Food Science & Technology · Foundations · also Statistical Methods, Biostatistics, Design of Experiments, Applied Statistics for Food Science, Chemometrics
Probability, inference and the designs used in food research and quality work: factorial and response-surface designs for formulation and process optimisation, mixture designs, the statistics of sensory panels, and multivariate methods (PCA, PLS) used with spectra and chromatograms. Goes further than the general statistics subjects in Agricultural Engineering (probability-statistics-data) and Agribusiness (business-statistics) by focusing on food formulation, sensory and analytical data.
In the exams: Applied throughout ICAR AICE-JRF/SRF research units and every PG thesis; probability and statistics also appear in GATE XE-A
Food science is full of variation. No two batches of flour, no two cuts of meat and no two assessors on a sensory panel are identical, and every measurement carries error. Statistics gives food scientists the tools to describe that variation, to decide whether an observed difference is real or chance, to plan experiments that give clear answers with the fewest trials, and to extract information from large data sets such as spectra and chromatograms. Without it, a product developer may chase a 'better' formulation that is better only by luck, a quality manager may reject good batches or accept bad ones, and a researcher may publish findings that cannot be repeated.
This subject covers the statistics most used in food research, quality control and product development: descriptive statistics and probability distributions; confidence intervals and hypothesis tests; correlation and regression; analysis of variance; the classical experimental designs; factorial, response surface and mixture designs for formulation and process optimisation; non-parametric tests for sensory data; multivariate methods for spectra, chromatograms and sensory profiles; the statistics of analytical methods, such as calibration, limits of detection and measurement uncertainty; and reproducible data analysis with software.
The subject is applied in every postgraduate thesis and research project, and it is examined worldwide: probability and statistics appear in the engineering mathematics sections of graduate engineering entrance tests, experimental design and data analysis in the research-technique papers of postgraduate entrance and fellowship examinations in food science, and the statistics of analytical results in food analyst and laboratory qualification examinations. The general theory is taught in depth in the Agricultural Statistics track, in subjects such as experimental-design-basic, factorial-split-plot-designs, response-surface-designs, hypothesis-testing, nonparametric-methods and multivariate-analysis; this manual keeps the food angle, using food examples throughout, and links to food-product-development for formulation optimisation and to sensory-science for panel data.
Descriptive statistics summarise data. Measures of location include the arithmetic mean, the median (the middle value, robust to outliers) and the mode. Measures of spread include the range, the variance (the average squared deviation from the mean, using n − 1 in the denominator for a sample), the standard deviation (its square root, in the units of the data) and the interquartile range. The coefficient of variation, the standard deviation as a percentage of the mean, compares variability between measurements on different scales: a fill weight with a mean of 10 g and a standard deviation of 0.5 g has a coefficient of variation of 5 per cent. For the moisture contents 12, 14, 11, 13 and 15 per cent, the mean is 13, the sample standard deviation is 1.58 and the standard error of the mean, s/√n, is 0.71.
Graphs reveal what numbers hide: histograms and box plots show distributions, skewness and outliers; scatter plots show relationships; control charts show behaviour over time. Data should always be plotted before formal analysis, because outliers, skewed distributions and non-linear relationships change which methods are appropriate. Food data are often skewed: microbial counts, for example, follow approximately log-normal distributions, so they are usually transformed to logarithms before averaging and testing.
Three distributions recur in food work. The binomial distribution describes the number of successes in n independent trials, each with probability p: the number of defective packs in a sample, or the number of correct answers in a triangle sensory test, where p = 1/3 by chance. Its mean is np and variance np(1 − p). The Poisson distribution describes counts of rare, independent events in a fixed interval or volume, such as colonies on a plate, foreign bodies per batch or customer complaints per week; its mean and variance are equal, λ, and the probability of zero events is e^(−λ), so with an average of 2 events the probability of none is e^(−2) = 0.135. The Poisson distribution underlies the statistics of plate counts and of detecting low-level contamination, as in the sampling plans of food-microbiology.
The normal distribution, the familiar bell-shaped curve, describes many continuous measurements such as fill weights, composition and instrument readings, and, by the central limit theorem, the distribution of sample means even when individual values are not normal. Any normal variable can be converted to the standard normal z = (x − μ)/σ: a pack weight of 110 g from a process with mean 100 g and standard deviation 5 g has z = 2, and about 2.3 per cent of packs exceed it. About 68, 95 and 99.7 per cent of values lie within one, two and three standard deviations of the mean. Related distributions, t, chi-square and F, arise in tests on sample means and variances (chapter 2).
