Examination of the Relationships Between

Original, Real and Apparent Extracts,

and Alcohol in Pilot Plant
and Commercially Produced Beers
Anthony J. Cutaia1, Anna-Jean Reid2 and R. Alex Speers2,3

ABSTRACT As was the practice of the period, little discussion,

justification or assumptions were presented in these
J. Inst. Brew. 115(4), 318–327, 2009
reports. Nor was any analysis of possible errors inherent
The historical development of equations used to relate alcohol in these formulas undertaken. Viewed from our current
and real extract to apparent extract and original gravity, as well perspective, these omissions may be considered serious
as ratios between the corrected Real (RDF) and Apparent De-
criticisms. One should remember, however, that statistical
grees of Fermentation (ADF), were examined in light of modern
polynomial and non-linear regression techniques. Comparisons routines and calculators were essentially non-existent. For
were performed using an extensive data set of 532 brews ob- instance, regression techniques were most certainly not
tained from commercial and pilot fermentations with statistical well known until some time after Pearson’s publication of
error analysis of these empirical relationships. New predictions 189623. In any case, without access to a personal com-
of apparent extract were calculated as a function of alcohol and puter, any relationships would have been derived via tedi-
real extract analogous to the Improved Tabarie’s formula. In ous graphical techniques. The tables produced by authors
addition, attempts at improving Balling’s original equation such as Balling2, Wahl and Henius27, Holzner13, and
model estimating original gravity from alcohol and real extract Pawloski and Doemens22 relating OE, AE, RE and Aw/w
are detailed and discussed. The statistical analyses of relation- involved an almost unbelievable amount of calculation. In
ships between Aw/w (alcohol by weight) and functions of OE
the course of this study we examined Table III of
(Original Extract), AE (Apparent Extract) and RE (Real Extract)
as well as ratios between the corrected Real and Apparent De- Holzner13 which tabulates 6252 predictions of real extract
grees of Fermentation (RDF/ADF) are reported. It is expected and alcohol content for given original and apparent ex-
that this paper will be useful for brewers to more accurately tracts. This massive effort was obviously done by hand!
estimate Aw/w and real extract values. The most well known expression relating OE, RE and
Key words: ADF, alcohol, apparent extract, brewing calcula- Aw/w is the Balling equation2 which relates the original
tions, original extract, RDF, real extract. extract to the real extract and alcohol (% w/w):
OE = 100 * (2.0665 * Aw/w + RE) /
INTRODUCTION (100 + 1.0665* Aw/w) (1a)
As early as 178425, almost a century before the time of This relationship is an approximate one as noted by
Balling2, brewers were interested in and reported on the Cutaia4:
density relationships between wort and fermented Two major assumptions are used in the deri-
beer12,13,22,27. Most of these researchers were concerned vation of Balling’s formula. The first assump-
with the relationships between wort and beer density and tion is that 0.11 g of carbohydrate is converted
alcohol content and published their work in German prior to yeast mass for each gram of ethanol pro-
to 1933. One notable exception was a chapter in the first duced in fermentation. The second assump-
edition of the Handbook of Brewing by Weissler28 that tion is that all fermentable dissolved wort sol-
reported the approximate relations between original ex- ids are monosaccharides. This assumption is
tract (OE, °P), apparent extract (AE, °P), real extract (RE, implicit in the composition of the Balling con-
°P) and alcohol, on a weight/weight basis (Aw/w)22. stant…
While relied upon by the brewing industry, it should be
emphasized that the Balling equation is an approximation
1 Science Source Consulting, Ballwin, MO.
violating various theoretical assumptions and is a “work-
2 Food Science and Technology, Dalhousie University, 1360 Barring- ing estimate” only. For instance, it is a common fallacy
ton Street, D401, Halifax, NS B3J 2X4 Canada. that Balling’s formula can be used to back-calculate a
3 Corresponding author. E-mail: [email protected]. diluted beer’s original (undiluted) extract. In fact, a back-
calculation of the original extract knowing dilution rates,
Publication no. G-2009-1211-1042 Aw/w and RE always gives a larger estimate of OE than the
© 2009 The Institute of Brewing & Distilling true undiluted value.

