a good estimate of w. A glucose analyzer Spreadsheet Summary In Chapter 2 of Applications of Microsofi® Excel in Anagtical Chemistry, 2nd ed, we develop a worksheet to calculate the pooled standard deviation ofthe data from Example 6-2, The Excel function DEVSQ 0 is introduced to find the sum of the squares ofthe deviations. As extensions of this exercise, you may use the worksheet to solve some of the pooled standard devia~ tion problems atthe end of this chapter. You can also expand the worksheet to accom~ rmodate more data points within datasets and larger numbers of sets 6B-5 Variance and Other Measures of Precision Although the sample standard deviation is ually used in reporting the precision of analytical data, we often find chee other terms. Variance (5!) “The variance, sequal w dhe squase The variance is just the square of the standard deviation. The sample variance # is of the standard deviation an estimate of the population variance 0? and is given by Zen w Bw N-1” N=1 es Note that the standard deviation has the same units as the data, while the variance has the units of the data squared. Scientists tend co use standard deviation rather than variance because itis easier to relate a measurement and its precision if theySB Statistical Treatment of Random Errors 109 both have the same units. The advantage of using variance is that variances are addi- tive in many situations, as we discuss laer in this chapter. Relative Standard Deviation (RSD) and Coefficient of Variation (CV) Frequently standard deviations are given in relative rather than absolute terms. We ealeu- late the relative standard deviation by dividing the standard deviation by the mean value ‘of the dataset. The relative standard deviation, RSD, is sometimes given the symbol, ‘The International Union of Pure and Applied Chemistry recommends that the symbol 5, be used for zelative sample standard deviation and for relative population standard ‘The resule is often expressed in parts per thousand (ppt) or in percent by multiplying this ratio by 1000 ppt or by 100%. For example, deviation In equations where it : is cambersome to use RSD, we RSD in ppt 1000 ppt will uses, and o,, ‘The relative standard deviation multiplied by 100% is called the coefficient of “the coefficient of variation, CV. variation (CV). the percent relative standatd deviation CV = RSD in percent = = X 100% 69 Relative standard deviations often give a clearer picture of data quality than do abso- luce standard deviations. As an example, suppose chat a copper determination has a scandard deviation of 2 mg, If che sample has a mean value of 50 mg of copper, che CV is 20%, Spread or Range (w) "The spread, o: range, us another term that is sometimes used to describe the precision ‘ofa set of replicate results I isthe difference beoween the largest value in the set and the smallest. Thus, the spread of the data in Figure 5-1 is (20.3 ~ 19.4) = 0.9 ppm Fe. ‘The spread inthe results for month 1 in Example 6-2 is 1122 ~ 1075 = 47 mg/L glucose. EXAMPLE 6-3 For the set of data in Example 6-1, calculate (a) the variance, (b) the relative standard deviation in parts per thousand, (c) the coefficient of variation, and (d) the spread, Solution In Example 6-1, we found = 0.754 ppm Pb ands = 0.0038 ppm Pb (a) 2 = (0.0038) = 14 x 10° 0.0038 (&) RSD = SOEF = x 1000 ppr = 5.0 ppe to cy = 2238 x 100% = 0.50% O754 () w = 0760 ~ 0751 = 0.009 ppm Pbm0 CHAPTERS ‘The variance of a sum or difference is equal tothe sum of the variances of the rpumbers making up that ‘sum oF difference > Random Exrors in Chemical Analysis [Gia STANDARD DEVIATION OF CALCULATED RESULTS Often we must estimate the standatd deviation of a result that has been calculated from two or more experimental data points, each of which has 2 known sample stan- dard deviation. As shown in Table 6-4, the way such estimates are made depends on the types of calculations that are involved. The relationships shown in this table are derived in Appendix 9 TABLE 6-4 Error Propagation in Arithmetic Calculations a, band care experimental variables with standard deviations off tp and 1 eepecively "These relationships are derived in Appendix 9. The values lor yy ae absolute values fy ea negative number 6C-1 Standard Deviation of a Sum or Difference Consider the summation: +050 (+ 0.02) +410 (= 0.03) =197 (= 0.05) 2.63 where the numbers in parentheses are absolute standard deviations. If che thtee in- dividual standard deviations happen by chance to have the same sign, the standard deviation of the sum could be as large as +0.02 + 0.03 + 0.05 = +0.10 or ~0.02 — 0.03 = 0.05 = =0.10. On the other hand, iis possible that the three standard devia- tions could combine o give an accumulated value of zero: ~0.02 ~ 0.03 + 0.05 = 0 or +0.02 + 0.03 ~ 0.05 = 0. More likely, however, the standard deviation of the sum, wil lic berween these two extremes. The variance ofa sum or difference is equal to the sum of the individual variances.’ The most probable value fora standard deviation ofa sum or difference can be found by taking the square root of the sum of the squares of the individual absolute standard deviations. So, for the computation y= als) + BH) — ees) the vatiance ofp, 2, is given by ‘see R, Bevington and D,K Robinion, Date Redaction and Eror Ana fr the Pel See, Sed ed, New Yorks McCraw Hil, 2002, ch. 36C Standard Devistion of Caleulated Results 132 Hence, the standard deviation of the rer is ga EtG eS @19) wheres, sj, and s,are the standard deviations of the three terms making up the result Substituting the standard deviations from the example gives, 5, = Ve0.077 + (=0.03" + (£0.05 = =0.06 and the sum should be reported as 2.64 (£0.06). 6C-2 Standard Deviation of a Product or Quotient Consider the following computation where the numbers in parentheses are again absolute standard deviations: 4.10(2£0,02) x 0,0050(0.0001) 1970.0) = greases) In this situation, the standard deviations of two ofthe numbers in the calculation are larger than the result isle. Evidenly, we needa different approach for multiplication and division. As shown in Table 6-4, the relative standard deviation of a product ot ‘quotient is determined by the relative tandand deviations ofthe numbers foreaing the computed result, For exaople in the cate of axb ou) ‘we obtain the relative standard deviation s,/yof the result by surnming the squares of the relative standard deviations of a, 6, and c and then calculating the square root of the sum: 6 Applying this equation co the numerical example gives y G cy (= + +(2u 0.0050 197 = V(0.0049)? + (0.0200)? + (0.0203)* = =0.0289 ‘To complete the calculation, we must find the absolute standard deviation of the resul, s, = y X (40.0289) = 0.0104 X (=0.0289) = +0,000301 and we can write the answer and its uncertainty as 0.0104 (0.0003). Note that ify isa negative number, we should creat s,/y as an absolure value Example 6-4 demonstrates the calculation of the standard deviation of the result for a more complex calculation. For a sum ora difference, the standard deviation ofthe antwer's the squate root of the sum of the squares of the tandard deviationt of the numbers sed inthe ealeulation, C thoughout hers of thi chapter, we highlight the uncertain digit, by showing it in a recond colon For mulplicaion or division, the relative standard deviation of the anaver i the square 00% of the sum fof the squares ofthe relative standard deviavons of tne numbers that ate multiplied or divided, Cipiind the absolute standard deviation in a product or a quotient, first find the elative standacd deviation in the result tnd then multiply it by the reculsuz CHAPTERS Random Ezzors in Chemical Analysis EXAMPLE 6-4 Caloulate the standard deviation of the result of [14.5(40.2) — 11.6(¢0.2)] X 0,050(+0,001) [820 10) + 1030(#5)] x 42.3(£0.4) 1.725(2) x 10-6 Solution First, we must calculate the standard deviation of the sum and the difference. For the difference in the numerator, 4, = Vi=02) + (202) = +0.283 and for the sum in the denominator, = Vieloy + 57 = 11.2 ‘We may then rewrite the equation as = 1725 x 10-* “The equation now contains only products and quovients, and Equation 6-12 applies. Thas, 5 (ey y+ 6 Hay 4 ay y NW 27 0.050, 1850, 423, ‘To obtain the absolute standatd deviation, we write 0.107 y X 0.107 = 1.725 x 10-€ x (£0,107) = 0.185 x 10-¢ and round the answer to 1.7(£0.2) 10+, 6C-3 Standard Deviations in Exponential Calculations Consider the relationship =a where the exponent x can be considered so be free of uncereainty. As shown in Table 6-4 and Appendix 9, che relative standard deviation in y resulting ftom the £4) os ‘Therefore, the relative standard deviation of the square of a number is twice the rela- tive standard deviation of the