A sample statistic such as a mean varies from sample to sample; its sampling distribution describes that variation, and its standard deviation is the standard error. A confidence interval gives a range of plausible values for the population parameter. For a mean with unknown population standard deviation, the interval is x̄ ± t × s/√n, where t is taken from the t distribution with n − 1 degrees of freedom. For the five moisture values in chapter 1 (mean 13, s = 1.58, n = 5), t for 95 per cent confidence with 4 degrees of freedom is 2.776, so the interval is 13 ± 2.776 × 0.71, about 11.0 to 15.0 per cent. With large samples or known σ, z values (1.96 for 95 per cent) replace t. A 95 per cent confidence interval means that, over many repeated samples, 95 per cent of intervals constructed this way would contain the true value.
Confidence intervals are often more informative than a bare test result, because they show both whether an effect is likely to be zero and how large it may be. Sample size planning uses the same logic: to estimate a mean within a margin E with confidence given by z, the required n is about (zσ/E)².
A hypothesis test assesses evidence against a null hypothesis, usually of no difference or no effect. The p-value is the probability of obtaining a result at least as extreme as that observed if the null hypothesis were true; if it is below the chosen significance level α, commonly 0.05, the null hypothesis is rejected. A type I error rejects a true null hypothesis (probability α); a type II error fails to reject a false one (probability β); the power of a test, 1 − β, is its ability to detect a real effect of a given size, and rises with sample size. A statistically significant result is not necessarily important in practice, and a non-significant result is not proof of no effect.
Common tests include the z and t tests for means: the one-sample t test compares a mean with a target, such as declared fat content; the two-sample t test compares two independent groups, such as two suppliers; and the paired t test compares two measurements on the same units, such as the same samples analysed by two methods. The chi-square test compares observed with expected counts, testing goodness of fit or independence in contingency tables: if 30 of 40 consumers choose product A in a paired preference test, where 20 each would be expected by chance, chi-square = (30 − 20)²/20 + (10 − 20)²/20 = 10, well above the critical value of 3.84 for one degree of freedom at α = 0.05. The F test compares two variances and is the basis of analysis of variance. Tests assume random sampling and, for t and F tests, approximately normal data and, for some, equal variances; checking assumptions is part of the analysis.
Correlation measures the strength of a linear relationship between two variables. Pearson's correlation coefficient r ranges from −1 to +1; values near ±1 indicate strong linear relationships, values near 0 little linear relationship. Spearman's rank correlation measures monotonic relationships using ranks and is robust to outliers and non-normality. Correlation does not imply causation: hardness and sweetness of a set of biscuits may correlate because both depend on sugar content, not because one causes the other. A strong correlation can also be produced by one or two extreme points, so scatter plots are essential.
Simple linear regression fits a straight line, y = a + bx, by least squares, minimising the sum of squared vertical deviations. It is used for calibration curves in food analysis (chapter 11), for predicting a sensory attribute from an instrumental measurement, and for shelf-life kinetics. The coefficient of determination, R², is the proportion of the variation in y explained by the model; in simple regression it equals r², so a correlation of 0.9 explains 81 per cent of the variation. The significance of the slope is tested by a t test, and the residuals, the differences between observed and fitted values, are plotted to check for non-linearity, unequal variance and outliers.
Multiple regression models a response as a function of several predictors, y = b0 + b1x1 + b2x2 + …, for example predicting cake volume from flour protein, sugar level and baking temperature. Coefficients give the effect of each predictor with the others held constant. Problems include multicollinearity, when predictors are correlated with each other, which makes coefficients unstable and hard to interpret; overfitting, when too many predictors are fitted to too few observations; and extrapolation beyond the range of the data. Adjusted R² penalises extra predictors, and validation on separate data checks predictive ability. When predictors are many and highly correlated, as with spectral data, partial least squares regression is preferred (chapter 10).
