Dairy Science & Technology · Postgraduate & Research · also Research Methodology, Experimental Design, Technical Writing
Designing, analysing and publishing dairy research, and the ethics, IPR and grant writing that a research career needs.
In the exams: ICAR AICE Part A (general) and all dairy subjects; ASRB NET; required in every Indian PG programme (PGS courses)
Research turns questions about milk, microbes, processes and people into reliable knowledge. A master's or doctoral student in dairy science must be able to formulate a researchable problem, find and evaluate what is already known, design experiments that can answer the question with the least effort and the greatest confidence, analyse multivariate and sensory data, write theses, papers and grant proposals, act ethically, protect and transfer the results, and work safely in laboratories handling hazardous chemicals and micro-organisms. These skills are taught in research methodology courses in postgraduate programmes worldwide and are tested in doctoral entrance and research-service examinations.
This manual covers ten topics: the scientific method and research problem formulation; literature search and systematic review; completely randomised, block and factorial designs; response surface and mixture designs; multivariate analysis by principal component analysis and clustering; sensory and consumer data analysis; writing theses, papers and grant proposals; research ethics and data integrity; intellectual property, patents and technology transfer; and laboratory safety and biosafety. It builds on the statistics, sensory evaluation and product development subjects, applying them to the design and conduct of research rather than repeating them.
Examples are drawn from dairy research: comparing starter cultures, optimising a cheese process, profiling milk from different regions, evaluating a functional yoghurt with consumers, and filing a patent on a new process. International guidelines are cited where they exist, such as those on reporting systematic reviews, research ethics, animal welfare in research, intellectual property treaties and laboratory biosafety; national rules and institutional policies apply in addition and must be checked by each researcher.
Science advances by proposing explanations and testing them against evidence. A typical cycle starts with an observation, such as that some cheeses made in winter develop bitterness; leads to a question, why winter cheeses are bitter; proposes a hypothesis, a testable explanation, such as that higher plasmin activity in late-lactation milk causes more bitter peptides; derives a prediction, such as that adding a plasmin inhibitor to winter milk will reduce bitterness; and tests the prediction by experiment or observation. Results support, refine or refute the hypothesis, and new questions follow.
Philosophers of science have emphasised that a scientific hypothesis must be falsifiable: it must be possible, in principle, for evidence to show it wrong. Karl Popper argued that no number of confirming observations proves a general hypothesis, while a single well-established contradiction can refute it. In practice, scientists rarely abandon a theory on one contrary result, because experiments have errors and auxiliary assumptions; but designing tests that could fail is the core of good research. Deductive reasoning derives specific predictions from general principles; inductive reasoning generalises from observations; both are used.
Statistical hypothesis testing formalises the test: a null hypothesis of no effect is compared with the data, and a p-value expresses how surprising the data would be if it were true. Statistical significance is not the same as importance, and a failure to reject the null hypothesis is not proof of no effect, especially with small samples. Estimation of effect sizes with confidence intervals, and replication by independent studies, give stronger evidence than single p-values. Concerns about irreproducible research across many fields have led to practices such as pre-registration of hypotheses and analysis plans, sharing data and code, and reporting all outcomes.
A good research problem addresses a real gap in knowledge or practice, matters to someone, and can be answered with the resources available. One widely used checklist, FINER, asks whether the question is feasible (enough samples, time, skills, equipment and money), interesting, novel, ethical and relevant. Gaps are found by reading reviews and recent papers, noting their stated limitations and future directions, talking to industry and farmers about their problems, and noticing inconsistencies between studies.
The broad problem is narrowed into specific research questions and objectives. Objectives should be specific, measurable, achievable, relevant and time-bound, and each should correspond to a planned experiment or analysis. For example, a broad aim to improve the texture of reduced-fat cheddar-type cheese might become specific objectives: to determine the effect of an exopolysaccharide-producing adjunct culture on moisture, firmness and meltability over six months of ripening; to identify the optimum combination of culture level and cooking temperature by a response surface design; and to evaluate consumer acceptance of the optimised cheese.
Formulating the problem also means defining variables precisely. Independent variables are those the researcher manipulates or classifies, such as culture type; dependent variables are the outcomes measured, such as firmness; controlled variables are kept constant; and confounding variables are uncontrolled factors associated with both the independent and dependent variables that can create spurious associations. Operational definitions state exactly how each variable is measured, such as 'firmness measured as peak force in a uniaxial compression to 30% strain at 10 °C'.
The experimental unit is the smallest unit to which a treatment is independently applied. In a cheese study, if each treatment is applied to one vat and several cheeses are made from each vat, the vat is the experimental unit, and the cheeses are subsamples. Treating subsamples as independent replicates, an error called pseudoreplication, described influentially by Hurlbert in ecology, inflates the apparent precision and produces false positives. True replication requires independent applications of each treatment, such as separate vats on different days with different milk batches, and statistical analysis must use the experimental unit, or a model that accounts for the nesting of subsamples.
Research designs differ in their ability to establish causation. Experimental studies, in which the researcher assigns treatments at random, provide the strongest evidence of cause and effect, because randomisation balances unknown confounders on average. Observational studies, in which the researcher measures existing differences, such as comparing milk from farms using different feeding systems, can show associations but are open to confounding, since farms differ in many ways; careful design, measurement of potential confounders and statistical adjustment help but cannot fully replace randomisation. Descriptive studies characterise what exists, such as surveys of milk quality in a region; exploratory studies generate hypotheses; and confirmatory studies test them.
Research is also classified as basic, seeking understanding without immediate application, such as the structure of casein micelles; applied, aimed at a practical problem, such as reducing fouling in a pasteuriser; or translational and developmental, moving knowledge into products and practice. Dairy science spans all of these.