number, the relative standard deviation of the cube root of a number is one third that of the number, and so forth, Example 6-5 illustrates this type of calculation,6C Standard Deviation of Caleulated Results 113 EXAMPLE 6-5 The solubility product K,, for the silver salt AgX is 4.0 (+0.4) X 10%, and the molar solubility is solubil = 20 x 10-"M (kK)! = (0 x 107 What isthe uncertainty in the calculated solubility of AgX? Solution Substicuing y = solubility, a= Ky, and x = into Equation 6-13 gives 4 _ 04 x 107% 2) fox10 04 4x4 A 00: ao °° 4, = 2.0 x 1074 X 0005 = 0.1 x 1074 solubility = 2.0 (0.1) x 104M Ie is important to note that the error propagation in taking a number to a power is dfferenc from the error propagation in multiplication. For example, consider the tuncertaingy in the square of 4.0 (+0.2). The relative error in the result (16.0) is given by Equation 6-13; ‘The result is then y = 16 (£2) Consider now the situation where y is the product of two independently measured numbers that by chance happen to have identical values of a, = 4.0 (£0.2) and ay = 4.0 (40.2). The relative ertor of the product 4:4) = 16.0 is given by Equation 6-12: ‘The result is now y = 16 (1), The reason for the difference between this and the the sign associated with one error can be the same as or different from that of the other previous result is that with measurements that are independent of one anothe error. If they happen to be the same, the error is identical to that encountered in the firsc case where the signs mur be the same, On the other hand, if one sign is positive and the other negative, the relative errors tend to cancel. Thus the probable error for the case of independent measurements lies somewhere between the maximum (10%) and zero, ‘The relative standard deviation ofy =a? isnot the same as the relative standard deviation of the product of three independent measurements y= abc, where414 CHAPTERS Random Ezzors in Chemical Analysis 6C-4 Standard Deviations of Logarithms and Antilogarithms ‘The las two entries in Table 6-4 show that for y= log a 9 and far y = antilog a 19) ‘As shown, the absolute standard deviation of the logarithm of a number is determined by the relative standard deviation of the number: conversely, the relative scandard deviation of the antilogarithm of a number is determined by the absolute standard deviation of the number. Example 6-6 illustrates chese calculations. EXAMPLE 6-6 Calculate the absolute standard deviations of the results of the following calculations. The absolute standard deviation for each quantity is given in parentheses. G@) y= logl2.00(+0.02) x 10+ = -3.6990 +7 ()_ y= antilog(1.200(+0.00;)] = 15.849 =? (© y= antilogl45.4(=0.3)] = 2.5119 x 108 + Solution (@) Referring to Equation 6-14, we see that we must multiply the relative standard deviation by 0.434: 0.02 x 10-4 0.434 x SO = 40,004 2,00 X 10 Thus, y= logl2.00(:20.02) x 10 3.699 (40.008) (©) Applying Equation 6-15, we have 2.303 X (0.003) = 0.0069 a 0.0069y = 0.0069 x 15.849 = 0.11 ‘Therefore, ‘y= anslog|1.200(0.005)] = 15.8 * 0.1= 2.303 X (0.3) = 0.69 4, = 065y = 0.69 x 25119 X 10" = 1.7 x 10% Thus, y= anilog(45.4(40.9)] = 25(21.7) x 104 = 3 (22) x 108 Example 6-6 demonstrates that a large absolute ertor is associated with the an logarithm of a number with few digits beyond the decimal poins. This large uncer- tainty is due co che fact chat the numbers co the lef ofthe decimal (che characteristic) serve only to locate the decimal point. The large error in the antilogarithm results from the relatively large uncertainty in the mantissa of the number (that is, 0.4 * 0.3) [EB] REPORTING COMPUTED DATA. A numerical result is worthless (o users of the data unless they know something, about its quality. Therefore, itis always essential to indicate your best estimate of the reliability of your data, One of the best ways of indicating reliability isto give a confidence interval at the 90% or 95% confidence level, as we describe in Section 7K-2, Another method is to report the absoluce standard deviation or the coellicient ‘of variation of the data. If one of these is reported, itis a good idea to indicate che number of data points that were used to obtain the standard deviation so thas the user has some idea of the reliability of s. A much less satisfactory but more common indicator of the quality of data is the significant figure