318 JOURNAL OF THE INSTITUTE OF BREWING

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Neilson et al.21 provided a critique of Balling’s formula and Real Degree of Fermentation (RDF)1:
and listed six effective but minor objections to the for-
RDF = [100 * (OE – RE) / OE)] *
mula. Surveys by the Carlsberg brewery in 1942 and 1971
[1 / (1 – .005161 * RE)] (8)
showed the formula overestimated OE by 0.221. Similarly,
with a limited range of data, Navarro et al.18 reported the The correction term in equation 8 ‘[1 / (1 – .005161 *
Balling formula fit the data with a coefficient of determi- RE)]’ was introduced in 1979 and corrects for mass lost
nation of 0.995 and a residual mean square of 0.032. They by CO2 and yeast uptake during fermentation5. The ratio
also reported on alternate formulas, which improved the of RDF to ADF is also mentioned in many texts:
prediction of original extract with their limited 36 fermen-
tation dataset. A recent survey by Neilson et al.21 of 50 RDF = ADF * Q† (9a)
fermentations (with a OE specification of 14.3°P) indi- †
Note the factor Q should reflect the corrected RDF
cated the Balling formula underestimated the true OE
value. Of course the original definitions of RDF (and thus
when applied to fermenting worts with developing Aw/w equation 9a) did not account for the ‘[1 / (1 – .005161 *
levels of up to ~3.66 and overestimated the true OE by RE)]’ correction term. It was noted that Balling first pro-
0.1°P at the cessation of fermentation.
posed calculating the ratio of RDF/ADF20 and it is termed
There are also other calculations in the literature that ‘Balling’s shorter formula’. The uncorrected form of fac-
include22: tor Q† (i.e., [(OE – RE) / OE] / [(OE – AE) / OE] or RDF
uncorrected/ADF) has been often cited in the past as
OE = [(RE – AE) / q] + RE (2a)
0.8196 or 0.8128. Kunze16 and Neilson and Erdal20 both
reported a value of 0.81 and attributed this value to a pub-
Aw/w = a * (OE – AE) (3a)
lication by Balling. It is noteworthy that the uncorrected
value of Q† (i.e., RDF uncorrected/ADF) is equal to “1 /
Aw/w = b * (OE – RE) (4a)
(q + 1)”, q being a constant introduced in equation 2a that
was reported to be dependent on original extract as shown
Aw/w = c * (RE – AE) (5a) in Table I22.
Given the development in brewing sciences and ad-
These equations were apparently taken from Balling’s
vances in our chemical and statistical analyses, one might
original work2 as they use the German phrase “nach argue it is well overdue to re-examine the brewing values
Balling” that can be translated as “according to Balling”. and relationships reported in the 1800s. This paper will
Confusingly, however (at least to the English reader), the
report on the relationships between these brewing values
coefficients q, a, b and c of equations 2a–5a (hereafter and various brewing parameters and comment on the in-
referred to as “brewing values”) in Balling’s 1865 paper herent error when using these brewing formulas.
(Table IX p. 243) are different than those reported by later
authors6,12,22. Interestingly, both the brewing values listed
by Balling2 and Pawloski and Doemens22 were reported to Table I. Empirical values published by Pawloski and Doemens22 and
Weissler28.
vary linearly with OE (Table X, p. 251). Table I lists the
Pawloski-Doemens values which agree with Weissler28. OE (°P) q (%/°P) a (%/°P) b (%/°P) c (%/°P)
Curiously, the values for q published by Pawloski and 01 0.221 0.3983 0.4864 2.2010
Doemens in 193222 were first presented by Holzner in 02 0.222 0.4001 0.4889 2.2024
187712. It is worth noting that that small brewers and 03 0.223 0.4018 0.4915 2.2041
04 0.224 0.4036 0.4941 2.2058
home brewers can seldom measure real extract or alcohol 05 0.225 0.4054 0.4967 2.2076
content easily and thus formulas containing only OE and 06 0.226 0.4073 0.4993 2.2096
AE values are more than curiosities. 07 0.227 0.4091 0.5020 2.2116
Furthermore, an expression analogous to equation 3, 08 0.228 0.4110 0.5047 2.2137
09 0.229 0.4129 0.5074 2.2160
but using specific gravity rather than extract values, is leg- 10 0.230 0.4148 0.5102 2.2184
islated for use by small brewers in Great Britain to calcu- 11 0.231 0.4167 0.5130 2.2209
late alcohol levels on a volume/volume basis (Av/v) for 12 0.232 0.4187 0.5158 2.2234
excise purposes11: 13 0.233 0.4206 0.5187 2.2262
14 0.234 0.4226 0.5215 2.2290
Av/v = f * (OG – PG) (6) 15 0.235 0.4246 0.5245 2.2319
16 0.236 0.4267 0.5274 2.2350
where OG and PG are the original and present (apparent) 17 0.237 0.4288 0.5304 2.2381
gravities multiplied by 1000 and f is a factor ‘mandated’ 18 0.238 0.4309 0.5334 2.2414
19 0.239 0.4330 0.5365 2.2448
to vary from 0.125 to 0.135 as (OG – PG) varies from 6.9 20 0.240 0.4351 0.5396 2.2483
to 100.7. Other estimations using specific gravity units 21 0.241 0.4373 0.5427 2.2519
have been reported by early British researchers10,26. One 22 0.242 0.4395 0.5458 2.2557
simple estimation of Av/v is reported by multiplying by the 23 0.243 0.4417 0.5490 2.2595
24 0.244 0.4439 0.5523 2.2636
drop in specific gravity by 10214. 25 0.245 0.4462 0.5555 2.2677
Aside these relationships, there are the relatively well- 26 0.246 0.4485 0.5589 2.2719
known factors of Apparent Degree of Fermentation 27 0.247 0.4508 0.5622 2.2763
(ADF): 28 0.248 0.4532 0.5656 2.2808
29 0.249 0.4556 0.5690 2.2854
ADF = 100 * (OE – AE) / OE (7) 30 0.250 0.4580 0.5725 2.2902

VOL. 115, NO. 4, 2009 319

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table IIa. Polynomial fit of specific gravity function (SG-1) and apparent extract (AE) values of ASBC Table 1a.
Statistical value Eq. 10 prediction of SG Eq. 11 prediction of AE Eq. 12 prediction of AE
Dependent variable SG-1 AE AE
Coefficients1
b0 1.3084 × 10–5 –4.63374 × 102 -
b1 3.86777 × 10–3 6.68723 × 102 2.56899 × 102
b2 1.2745 × 10–5 –2.05349 × 102 6.7126
b4 6.3 × 10–8 - –1.44816 × 104
b5 - - 5.17578 × 105
b6 - - –1.07464 × 107
b7 - - 1.3011 × 108
b8 - - –8.50786 × 108
b9 - - 2.3231 × 109
Adjusted multiple R2 0.999 0.999 0.999
Standard error of estimate 2.36 × 10–6 1.59640 × 10–3 2.0098 × 10–4
ab are the intercept, linear, squared, and cubed terms, and so on. n = 1661.
0,1...9

Table IIb. Polynomial fit of specific gravity (SGOH) and alcohol values
DATA COLLECTION AND ANALYSIS of ASBC Table 2a.