A literature search finds what is already known about a question, avoids duplicating work, identifies methods and gaps, and provides the context for interpreting results. Sources include peer-reviewed journals, conference proceedings, books, theses, patents, standards and grey literature such as government reports, technical bulletins and industry publications. Major sources for dairy science include multidisciplinary citation databases, biomedical databases such as the public PubMed database for nutrition and health topics, food science and technology abstracting databases, agricultural databases including the FAO's AGRIS, and patent databases such as WIPO's PATENTSCOPE and other public patent search services. General academic search engines are useful for a first look but are not reproducible enough for systematic searches.
Effective searching uses Boolean operators: AND narrows by requiring all terms, OR broadens by accepting synonyms, and NOT excludes terms. Phrase searching with quotation marks, truncation with symbols such as an asterisk to capture word variants (lact* for lactose, lactase, lactic), and field searching in titles, abstracts or keywords refine results. Controlled vocabularies, such as subject headings assigned by indexers, retrieve papers regardless of the words authors chose. A search for studies on lactose-free milk shelf life, for example, might combine (lactose-free OR "lactose free" OR "lactose-hydrolysed" OR "lactose-hydrolyzed") AND (milk) AND ("shelf life" OR storage OR browning OR Maillard), noting both British and American spellings.
Citation chaining, following references backward and tracing forward citations of key papers, finds work missed by keyword searches. Critical evaluation of each source considers its peer review, methods, sample size, controls, statistical analysis, funding and conflicts of interest; journals that charge fees without genuine peer review, often called predatory journals, should be recognised and avoided as publication venues and treated cautiously as sources.
A narrative review summarises literature selected by the author, which is useful for overviews but vulnerable to bias in selection and interpretation. A systematic review answers a defined question using explicit, reproducible methods to identify, select, appraise and synthesise all relevant studies. It begins with a protocol, ideally registered publicly before the review starts, that states the question, eligibility criteria, search strategy, screening and data extraction procedures, risk-of-bias assessment and synthesis methods.
Questions are often structured with frameworks such as PICO, which defines the population, intervention, comparator and outcome; for example, in healthy adults (population), does daily consumption of fermented milk (intervention), compared with non-fermented milk (comparator), reduce blood pressure (outcome)? Searches cover several databases and grey literature, with full search strings reported. Two reviewers independently screen titles and abstracts and then full texts against eligibility criteria, resolving disagreements, and extract data with standard forms. Each study's risk of bias, the likelihood that its design or conduct distorted its results, is assessed with appropriate tools, such as those for randomised trials or observational studies.
The PRISMA 2020 statement, Preferred Reporting Items for Systematic reviews and Meta-Analyses, provides a 27-item checklist and a flow diagram showing the numbers of records identified, screened, excluded with reasons, and included. Following it makes reviews transparent and reproducible. Scoping reviews map the extent and nature of research on a broad topic without detailed synthesis of effects, and use their own reporting extension. Certainty of evidence for each outcome can be graded with frameworks such as GRADE, which considers risk of bias, inconsistency, indirectness, imprecision and publication bias.
Meta-analysis combines the quantitative results of several studies to estimate an overall effect with greater precision than any single study. Each study contributes an effect size, such as a mean difference, standardised mean difference, risk ratio or correlation, with its variance, and studies are weighted by the inverse of their variance, so larger, more precise studies count more. A fixed-effect model assumes that all studies estimate one true effect and that differences are due only to sampling error; a random-effects model assumes that true effects vary between studies, for example because of different products, populations or methods, and estimates the mean of their distribution, giving wider confidence intervals.
Heterogeneity, variation in true effects between studies, is assessed by Cochran's Q test and quantified by the I² statistic, the percentage of total variation due to heterogeneity rather than chance; values around 25%, 50% and 75% are often described as low, moderate and high. When heterogeneity is substantial, subgroup analyses and meta-regression explore its sources, such as dose, duration or product type, though such analyses are observational and should be pre-specified.
Results are displayed in a forest plot, showing each study's estimate and confidence interval and the pooled estimate as a diamond. Publication bias, the tendency for studies with significant or favourable results to be published more often, can distort meta-analyses; funnel plots, which plot effect size against precision, and statistical tests for asymmetry help detect it, though asymmetry has other causes too. Sensitivity analyses check whether conclusions depend on particular studies or choices. A meta-analysis is only as good as the studies it combines: pooling biased studies produces a precise but biased answer.
Experimental design applies three principles set out by R. A. Fisher: replication, to estimate experimental error and increase precision; randomisation, to avoid systematic bias and justify statistical tests; and local control or blocking, to remove known sources of variation from the error. A fourth practical principle is balance, giving each treatment the same number of replicates where possible, which simplifies analysis and maximises efficiency.
In a completely randomised design (CRD), treatments are assigned to experimental units entirely at random. It suits homogeneous experimental material, such as aliquots of one well-mixed bulk milk tested under different laboratory heat treatments, or identical fermenters in a controlled room. The analysis is one-way analysis of variance; with t treatments and r replicates each, the total degrees of freedom are tr − 1, treatment degrees of freedom t − 1, and error degrees of freedom t(r − 1). For example, comparing four starter cultures with five replicate fermentations each gives 3 treatment and 16 error degrees of freedom.
