What each analysis method can actually answer, the steps it needs to hold, and where it falls apart if you leave them out

Most lists of analysis types are catalogues. They put a descriptive mean next to a SWOT (Strengths, Weaknesses, Opportunities, Threats), a regression next to a gravimetric titration, and call the result comprehensive. The mixing is not harmless. It lets a correlation stand in for a cause, a map stand in for a mechanism, and a workshop adjective stand in for evidence. Once the label is chosen, the work often stops.

This is not a catalogue. Each method appears once. For each, three things: what question it can answer, the minimum steps that make the result defensible, and where it breaks if a step is skipped. The steps are not a laboratory protocol for every field. They are the floor under which the result cannot be defended.

The same analysis can serve more than one purpose. The criterion for choosing is the question, the data that exist, and the decision the result has to support. Methods that measure a physical signal do not belong in the same class as methods that support a decision. Treating them as peers is how a spectral decomposition ends up next to a cost-benefit ratio, as if both were “just analysis.”

By purpose

Purpose is the first cut, because it decides what the other methods are allowed to claim.

Descriptive

It answers what happened and what the data look like, without explaining why and without saying what will follow. It is the base the other analyses stand on. If the description is wrong, the diagnosis and the prediction start from a false picture. It is not a lesser stage. In many decisions, a correct description of volume, composition, and dispersion is exactly the result required.

Define the population, the period, and the unit of observation. Check what each variable means and how it was measured. For categories, calculate frequencies and proportions. For numbers, calculate mean, median, spread, percentiles, and extremes. Draw the distribution, not only a table of means. Separate the subgroups that can hide different averages. Note how much data is missing and whether the missingness collects in one subgroup. The result is a description that can be read without a model: how many, how much, how they are distributed, what is typical, and what is rare.

It does not establish cause. A mean can hide two groups. Comparing two means without an interval or without the size of the difference leaves an impression of certainty the data do not have.

Exploratory

It looks for patterns, anomalies, and possible relations when no hypothesis was fixed in advance. Its role is to generate good questions, not to confirm them. It is appropriate at the start of a new data set, after an incident that is not understood, or when existing theory does not say what to look for.

Inspect each variable, then relations two at a time, with graphs before coefficients. Look for subgroups, sudden changes, values that dominate the result, and variables that move together. Note the pattern, on which cases it appears, and what could produce it other than a real cause: measurement error, selection, a third variable. Separate discoveries from tests. If a pattern is worth keeping, formulate a hypothesis and check it on data that were not used in the search.

If enough relations are searched, some will appear by chance. A pattern found and confirmed on the same data is not a confirmation.

Confirmatory

It checks a hypothesis written before the result was seen. The difference from exploration is not the formula. It is the moment the question was chosen. This is the analysis that allows you to say the data support or do not support a pre-specified claim.

Write the hypothesis, its direction if there is one, the measured outcome, the threshold, and the stopping rule before analysis. Check that the sample and the measurement match the hypothesis. Apply the announced test. Report the estimate, the interval, and the effect size, not only whether the threshold was crossed. Analyses that appear after the result was seen are labelled separately, as explorations.

Changing the outcome, the subgroup, or the test after seeing the data turns a confirmatory analysis into an exploratory one, even if the table looks the same.

Diagnostic and causal

It explains why a result occurred. In operations it takes the form of root-cause work: 5 Why and the Ishikawa diagram. In research it requires a design that separates cause from mere association. A good diagnosis closes on the cause that, once removed, makes the problem disappear, not on the list of every factor that was present.

Describe the problem in observable fact: what, where, when, how large, against what benchmark. Separate the symptom from the cause. On the Ishikawa, list candidate causes across people, method, equipment, material, environment, and measurement. Apply 5 Why only to a chain that can be checked, not to opinions. Test each cause: was it present when the problem appeared and absent when the problem did not? In research, add the comparison with a group or a period in which the cause did not act, and control the confounders. Stop at the cause that can be acted on and whose removal is followed by the disappearance of the problem.

It breaks when a correlated cause is taken as the cause. The chain of “why” can stop too early, at an intermediate symptom, or it can invent a culprit when the data show a system.

