O UN'VERS'TY F QUEENSIAND DEPARTMENT CHEM'SIRY OF

PRACTICAL PHYSICAL CHEMISTRY

T H ET R E A T M E N T OFDATA
(REVISED) EDITION SECOND Lgg6

By: T. t. QU,CI(ENDEN G. T. BARNES D. .!. T. H'LL

INTRODUCTION THE TREATMENT OF DATA

Introduction In physical chemistry most experiments involve quantitative measurementsusually intendedto yield the magnitudeof some property of a system and/or the quantitative relationshipbetweentwo (or more) properties.In order to extract the maximum amount of useful information frorn such measurements is often desirableto treat the data by it statisticalmethods,such as averaging determining line of best fit. At the sametime it or a is essentialo make an estimate the reliability of the result. t of In order to remove any possible misconceptions the following statementsshould be considered carefully. quantities 1. AU measured containerrors. 2. The bestvalueof everyphysicalconstant subjectto error. is 3. Every instrument,no matterhow precise, lwaysleadsto someerror of measurement. a 4. Every personwho operatesan instrumentis subjectto experimental error. Such errors are usually classifiedeither as errors of reading,as errors of setting, or as blunders, although thelatterdo not reallyform a distinctclass. 5. Thereare no such things as correct values,exact values,or true values,of physical quantities. constants r of any measured o 6. Repeated attempts to measure the same quantity always give different values. (However,you will notethat you,or the instrument, may truncate answers that there so are too few significantfiguresto show thesedffirences. This does not meanthat there are no differences.)The more preciselycontrolled and careful are the measurements, the smallersuchdifferences become. is, of course, aim of a scientisto reduce will It the t these differences (or errors) to a minimum, but they can never be completely eliminated. It is possible become quitephilosophical to aboutthe abovepoints.For instance, you may argue that a precise, error-free value must exist for each physical constant, quite independently our ability or inabilityto measureit. However, evenif precisevaluesdo of exist,thev are not determinable and arethereforenon-existent us. to Items J to 6 in this introduction have considerable relevanceto the performanceand presentationof a laboratoryexperiment.For example,as your experimentalresults will invariablybe subjectto error, some estimateof this error must always accompanyyour y final results.Withoutsuchan estimate. our resultsare worthless.

Anotherpoint to consider the phraseologyou usewhen writingthe Conclusion your is y to "My experiment.You will now realisethat phrases resultdiffers from the correct such as value";"My results arewrong"; "The correctvalueof -r was obtained"are all meaningless. Examplesof valid statementsare: "The results agree with literature valueswithin the estimatedexperimental errors"; "Repeated measurements ;r did not lie within the range of of the estimated error"; "Experimentaland theoretical values of x agreed within the estimated experimental error". Classification of Errors Errors are often classifiedas either random or systematic. 1. Random errors (sometimes termednoise)producea scatterin the experimental results. They vary both in magnitude in direction. and '2. Systematic errors can be subdivided into: (a) Constant rrorswhich alwayshavethe samemagnitude sign; e and (b) Drift errorswhichalwayshavethe samesignbut change slowlyin magnitude. These types of error can be illustrated by measurements the mass of water in a of pycnometer. ven under optimum conditionsreplicatedmeasurements show some E will variation. These variations are random errors. They could be due to air currents in the balancecase, slight variationsin the filling of the pycnorneter, mall variationsin the s temperature the waterand of the air in the balance of case,and so on. The meanreadingis also subject to a constant error due to the effect of buoyancy.This error is usually recognizedand an appropriatecorrectioncan be applied.Furthermore,if the temperature of the pycnometeris not the sameas that of the balancecasethere will be a systematic error which will changeas the two temperatures equalize. This would be a drift error. Sourcesof Error Random errors are readily apparentafter a number of measurements have been made. They can arise from lack of skill in the experimenter, imperfectionsin the apparatusor procedures, in other ways.(A commonexample sampling and would be errorscausedby small temperaturefluctuationscausedby an imperfect thermostat).Random errors call sometimes reducedby improvements the experiment, be in but they can never be entirell, removed.Their effect can be reduced taking a largenumberof measurernents using by and one of the variousstatistical methods averaging. of Systematicerrors are often very difficult to detect. They can arise from such factors as incorrectcalibration(e.g.of a pipette,a balance, a standard or solution);an unjustified theoreticalassumption(e.g. that a gas behavesideally); and incorrect use of apparatus (e.g. use of volumetricglassware a temperature at differentfrom the temperature which at it was calibrated).Sometimesa systematic error can be detectedif the samequantity is measured a differentexperimental by method.