The design is simple and flexible, allows unequal replication, and gives the maximum error degrees of freedom for a given number of units. Its weakness is that all variation between units enters the error, so if units differ systematically, as cheese vats made on different days with different milk do, the error is large and differences between treatments are harder to detect. Power analysis before the experiment, using an estimate of error variance from pilot data or literature and the smallest difference worth detecting, determines how many replicates are needed; underpowered experiments waste resources because they are unlikely to detect real effects and, when they do find significance, tend to overestimate effect sizes.
When experimental units fall into groups that differ, such as batches of milk collected on different days, different cheese vats, different panellists or different seasons, blocking assigns every treatment once within each group, or block, at random. In a randomised complete block design (RCBD) with t treatments and b blocks, the analysis of variance separates block variation from the error, so the error degrees of freedom are (t − 1)(b − 1). In dairy research, the milk batch is often the most important block: comparing three coagulants using milk from each of six days gives 2 treatment, 5 block and 10 error degrees of freedom, and removes day-to-day variation in milk composition from the comparison.
Blocking pays off when blocks differ substantially; if they do not, it costs error degrees of freedom for little gain. The efficiency of blocking relative to a CRD can be estimated from the analysis. When blocks are too small to contain all treatments, such as when only three vats can be made per day but five treatments are compared, incomplete block designs, such as balanced incomplete block designs, arrange subsets of treatments in each block so that every pair of treatments appears together equally often.
The Latin square design controls two sources of variation at once by arranging t treatments in a t × t grid, with each treatment once in every row and once in every column. It is used, for example, when four cheese-making treatments are applied in four vats over four days, with rows as days and columns as vats, or in sensory trials with panellists and serving positions. Its error degrees of freedom are (t − 1)(t − 2), small for few treatments, so it is often replicated. Crossover designs, in which each subject receives several treatments in sequence, as in human nutrition studies, are related designs that must account for carry-over between periods.
Factorial designs study two or more factors together, with all combinations of their levels, so that main effects and interactions are estimated from the same experiment. A 3 × 2 factorial of three culture levels and two cooking temperatures has six treatment combinations; with four replicate vats each, the analysis separates the main effect of culture (2 degrees of freedom), the main effect of temperature (1), their interaction (2) and the error. Factorials are efficient because every run contributes to estimating every main effect, and they reveal interactions that one-factor-at-a-time experiments miss. Two-level factorials, 2ᵏ designs, and fractional factorials screen many factors economically.
Split-plot designs arise when one factor is applied to large units and another to subdivisions of them. In cheese research, milk pasteurisation temperature might be set for a whole day's milk (the whole plot), while different ripening temperatures are applied to cheeses from that milk (the subplot). The two factors have different error terms: the whole-plot factor is tested against variation between whole plots, which usually has few degrees of freedom, and the subplot factor and the interaction are tested against the within-whole-plot error. Analysing a split-plot experiment as a simple factorial uses the wrong error for the whole-plot factor and exaggerates its significance.
Nested or hierarchical designs arise when units are sampled within units, such as several cheeses within each vat, several samples within each cheese and duplicate analyses within each sample. Variance components analysis estimates how much variation comes from each level, which guides future sampling: if most variation is between vats, analysing more cheeses per vat adds little precision. Mixed-effects models, which treat some factors as fixed and others, such as blocks, vats, panellists or animals, as random, are now the standard tool for analysing such structured data, including repeated measurements over time.
Response surface methodology (RSM) models how a response depends on several continuous factors near an optimum, usually with a second-order polynomial including linear, quadratic and interaction terms. It is used in dairy research to optimise processes, such as the time, temperature and enzyme dose of lactose hydrolysis, and formulations, such as the levels of protein, stabiliser and sugar in a dessert. The research sequence is typically screening with two-level factorial or fractional factorial designs to identify important factors, moving toward the optimum region by steepest ascent if needed, and then fitting a second-order model with a central composite design or a Box–Behnken design.
A central composite design for k factors has 2ᵏ factorial points, 2k axial points and several centre points; for three factors, 8 + 6 + 6 = 20 runs is common. Choosing the axial distance α = (2ᵏ)^(1/4) makes the design rotatable, giving equal prediction precision at equal distances from the centre; face-centred designs with α = 1 keep all runs within the factor ranges. A Box–Behnken design for three factors has 12 edge-midpoint runs plus centre points, often 15 runs in total, and avoids extreme corners. Replicated centre points estimate pure error and allow a lack-of-fit test.
In research, RSM results should be reported with the coded and actual factor levels, the full model with coefficients and their significance, analysis of variance including lack of fit, adjusted and predicted R², residual diagnostics and verification experiments. Claims of optimisation should be confirmed by running the predicted optimum and comparing results with prediction intervals. RSM describes empirical relationships within the studied region and should not be extrapolated beyond it or interpreted as mechanistic.
When the factors are proportions of a mixture that must sum to a constant, such as the fractions of skim milk powder, whey protein concentrate and sodium caseinate in a protein blend, or the proportions of milk, cream, sugar and stabiliser in an ice cream mix, the factors are not independent, and ordinary response surface designs cannot be used. Mixture designs place experimental points within the simplex, the geometric region where all proportions are non-negative and sum to one; for three components, the simplex is a triangle.
The simplex lattice design {q, m} for q components takes each proportion at the values 0, 1/m, 2/m, …, 1, in all combinations that sum to one; the number of points is (q + m − 1)!/(m!(q − 1)!). A {3, 2} lattice has six points: the three pure components and the three binary 50:50 blends. The simplex centroid design includes the pure components, all binary 50:50 blends, all ternary equal blends and so on up to the overall centroid, giving 2^q − 1 points, seven for three components.