convention, 6D-1 Significant Figures ‘We often indicate che probable uncertainty associated with an experimental mea- surement by rounding the result so chac ic contains only significant figures. By definition, the significant figures in a number are all of the certain digits plus the first uncertain digit. For example, when you read the 50-ml. buret section shown in ‘Figure 6-5, you can easily tell char che liquid level is greater than 30.2 mL and less than 30.3 mL. You can also estimate the position of the liquid between the gradu- ations to about 0.02 mL. So, using the significant figure convention, you should report the volume delivered as 30.24 mL, which has four significant figures. Note thar the firs three digits are certain, and che last digit (4) is uncertain, ‘A 2er0 may oF may not be significant depending on its location in a number. A sero that is surrounded by other digits is always significant (such asin 30.24 mL) be- «cause i is read directly and with certaincy ftom a scale or instrument readout. On the ‘other hand, zeros that only locate the decimal point for us are not. Ifwe write 30.24 Las 0.03024 L, the number of significant figures is the same, The only function of the zero before the 3 is to locate the decimal point, so itis not significant. Terminal, ‘or final, 2eros may oF may not be significant. For example, if the volume of a beaker is expressed as 2.0 L, the presence of the zero tells us chat the volume is known afew tenths ofa liter so that both the 2 and the zero ate significanc figures. If this same volume is reported as 2000 mL, the situation becomes confusing. The last «wo zeros ate not significant because the uncertainty is sil a few tenths ofa liter ora few SD Reporting Computed Data 115, “The significant figures in a number are all of the certain digite pls che fre uncertain digit Rules for determining the umber of signifieant figures: 1, Disregard all initial zeros. 2, Disregard all final zeros unless they follow a decimal point 3, Allemaining digits including zeros between nonzero digits are significant,416 CHAPTERS 20 i 6 Figure 6-5 Aburet secion showing the iguid level and meniscus Express data in scientific notation to avoid confusion in determining whether terminal zeros are significant. Asarule of thumb, for addition and subtraction, the result should contain the same mamber of decimal places ae the number with the smallest number of decimal places, When adding and subtracting rnumbers in scientific notation, express the numbers tothe same power of ten For example, ‘nage X 108 = 2492.0! +6532 X 10! = Fa.065:2 X20" “1227 108 = Toaaz) Xo 24a X1 = 2374 10 Goundes) ‘The weak link for multiplication and division is the number of significant figures in the number with the emallest number of significant figures. Use this rule of thumb with caution. > > > Random Ezzors in Chemical Analysis hundred milliliters. In order to follow the significant figure convention in a case such as this, use scientific notation and report the volume as 2.0 X 10° mL. 6D-2 Significant Figures in Numerical Computations [Determining the appropriate number of significant figures in the result ofan arithmetic combination of ewo or more numbers requires great care Sums and Differences For addition and subtraction, the number of significant figures can be found by visual inspection. For example, inthe expression 3.4 + 0,020 + 7.51 = 10.730 (round to 10.7) = 10.7 (rounded) the second and third decimal places in the answer cannot be significant because 3.4 is uncertain in the frst decimal place, Hence, the result should be rounded to 10.7 ‘We can generalize and say that, for addition and subtraction, the result should have the same number of decimal places as the number with the smallest number of deci- sal places. Note that the result contains three significant digits even though two of the numbers involved have only two significant figures Produets and Quotients Sometimes itis suggested for multiplication and division that the answer should bbe rounded so that it contains the same number of significant digits as the original number with the smallest number of signficanc digits, Unforcunately, this procedure sometimes leads to incorrect rounding, For example, consider the wo calculations WX 452s og yg KAO _ 9 agg 100.0 ae 100.0 If we follow the suggestion, the first answer would be rounded to 1.1 