Data collection Statistical value Eq. 13 prediction of SGOH

Dependent variable SGOH
In the course of this investigation, five brewing data- Coefficients1
sets were obtained. The first two data collections were b0 1.000011541
from a pilot brewery (sets RPB-I and RPB-II) of 92 and b1 –1.92647 × 10–3
112 fermentations respectively. A third dataset was a ‘pro- b2 3.0266 × 10–5
prietary’ set of 71 commercial lager fermentations. The Adjusted multiple R2 0.999
Standard error of estimate 1.0165 × 10–5
first three data collections contained directly measured
ab , b1 and b2 are the intercept, linear, and squared terms. n = 501.
values of original extract, apparent extract and alcohol (% 0

w/w). A fourth ‘CMBTC’ dataset consisted of 221 directly

measured values of original extract, (apparent) specific tions, colinearity and variance inflation factors were not
gravity (SG), and alcohol (% v/v) from pilot scale fermen- considered29.
tations. These were obtained from the Canadian Malting The empirical relationship between original gravity, al-
and Brewing Technical Centre, Winnipeg, MB. This labo- cohol (w/w) and real extract was determined for equations
ratory routinely produces pilot beers in the course of eval- in Balling’s original format and related variations for the
uating malting barley. While measurement methods were combined datasets and each subset individually. The non-
proprietary, RPB-I and RB-II dataset values were derived linear curve-fitting statistical package used was STATIS-
with a SCABA™ instrument and an Anton-Parr™ appara- TICA (StatSoft, Inc., Tulsa, OK). Comparisons of statisti-
tus was used to collect the Proprietary and CMBTC data- cally significant differences between data subset parame-
sets. A fifth dataset was formed from the published values ters were performed using a Monte Carlo bootstrapping
of Navarro et al.18 (original, apparent and real extracts and technique7 implemented on Microsoft’s EXCEL spread-
alcohol levels % w/w). While Navarro et al. reported 142 sheet package and the @RISK package (Palisades Corp.,
data points, they were in fact replicates of 36 fermenta- Ithaca, NY). This method was employed to eliminate spu-
tions measured using European Brewing Convention pro- rious results due to non-normal distributions of the esti-
cedures. Thus, the replicate values were averaged for each mated subset parameters. In this instance, distributions of
of the 36 fermentations resulting in 36 values for the ori- 10,000 differences between all possible subset pairs of
ginal, apparent and real extracts, as well as the alcohol parameter estimates were generated by random sampling,
level (assumed to be % w/w). All the values collected with replacement of all subset dataset observations. By
were measured prior to any post-fermentation dilution. inspection of the resulting distributions, it was determined
It is noteworthy that during the course of data collec- whether zero (indicating no difference in parameters) was
tion, it became apparent that the direct measurement of contained between the 2.5% and 97.5% region of the dis-
the real extract is now very rarely undertaken. Only one tribution of differences. If indeed it was not, then a statis-
dataset18 contained measured real extract values8. The tical difference at the 95% confidence level was deter-
correct calculation of the real extract values from directly mined for the indicated parameter estimate comparison.
measured alcohol and apparent extract levels is discussed
later in this paper. Relations between specific gravity and extract
Data analysis (as °Plato, ASBC Table 1)
In addition to the analysis of the datasets discussed In order to compare and recalculate constants in the
above, the relationships of °Plato (AE) versus specific brewing equations, accurate relationships between °Plato
gravity (sugar solutions) and specific gravity versus alco- (extract) and specific gravity were required. Relationships
hol (w/w) in ASBC Tables 1 and 21, real extract and fi- between these two values have been reported in the past
nally alcohol estimations were determined using the Sys- by Seibert24, Lincoln17 and Hackbarth9. Seibert24 reported
tat 11 package (Systat Software Inc., San Jose, CA). As a third degree polynomial that related specific gravity to
the purpose of these fits was to develop predictive equa- Plato values using selected data from ASBC Table 11.