Many food relationships are non-linear: microbial growth curves, Arrhenius temperature dependence of reaction rates in shelf-life studies, moisture sorption isotherms and texture-composition relationships. Some can be linearised by transformation, such as taking logarithms for first-order kinetics; others are fitted directly by non-linear regression. Polynomial terms (x², x1x2) extend linear models to curvature and interactions and are the basis of response surface models (chapter 7).
Analysis of variance (ANOVA) tests whether the means of several groups differ, by partitioning the total variation in the data into components due to the factors studied and to random error. In one-way ANOVA, for example comparing the firmness of bread made with four flour types, the total sum of squares is split into a between-treatment sum of squares and a within-treatment (error) sum of squares; each is divided by its degrees of freedom to give mean squares, and their ratio is the F statistic. If F exceeds the critical value for the degrees of freedom, at least one mean differs. With t treatments and r replicates each, the treatment degrees of freedom are t − 1 and the error degrees of freedom t(r − 1).
Two-way ANOVA studies two factors together, for example flour type and baking temperature, and estimates their main effects and their interaction, which tests whether the effect of one factor depends on the level of the other. Interactions are common in foods and change interpretation: if the best flour depends on the baking temperature, it is meaningless to report the 'best flour' overall. Assumptions are independence of observations, approximately normal residuals and equal variances among groups; residual plots and tests check them, and transformations such as logarithms of microbial counts help when they fail. Sensory data, with assessors as a random factor, use mixed models, as described in sensory-science.
A significant F tells only that some means differ. Multiple comparison procedures identify which. Fisher's least significant difference (LSD) compares pairs of means with t tests using the ANOVA error, and is sensitive but inflates the chance of false positives when many comparisons are made. Tukey's honestly significant difference (HSD) controls the error rate across all pairwise comparisons and is widely recommended. Duncan's multiple range test, still common in food science papers, is less conservative than Tukey's and is criticised for not controlling the overall error rate adequately. Dunnett's test compares each treatment with a single control, as when testing several new formulations against the current product. Bonferroni correction divides α by the number of comparisons and is simple but conservative.
Results are often shown with letters, means sharing a letter not differing significantly, but the comparison method should be stated, and planned comparisons, decided before the experiment from the research question, are more powerful and more meaningful than testing every possible pair after the fact. When treatments are levels of a quantitative factor, such as sugar concentration, fitting a regression or response curve is usually more informative than comparing every pair of levels.
Experimental design rests on three principles set out by R. A. Fisher: replication, which allows experimental error to be estimated and reduces it; randomisation, which allocates treatments to experimental units by chance so that unknown sources of bias are spread randomly; and local control or blocking, which groups similar units so that comparisons are made among like units, reducing error. The general theory is treated in experimental-design-basic in the Agricultural Statistics track; here the emphasis is on food experiments.
In a completely randomised design (CRD), treatments are assigned to units entirely at random. It is the simplest design, gives the maximum degrees of freedom for error, and is appropriate when units are homogeneous, as with portions of a well-mixed batch tested in a laboratory under uniform conditions. It is analysed by one-way ANOVA. Its weakness is that any systematic variation among units, such as day-to-day differences in raw materials or equipment, inflates the error and hides treatment effects.
In a randomised complete block design (RCBD), units are grouped into blocks that are as uniform as possible within themselves, and each treatment appears once in each block, randomised within it. In food experiments, blocks are often days, batches of raw material, production runs, ovens or assessors: if four formulations are each made once on each of five days, days are blocks, and day-to-day variation is removed from the error. The analysis is a two-way ANOVA without interaction, with treatment, block and error sums of squares; with t treatments and b blocks, error degrees of freedom are (t − 1)(b − 1). Blocking that removes a large source of variation makes tests much more sensitive.
The Latin square design controls two sources of variation at once, arranged in rows and columns, with each treatment appearing once in each row and each column; it requires the number of treatments to equal the number of rows and columns. In sensory testing, for example, rows can be assessors and columns the order of tasting, so that each sample is tasted once by each assessor and once in each position, balancing order effects. With n treatments, error degrees of freedom are (n − 1)(n − 2), so a 5 × 5 square has 12. Latin squares give few error degrees of freedom when small and assume no interactions between rows, columns and treatments. Other designs used in food work include incomplete block designs, when not every treatment can appear in every block, as when an assessor can taste only a few samples per session, and split-plot designs, when one factor is applied to large units and another to subunits, as with oven temperature applied to whole bakes and formulations to individual pans, covered in factorial-split-plot-designs and incomplete-block-designs.