and the second {0 0.96. A better procedure is to assume unit uncertainty in the last digit of each number. For example, in the first quotient, the relative uncertainties associated with each of these numbers are 1/24, 1/452, and 1/1000. Because the first relative uncer tainty is much larger shan the other two, the relative uncertainty in the result is also 1/24; the absolute uncertainty is then. 108 x 4 ui 045 = 0.04 By she same argument, the absolute uncertainty ofthe second answer is given by 1 0.965 X 37 = 0.040 = 0.08 ‘Therefore, che fist result should be rounded to three significant figures or 1.08, bus the second should be rounded co only «wo, that is, 0.96. “ran extensive ducuion of propagation of igafcaneSigates see TM. Sehwar J Chem Bd, 1985, 62,693, DOK. 10.1021/ed062p693.Logarithms and Antilogarithms Be especially careful in rounding she results of calculations involving logarithms. The following rules apply co most situations and are ilustrated in Example 6-7: 1, Ina logarithm of a number, keep as many digits to the right of the decimal point as there are significant figures in the original number. 2, In an antilogarithm of a number, keep as many digits as there are digits to the he of the decimal poinc in she original numbes? EXAMPLE 6-7 Round the following answers so that only significant digits are retained: @) log 4.000 x 10 = ~4.3979400, and (b) antilog 125 = 3.162277 X 10" Solution (a) Following rule 1, we retain 4 digits to the right of the decimal point Jog 4.000 10 = -4.3979 (b) Following rule 2, we may retain only 1 digit antilog 12.5 = 3 X 10 6D-3 Rounding Data Always round the computed results of a chemical analysis in an appropriate way. For example, consider the replicate results: 41.60, 41.46, 41.55, and 41.61. The mean of this data set is 41.555, and the standard deviation is 0.069. When we round the mean, do we take 41.55 or 41.56? A good guide to follow when round- ing a 5 is always to round to the nearest even number. In this way, we eliminate any tendency to round in a fixed direction. In other words, there is an equal likelihood that the nearest even number will be the higher or the lower in any given sicua- Lion, Accordingly, we might choose to report the result as 41.56 + 0.07. I we had reason to doubs the reliability of the estimated standard deviation, we might report the result as 41.6 + 0.1 ‘We should note that é is seldom justifiable co keep more than one significant figure in the standard deviation because the standard deviation contains ervor as well. For certain specialized purposes, such as reporting uncertainties in physical constants in research article, it may be useful to Keep two significant figures, and there is cer- tainly nothing wrong with including a second digit in the standard deviation. How- ‘ver, ic is important to recognize that the uncertainty usually lies in the fist digi 6D-4 Expressing Results of Chemical Calculations ‘Two cases are encountered when reporting the results of chemical calculations. If the standard deviations of the values making up the final calculation are known, ‘we then apply the propagation of error methods discussed in Seetion 6C and round SDE Jones F Chem lu, 197, 49.758, DOL 101021fed0499753, “foe mare denon shi topic, see hepllwww.chem lyeda/caurses/che226/dawalead! (CL for_ signal SD Reporting Computed Date 117 {The number of significant figures inthe mantissa orthe digits to the right ofthe decimal point of a logarithm, isthe same as the rpumber of significant figures in the original number. Thus, log (a57 X 104) = 4.981 Since 957 hhas a significant figures there are digits tothe right of the decimal point in the result. In rounding a number ending in 5, always round so that the result fence with an even number, Thus 0.835 rounds to 0.64 and (0625 rounds to 062,48 CHAPTERS Random Ezzors in Chemical Analysis the results to contain significant digits. However, if we are asked to perform ealeu- lations where the precision is indicated only by the significant figure convention, common sense assumptions must be made as to the uncertainty in each number. ‘With these assumptions, the uncertainty of the final result is then estimated using Section 6C. Finally, the result is rounded s0 th the methods presence con- tains only significane digi Ie is especially important to postpone rounding until the calewlation is completed. ‘Ac