320 JOURNAL OF THE INSTITUTE OF BREWING

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Lincoln reported a calculation for Plato values as a Calculation of real extract
function of specific gravity equivalent to a second-degree
polynomial17. Hackbarth9 reported the use of an eighth While the measurement of real extract is simple, it is
degree polynomial to also calculate for Plato values as a tedious and now rarely undertaken in brewing laborato-
function of specific gravity. ries. Rather, it is calculated by formulas derived by either
To update the calculations of Seibert24 and Lincoln17, Balling or Tabarie3. Recently, a new calculation was re-
third and second degree polynomials were fit to all the ported of an ‘Improved Tabarie’s formula’ which relates
values from ASBC Table 1, using a forward stepwise pro- the specific gravity of beer to the specific gravity of the
cedure (P > 0.15 to include). A form of Hackbarth’s for- real extract and the alcohol19:
mula, an eight-degree polynomial relating the apparent SG = SGRE + [SGOH – 1 +2.96 * (1 – SGOH)2 – 381 *
extract to the specific gravity less 1.0 (i.e., [SG-1], [SG- (1 – SGOH)3] (15a)
1]2 ... [SG-1]8) was also examined using a polynomial
regression model. The best-fit results are shown in Table or
IIa and can be represented by the following equations: SGRE = SG – [SGOH – 1 + 2.96 * (1 – SGOH)2 – 381 *
SG-1 = 1.308 × 10–5 + 3.868 × 10–3 * extract (1 – SGOH)3] (15b)
+ 1.275 × 10–5 * extract2 The SGReal value can first be converted into gextract per
+ 6.300 × 10–8 * extract3 (10) 100 g (of water) via equation 12 and then to RE via ASBC
Extract = –4.6337 × 102 + 6.6872 × 102 * SG method 5a:
– 2.0535 × 102 * SG2 (11) RE = [(gextract / 100 gwater) * SGReal] / SG (16)
Extract = 2.569 × 10 * (SG-1) + 6.7126 × 10 *
2 0
It is worth noting that the Tabarie or ‘Improved’
(SG-1)2 – 1.4482 × 104 * (SG-1)3 Tabarie formula involves conversion of Av/v (not Aw/w!) to
+ 5.1758 × 105 * (SG-1)4 – 1.0746 × 107 * SGOH and that SGRE calculated from that value is by vol-
(SG-1)5 + 1.3011 × 108 * (SG-1)6 ume.
– 8.5079 × 108 * (SG-1)7 + 2.3231 × 109 * However, importantly and most recently, it was noted9
(SG-1)8 (12) that neither of these relationships account for ethanol-su-
The coefficients for the specific gravity prediction dif- crose interactions and a routine was developed that ac-
fer slightly from those of Seibert24 presumably due to ma- counted for these interactions. Using this routine9, the
chine error of the number of significant figures that could dependency of real extract on Aw/w and AE was deter-
be carried by the machine calculators of the time. Our mined by varying these values at 0.5 intervals from 0–7%
regression of apparent extract as a function of specific Aw/w and 0–10°P respectively. A response surface-type
gravity confirms previous equations. The standard error of function was then developed to predict the real extract
the estimate of these equations equal 2.36 × 10–6 for equa- when the alcohol and apparent extract values are known:
tion 10, 1.59 × 10–3 for equation 11 and 2.01 × 10–4 for RE = 0.496815689 * Aw/w + 1.001534136 * AE
equation 12. Considering the adjusted multiple R2, the – 0.000591051 * Aw/w * AE – 0.000294307 * AE2
standard error of the estimate and the residual ranges, the – 0.0084747 * Aw/w2 + 0.000183564 * Aw/w3
most exact formula to estimate extract from specific grav- + 0.000011151 * AE3 + 0.000002452 * Aw/w2 *
ity is obtained with equation 12 as noted in Table IIa. AE2 (17)
These relations allowed calculation of apparent extract
values from the specific gravity values for the CMBTC The function estimated the prediction of the real ex-
dataset using equations 12 and 10 respectively. tract9 over the range of Aw/w 0–7% and AE 0–10°P. The
Another quadratic expression relating the specific residual error of the function was below ±0.0015°P for all
gravity of ethanol solutions (SGOH) to a given alcohol but 6 (of 278) cases and was never greater than
level (Aw/w) from Table 21was developed: ±0.0035°P. Since one of the objectives of this study was
to examine Balling’s formulas (i.e., equations 1–5 and 9)
SGOH = 1.1541 × 10–5 – 1.9265 × 10–3 * Aw/w equation 17 was applied to calculate real extract values in
+ 3.0266 × 10–5 * Aw/w2 (13) appropriate datasets.
The residual error arising from the use of this equation Aside from the measurement of real extract leading to
ranges from –0.00003 to 0.00003 for prediction of SGOH Eq 17 by Hackbarth9, the authors were unable to find lab-
(equation 13). The corresponding regression analysis is oratories that currently measure real extract by distilla-
presented in Table IIb. tion. In our collection of data only one report18 presented
measured OE, Aw/w, AE and RE values. A comparison of
Calculation of alcohol (w/w basis) these 36 measured and predicted real extract values18 (via
In order to convert alcohol to a percent weight/weight equation 17) gave an error range from –0.02 to 0.05°P and
basis from alcohol in a percent volume/volume basis for a standard deviation of these residuals of 0.019°P.
future calculations, we used (beer) specific gravity and Final datasets
alcohol level values (Av/v) and the formula of Weisler28
and the exact conversion factor of the International Bu- To summarize, the use of the above calculations re-
reau of Legal Metrology15: sulted in four datasets of 92, 112 and 71 fermentations
and a CMBTC dataset containing the results of 221 fer-
Aw/w = Av/v * 0.7907 / SG (14) mentations with measured original extract, apparent ex-

VOL. 115, NO. 4, 2009 321

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
tract and alcohol (w/w) values and calculated real extract RESULTS AND DISCUSSION
amounts. A fifth dataset of Navarro of 36 fermentations
contained the means of four measured original, apparent Functional modelling of original wort extract
and real extract and alcohol (Aw/w) levels. Table III shows estimation
the minima, maxima and average values of original and Using the first five datasets, it was possible to deter-
apparent extracts, alcohol (Aw/w) and real extract values mine the accuracy of the Balling equation constants.
for the five datasets. Figure 1 shows the distribution of the Balling’s original equation (equation 1a) was based on an
datasets as a function of their real extract and alcohol lev- implied functional relationship between alcohol (Aw/w)
els. and unfermented solids (RE or real extract) remaining in

Table III. Original extract, alcohol (% w/w), calculated real extract and apparent extract values used in this study.
Dataset Original extract (°P) Alcohol (% w/w) Real extracta (°P) Apparent extract (°P)
RPB-I, n = 92
Minimum 11.62 3.04 3.92 1.24
Maximum 15.11 6.04 9.28 7.86
Mean 14.08 4.72 5.28 3.11
RPB-II, n = 112
Minimum 12.14 4.09 1.99 –0.42–
Maximum 15.16 6.27 5.59 3.49
Mean 14.54 5.14 4.88 2.53
Proprietary, n = 71
Minimum 17.09 6.24 5.07 2.06
Maximum 18.03 6.79 5.67 2.77
Mean 17.67 6.56 5.38 2.45
CMBTC, n = 221
Minimum 09.44 2.94 3.10 1.26
Maximum 17.12 5.99 6.42 4.08
Mean 11.34 3.82 3.80 2.01
Navarro, n = 36
Minimum 09.51 3.16 3.27 1.79
Maximum 16.49 5.59 6.14 3.85
Mean 12.75 4.21 4.74 2.79
a Calculated by equation 17.