Minirnising Errors As errors of measurement unavoidable, are two important decisionsmust be made when is Firstly,one must decidejust what magnitudeof any seriesof rneasurements planned. error is acceptable,and secondly what experimentaltechniquesand methods of data level. treatment will reduce errorto this acceptable the This principle is applicableto real-life situations, but in most laboratorycoursesthe situation is somewhat artificial in that the method of measurernent and the particular with the knowledgethat the experiment apparatus have alreadybeenselected beforehand yieldsan acceptable level of emor. As the random error in a mean decreasesby the square root of the number of (see measurements equation[4]), it is alwayspossible reducea randornerror to any to must be assessed requiredlevelby repetitive measurements. valueof this approach The by option of improving the considering the amount of time involved and the alternative method or apparatus. Although in the laboratorycourse you do not have the option of you improvements which would reduce improvingthe apparatus, can,and should,suggest the major errors in the experiment.Furthermore,measurements should be replicated whenever this seems advantaseous time allows. and Calculation of Best Values and Errors There are two quite different approaches the estimationof errors. In one approach,the to into a seriesof unit total error in a final result rs predictedby analysing experiment the operationssuch as the measurement mass,temperature, of current, etc. The greatest possibleerror in each operationis then estimated and theseerrors are propagatedthrough possible error in the final results.We will call this the calculations give the greatest to predictederror. error the The secondapproachis to determinethe observederror rn the result by examiningthe variability of the final resultover several replications. In the following sectionswe discussmethodsof determiningthe best value of a required quantity and of estimatingthe error in this value. The sectionsare arrangedin order of increasing complexityof thecalculations. SingleMeasurement Even though a quantitymay havea continuousrangeof valuesthe measuring apparatus is only capableof measuringit to the nearestscaledivision or estimatedsub-division.The interval betweenscaledivisionsor estimated sub-divisions known as the least count. is The accuracywith which a readingmay be made in thus determined the leastcount. It by is assumed that there is usuallyno difficulty in decidingwhetherthe readingis nearerthe divisionaboveor the divisionbelow (exceptnearthe midpointbetween divisions) that so + half the leastcount. possible the greatest error in the readingwould be

Repeated Measurements The Mean Value If the experimentinvolvesthe determination a singlequantity,the meanof a numberof of separate determinations will provide a better estimate than any of the individual is measurements. the numberof measurements small and each is equally reliablein the If opinion of the experimenter arithmeticmeanis usuallytaken: the
s N

+\

+ irj +.... "trN

Lxi
i-l

=-

tll

The existenceof other measuresof central where N is the number of measurements. tendencysuchas the median, the mode,and the geometricandhannonic meansshouldbe than the arithmeticmean. noted.In certaincasesone of thesemight be more appropriate The Range,w valueand the lowest.It givesthe simplest The rangeis the difference between highest the measurements may be used when up to and method of estimatingan error from repeated When large numbersof replicationsare availablethe five or six replicationsare available. A f'actor, should c, than other methods. conversion use of the rangeis much lessaccurate as: be appliedto the range(Table I ), andthe resultwould thenbe expressed i-tcw level) factors,c, for calculating error in a mean(95Vocanfidence the Table 1: Conversion ( from the range,w, of N readings I ).