Comparative

It establishes similarities and differences between cases, groups, periods, places, or policies. It is not a parallel placement of two descriptions. It is useful when the decision depends on what differs, not only on how each side looks.

Choose the cases and the dimensions before reading the conclusion. The dimensions must be observable in all cases, under the same definition. Align units, periods, and inclusion rules. For each dimension, report the value, the difference, and whether the difference is large relative to the variation inside the case. In qualitative comparisons, say what is shared, not only what contrasts. Keep the negative cases, in which the expected result does not appear.

Cases chosen because they resemble the desired conclusion are not a comparison. Different definitions of the same indicator produce false differences.

Predictive

It estimates what is likely to happen in a declared horizon, on the basis of patterns already observed. Prediction does not explain the mechanism. A model can forecast well and still not say why.

Fix the horizon, the unit, and what happens if no new decision is taken. Choose a model to the measure of the data: moving average, regression, time series, or a more flexible model. Training uses only the past available at the moment of prediction. Validation is done on a later period or on later cases, untouched during training. Report the typical error, the cases in which it errs badly, and the interval, not a single figure. If the world has changed, the model is re-evaluated, not extended mechanically.

Fit on the training data overestimates performance. A precise short-term prediction is not extended to a long horizon without data.

Prescriptive and decisional

It recommends an action. It uses description, diagnosis, and prediction, but adds alternatives, a criterion, and constraints. Without an explicit criterion, the recommendation is a preference, not an analysis.

Write the real alternatives, including doing nothing. Choose the criterion: cost, time, risk, equity, or a combination with weights. Estimate the result of each option under uncertainty, not in a single scenario. Check the constraints that eliminate options before optimisation. Do a sensitivity check: at what change of parameter does the recommendation change. Say what information, if known, would overturn the decision.

Optimising an unimportant criterion, or ignoring a legal or capacity constraint, produces a recommendation that cannot be executed.

Evaluative

It establishes whether an intervention worked, for whom, and with what effect. It differs from monitoring, which only tracks indicators. Evaluation needs a comparison with what would have happened without the intervention.

Fix the outcome and the definition of success before the intervention. Describe what was actually delivered, not only what was in the plan. Build the comparison: an unexposed group, a prior period, or a stronger design. Estimate the difference and the interval. Check unintended effects and whether the result holds after the intervention stops. Separate failure of implementation from failure of the idea.

Before-after, without a comparison group, confuses the intervention with the trend that would have existed anyway.

Impact

It estimates the consequences of a decision, policy, or project, including those that were not the aim. It can be done before, as an estimate, or after, as a measurement. It includes who gains, who loses, and which effects appear late.

Draw the chain from decision to direct, indirect, and distant effects. Identify the affected groups. Quantify the effects that can be measured, with source and uncertainty. Name the effects that cannot be measured; do not set them to zero. Separate the short term from the long term, and the average effect from the effect on the most exposed. Compare with the alternative, not with an imaginary state without costs.

Counting only the announced benefits, or treating unmeasured effects as nonexistent, underestimates the consequences.

Synthesis

It brings together the literature or the results of several sources to say what is known, where the sources converge, and where they contradict. It does not combine effects statistically. That is meta-analysis, treated separately.

Write the question and the inclusion rules before searching. Search in several places and note what was excluded. From each source extract the design, the result, the size, and the limits, not only the author’s conclusion. Group by question and by quality of evidence. Keep the contradictions and try an explanation: different populations, different measures, different interests. Say what has not been studied.

A synthesis that cites only the works supporting the desired conclusion does not summarise the literature.

Quantitative and statistical

These methods share a dependence on numbers and on rules for moving from numbers to claims. They do not share a right to causal language.

Univariate

It describes a single variable, before any relation. It shows whether the variable is usable: what values it takes, how often, whether it has errors, and whether a mean makes sense.

Identify the scale: category, order, interval, or ratio. For categories, make the frequency distribution, including the missing category. For numbers, calculate mean, median, spread, minimum, maximum, and a few percentiles, then a histogram or a diagram. Look for impossible values, duplicates, and coding jumps. Only after that decide whether the variable can enter an analysis with others.

A variable with almost all values identical, or with an uncorrected coding error, ruins any model it is entered into.