Fig. 1. Distribution of alcohol and real extract values for the five datasets used in this study.

322 JOURNAL OF THE INSTITUTE OF BREWING

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
the product after fermentation, the mass lost to carbon incorporated. This coefficient is, in itself, an approxi-
dioxide and yeast, and the solids content of the original mation since it strictly should be corrected to a wort basis.
wort (OE): The fixed factor of 0.9565 in the equation denominator is
the molar ratio of CO2 produced (and lost) for each gram
OE = 100 * (2.0665 * Aw/w + RE) / of ethanol generated. This molar factor was used since the
(100 + 1.0665* Aw/w) (1a) fermentable mass converted to alcohol is strictly used for
His empirically determined conversion coefficient of this purpose and fermentable mass not producing alcohol
2.0665 used in the numerator of his equation indicated the is accounted for by the second coefficient.
grams of fermentable material necessary to produce one In equations 1a–1c, it was implicitly assumed that eth-
gram of ethanol. The 1.0665 coefficient used in the de- anol and other losses such as moisture were non-existent,
nominator of equation 1a represented the mass in grams that the utilization of mass to produce other volatile me-
of fermentable material lost to carbon dioxide and yeast tabolites is insignificant, and that the product is measured
mass increase (2.0665 grams of fermentable material in undiluted.
original wort minus 1 gram of ethanol remaining in fer- In these fitted models, the data were initially subjected
mented product). to non-linear curve fitting procedures with all sets pooled
Two other variations on the Balling equation as dis- as a single population. Calculation of root mean squared
cussed below were also examined. Table IV shows the fit error, mean deviation of model estimate from observed
of equations predicting original extract values (OE) using original extract and explained variation (r2) was per-
the original Balling equation and the two variations to the formed for the pooled data and for the individual datasets
five datasets. The models shown in Table IV were derived where appropriate. In Eq 1c, however, the non-linear fit
from the following equations: analysis for the individual datasets determined individual
dataset yeast mass uptake coefficients, (k2), using a
OE = 100 * ( k1 * Aww + RE) / dummy variable technique while simultaneously employ-
(100 + (k1 – 1) * Aww) (1b) ing all datasets for conversion coefficient estimation, (k1).
OE = 100 * (k1 * Aww + RE + k2) / This approach was based on the assumption that the con-
(100 + .9565 * Aww+ k2) (1c) version factor was relatively stable across disparate fer-
mentations, while the apportionment of solids to yeast
where the variables are as defined above and k1 and k2 are mass, being highly dependent on process factors such as
coefficients estimated by non-linear regression. aeration and wort nutrients, would be expected to be less
Equation 1a is the original Balling equation with his stable across various processes.
familiar coefficients inserted. Figure 2 shows the residual Referencing Table IV, equation 1a (i.e., the original
fit of equation 1a. Equation 1b uses Balling’s assumptions Balling model), is seen to possess a root mean square
as stated previously, but with its conversion coefficient error (RMSE) of 0.237°P and an average mean deviation
derived from non-linear curve fitting to observed data. of –0.043°P for the pooled datasets. This negative mean
Equation 1c is similar to equation 1b but was augmented deviation would indicate a consistent, though slight, over-
by an additional assumption that solids converted to yeast estimation of measured original extract by Balling’s for-
mass are constant and are not proportional to the amount mula. The elevated RMSE appears to be strongly influ-
of mass converted to ethanol and CO2. Thus, a second enced by the large RMSE associated with the largest
coefficient directly representing yeast mass uptake was dataset. If this dataset were eliminated, the precision

Table IV. Fit of equations 1a-c to estimate original extract values.

Statistical Statistical Obs-Est Explained
Population n k1 k2 grouping RMSE (°P) grouping deviation (°P) variance (°P)
Equation 1a: OE = 100 * (2.0665 * Aww + RE) / (100 + (2.0665 – 1) * Aww)
Pooled 532 2.0665 - - 0.237 - –0.043– -
RPB I 092 2.0665 - - 0.102 b –0.230– -
RPB II 112 2.0665 - - 0.077 a –0.153– -
Proprietary 071 2.0665 - - 0.164 c –0.031– -
CMBTC 221 2.0665 - - 0.276 d 0.098 -
Navarro 036 2.0665 - - 0.108 b –0.109– -
Equation 1b: OE = 100 * (k1 * Aww + RE) / (100 + (k1 – 1) * Aww)
Pooled 532 2.0518 - - 0.234 - 0.000 0.9901
RPB I 092 2.0052 - a 0.075 a 0.000 0.9961
RPB II 112 2.0294 - b 0.072 a 0.000 0.9898
Proprietary 071 2.0602 - c 0.164 c 0.000 0.3285
CMBTC 221 2.0951 - d 0.278 d 0.000 0.8570
Navarro 036 2.0329 - b 0.094 b 0.000 0.9982
Equation 1c: OE = 100 * (k1 * Aww + RE + k2) / (100 + .9565 * Aww + k2)
Pooled 532 1.9629 0.4219 - 0.221 - 0.000 0.9911
RPB I 092 1.9298 0.3847 a 0.064 a 0.000 0.9917
RPB II 112 1.9298 0.5374 c 0.068 a 0.000 0.9914
Proprietary 071 1.9298 0.8963 e 0.156 b 0.000 0.3917
CMBTC 221 1.9298 0.6497 d 0.272 c 0.000 0.8627
Navarro 036 1.9298 0.4605 b 0.065 a 0.000 0.9991