2 6.35

7 0.33

B 0.29 0.26

t0
0.23

r.30

0.12

r 0.5

0.40

The Standard Error, s If a large numberof replicatedmeasurements made it is usuallyfound that the values are are distributedabout the mean in a particularway, known as the normal distribution.The spreadof any distributionis characterized the standarddeviation of the set of readings, by and this standard of deviationis in turn an estimate the standarderror of a singlereading. from the expressions: The (estimated) standard error,s, of a reading can be calculated

or moreconcisely:

t=1-*"

fi.",-,r

Lzl

readings, is their arithmetic where xl, x2,. . .xi, -rrp the individual &re x mean,and N is the total numberof readinss. It is worth noting that, particularly when the number of readingsis large, it is more formulato calculate quantity It"--r)' the convenient use an alternative to appearing above. This formulais:

It"-')'= Z"'-q4 N

t3l

is, The standarderror of the meanof N measurements of course,smallerthan the standard error of an individual measurement, furthermore,it decreases more measurements and as from the standard are incorporated.T'hestandard error, .e,,,, the meancan be calculated of by error, s, of an individualmeasurement the expression:

l4l
resultsit is customary use + /qss,nlimits.Thus the result to When reportingexperimental would be: ,r * /.n.s,, where /q5 is the value of Student's/ at the 957o confidence level for N-l degreesof freedom.These limits have a definite statisticalmeaningif a large number of results have probability beentaken.Thereis then a 95Vo that any one determination x will lie within of * /e5sn, the meanof a largenumberof determinations x. Furtherdiscussion be of of can (1) found in references to (6). Student / valuesat the 95Voconfidence level are given in Appendix1. Linenr Relationshins When the object of an experiment to studythe relationship is between two properties an attempt is usuallymadeto find functionswhich will make the relationship linear.Thus, for example, the vapourpressure a liquid is simplyplottedagainst if of temperature, curve is a obtained.However, when the logarithmof the vapour pressureis plotted againstthe reciprocalof the absolutetemperature, straightline results.This latter plot is much more a valuablethan the former for severalreasons: firstly it establishes functionalrelationship the

betweenvapour pressureand temperature, and secondlyit permits the relationshipto be formulatedmathernatically the constants be evaluated. and to However, even when the appropriate functional relationship has been found, the experimentalpoints will not lie exactly on the line as random errors will inevitably be present. The problemthen is to find the straightline which bestfits the data.This line is by line.lt can be represented the equation: known as a line of best or as a regression fit _y=a*bx on wherea is the intercept the y axis,andb is the slope. line to determinethe value of a and/or b, or the value It is common to use the regression the appropriate of .yfor a given value of x. Wheneverone of thesequantitiesis evaluated errormust alsobe estimated. Fitting by Eye if Fitting a straight line by eye is often satisfactory the data points are not too scattered. of straightedgerule facilitatesthe balancing points on either side The use of a transparent on line of the line. An improvedversionof this methoddepends the fact that the regression must passthroughthe point (t, t) . With this point fixed, the fitting of a straightline by eye is niucheasier. The error in the slope can be found by drawing in those lines of maximum and minimum for be slopewhich could reasonably mistaken the line of best e.g.In Figure I the slopeof the line of bestfit rs:2.0 + 0.2 are and Howeverthesemethods approximate shouldonly be usedwhen the scatteris small line. andthereis little doubtaboutthe positionof the regression t5l

Figure 1: Graphical determination of the error in the slope.