Mixture models are usually Scheffé polynomials, which omit the intercept because proportions sum to one; the quadratic Scheffé model for three components has six terms, three linear and three binary interactions, and the special cubic model adds the ternary term. In practice, components have lower and upper limits, such as a stabiliser between 0.1 and 0.5%, which restrict the feasible region to an irregular polygon within the simplex; extreme-vertices designs and computer-generated optimal designs place points within such constrained regions. Mixture–process designs combine mixture components with process variables, such as homogenisation pressure, at the cost of more runs.
Standard designs do not always fit research constraints: some factor combinations may be impossible or unsafe, the budget may limit the number of runs, factors may be categorical as well as continuous, or blocking may be needed because runs span several days. Optimal designs, generated by algorithms, choose runs from a candidate set to best estimate a specified model. D-optimal designs maximise the determinant of the information matrix, minimising the joint confidence region of the coefficients; I-optimal designs minimise the average prediction variance over the region, which suits optimisation studies where prediction matters more than coefficients. Optimal designs depend on the assumed model, so including a few extra runs to check lack of fit is wise.
Practical issues often decide whether a designed experiment succeeds. Factor ranges must be wide enough to produce detectable effects but not so wide that the process fails or the response becomes discontinuous. Hard-to-change factors, such as pasteurisation temperature in a pilot plant, lead to split-plot structures that should be recognised in design and analysis. Runs should be randomised in order, or blocked by day or batch of raw milk, to avoid confusing time trends with factor effects. Measurement error should be small relative to the effects of interest, so methods are validated first.
Several responses are usually measured: yield, composition, texture, sensory scores, cost, shelf-life indicators. Multi-response optimisation by overlaid contour plots or desirability functions finds compromise conditions, and the choice of weights or targets should reflect the research or product objectives explicitly. Reporting the design table, the raw data and the analysis allows others to reproduce and build on the work.
Dairy research often measures many variables on each sample: dozens of fatty acids, hundreds of volatile compounds, thousands of spectral wavelengths, or many sensory attributes. Multivariate methods analyse these variables together, taking their correlations into account. Principal component analysis (PCA) transforms the original correlated variables into new uncorrelated variables, principal components, which are linear combinations of the originals, ordered so that the first component explains the largest possible share of the total variance, the second the largest share of the remainder, and so on. A few components often capture most of the information, reducing dimensionality and revealing structure.
When variables are measured in different units or have very different variances, such as fat content in percent and volatile compounds in micrograms per kilogram, they are usually standardised to zero mean and unit variance before PCA, so that the analysis is based on the correlation matrix and each variable has equal weight; otherwise variables with large variances dominate. Spectral data, measured on a common scale, may be analysed without standardisation, often after preprocessing such as smoothing, derivatives or scatter correction.
Results are interpreted through eigenvalues, the variance explained by each component; scores, the coordinates of samples on the components, plotted to show groupings, trends and outliers; and loadings, the weights of original variables on each component, showing which variables drive each pattern. A biplot shows scores and loadings together. The number of components to retain is chosen by the scree plot, looking for the point where eigenvalues level off, by the Kaiser criterion of retaining components with eigenvalues above one for standardised data, by the cumulative variance explained, or by cross-validation. PCA is exploratory: it describes the main patterns of variation but does not test hypotheses or guarantee that those patterns relate to the question of interest.
Cluster analysis groups samples so that those within a group are similar and those in different groups are dissimilar, without predefined classes. It is used to find natural groupings, such as types of farms by milk composition, cheeses by flavour profile, consumers by liking patterns, or microbial isolates by fingerprints.
Clustering depends on how similarity is measured. Euclidean distance, the straight-line distance between samples in multivariate space, is the most common; Manhattan distance sums absolute differences; correlation-based distances focus on profile shape rather than magnitude; and specialised measures such as Bray–Curtis dissimilarity are used for community data. Variables are usually standardised first, for the same reasons as in PCA.
Hierarchical agglomerative clustering starts with each sample as its own cluster and repeatedly merges the two closest clusters until one remains, recording the sequence in a tree diagram called a dendrogram. The distance between clusters is defined by a linkage method: single linkage uses the closest pair of members and tends to produce long chains; complete linkage uses the farthest pair and produces compact clusters; average linkage uses the mean distance; and Ward's method merges the pair that least increases total within-cluster variance, tending to produce clusters of similar size. Cutting the dendrogram at a chosen height gives the clusters. Partitioning methods such as k-means start from k initial centres, assign samples to the nearest centre, recalculate centres and iterate; the number of clusters must be chosen, using criteria such as the elbow in within-cluster variance or the average silhouette width, and results depend on starting points, so runs are repeated.
When samples belong to known classes, such as milk from different regions or cheeses of different ages, supervised methods build models that discriminate between them and classify new samples. Linear discriminant analysis finds linear combinations of variables that best separate classes, but requires more samples than variables; partial least squares discriminant analysis (PLS-DA) and its orthogonal variant handle many correlated variables, as in spectroscopy and omics data; and machine-learning classifiers such as random forests and support vector machines are increasingly used. For predicting continuous variables, such as composition from spectra, partial least squares regression is standard.
Supervised models with many variables can easily overfit, appearing to discriminate perfectly while capturing noise. Validation is therefore essential: models are tested by cross-validation and, preferably, on an independent test set of samples not used in building them, collected at a different time if possible; permutation tests, refitting models with randomly shuffled class labels, show whether apparent discrimination exceeds chance. Performance is reported as classification accuracy, sensitivity and specificity, or for regression as the root mean square error of prediction.