VOL. 115, NO. 4, 2009 323

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
would be significantly enhanced though it would strongly row range of both dependent and independent variables.
accentuate the overestimation of measured original ex- This low r2 value for the Proprietary dataset carried over
tract. Examination of individual dataset results shows that into the equation 1c analysis in Table IV.
the RMSE ranges between 0.077 and 0.276°P while the Monte Carlo bootstrap analysis was used to determine
mean deviation ranges between –0.230 and 0.098°P. Mea- grouping significance in order to circumvent multiple
sured original extract was overestimated in four of the five comparison problems associated with asymptotic vari-
datasets while two of the datasets showed significantly ances, widely disparate variance estimates, and the large
greater RMSE. Statistical grouping of RMSE values using differences in dataset sample sizes.
F-ratio testing is presented and indicates that they are not Equation 1c represents a departure from previous con-
equivalent. This further accentuates the elevated values cepts of original extract estimation, incorporating not only
for the Proprietary and CMBTC dataset. a mass conversion coefficient, but also a yeast mass up-
Non-linear fitting to obtain the conversion coefficient take coefficient, thus divorcing yeast proliferation propor-
using equation 1b for the pooled population resulted in an tionality from ethanol formation. The conversion coeffi-
insignificant improvement in the RMSE at 0.234 though, cient resulting from the fitting of the total population was
as expected, the mean deviation is zero. Notably, the fitted 1.9629 while the yeast uptake coefficient was 0.4219.
conversion coefficient of 2.0518 was remarkably close to Unfortunately, the RMSE was only slightly less than that
the 2.0665 derived by Balling over 150 years ago. of either Balling’s original equation equation 1a or the
Estimation of the conversion coefficient for the individual fitted equation 1b. Examination of the individual datasets,
datasets yielded the values presented in Table IV. Al- however, is more encouraging.
though these fitted coefficients appeared to be close in The original assumption that a pooled model for the
value, ranging from 2.0052 to 2.0951, statistical grouping conversion coefficient was warranted was justified in that
using Monte Carlo bootstrap analysis showed that the five the asymptotic variance-based error was only 0.0154 for
values formed four statistically separate groups with only an estimate of 1.9298. The individual dataset estimates for
the RPB II and Navarro datasets being statistically equiva- the yeast uptake coefficients ranged from 0.3847 to 0.8963
lent. Additionally, the RMSE values of the individual and all were shown to be statistically distinct from each
datasets for this model also grouped into four statistically other by Monte Carlo bootstrap analysis. The RMSE val-
unique sets, though in a different pattern. The “explained ues from the RPB I, RPB II and Navarro datasets were
variance” or r2 values were for the most part, quite reason- statistically identical and were the lowest of the three
able. The exception was the value for the Proprietary data- models. The Proprietary and CMBTC datasets, however,
set. An r2 of 0.3285 implies that the selected model would continued to exhibit the largest RMSE values. Elimination
be a very poor predictor. If one examines Table III, how- of these datasets and refitting the remaining data did not
ever, the reason for this may be attributed to the very nar- materially improve the already low RMSE values for the

Fig. 2. Distribution of error in calculation of original extract from the Balling equation.