Bivariate

It links two variables. It is the first place a relation appears, and the first place the relation can be misleading. It serves to choose the later model and to check that an association exists at least in the raw data.

Choose the method by scales. Two categories: a contingency table and a measure of association. Category and number: means or distributions by group. Two numbers: a scatter plot and a coefficient. Look at the shape, not only the coefficient: a curved relation, separate groups, a point that pulls the line. Note whether a known third variable could produce the picture.

A two-variable relation disappears or changes sign when another factor is held constant. The graph is obligatory.

Multivariate

It analyses several variables at once in order to separate their effects. It is necessary when an outcome has several possible causes and when a raw comparison mixes different groups. It does not create causality merely because there are several predictors.

Write in advance which variable is the outcome and which are the predictors, including those entered only to hold a confounder constant. Check how many observations there are per parameter, collinearity, and values that change the coefficients. Estimate the model. Read each coefficient as the difference associated with that predictor when the others in the model are held constant. Check whether adding or removing an important predictor changes the conclusion.

A model with too many predictors on too few cases produces unstable coefficients. The omitted variable remains inside the coefficients of the others.

Descriptive statistics

It summarises a set without generalising to a wider population. It is the correct analysis when you have the whole universe — all employees of a firm, all transactions of a month — and not a sample from which you want to infer.

Choose summaries that do not distort: median and percentiles if there are extremes, mean and spread if the distribution is roughly symmetric, proportions for categories. Give the denominators. Separate periods and segments that are not under the same regime. Graphs keep the scale and do not cut the axis to exaggerate a small difference.

On a complete universe, a significance test answers nothing: the difference is already the real one. The question is whether it is large.

Inferential

It says how plausible it is that a sample result reflects the population from which the sample was drawn. It produces intervals and tests, not the truth of the population. It makes sense only if the sample and the population are defined.

Define the population and the sampling rule. Check non-response. Choose the estimator and the test by design: simple, stratified, or clustered sample. Report the estimate, the interval, and the effect size. Read the p-value together with the size, not instead of it. If the design is clustered, calculate the error at the correct level.

A convenience sample does not become representative through a test. Significance on a huge sample can accompany a negligible difference.

Correlation

It measures how tightly two variables move together and in which direction. Pearson assumes a linear relation on numbers. Spearman uses ranks. Coefficients for categories are different. None of them says which variable produces the other.

Make the graph. Choose the coefficient by scale and shape. Check whether a point or a subgroup produces almost the whole coefficient. Report the coefficient, the interval, and the number of complete pairs. If the relation is curved, declare the linear coefficient unsuitable and describe the shape.

Correlation is symmetric. A variable and a transformation of it, or two effects of a common cause, can be strongly correlated without one determining the other.

Regression

It estimates how much an outcome changes on average when a predictor changes, holding the predictors in the model constant. Linear for numeric outcomes, logistic for yes or no, Poisson or negative binomial for counts, multilevel when observations sit in groups, quantile when what matters is not the mean but a part of the distribution.

Specify the outcome, the predictors, and why each is in the model. Check the shape, influential values, collinearity, and, in logistic regression, separation. Estimate. Report the coefficient in intelligible units, the interval, and the fit. For prediction, measure the error on data held out. For explanation, say what might remain uncontrolled. Put interactions only if there is a reason, then interpret the conditional effect, not the coefficient alone.

A coefficient is not the causal effect merely because the model adjusted for it. Extrapolation outside the observed data is not supported by the regression.

Mediation and moderation

Mediation decomposes a relation into a direct effect and an effect that passes through an intermediate variable. Moderation shows that the size or the sign of a relation differs by a third variable. They are different questions and do not replace each other.

In mediation, measure the cause, the intermediate, and the outcome in the correct temporal order. Estimate the effect on the intermediate and the effect of the intermediate on the outcome, then the product, with an interval. In moderation, include the interaction and report the effect at a few values of the moderator, not a single coefficient. In both, look for confounders of each link.

An intermediate measured after the outcome, or chosen because it is correlated, does not demonstrate the path. An interaction not specified in advance needs replication.