Factorial designs study several factors simultaneously, each at two or more levels, with every combination of levels included. They are far more efficient than changing one factor at a time and estimate interactions, which one-factor-at-a-time experiments cannot. In a 2^k factorial, each of k factors is studied at a low and a high level, coded −1 and +1, requiring 2^k runs per replicate: 2^3 = 8 runs for three factors such as sugar, fat and baking time. Main effects are estimated as the difference between the mean response at the high and low levels of a factor; interaction effects measure how much the effect of one factor changes with the level of another. Because every run contributes to every effect estimate, factorial designs use data efficiently.
Centre points, runs at the middle level of all factors, are added to estimate pure error without replicating the whole design and to detect curvature: if the mean of the centre points differs from the mean of the factorial points, a straight-line model is inadequate and a response surface design (chapter 7) is needed. Results are analysed by ANOVA or regression and shown in effect plots, interaction plots and normal or half-normal plots of effects, on which important effects stand out from the line formed by negligible ones.
The number of runs in a full factorial grows quickly: seven factors at two levels need 128 runs. When many factors might matter, as in early formulation work, fractional factorial designs use a fraction of the runs, such as a half (2^(k−1)) or quarter (2^(k−2)), chosen so that main effects and important interactions can still be estimated. The cost is aliasing or confounding: some effects cannot be separated from others. The resolution of a design describes how serious aliasing is: in resolution III designs, main effects are aliased with two-factor interactions; in resolution IV, main effects are clear of two-factor interactions, but two-factor interactions are aliased with each other; in resolution V, main effects and two-factor interactions are clear of each other. A 2^(5−1) half fraction with 16 runs is resolution V.
Plackett-Burman designs, with run numbers that are multiples of four, screen many factors, for example 11 factors in 12 runs, assuming interactions are negligible. Screening experiments identify the few factors that matter most, according to the sparsity-of-effects principle that most systems are dominated by a few main effects and low-order interactions; these are then studied in detail. Factorial designs are used in food for formulation, process optimisation, ingredient interactions, shelf-life studies and the robustness testing of analytical methods. Their detailed theory, including confounding in blocks and analysis of unreplicated designs, is in factorial-split-plot-designs.
Response surface methodology (RSM), introduced by Box and Wilson in 1951, finds the combination of factor levels that gives the best response, by fitting an empirical model and exploring its surface. In product and process development it is used after screening has identified two to four important factors, to optimise, for example, the levels of sugar, fat and leavening for maximum cake volume and liking, or the temperature and time of a process for maximum yield and minimum colour change. The usual model is a second-order polynomial: y = b0 + Σ bi xi + Σ bii xi² + Σ bij xi xj, which includes linear, quadratic and interaction terms and can describe maxima, minima, ridges and saddle points. Its adequacy is judged by ANOVA, R² and adjusted R², the lack-of-fit test (comparing model error with pure error from replicated centre points) and residual analysis. The optimum is located by examining contour and three-dimensional surface plots, by solving for the stationary point, or by numerical optimisation. Detailed theory is in response-surface-designs in the Agricultural Statistics track.
The central composite design (CCD) is the most widely used RSM design. It combines a two-level factorial or fractional factorial (2^k points), axial or star points located at ±α along each factor axis with the others at their centre (2k points), and replicated centre points. With three factors and six centre points it has 8 + 6 + 6 = 20 runs; with two factors and five centre points, 4 + 4 + 5 = 13 runs. The axial distance α determines the design's properties: α = (2^k)^(1/4) makes the design rotatable, so that prediction variance depends only on distance from the centre; this gives α = 1.414 for two factors and 1.682 for three. A face-centred design (α = 1) keeps all points within the factorial range, useful when factor levels beyond it are impossible or unsafe.
The Box-Behnken design places points at the midpoints of the edges of the design space plus centre points, with no corner points, so extreme combinations of all factors, which may be impractical or give unacceptable products, are avoided; it needs three levels of each factor. For k factors it has 2k(k − 1) edge points plus centre points: three factors need 12 + centre points, commonly 15 runs in total; four factors need 24 + centre points, commonly 27. When several responses must be optimised together, such as maximising liking and yield while minimising cost, overlay contour plots or the desirability function approach, described in food-product-development, find a compromise. Optimal predicted conditions must be confirmed by verification experiments.