324 JOURNAL OF THE INSTITUTE OF BREWING

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
RPB I, RPB II and Navarro datasets. Additionally, the es- as the applicability of equations 3a–5a and 9a to the total
timates for the yeast uptake coefficients also changed little range of brewing measurements collected and (2) cogni-
and remained statistically distinct. Again, r2 values were sant of the need by industrial, craft and home brewers for
quite good, although the Proprietary dataset still had a a method to estimate alcohol levels given OE, AE and RE
poor value at 0.3917. values. Table V shows a calculation of the brewing values
The conclusion one can obtain from these results is a, b, c, and Q† from equations 3a–5a and 9a using the
that a general equation, which is solely a function of etha- combined datasets. They did not vary substantially with
nol and real extract in a fermented product, seems to be coefficients of variation less than 4%.
unable to reliably predict the measured original wort ex- Given previous reports of these values correlations
tract with sufficient accuracy. Examination of the dataset with original extract, further relationships were examined
coefficients for equation 1c in Table IV offers some in- with the Systat statistical analysis software. A difficulty in
sight as to why this may be the case. The amounts of sol- examining these datasets was their distribution with re-
ids incorporated into yeast were significantly different for spect to the original gravity. While the datasets were ran-
each dataset, possibly reflecting process differences that domly selected, the distribution of data with respect to the
were not apparent from measuring only ethanol and real original extract was not. For this reason we considered
extract. If that were the case, equations incorporating only weighting the data to account for uneven sampling (noted
those measures or any measures derived from them, such in Fig. 1) However, weighting datasets is a debatable and
as apparent extract, would be incomplete and would not non-trivial procedure. As a first estimate, ordinary least
be reliable predictors of original wort gravity. squares regression was employed to estimate the brewing
Nonetheless, it is tribute to the analytical skills of values a, b, c, and Q†. Specifically, the dependence of
Balling that he could determine the Balling constants of brewing values on original extract was examined by linear
2.0665 and 1.0665 in the mid-1800s. The original Balling regression using variations of equations 3a–5a and 9a
values are embedded in the brewing literature and their where the terms “bo + b1 * OE” were substituted for a, b, c
use, as shown in equation 1a, is reported to slightly over- or Q† as appropriate:
estimate21 the original extract. Application of Balling’s ori-
Aw/w / (OE – AE) = bao + ba1 * OE (3b)
ginal equation to the datasets in this paper similarly over-
estimates original wort extract in all datasets but one. Aw/w / (OE – RE) = bbo + bb1 * OE (4b)
However, OE, RE and OHww measurement errors may
Aw/w / (RE – AE) = bco + bc1 * OE (5b)
overshadow differences in the Balling constants. For these
reasons and since the Balling equation is embedded in the RDF / ADF = b Q†o + b Q†1 * OE (9b)
brewing literature (and culture!) we do not expect (nor
Table VI also lists the linear dependence of the same
recommend) adaptation of a new Balling equation.
brewing parameters cited by Pawloski and Doemens22 in
Modelling of traditional brewing values Table I. The results of the linear regressions of equations
3b–5b and 9b are also shown in Table VI. While the origi-
Notwithstanding the finding of differences in various
nal extract significantly (p < .0001) affects the dependent
datasets reported above, the authors were both (1) curious
variable of equations 3b–5b (Aw/w / (OE – AE), etc.) the
left-hand terms are only weakly dependent on OE.
Table V. Average brewing values (a, b, c and Q†) of equations 3a–5a and Our findings are in general agreement with previous
9aa. reported brewing values. The initial literature on this
a b c Q† topic2,12 is over 100 years old, in German, and apparently
Minimum 0.371 0.449 2.05 0.805 does not report on how the relationships were developed.
Maximum 0.455 0.579 2.24 0.841 We note that equation 4 is a transformation of the Balling
Mean 0.421 0.522 2.17 0.825 equation. If one assumes the Balling equation is correct,
Standard deviation 0.013 0.019 0.034 0.005 then the b value is equivalent to 1 / (2.0665 – 0.010665 *
a n = 532. OE) and agrees with values in Table I.

Table VI. Estimation of brewing values (a, b, c and Q†) as a function of original extract.
Brewing values
Statistical terma a b c Q†
Coefficientsa
bo 3.72 × 10–1 4.59 × 10–1 1.99 × 10–1 8.14 × 10–1
b1 3.57 × 10–3 4.69 × 10–3 1.34 × 10–2 8.62 × 10–4
Standard error of the estimate 1.00 × 10–2 1.52 × 10–2 1.26 × 10–2 4.84 × 10–3
Adjusted r2 0.410 0.343 0.860 0.148
Pawloski and Doemens coefficientsb
bo 3.95 × 10–1 4.81 × 10–1 2.19 8.20 × 10–1c
b1 2.1 × 10–3 3.0 × 10–3 3.1 × 10–3 –6.6 × 10–4c
a Coefficientsbo and b1 are estimates of intercept and slopes for a, b, and c and Q† as shown in equations
3b–5b and 9b. All regressions are significant p < 0.0001, n = 532.
b Linear dependence of previously published brewing values from the 1932 Table I of Pawloski and

Doemens22.
c An uncorrected estimate of Q† calculated as 1 / (1 + q).

VOL. 115, NO. 4, 2009 325

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Prediction of alcohol levels from original ACKNOWLEDGEMENTS
extract and apparent extract levels The authors thankfully acknowledge the donation of the con-
Small brewers and home brewers almost always know fidential datasets and the CMBTC data from R. McCaig, Winni-
their original and apparent extract values, but find meas- peg, MB. As well, the most useful discussions with J.
urement of alcohol levels difficult. While equation 3a or Hackbarth, The Gambrinus Company, San Antonio, TX are
greatly appreciated. Patrick Speers is thanked for his assistance
3b can be used with original and apparent extract values
in data entry.
to predict Aw/w, an attempt was made to find a more accu-
rate and precise expression to predict Aw/w. Using all five
datasets (n = 532, for ranges of Aw/w, OE and AE values REFERENCES
see Table III), a number of transforms of OE and AE 1. American Society of Brewing Chemists. Tables related to deter-
terms were regressed on the alcohol level Aw/w to attempt minations on wort, beer, and brewing sugars and syrups. Beer-
to find a prediction equation. The best prediction as evi- 5A, Beer-6A, Extract of Wort: Tables I and II. The Society: St.
denced by high correlation coefficients and no trend in Paul, MN, 2004.
residual errors was found using a forward stepwise regres- 2. Balling, C. J. N., Die Bierbrauerei. Verlag von Friedrich Temski:
sion technique (Systat software; p = 0.15 to enter terms). Prague, CHZ, 1865.
3. Beckett, M. H. and Pavelsek, V. L., Automatic alcohol and real ex-
The Aw/w value was used as the dependent variable and
tract determination. J. Am. Soc. Brew. Chem., 1986, 44, 78-80.
OE – AE and (OE – AE)2 as independent terms. The re- 4. Cutaia, A. J., Estimation of yeast mass increase in production
gression analysis is presented in Table VII. brewing fermentations by calculations of carbohydrate utiliza-
tion. J. Am Soc Brew Chem., 2007, 65, 166-171.
Aw/w = 0.38726 * (OE – AE) 5. Cutaia, A. J. and Munroe, J. H., A method for the consistent
+ 0.00307 * (OE – AE)2 (18) estimation of real degree of fermentation. J. Am Soc Brew
Using this expression the standard error of the estimate Chem., 1979, 37, 188-189.
6. DeClerck, J., A Textbook of Brewing. Vol. 2. Translated by K.
was 0.106. Figure 3 shows the ability of both equations 3b Barton-Wright. Chapman and Hall Ltd: London, 1958.
and 18 to predict final alcohol levels. Examination of these 7. Efron, B. and Tibshirani, R., An Introduction to the Bootstrap.
figures as well as statistics from tables indicates that equa- CRC Press: Boca Raton, FL, 1993.
tion 3b fits the data slightly better. Thus, it is hoped that 8. European Brewery Convention, Original Extract of Beers - Real
function 3b will be utilized by small and craft brewers Extract. Analytica - EBC. 3rd Ed. Method 7.1. P. E56. Schwei-
who do not, or are not able to determine the alcohol or zer Brauerei-Rundschau: Zurich, 1975.
real extract content of every brew. 9. Hackbarth, J. J., The effect of ethanol-sucrose interactions on
specific gravity. J. Am. Soc. Brew. Chem., 2009, 67, 146-151.
10. Her Majesty’s Stationary Office. The Alcoholic Liquor Regula-
Table VII. Prediction of Aw/w as a function of original and apparent tions. 1979. Statutory Instrument, No. 1146. HMSO: London,
extract levels by equation 18a. 1979, pp. 2908-2913.
Independent variable Aw/w 11. Her Majesty’s Revenue and Customs. Notice 226, Beer Duty.
Plymouth: UK, 2005, pp. 85-86.
Coefficients 12. Holzner, G., Über den Zusammenhang Zwischen dem Schein-
OE – AE 3.8726 × 10–1 bargen und Wirklichen Vergahrungsgrande. Zeitschrift für das
(OE – AE)2 3.07 × 10–3 Gesammte Brauwesen, 1877, 12, 104-107.
Adjusted r2 0.999
13. Holzner, G., Tabellen zur Berechnung der Ausbeute aus Malze
Standard error of estimate 1.06 × 10–1
und zur Saccharometririchen Bieranalyze. Druck und Verlag
an = 532. von R: Oldenburg, Munich, 1904.