ANOVA (Analysis of Variance), ANCOVA (Analysis of Covariance), MANOVA (Multivariate Analysis of Variance)

ANOVA compares means between two or more groups and separates the variation between groups from the variation inside them. ANCOVA removes a numeric covariate from the comparison. MANOVA compares the groups on several correlated outcomes at once, so that each outcome is not tested in isolation.

Check independence of observations, the shape of the distribution, and whether the spread is comparable between groups. Apply the global test. Only if there is a difference do pairwise comparisons, with a correction for the number of comparisons. In ANCOVA, check that the relation of the covariate to the outcome is similar across groups. Report the means, the differences, and the effect size, not only the p-value.

Groups of very different sizes and different spreads make the classical test misleading. ANCOVA does not repair a group that was formed by the covariate itself.

Nonparametric tests

They compare groups or associations without requiring a particular shape of the distribution. Mann-Whitney and Kruskal-Wallis work with ranks. Wilcoxon compares paired measurements. Chi-square compares frequencies. They are appropriate for ordinal scales, very asymmetric distributions, and small samples.

Check the design: independent groups or pairs, two or more groups. Calculate the test and an effect size on ranks or on frequencies. Report medians, interquartile ranges, or proportions, because the test does not compare means. For chi-square, check whether expected frequencies are too small and, if so, use an exact variant.

Nonparametric does not mean without assumptions. Observations must be independent, and a test on ranks does not answer a question about the mean.

Factor analysis and principal components

Both reduce many correlated variables to a few dimensions. Principal components summarise the observed variance. Factor analysis looks for latent factors that would explain the correlations, with a unique variance for each variable. They are used to build an index or to see the structure of a questionnaire.

Keep variables on a suitable scale and check whether the correlations justify the reduction. Choose the number of dimensions by multiple criteria, not by a single rule. Rotate the solution so that each variable loads clearly on few dimensions. Name the dimension by the content of the items, not by the initial hope. Check stability on a half-sample or on a new sample. Build the index only after this structure.

An attractive name does not make the factor real. Items phrased almost identically create an artificial factor.

Correspondence analysis

It shows which categories of nominal variables appear together more often than by chance and projects them onto a plane. It is useful for large contingency tables, where reading cell by cell is no longer possible.

Build the frequency table, with categories that are not almost empty. Extract the axes and keep as many as explain a clear part of the association. Categories close on the plane are read as having similar profiles. The distance between a row and a column is interpreted cautiously, by the rule of the method used. Check the solution by removing rare categories, which can pull the map.

The plane is a projection. Two categories close on the drawing can be far apart on a dimension that is not displayed.

Cluster analysis

It groups similar observations without pre-existing groups. It serves segmentation, the finding of regimes, and the reduction of a heterogeneous set. The groups are a useful partition, not a discovery that nature is cut exactly that way.

Choose the variables that should define similarity and leave out those that would dominate because of scale. Standardise. Choose the distance and the algorithm. Judge the number of groups by stability on repeated samples and by the fact that the groups can be described differently. Describe each group by means or proportions and by size. Mark a boundary observation; do not force it into a label.

Changing the scale or a variable can remake the groups. A cluster is not valid merely because the algorithm produced one.

Discriminant analysis

It classifies observations into already known groups, on the basis of measured variables. Unlike clustering, the labels exist. It is used when you want a rule that says, for a new case, which group it falls into.

Split the set into a training part and a test part. Build the rule on the training part. Measure on the test part how many classifications are correct in each group, not only overall. Check the cost of a mistake: missing one group can be more serious than missing another. See which variables separate the groups and whether the rule remains stable without one of them.

Accuracy on the data from which the rule was built is optimistic. Groups of very different sizes make a rule that classifies almost everything into the large group and looks good.

Decision trees and classification

They split the data through successive questions, each cutting on a variable, until they reach groups that are as pure as possible. They are used when the rule must be readable by a person or when the relations are steps and interactions, not straight lines.

Limit the depth, the minimum number of cases in a node, and prune branches that do not improve prediction on unused data. Report the error on those data, separately by class. Extract the rules that change the decision and check whether they make sense in the domain. An ensemble of trees can predict better, but then the single rule disappears and variable importance is reported with caution.

A deep tree memorises the sample. A cut on a noisy variable can look decisive and not repeat.