Fitting by the Method of Least Squares The best linear representation a set of experimental of data is obtainedby the method of least squares.In this method,it is assumed that the line of bestfit is the line for which the sum of the squaresof the deviations of the points from the line is a minimum. In the that the x valuesare without error, and that only treatmentgiven here it will be assumed from the line (seeFigure2). efforsin the y values cause deviations

"\ Figure 2: Deviations of experimental y values from the line of best fit. (xz, yz),(x:, .y:), (x', we Suppose haveon the grapha set of N pointswith coordinates _y'), (xN,)N). If we assume that the straightline: ... -I=a*b"r is the line of bestfit, thenit canbe shownby the methodof leastsquares that:

b-

NIrry-I"I,
N)r'()r)'

l6l

and that: 1 ,-. a=;(I." - uI")

where: It= xt+x2*...JN

UI

I,
Io

= ) r * ) z* . . . ) N
= x r . y-rf x z r z i . . . x N l N

IrI.n

= ( x , x z * . . . x " ) ) r* ! z * . . .- ) * ) * (

Z"'=xl+xl+....ri
(Ir)t = ( x ,* x , + . . . x , u ) 2

data can be determinedfrom the correlation How well a straightline fits the experimental cofficient, r, which is givenby:

_ ..

Jt*I,'- (I")'l INIy'-(Iy),]

NI".v_ I,Iy

t8l

The correlation coefficientmust lie between+l and -1. If all of the data lie exactly on a or straight line, r will havea valueof + I if the slopeis positive, - 1 if the slopeis negative. of A correlationcoefficientof 0 indicates that the data exhibit no linear dependence y on x. In practice,r usuallyhasfractionalvaluesand the closertheseare to +1, the betterthe dataflt a linear relationship. valueof y is given by: error, sr, of a measured The (estimated) standard

tel
where:

I(ay)'=I(y-,y"nr")2 = Ity-(a+bx)12
= [ ] , - ( a + b x , ) l t+ [ , y ,- ( a + b x r ) ) t+ . . . + [ y " - ( u * b x * ) l t purposes following formulais convenient: For computational the

= I(oy)' I),'-

(Ir)' LI'-t -- I
N

\i rl.y

I,
tl0l

line The standard error.su. theslope the least-squares is givenby: of of

t - N sh:sr!.t;{tT
The standarderror, ,ro, the interceptis: of

[11]

112)
The standarderror, sr, of a value,y0, calculated fiom the fitted line as the value of y when
,t = -fn iS:

s"
Jyu =

N-+

s;(/o - x ) -

1.

t13l

As before,the error limits areusually set + /.ss + /.es or t /.e5 where /.e5 the value J'u, sa, syo is of Student's/ at the 95Voconfidence level for N-2 degreesof freedom (note N-2 instead of N- I when usinglinearleastsquares). N on-Linear Relationship s If a relationship into a linearform it is nevertheless possibleto fit a cannotbe transformed curve to the experimentaldata by methods of non-linearleast squaresanalysis.These techniques beyondthe scope the present are of study. Propagated Errors In order to predict the error in a final result it is often necessary carry the errors in to individual quantities through a complex calculation. This process is termed the propagation of errors. e.g. The pressure exertedby n moles of an ideal gas at temperatureZ and volume V is given by: P = nRTN. The error in P can be obtained by propagatingthe individual errorsin T, V, and R throughthis equation. The rules for propagationare given below, and are followed by a brief discussion their of theoretical basis. Stricfly speaking,these rules are valid only for propagating randun errors. They can be used for propagatingsystematicerrors,but great care must then be taken with signs,so as to ensurethat systematic errors which should cancel are not accumulated. Theserules are rigorous only for the propagationof maximumerrors. If statisticalerrors, such as standarderrors. are propagated theserules should be modified. However for the presentpurposesthey will be consideredadequatefor propagatingboth types of error. The processof propagationcan often be greatly simplifiedif someof the errors are small in comparison with the others. suchcases In only the major errorsneedbe propagated. Rulesfor Error Propagation (a) Addition and Subtraction:The absolute error in the sum or differenceof two or more quantities the sum of the absolute is in the quantities. errors