In many food formulations the response depends on the proportions of ingredients rather than their absolute amounts, and the proportions must add to one (or 100 per cent): a fruit beverage blend of mango, orange and pineapple juices, a flour blend of wheat, millet and pulse flours, or a fat blend for a spread. Because increasing one component necessarily decreases others, the components cannot be varied independently, and ordinary factorial or response surface designs do not apply. Mixture designs and models handle this constraint. For three components, the experimental region is a triangle (a simplex), with pure components at the vertices, binary blends along the edges and ternary blends inside; for four components, it is a tetrahedron.
The simplex lattice design {q, m} for q components places points at proportions 0, 1/m, 2/m, …, 1 of each component, in all combinations that sum to 1; the number of points is the combination C(q + m − 1, m), so a {3, 2} lattice has 6 points (the three pure components and the three 50:50 binary blends), and a {3, 3} lattice has 10. The simplex centroid design includes the pure components, all equal binary blends, all equal ternary blends and so on up to the overall centroid, giving 2^q − 1 points, or 7 for three components. Checkpoints inside the region help validate the model.
Mixture data are fitted with Scheffé polynomials, which have no intercept because of the constraint. The linear model, y = Σ bi xi, describes simple blending, in which the response is a weighted average of the pure components; the quadratic model adds terms bij xi xj that capture non-linear blending: a positive coefficient indicates synergism, the blend performing better than the average of its components, and a negative coefficient antagonism. The special cubic model adds a three-component term. Fitted models are shown as contour plots on the triangle, from which the best blend is chosen.
In practice, components rarely range from 0 to 100 per cent; a biscuit flour blend may need at least 60 per cent wheat flour and no more than 25 per cent millet flour. Constrained mixture designs restrict the region with lower and upper bounds, giving irregular regions in which design points are chosen by extreme vertices or optimal design algorithms. When process variables, such as baking temperature, also matter, mixture-process designs combine mixture and factorial designs. Mixture designs are powerful tools for the blending and reformulation tasks described in food-product-development, including optimising nutrient-dense blends, sensory acceptability and cost together.
Many tests assume normally distributed data measured on interval scales. Sensory data often violate these assumptions: ranks are ordinal, category scale data may be skewed or bounded, consumer data may be bimodal when segments differ, and small panels give too little data to check normality. Non-parametric tests make fewer assumptions, usually working with ranks or signs, and are robust to outliers. They are somewhat less powerful than parametric tests when parametric assumptions hold, but more reliable when they do not. The general theory is in nonparametric-methods in the Agricultural Statistics track.
The binomial test, used for difference and preference tests as described in sensory-science, is itself a non-parametric test on counts. The chi-square test analyses counts in categories, such as the frequencies with which attributes are ticked in check-all-that-apply questions or the distribution of responses across categories for different products. Cochran's Q test compares several related proportions, such as CATA frequencies for several products evaluated by the same consumers, and McNemar's test compares two related proportions.
The sign test and the Wilcoxon signed-rank test compare two related samples, such as each assessor's scores for two products: the sign test uses only the direction of differences, the Wilcoxon test also uses their ranked size and is more powerful. The Mann-Whitney U test (Wilcoxon rank-sum test) compares two independent groups, such as liking scores from two different consumer groups. The Kruskal-Wallis test extends this to three or more independent groups, as a rank-based analogue of one-way ANOVA. The Friedman test compares three or more related samples, such as products each ranked or rated by every assessor, and is the standard analysis for ranking tests; it is the rank-based analogue of the randomised block ANOVA, with assessors as blocks. Significant Kruskal-Wallis or Friedman results are followed by multiple comparisons on rank sums. Spearman's rank correlation and Kendall's coefficient of concordance (W), which measures agreement among assessors' rankings, are used to examine panel consistency.
The choice between parametric and non-parametric analysis should follow the data type and the purpose. Hedonic data from large consumer samples are routinely analysed by ANOVA, which is robust with large samples, while rankings and small-panel ordinal data call for non-parametric methods. In either case, plotting the data and reporting effect sizes with medians or means helps readers judge practical importance.