Path analysis and structural equation models (SEM)

They model several relations at once. Path analysis links observed variables in a scheme. Structural equation models add latent variables measured through indicators. They serve when theory says that effects pass through several links and that some concepts are not measured directly.

Draw the scheme before estimation, with each arrow justified. Check whether the model is identified, meaning the data can estimate the parameters. Evaluate the latent measurement before the relations between concepts. Report global fit and each coefficient with an interval. Mention alternative models that would produce the same correlations. Avoid adding links only to increase fit.

A good fit does not prove that the scheme is the true one. Several schemes can generate the same correlations.

Time series, panel, and cohorts

A series follows an indicator over time. A panel follows the same units at several moments. A cohort follows a group defined by a common event, for example hiring or diagnosis. All of them separate change over time from stable differences between units.

Draw the series before the model. Decompose trend, season, and remainder. Check whether the mean and the variation change over time, and difference or transform if the model requires it. In a panel, say how many units are missing at some moments. Fixed effects remove what is stable in the unit. Define the cohort by the event, not by age mixed with generation. Validate the forecast on the last periods, held out.

A series with a change of definition in the middle is not a series. The loss of units from a panel can be the result being studied.

Longitudinal and cross-sectional

Longitudinal means the same units measured repeatedly, so it can see change. Cross-sectional means one measurement at a moment, so it sees differences between units, not the trajectory of one. The choice depends on the question and on how long you can follow.

In longitudinal work, fix the moments, measure in the same way each time, and describe who leaves the follow-up. The analysis can compare average change or trajectories. In cross-sectional work, sample once, describe the differences, and do not interpret a difference of age or seniority as the effect of the passage of time. If both are possible, a question about change requires the longitudinal.

A repeated cross-section on different people is not longitudinal. Selective attrition means those who remain no longer represent the initial group.

Survival analysis

It measures time until an event: failure, return, death, departure. Some cases have not had the event by the end of follow-up and cannot be treated as if the event would never occur. That is censoring, and the method exists so as not to discard it.

Record the start of follow-up, the time, and whether the event was observed. Estimate the curve: the proportion that has not yet had the event, at each moment. Compare groups by curves and by a model that also allows other factors. Check whether the relative risk remains stable over time. Say what kind of censoring exists and whether exit from follow-up is related to risk.

If people leave the study precisely because they are close to the event, the curve is shifted. The mean time is not calculated by ignoring the censored cases.

Experiment and quasi-experiment

An experiment allocates the treatment at random, so that the groups differ only by the treatment and by chance. A quasi-experiment cannot allocate at random and approximates the comparison: difference-in-differences, regression discontinuity at a threshold, an instrumental variable that moves the treatment without moving the outcome directly.

Define the treatment, the outcome, and the moment of measurement. In an experiment, check the allocation and whether the groups were similar at the start. In difference-in-differences, show that the groups were moving in parallel before. In regression discontinuity, show that around the threshold only the treatment jumps, not other characteristics. For an instrument, justify why the instrument does not affect the outcome except through the treatment. Estimate the effect on those whose exposure was moved, and report the interval.

A group that chose the treatment itself is not an experiment. A weak instrument, or one linked directly to the outcome, does not identify the effect.

Bayesian analysis

It treats the parameter as having a distribution of credibility. It starts from a prior distribution, incorporates the data, and obtains the posterior. It is useful when prior information is real, when the sample is small, or when the model is too complicated for a closed formula.

Write the model and justify the priors: informative if there is prior evidence, or weakly informative if not. Sample the posterior and check that the sampling has stabilised. Report the mean or median and a credible interval. Redo the analysis with reasonably different priors. Check predictions on new data, as with any model.

A strong and undisclosed prior can dominate a small sample. A Bayesian interval is not read as a frequentist interval without stating the difference.

Monte Carlo

It repeats a calculation many times in which uncertain inputs are drawn from distributions. The result is the distribution of the possible outcome, not a point estimate. It is used for project risk, for the propagation of errors, and when the final formula has no simple solution.

Write the relations between inputs and outcome. For each uncertain input, choose a distribution on the basis of data or justified limits, including dependencies between inputs. Draw thousands of sets, calculate the outcome each time, and report the mean, the percentiles, and the probability of crossing a threshold. Check which input moves the outcome most.