least one extra digit beyond the significant digits should be carried through all of the computations in order to avoid a rounding error. This extea digit is sometimes called a “guard” digit. Modern calculators generally retain several extra digits that are noc significant, and the user must be careful co round final results properly so that only significant figures are included. Example 6-8 illus trates this procedure. EXAMPLE 6-8 A 3.4842-g sample of a solid mixture containing benzoic acid, CH,COOH 422.123 g/mol), was dissolved and titrated with base to a phenolphthalein end point. The acid consumed 41.36 ml. of 0.2328 M NaOH, Calculate the percent benzoic acid (HBz) in the sample. Solution [As shown in Section 13C-3, the caleulation takes the following form: mmol NaOH | I mmol HBr | 122123g HB feral) 5C 2228 el NSOH |, mmol Bey pZaIL Ze MEE GHB: see ae mL NaOH mmol NaOH “1000 mmol HBz _ 3.842g ample x 100% = 33.749% Since all operations are cither multiplication or division, the relative uncertainty ‘of the answer is determined by the relative uncertainties of the experimental data, Let us estimate what these uncertainties ate 1, The position ofthe liquid level in a buret can be estimated to 0,02 mL. (Figure 6-5) In eading the buret, (wo readings (initial and final) must be made so thatthe standard deviation ofthe volume s will be sy = V(0.02)? + (0.02)? = 0.028mL ‘The relative uncertainy in volume sy /Vis then su _ 0.028 1 = 9028 x 000 pp Vo 41.36 tales 68ppr 2. Generally, che absolute uncertainty of a mass obtained with an analytical balance will be on the order of 0.0001. Thus the relative uncertainty of she denominator ip/D is 0.0001 3.4842 X 1000pp¢ = 0.029 ppQuestions and Problems 119 3, Usually we ean assume that the absolute uncertainty in the concentration of a teagent solution is 0.0001, and so the relative uncertainty in the concenttation of NaOH, seis & _ 0.0001 fm ane % 1000pP¢ ABppe 4, The relative uncertaingy in the molar mass of HBz is several orders of magni- ‘ude smaller than any of che three experimental values and will not be significant. Note, however, that we should retain enough digit in the calculation so that the ‘molar mass is given co at least one more digit (the guard digit) than any of the experimental data. Thus, in the calculation, we use 122.123 for the molar mass (ore ae carrying two extra digits in this inseance) 5. No uncertainty is associated with 100% and the 1000 mmol HBz since these are exact numbers Subscicuting the chree relative uncertainties into Equation 6-12, we obtain 5 (eey «(ey (omy y \\ar36) © \3asiz) © \o2308 -V/(0.00068)" + (0.000029) + (0.00043)* = 802 X 10 x y=. 8,02 x 10 x 33.749 = 0.027 02 x 10-4 ‘Therefore, che uncertain inthe calculated result is 0.039% HB2, and we should sepor the resul a8 33.75% Hx, or beter 33.75 (+ 0.03)% HB fe must emphasize that rounding decisions are an important part of every cal & There's no relationship between, x phasias th ing ds portant pat of every ea the number of digits displayed ona computer screen or calculator and the true number culation. These decisions cannot be based on the number of digits displayed on an instrument readout, on the computer sereen or on a calculator display. of significant figures. WEB ‘The National Institue of Standards and Technology maintains Web pages of statistical data for WORKS | esting sftwate. Go 10 wwwccengage-com/chemistry/skoog/fue9, choose Chapter 6, and goto the Web Works, Here you will find a link wo dhe NIST Statistical Reference Datasets ste, Browse the ste co see what kinds of data ae available for testing, We use two of the NIST datasets in problems 6-22 and 6-23 a the end ofthis chapter. Under Databases, Scientific, choose Standard Reference Data Find the Analytical Chemisty databases, Ente the NIST Chemisty WebBook site, Find the gas chromatographic retention index data fo chlorobenzene. Find the four values for the retention index () of chlorobenzene on an SE-30 capillary column a a temperature of. 160°C. Decemine the mean setetion index and its standard deviation at this cempetatu. QUESTIONS AND PROBLEMS 6-1, Define 6-2, Differentiate beeween *(a)_ standard error of the mean, *() parameter and statistic (©) coefficient of variation, () population mean and sample mean, *(e)_vasiance *(0) random and systematic ertor. (2) sigoiticans figures. (@) accuracy and precision