Good practice in multivariate analysis includes checking data quality and outliers before analysis, documenting preprocessing choices, reporting the variance explained and validation results, interpreting patterns with knowledge of the samples rather than mechanically, and avoiding claims of causation from correlations. Raw data and code should be shared where possible so that others can reproduce analyses. Multivariate methods are tools for understanding complex data, not substitutes for sound experimental design: confounded sampling, such as collecting all samples of one class in one season, will produce patterns that reflect the confounder rather than the class.
Sensory and consumer data are central to dairy product research, but they are noisy, because people differ and vary from day to day, and they are vulnerable to bias. Research-grade sensory studies therefore apply the experimental design principles of earlier chapters to people. Samples are coded with random three-digit numbers, presented in balanced orders that control first-order carry-over effects, such as Williams Latin square designs, and served under controlled conditions. Trained panels are replicated, typically in two or three sessions, so that assessor performance and product differences can be separated from noise. Consumer studies recruit target users in sufficient numbers, often 100 or more per segment for liking studies, with sample sizes justified by power analysis.
The research question determines the method. Discrimination tests ask whether a perceptible difference exists; descriptive analysis characterises the sensory profile; hedonic tests measure liking; and rapid methods such as check-all-that-apply questions, projective mapping and temporal dominance of sensations offer cheaper or faster alternatives with different strengths. Combining methods, for example descriptive profiles from a trained panel with liking from consumers, allows researchers to identify the sensory drivers of liking.
Ethical requirements apply: participants give informed consent, are screened for allergies and dietary restrictions, are told what they will taste, and test only products that are safe; studies are reviewed by an ethics committee where institutional policy requires it, which is common for any research involving human participants.
Discrimination test results are analysed with the binomial distribution, comparing the number of correct responses with the number expected by chance, such as one third for triangle tests; for similarity testing, the design must control the risk of missing a meaningful difference, often specified as a maximum proportion of discriminators. Thurstonian models convert results into d′, a measure of perceptual difference comparable across test methods, and replicated discrimination data are analysed with models that account for differences between assessors, such as beta-binomial models.
Descriptive analysis data are analysed attribute by attribute with mixed-model analysis of variance: product and session as fixed effects, assessor as a random effect, and the product × assessor interaction as a random effect that serves as the error term for testing products when results are to be generalised beyond the particular assessors. Panel performance is evaluated by each assessor's ability to discriminate products, their repeatability across replicates and their agreement with the panel; graphical tools and indices summarise these, and poorly performing assessors are retrained. Multivariate analyses, such as PCA of product means and generalised Procrustes analysis of individual configurations, summarise product spaces.
Time-intensity and temporal dominance data are analysed as curves, with parameters such as maximum intensity, time to maximum and duration, or as dominance rates over time, which require methods that respect their temporal and compositional nature.
Hedonic ratings, usually on the nine-point hedonic scale or similar, are analysed by analysis of variance with consumers as a random blocking factor, followed by multiple comparisons. Because consumers differ in what they like, averages can hide segments with opposite preferences; cluster analysis of individual liking patterns identifies such segments, which are then described by demographics, consumption habits and attitudes.
Preference mapping links liking to sensory characteristics. Internal preference mapping performs PCA on the consumer-by-product liking matrix, showing the main dimensions of preference and the direction each consumer prefers; external preference mapping regresses each consumer's liking on the sensory dimensions obtained from descriptive analysis, indicating which sensory directions increase liking and where an ideal product might lie. Partial least squares regression can relate sensory attributes to liking directly.
Diagnostic questions add information. Just-about-right scales are analysed by penalty analysis, which calculates the mean drop in liking among consumers who rate an attribute too weak or too strong, weighted by the proportion of consumers in each group, to prioritise reformulation. Check-all-that-apply data are analysed by Cochran's Q test for differences in citation frequencies between products, by correspondence analysis to map products and attributes, and by penalty-lift analysis relating attribute citation to liking. Ranking data are analysed with the Friedman test. Purchase intent and willingness-to-pay measures connect liking to market behaviour, with known biases toward overstatement. Reporting should include sample sizes, recruitment criteria, test conditions, statistical models and effect sizes, not just significance.
Most scientific papers and theses follow the IMRaD structure: Introduction, Methods, Results and Discussion, preceded by a title and abstract and followed by conclusions, acknowledgements, references and supplementary material. The title should be specific and informative, the abstract a concise summary of the problem, methods, main results with key numbers and the conclusion, because many readers read only these.
The introduction establishes the context and the gap, moving from the general problem to the specific question, cites the most relevant literature, and ends with the aims or hypotheses. The methods section describes materials, experimental design, procedures, analytical methods with their references and validation, and statistical analysis in enough detail for another researcher to repeat the work: the source and composition of milk, the culture strains and their origin, equipment settings, numbers of replicates and what constitutes the experimental unit, and the software used. The results section presents findings logically, using tables and figures that are self-explanatory, with means, measures of variability and statistical results, without interpretation. The discussion interprets the results, compares them with previous work, explains mechanisms, acknowledges limitations and states implications, without overstating conclusions.
A thesis follows the same logic at greater length, often with a comprehensive literature review chapter and several research chapters, sometimes written as publishable papers, plus a general discussion. Clear, plain, precise writing, consistent use of units and terms, and careful citation are essential. Reporting guidelines, such as those for randomised trials (CONSORT), observational studies (STROBE) and animal research (ARRIVE), specify what must be reported and are increasingly required by journals.
Choosing a journal involves matching the paper's topic and audience to the journal's scope, considering its readership, reputation, review speed, open-access options and costs. Journal impact metrics measure average citations but say little about the quality of individual papers; many research assessment initiatives encourage evaluating research on its own merits rather than by the journal where it appears.