Food data are increasingly multivariate: a near-infrared spectrum has hundreds of absorbance values, a chromatogram many peaks, a sensory profile dozens of attributes, and a consumer test hundreds of individual liking scores. Chemometrics applies multivariate statistics to such chemical and sensory data. Principal component analysis (PCA) transforms many correlated variables into a few uncorrelated principal components, linear combinations ordered by the amount of variance they explain. A scores plot shows how samples relate, grouping similar samples and revealing outliers; a loadings plot shows which variables drive each component. Variables are usually standardised (autoscaled) first when they are on different scales, so that each contributes equally. If the first two components explain, say, 75 per cent of the variance, a two-dimensional plot captures most of the structure. PCA is used to map sensory profiles, classify oils or honeys by origin, detect adulteration and explore process data.
Cluster analysis groups objects so that those within a group are similar and those in different groups dissimilar. Hierarchical clustering builds a tree (dendrogram) by successively merging the closest objects or groups, using a distance measure such as Euclidean distance and a linkage method such as Ward's; k-means clustering partitions objects into a chosen number of groups. In food, clustering segments consumers by liking pattern, as noted in sensory-science, and groups samples by composition or origin.
Partial least squares (PLS) regression relates a block of predictor variables X to one or more responses Y by extracting latent variables that explain the covariance between them. It handles many highly correlated predictors and more variables than samples, which ordinary multiple regression cannot, so it is the standard method for calibrating near-infrared and other spectroscopic methods to predict composition, such as protein, fat and moisture in grains, milk and meat, and for relating sensory profiles to instrumental data or liking. The number of latent variables is chosen by cross-validation to avoid overfitting, and models are judged by the root mean square error of calibration, cross-validation and, most importantly, prediction on an independent test set, together with R².
For classification, PLS discriminant analysis (PLS-DA), linear discriminant analysis and soft independent modelling of class analogy (SIMCA) assign samples to classes, such as authentic or adulterated, or geographical origins, and are central to food authenticity work. Machine learning methods, such as random forests, support vector machines and neural networks, are increasingly applied to the same problems, with the same need for independent validation. Pre-processing of spectra, such as smoothing, derivatives, standard normal variate and multiplicative scatter correction, removes physical effects before modelling. A common error is to report excellent fits that do not hold on new samples; honest validation on samples not used in model building is essential.
Most instrumental food analyses depend on a calibration curve relating instrument response to known concentrations of standards, usually fitted by linear regression. A good calibration covers the range of sample concentrations, uses enough standards (commonly at least five levels), and is checked for linearity by residual plots rather than only by a high correlation coefficient, which can be high even for curved relationships. Matrix effects, where other components of the food change the response, are addressed by matrix-matched standards, standard addition or internal standards. Unknown concentrations are read from the curve within its range; extrapolation is avoided.
The limit of detection (LOD) is the lowest concentration that can be reliably distinguished from zero, and the limit of quantification (LOQ) the lowest that can be measured with acceptable precision and accuracy. A common approach estimates them from the calibration: LOD = 3.3 σ/S and LOQ = 10 σ/S, where σ is the standard deviation of the response at low concentration or of the blank or intercept, and S is the slope of the calibration curve. With σ = 0.02 absorbance units and S = 0.5 units per mg/L, LOD = 3.3 × 0.02/0.5 = 0.13 mg/L and LOQ = 10 × 0.02/0.5 = 0.40 mg/L. Other definitions, such as three times the standard deviation of blanks, are also used, so the approach must be stated.
Method validation shows that an analytical method is fit for its purpose, by measuring its performance characteristics: selectivity (specificity), linearity and range, LOD and LOQ, accuracy (trueness, assessed by recovery of spiked analyte or analysis of certified reference materials), precision (repeatability within a laboratory under the same conditions, and reproducibility between laboratories, usually expressed as relative standard deviations) and robustness (insensitivity to small changes in conditions). Laboratories accredited to ISO/IEC 17025 must validate or verify methods and estimate measurement uncertainty.