10

e . s . ( 1 0 0 1l ) + ( 5 0t 2 ) - ( 1 2 +2 ) = 1 3 8 + 5 (b) Multiplication and Division: Here it is proportional (not absolute) error that is propagated; by by The percentage this is measured a fraction or equivalently a percentage or fractionalerror in the product or quotient of two or more quantitiesis the sum of the percentage fractionalerrorsin the original quantities. or e.g. * = ( 1 0 0 +1 ) x ( 5 0 + ) 2 (nt02)

Becansepercentageerrors are required by this rule, the absoluteerrors given are first errors: convertedto percentage

(qetx6.7) otu.,-t28=(4.2t0.3) 10' x e.s. x = 416.7tlff )=

that there is no point in of Note that the uncertainty 0.3 in the answerof 4.2 indicates retaining figures beyond the second significant figure. You will note also that the experimentalerror need not be estimatedto better than one or at most two significant figures. (c) Logarithms: The absoluteerror in the logroof a numberequalsthe fractionalerror in the numberitself,dividedby 2.303. l o e . g . y = 1 6 9 ( 2 0 + 0 . 8 ;l= g ( 2 0 + 4 V o ) =. 3 0 t O t I f

( +

" OO Lffi)

| )

which we can truncateto sive: + x = 1 . 3 0 -O . O 2 (ln) areused,the divisor2.303is omitted. If NAPIERIAN logarithms TheoreticalBasis of the PropagationRules If The above rules may be derivedby meansof partial differentiation. ttre final result, z, is quantities,.tr, 4, . . ., then: calculated from severalindependent 1y', u: f (x,Y,2,"') For small changes x,y, Z, . . ., represented 6x, 61r, . . . the correspondir-rg in by 6i, change i n u , 6 u ,i s g i v e n y : b

- (E,r)^ far)^ (ar) 8 r r = l -l 6 x +. l l 6 u + l . 1 6 z + . . .

\drl ldy/ \dzl

[l4l

1l If thesesmallchanges,6x,5y,5z, . ., areequated the maximumerrorsin the respective . to quantities, and the sign of each error is chosenso that each term is positive,then 6a becomes maximumerrorin a. Thus: the

,d,'.' l ( a ro*.* l ( a ,or.'-l ^ ,= )^I ) ^ |*

| |l;J

I.l[a,1

Il5l

give thepropagation ruleslistedabove, Simpleapplications this relationship of h) Addition and Subtraction:If: u=x+y-z then:

(au'\ /drz\ /ar) l.l=lll.l=lll.l=-l

\dx / \dyl \dz / hence,for the maximum error | 6,l: r - r r ^ t t ^ t t - l

l6rrl=l6xl+l6vl+l6zl

[16l

i.e. the absoluteerror in the final result is equal to the sum of the absoluteerrors in the quantities. measured (b) Multiplication and Division: If: -{x y

then: y^ x^ -ry^ ^ drl=-dr+-dv,dz z z z Dividing throughoutby z givesthe fractionalerror in the final result:

6rz l s x ll a v ll s z l l - l + l : y + l - lr l u t x tI | | I

tlTl

The samerule for percentage errors follows on multiplyingby 100 on both sides,Thus the fractionalor percentage error in u is equalto the sum of the fractionalor percentage errors quantities. in thecomponent

12

(c) Logarithms:If: u =lnx then:

^ 6 du= -

tl8l

Thus the absoluteerror in a ecualsthe fractionalerror in x. Alternative Method of Propagctting Errors When errors must be propagatedthrough a complicatedexpression is often simpler to it use an alternativeprocedure.This involves performing the calculationof the final result two additional times:onceusingthoseextreme values eachvariable of which will produce the highestresult, and once using those extremevalueswhich will produce the lowest result. Presentation of Results Nurnerical results In presenting numericalresultsit is necessary the correct numberof significantfigures that be used, that rounding off be carried out correctly, that the reliability of the result be given,and that the correctunits be used. SignificantFigures If the reliability of the resulthasbeencalculated one of the methodsoutlined aboveit is by a simple matter to decidewhich figures are doubtful. Llsuallyonly one doubtful figure is retained, occasionally retention two is justified.For example, result8.1 + 0.2 but the of the cm meansthat the best value most probablylies between8.5 and 8.9 cm. Thus even thoughthe 7 is doubtfulit mustbe retained. the result8.75 + 0.2 cm the 5 alsoserves In a purpose. resultcould, with equaljustification, roundedoff to either8.7 or useful This be 8.8 cm, but neither8.1 + 0.2 cm nor 8.8 + 0.2 cm properlyexpresses range within the which the bestvalueprobablylies. Where two doubtful figures are retainedthe secondis often placed in smaller print or dropped belowthe line. + + e . g . 8 . 7 s 0 . 2 c m ,o r 8 . 7 s 0 . 2c m . It is importantto rememberto includea zero wherethe zero is a significantfigure. e.g.8.0 + 0.2 cm, not 8 t 0.2 cm.

l3 When a result is to be used in a later calculationit is advisable retain one more to roundingoff errors. significantfigure than usualto avoid accumulating Rounding Off Numbers When a number is being rounded off the last retaineddigit should be increased one by when the succeeding digit is greaterthan 5 or when it is 5 and there are other non-zero digit is 5 with no otherdigitsto its right or with digits to the right of the 5. If the discarded by only zeros to its right the retaineddigit is increased one if that would make it an even number,or left unalteredif it is alreadyan evennumber.If the discarded digit is lessthan 5 the lastretained digit is unaltered. e.g. 8.76 8.751 8.15 8.65 8.651 8.13 becomes becomes becomes becomes becomes becomes 8.8 8.8 8.8 8.6 8.7 8.1

The rule for rounding off fives is designedto avoid the accumulationof rounding off calculations. numbers usedin subsequent are errorswhen rounded-off Graphical Results of Results which show the functional dependence one variable on another should be presented graphically. The original data should be plotted and if this data can be a datashouldalsobe given. transformed into a linear relationship plot of the transformed pointson the graph, It is importantthat error barsbe placedon at leasttwo representative points may be and preferablyon all points (seeFigure l). The error bars on intermediate estimated interpolation. by Graphsshouldbe labelled, curvesandpointsproperlyidentified, the axeslabelled and with the nameof the variableand its units. Scales shouldbe selected that it is possibleto so show the error barsor the scatterof the points.The origin of the graphdoesnot haveto appear the page. on References (McGraw-Hill, New York, 1. WLSON, E.B.: An Introduction to ScientificResearch, 1952),Chapters8 and 9. 2. BAIRD, D.C.: Experimentation: An Introduction to Measurement Theory and Experiment D esi gn, (Prentice-Hal EnglewoodCl iffs, 19 62). I, 3. LARK, P.D., CRAVEN. B.R., and BOSWORTH, R.C.L.: The Handling of Chemical Data, (Pergamon. London, 1968),Chapter and4. 3

t4 4. DANIELS, F., et al.: Experimental PhysicalChemistry, ed., (McGraw-Hill, New 6th York, 1962),Chapter17. 5. SCHOEMAKER, D.P., and GARLAND, C.W.: Experimentsin Physical Chemistry, (McGraw-Hill,New York, 1961),Chapter. 2. ( 6 . D E A N , R . B . ,a n dD I X O N , W . J . :A n a l .C h e m . , 2 3 , 6 3 61 9 5 1 ) .

t5

Appendix I
Student / Values at the 95%oConfidenceLevel

/ Value 1 2 .1 1 4.303 3 .1 8 2 2 .1 1 6 2.51| 2.441 2.365 2.306 2.262 2.228 2.201 2.119 2.160 2.145 2.131 2.086 2.042 2.009 1.984

Degrees Freedom of 1 2
-t

4 5 6 1 8 9 10
ll

t2

r3
t4 l5 20 30 50 100