Authorship should reflect substantial contribution. Widely used criteria, such as those of the International Committee of Medical Journal Editors, require substantial contributions to conception, design, data acquisition or analysis; drafting or critically revising the work; approval of the final version; and agreement to be accountable for the work. People who contributed in other ways, such as providing samples or technical help, are acknowledged rather than listed as authors. Gift, guest and ghost authorship are unethical.
Manuscripts undergo peer review, in which experts evaluate the work for the editor; review may be single-blind, with reviewers anonymous, double-blind, with authors also anonymous, or open. Authors respond to reviewers point by point, revising the manuscript where comments are valid and explaining politely where they disagree. Rejection is common and not necessarily a judgement on quality; many papers are published after revision or in another journal. Preprints, manuscripts posted publicly before peer review, allow rapid sharing, with the caveat that they have not yet been reviewed. Open-access publishing makes papers freely available, funded by article charges, institutional agreements or other models; many funders require it.
Grant proposals request funding for research from public agencies, foundations, industry or international programmes. Although formats vary, most proposals contain a summary; background and justification establishing the importance of the problem and the gap; aims or objectives, often two to four specific aims; preliminary data showing feasibility; the research plan or approach, describing design, methods, analysis and expected outcomes for each aim; innovation, explaining what is new; a timeline or work plan with milestones and deliverables; a description of the team and its expertise; a budget with justification; risk assessment with alternative approaches if methods fail; and expected impact, including pathways to benefit for industry, farmers, consumers or policy. Logical frameworks or theories of change are often required for development-oriented proposals.
Successful proposals are clear, focused and realistic. Reviewers look for an important question, a feasible and rigorous plan, a capable team, value for money and impact. Common weaknesses include aims that are too broad or interdependent so that failure of one dooms the rest, insufficient detail on methods or statistics, no power analysis, unrealistic timelines and budgets, and failure to address the funder's priorities and criteria. Reading the call carefully, following its format, and having colleagues review drafts improve chances.
Many funders now require a data management plan describing what data will be generated, in what formats, how they will be documented, stored, backed up and protected, who will have access, how they will be shared and preserved after the project, and how ethical and legal requirements, such as consent and confidentiality, will be met. Planning data management at the start makes research more efficient and reproducible.
Research integrity means conducting research honestly, accurately, transparently and responsibly. Research misconduct is commonly defined as fabrication, making up data or results; falsification, manipulating materials, equipment, processes, data or images so that the research is not accurately represented; and plagiarism, appropriating another person's ideas, processes, results or words without credit. Misconduct damages science, misleads other researchers and the public, wastes resources and, in food and health research, can harm people. Institutions investigate allegations under formal procedures, and journals retract papers found to be unreliable, guided by publication ethics bodies such as the Committee on Publication Ethics.
Beyond outright misconduct, questionable research practices undermine reliability. They include selective reporting of outcomes or experiments that gave favourable results; p-hacking, trying many analyses until a significant result appears; HARKing, hypothesising after the results are known and presenting the hypothesis as if it had been formed in advance; excluding outliers without pre-specified criteria; inadequate replication; and failure to disclose conflicts of interest. Pre-registration of hypotheses and analysis plans, reporting all outcomes, sharing data and code, and clear distinction between confirmatory and exploratory analyses guard against them.
Conflicts of interest arise when personal, financial or institutional interests could influence, or appear to influence, research. Funding by dairy, ingredient or competing industries is common and legitimate, but must be disclosed, and contracts should protect researchers' freedom to publish results whatever they show.
Research involving human participants, including sensory panels, consumer studies, nutrition trials and surveys of farmers, must respect their autonomy, welfare and privacy. Principles set out in international documents, such as the World Medical Association's Declaration of Helsinki for medical research, require that research be reviewed and approved by an independent ethics committee, that participants give voluntary informed consent based on clear information about the purpose, procedures, risks, benefits and their right to withdraw at any time without penalty, that risks be minimised and justified by potential benefits, that vulnerable groups receive special protection, and that personal data be kept confidential. Food studies must also ensure the safety of test products and screen participants for allergies and intolerances.
Research involving animals, such as dairy cows in feeding or health trials, or laboratory animals in studies of probiotics or bioactive compounds, must follow the principles of the 3Rs, first articulated by Russell and Burch in 1959: replacement of animals with alternatives where possible, such as in vitro models; reduction of the number of animals to the minimum needed for valid results, through good design and power analysis; and refinement of procedures to minimise pain and distress and improve welfare. Animal studies require approval by an animal ethics committee under national law, and should be reported following guidelines such as ARRIVE so that their methods and results can be evaluated and reproduced.
Research with farmers and communities, especially in low-income settings, raises additional ethical issues: fair benefit sharing, avoiding exploitation of local knowledge or genetic resources, returning results in useful form, and respecting local customs and data ownership.
Reliable research rests on reliable data. Laboratory notebooks, paper or electronic, should record what was done, when, by whom, with which materials and instruments, and what was observed, contemporaneously and permanently, with corrections made so that the original entry remains legible and the change is dated and initialled. Raw data, such as instrument files, chromatograms, images and plate counts, should be retained in their original form, backed up, and linked to the processed results and analyses derived from them. The ALCOA+ principles, data that are attributable, legible, contemporaneous, original and accurate, as well as complete, consistent, enduring and available, summarise the expectations widely applied in regulated and research laboratories.
Image manipulation is a particular risk. Adjusting brightness or contrast uniformly across a whole image is generally acceptable, but enhancing, obscuring, moving or removing features, or splicing images without indication, is falsification. Journals increasingly screen images and ask for originals.