Measurement uncertainty is a parameter characterising the dispersion of values that could reasonably be attributed to the measured quantity. In the approach of the Guide to the Expression of Uncertainty in Measurement, each source of uncertainty, such as weighing, volumetric measurement, calibration, recovery and repeatability, is expressed as a standard uncertainty, and independent components are combined by the root sum of squares: with components of 0.3 and 0.4 units, the combined standard uncertainty is √(0.3² + 0.4²) = 0.5. The expanded uncertainty U = k × uc, with a coverage factor k = 2 giving about 95 per cent confidence, here U = 1.0, and results are reported as value ± U. Uncertainty matters in compliance decisions: when a result lies close to a legal limit, its uncertainty determines whether non-compliance can be concluded beyond reasonable doubt. Quality control in the laboratory uses control charts of reference materials, duplicates and blanks, and participation in proficiency testing.
Statistical analysis in food science is done with software ranging from spreadsheets, suitable for simple summaries and charts but error-prone for complex analysis, to general statistical packages and programming environments. Free, open-source environments such as R and Python, with libraries for designed experiments, mixed models, sensory analysis, chemometrics and graphics, are widely used in research and increasingly in industry; commercial statistical, design-of-experiments and chemometrics packages offer menu-driven interfaces and validated procedures used in many companies and regulated laboratories. Specialised software supports sensory data collection and analysis, spectroscopic calibration and statistical process control. Whatever the tool, the analyst must understand the method being applied, check its assumptions and interpret the output; software produces numbers for any data, including inappropriate analyses. In regulated settings, such as accredited laboratories and pharmaceutical-grade nutraceutical production, software used for calculations may need to be validated, with documented checks that it gives correct results for known data sets, and spreadsheets used for routine calculations are protected against accidental changes to formulas. Training in the chosen tools, shared templates for common analyses and peer review of analyses reduce errors.
Reproducible analysis means that another person, or the same analyst later, can obtain the same results from the same data. Good practice includes keeping raw data unchanged and separate from processed data; documenting every step of data cleaning, transformation and analysis, ideally as scripts rather than manual operations; recording software versions and settings; using version control for scripts; and archiving data and code with publications where possible. Data should be organised in tidy tables, with one row per observation and one column per variable, consistent units and clear labels, which makes errors easier to find.
Good statistical practice also covers planning and reporting. Experiments should be planned with a clear question, an appropriate design, randomisation, and a sample size or number of replicates justified by the effect size to be detected and the variability expected. Analyses should be planned in advance where possible, to avoid searching many tests until something is significant, a practice that produces false findings; when many tests are made, corrections for multiple comparisons are needed. Reports should give the design, the number of replicates and what was replicated (true replicates, not repeated measurements on one batch), the statistical methods and software, measures of variability and effect sizes with confidence intervals, not only p-values, and should avoid claiming more than the data support. In food science, where results guide formulations sold to millions and regulatory decisions, these habits are as important as the methods themselves.
A student uses this subject to design and analyse every laboratory exercise and thesis experiment properly: choosing a design, randomising, replicating, running ANOVA with an appropriate multiple comparison, and reporting means with variability. The recurring calculations are short: standard error, confidence intervals, chi-square, F, design run counts, α for rotatability, mixture design points, LOD and LOQ and combined uncertainty. Learning R or Python early makes analysis reproducible and saves time.
Five years into product development, quality or a laboratory, a technologist designs formulation and process optimisation studies, analyses sensory and consumer data, builds NIR calibrations, validates analytical methods and estimates uncertainty, runs control charts and capability studies, and explains statistical results to managers who need clear decisions rather than p-values.
A senior research or quality leader with twenty-five years' experience sets standards for experimental design and data analysis across the organisation, decides on investments in chemometrics and data infrastructure, reviews the statistical basis of claims, shelf-life and compliance decisions, and builds teams that combine food science with data science.
A researcher at this level develops and applies advanced methods: mixed models for complex sensory and consumer data, Bayesian and machine learning approaches to food authenticity and prediction, optimal designs for constrained formulations, multiblock data integration across sensory, instrumental and omics data, and reproducible research practice. Such work joins statistics, computing and food science.
This manual explains methods with food examples and short calculations but does not derive their theory or provide statistical tables, which are in the Agricultural Statistics track subjects (experimental-design-basic, factorial-split-plot-designs, response-surface-designs, hypothesis-testing, nonparametric-methods, multivariate-analysis) and standard texts. Mixed models, Bayesian methods, survival analysis for shelf life and statistical process control are introduced only briefly; statistical quality control is in statistical-quality-control and quality-assurance-lab-accreditation. Software is described generally without step-by-step instructions.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.