Fig. 3. Prediction of alcohol (w/w) levels using original and apparent extract values with equations 3b and 18.

326 JOURNAL OF THE INSTITUTE OF BREWING

 20500416, 2009, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/j.2050-0416.2009.tb00387.x by Cambodia Hinari NPL, Wiley Online Library on [03/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
14. Hough, J. S., Briggs, D. E., Stevens, R. and Young, T. W., Malting 24. Siebert, K. J., Routine use of a programmable calculator for
and Brewing Science. Vol. II, Chapman and Hall: London, 1971. computing alcohol, real extract, original gravity and calories in
15. International Bureau of Legal Metrology. International Alco- beer. J. Am. Soc. Brew. Chem., 1980, 38, 27-33.
holic Tables. The Society: Paris, 1973. 25. Teich, M., A chapter on the history of transfer of information on
16. Kunze, W., Technology of Malting and Brewing. 7th Ed. Trans- attenuation. The Journal of the Brewing History Society Online.
lated by T. Wainwright. VLB: Berlin, 1996, pp. 353-354. 2005, 121, 40-46. www.breweryhistory.com/journal/archive/121/
17. Lincoln, R. H., Computer compatible parametric equations for bh-121-040.htm (last accessed November 2009).
basic brewing computation. Tech. Q. Master Brew. Assoc. Am., 26. Thorpe, E. and Brown, H. T., Reports on the determination of
1987, 24, 129-132. the original gravity of beers by the distillation process. Part I. J.
18. Navarro, A., Carbonell, J. V., Pinaga, F. and Lopez, F., Study for Instit. Brew., 1914, 20, 569-593.
the accuracy of Balling’s formula and proposal for alternative 27. Wahl, R. and Henius, M., American Handy Book of the Brew-
equations to calculate the original extract of beer. Cerveza y ing, Malting and Auxiliary Trades. 3rd Ed. Vol. II. Wahl-Henius
Malta, 1987, 24, 44-48. Institute: Chicago, IL, 1908.
19. Neilson, H. and Aastrup, S., Improving Tabarie’s formula. Scan- 28. Weissler, H. E., Brewing calculations. In: Handbook of Brew-
dinavian Brewers Review, 2004, 61(4), 30-34. ing. 1st Edition. W. A. Hardwick, Ed. Marcel Decker Inc: New
20. Neilson, H. and Erdal, K., The degree of fermentation. Scandi- York, NY, 1995, pp. 643-706.
navian Brewers Review, 2005, 62(3), 34-35, 37-39. 29. Yu, C. H., An overview of remedial tools for collinearity in
21. Neilson, H., Kristiansen, A. G., Krieger Larsen, K. M. and SAS. Proceedings of 2000 Western Users of SAS Software Con-
Erikstrom, C., Balling’s formula - scrutiny of a brewing dogma. ference. 2000, pp. 196-201.
Brauwelt Int., 2007, 25, 90-93. 30. www.creative-wisdom.com/pub/collin.pdf (last accessed Novem-
22. Pawloski, R. and Doemens, A., Die Brautechnischen Untersu- ber 2009).
chungs-Methoden. Verlag von R: Oldenburg, Munch, 1932.
23. Pearson, K., Mathematical contributions to the theory of evolu-
tion. III. Regression, heredity and panmixia. Philosophical
Trans. Royal Soc. London, 1896. 187, 253-318. (Manuscript accepted for publication November 2009)

VOL. 115, NO. 4, 2009 327