Data sharing improves reproducibility and allows reuse. The FAIR principles state that research data should be findable, with persistent identifiers and rich metadata; accessible, retrievable by standard protocols; interoperable, using standard formats and vocabularies; and reusable, with clear licences and provenance. Data are deposited in institutional, subject-specific or general repositories, with sequence data in public sequence databases as many journals require. Restrictions apply to personal data, commercially confidential information and data subject to agreements, and must be planned in the data management plan. Retention periods, often several years after publication or as required by funders and institutions, ensure that data remain available for verification.
Intellectual property (IP) rights give creators control over the use of their creations for a period, encouraging investment in innovation. Patents protect inventions, new products or processes that solve technical problems. Trade secrets protect confidential information that gives a business advantage, such as recipes, process conditions and culture strains, for as long as secrecy is maintained, with no registration. Trademarks protect brand names and logos that distinguish goods. Copyright protects original works of authorship, such as texts, software code and images, from the moment of creation. Industrial designs protect the appearance of products, such as packaging shapes. Plant variety rights protect new plant varieties. Geographical indications protect names of products whose qualities are linked to a place, which is important for many traditional cheeses and other dairy products with protected designations of origin.
International agreements set common minimum standards. The Paris Convention for the Protection of Industrial Property establishes a priority right: an applicant who files a patent application in one member country has twelve months to file in other countries while keeping the original filing date. The Patent Cooperation Treaty (PCT), administered by the World Intellectual Property Organization, allows a single international application that preserves the right to seek patents in many countries, with national phase entry usually at 30 months from the priority date. The agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS) of the World Trade Organization requires members to grant patents for inventions in all fields of technology, subject to permitted exclusions, with a minimum term of twenty years from the filing date.
Patents are territorial: a patent granted in one country gives no rights in another, so applicants choose countries according to markets and manufacturing.
To be patentable, an invention must generally be new (novel), not part of the prior art, everything made available to the public anywhere before the filing date; involve an inventive step, or be non-obvious to a person skilled in the field; and be capable of industrial application, or useful. It must also fall within patentable subject matter: discoveries, scientific theories, natural phenomena and mathematical methods as such are excluded in most systems, and some countries exclude further categories, such as methods of medical treatment, plant and animal varieties or traditional knowledge; the exact rules differ.
Novelty is destroyed by prior disclosure, including the inventor's own publication, conference presentation, thesis or poster. Most patent systems use a first-to-file principle with absolute novelty, so publication before filing can make an invention unpatentable; a few systems allow a limited grace period for the inventor's own disclosures, but relying on it is risky. Researchers who may have patentable results should therefore consult their technology transfer office before publishing, and file first, which can usually be done quickly, then publish.
A patent application contains a description that discloses the invention in enough detail for a skilled person to carry it out, and claims that define the scope of protection; drafting claims well is a specialised skill. For inventions involving micro-organisms that cannot be described adequately in words, such as a new phage-resistant starter strain, the Budapest Treaty allows deposit of the strain in a recognised international depositary authority to satisfy disclosure requirements. Applications are examined against the prior art; granted patents are published, which makes the technology public in exchange for exclusive rights. Patent databases are rich sources of technical information and are searched before research, to avoid duplication, and before commercialisation, in freedom-to-operate analyses that check whether a product would infringe others' patents.
Technology transfer moves knowledge and inventions from research organisations to industry and society, where they create value. Universities and research institutes usually have technology transfer offices that evaluate inventions disclosed by researchers, decide whether to protect them, market them to potential licensees, negotiate agreements and manage income, often shared with inventors under institutional policies.
Licensing is the commonest route. A licence grants another party the right to use IP in exchange for payment, such as upfront fees, milestone payments and royalties on sales. Exclusive licences give one licensee sole rights, often in a field or territory, which may be needed to justify its investment; non-exclusive licences allow several users. Spin-off companies, founded to develop a technology, are another route, especially for platform technologies. Collaborative research agreements with industry define ownership of results, publication rights, confidentiality and licensing terms in advance; confidentiality or non-disclosure agreements protect information shared during discussions; and material transfer agreements govern the exchange of research materials, such as culture strains, cell lines or samples, specifying permitted uses and rights to results.
Access to genetic resources, including micro-organisms isolated from traditional dairy products in other countries, is governed by the Convention on Biological Diversity and its Nagoya Protocol, which require prior informed consent of the providing country and mutually agreed terms for sharing benefits arising from their use; researchers collecting or using such strains must comply with the access and benefit-sharing rules of the countries involved. Open science and IP are not incompatible: many institutions publish freely after filing, license non-exclusively for public benefit, or place some results in the public domain, choosing the approach that best serves the impact of the research.
Laboratory safety starts with risk assessment: identifying hazards, evaluating who could be harmed and how, and deciding on controls, recorded before work begins and reviewed when procedures change. Controls follow the hierarchy of controls, from most to least effective: elimination of the hazard; substitution with a less hazardous material or method; engineering controls, such as fume hoods and enclosures; administrative controls, such as procedures, training and signage; and personal protective equipment, such as laboratory coats, gloves and eye protection, which is the last line of defence.
Chemicals are classified and labelled under the United Nations Globally Harmonized System, with pictograms, signal words, hazard statements and precautionary statements, and each hazardous chemical has a safety data sheet with sixteen standard sections covering hazards, composition, first aid, firefighting, spill response, handling and storage, exposure controls, properties, stability, toxicology, ecology, disposal, transport and regulation. Dairy laboratories use several hazardous chemicals routinely: concentrated sulfuric acid and amyl alcohol in the Gerber fat test; diethyl ether and petroleum ether, highly flammable and volatile, in solvent extraction methods; concentrated sulfuric acid, catalysts and corrosive fumes in Kjeldahl digestion; sodium hydroxide; and various preservatives, reagents and solvents for chromatography. These are handled in fume hoods with appropriate protective equipment, stored by compatibility, with flammables in approved cabinets away from oxidisers and ignition sources, and disposed of through hazardous waste routes.
Other physical hazards include compressed gas cylinders, which must be secured; liquid nitrogen, which causes cryogenic burns and can displace oxygen in enclosed spaces; centrifuges, which must be balanced and have intact rotors; autoclaves, with risks of burns and pressure; ultraviolet light; and sharps. Emergency preparedness includes eyewash stations, safety showers, fire extinguishers, spill kits, first-aid training and clear procedures.
Biosafety protects people and the environment from biological agents handled in the laboratory. Micro-organisms are classified into risk groups, usually 1 to 4, according to their ability to cause disease in healthy humans, the severity of disease, the availability of prevention and treatment, and the risk of spread in the community. Most dairy starter and probiotic cultures and indicator organisms are in risk group 1. Many foodborne pathogens studied in dairy research, such as Listeria monocytogenes, Salmonella, pathogenic Escherichia coli, Staphylococcus aureus and Bacillus cereus, are in risk group 2. Some, such as Brucella species and Mycobacterium bovis, are in risk group 3 because they cause serious disease and are transmitted by aerosols, and require higher containment.
Laboratories are designed and operated at matching biosafety or containment levels, with increasing requirements for facilities, equipment and practices: restricted access, hand-washing, decontamination of surfaces and waste, use of biological safety cabinets for procedures that generate aerosols, and, at level 3, directional airflow, sealed rooms and respiratory protection. Biological safety cabinets protect people and materials: class I cabinets protect the user and environment but not the sample; class II cabinets, the most common, protect user, sample and environment with HEPA-filtered downflow and exhaust; and class III cabinets are sealed glove boxes for the highest-risk work. Laminar-flow clean benches protect only the sample and must never be used with pathogens.
The fourth edition of the World Health Organization's Laboratory Biosafety Manual, published in 2020, emphasises a risk-based approach, tailoring controls to the specific risks of each procedure rather than relying only on fixed levels. Work with genetically modified organisms is regulated separately in many jurisdictions, requiring risk assessment, notification or approval, and containment.
Rules and equipment protect people only when they are used consistently, which depends on safety culture: leaders who model safe behaviour, training before work begins, supervision of new researchers, open reporting of incidents and near misses without blame, and learning from them. Students should not work alone with hazardous materials, should know the emergency procedures, and should feel able to stop work they consider unsafe.
Waste management prevents harm beyond the laboratory. Biological waste, such as cultures, plates and contaminated materials, is decontaminated, usually by autoclaving, before disposal; sharps go into puncture-resistant containers; chemical waste is segregated by type and disposed of through licensed routes, never poured down drains unless permitted; and radioactive materials, if used, follow specific regulations. Records of waste disposal are kept.
Biosecurity complements biosafety by protecting valuable or dangerous biological materials from loss, theft, misuse or intentional release. It includes inventories of cultures and pathogens, controlled access to storage freezers and laboratories, secure transport of samples under international and national rules for infectious substances, and vetting of personnel where required. Culture collections and strain inventories in dairy research laboratories, which may include pathogens and commercially valuable proprietary strains, should be documented and secured. Finally, safety extends to field and plant work: researchers sampling on farms or in factories must follow site rules, use appropriate protective equipment, observe hygiene to avoid carrying contamination between sites, and respect animal welfare and biosecurity measures that protect herds from disease.
Write your research proposal as if for a funder: a precise question, objectives tied to experiments, a design with stated experimental units and a power analysis, and a data management plan. Run a small systematic search with a documented search string and a PRISMA-style flow. Analyse a real multivariate dataset with PCA and clustering and a sensory dataset with a mixed model. Keep a proper laboratory notebook from the first day, complete the safety training for every method you use, and ask about patentability before presenting results at a conference.
A research scientist five years after the doctorate designs studies that use resources efficiently, supervises students and technicians, writes papers and grant applications, reviews manuscripts, manages collaborations with companies under agreements that protect publication and IP, and serves on ethics or safety committees. Typical decisions concern the right design for a constrained budget, how to handle unexpected results honestly, whether an invention should be patented before publication, and what containment a new pathogen study requires before any laboratory work begins.
A research director or industry R&D head with 25 years' experience sets research strategy, allocates budgets, builds teams and infrastructure, negotiates large collaborations and licences, and is accountable for integrity, ethics and safety across the organisation. They decide which inventions to patent and where, how to balance open publication with commercial protection, how to respond to allegations of misconduct, and how to maintain a safety culture in which incidents are reported, investigated openly and learned from rather than hidden or punished.
A professor with 25 years' experience trains the next generation in research methods, examines theses, edits journals, chairs ethics and biosafety committees, and contributes to national and international research policy. They model integrity, insisting on pre-specified analyses, honest reporting and data sharing, and they mentor students in writing, in navigating intellectual property, and in building collaborations across disciplines and countries that keep dairy science relevant and trustworthy. Their examining and refereeing set the standard that younger researchers learn to meet.
National rules on research ethics review, animal research, intellectual property details (grace periods, excluded subject matter), and laboratory safety differ between countries and institutions; only international principles are given. Statistical methods are introduced at the level needed to choose and interpret them; detailed computation is covered in mathematics-statistics-dairy and specialist texts. Bayesian methods, qualitative research methods and advanced mixed-model theory are not covered. Database and software products are described generically and not taught.
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.