If the input distributions are invented, the percentiles have a false precision. Drawing independently inputs that move together underestimates the risk.

Sensitivity analysis

It shows how much the conclusion changes if a parameter, a definition, or an assumption changes. It is not an optional annex when the decision rests on a threshold. It is the way to see whether the result is robust or hangs on a single choice.

List the choices that can be doubted: a price, a rate, a case definition, the exclusion of some observations, a form of model. Vary them one by one in intervals that can actually occur. Mark the value at which the decision or the sign of the effect changes. If several parameters are uncertain together, combine with scenarios or with Monte Carlo.

Varying a parameter by a symbolic percent, far from the real uncertainty, is not a sensitivity analysis.

Scenario analysis

It builds a few internally coherent futures, starting from major uncertainties. It does not give the probability of a future and does not replace a short-term forecast. It serves when the horizon is long and a few forces can combine in different ways.

Choose two or three uncertainties that actually change the decision. Combine them into a few scenarios, each with its own logic, not into an exhaustive grid. For each, describe the consequences for the objective. Look for actions that work in several scenarios and actions that work in only one. Specify what signal would show that one scenario is becoming more plausible.

Three scenarios named pessimistic, base, and optimistic, with the same assumptions only scaled up or down, are not scenarios. They are a sensitivity check.

Meta-analysis and meta-regression

It combines estimates from several studies, weighting the more precise ones. Meta-regression looks for why the effects differ: population, dose, design. It is a quantitative synthesis, not an average of the conclusions written by authors.

Write the search and inclusion protocol. Extract the effect and its error, in the same metric. Combine with a model that allows different effects between studies if heterogeneity is large. Report the pooled effect, the interval, and how much of the variation is between studies. Check whether small studies systematically show larger effects. Do meta-regression only if there are enough studies and a pre-specified explanation.

Studies that cannot be compared are not combined merely to produce a number. Preferential publication of positive results pulls the pooled effect.

Power and effect size

Effect size says how large the difference or the relation is, in absolute or standardised units. Power says what chance a study of a given size has of detecting an effect considered important. Without them, a non-significant result can mean either the absence of an effect or a study that is too small.

Before collection, choose the minimum effect that would matter in practice. Calculate the sample for the desired power, with the expected variance and the real design. After the study, report the estimate and the interval. If the interval includes both important effects and null effects, the study has not decided. Avoid verbal labels for standardised sizes without the context of the domain.

A power calculation done after the result is known, to explain a non-significant result, does not replace the interval.

Extreme values and missing values

An extreme value can be an error, a rare real case, or the observation that changes the conclusion. A missing value can be random or linked to what you are studying. The analysis exists so that these cases do not decide in silence.

List the extremes, check them in the source, and classify them. Redo the analysis with and without those that are not clear errors. For missing values, describe how many there are and whether they collect in a subgroup or in an outcome. If you impute, do it as a function of the available information and compare with the analysis on complete cases. Say what assumption about the missingness is required.

Deleting extremes until the desired result appears is not an analysis. Imputation does not create information that was not in the data.

Fuzzy logic and fsQCA (fuzzy-set Qualitative Comparative Analysis)

Fuzzy logic treats membership in a category as a degree. fsQCA uses this to compare combinations of conditions on a medium number of cases and to see which combinations are sufficient for an outcome. It does not estimate an average effect. It estimates configurations.

Choose the conditions from theory, few of them. Calibrate each between 0 and 1 with thresholds justified in the domain. Build the configurations, keep those with enough cases, consistency, and coverage. Read the solution as a combination, not as a list of important variables. Return contradictory cases to the material. Check whether other calibration thresholds change the solution.

Calibration done so that a desired solution appears is not an analysis. Many conditions on few cases produce apparent configurations.

Qualitative and interpretive

These methods work with texts, practices, images, and sequences. Their discipline is not the absence of numbers. It is the traceability of the interpretation.

Content analysis

It systematises a corpus of texts, images, or discourses into categories that can be counted or compared. It can remain at the presence of a category or it can count frequencies. It is appropriate when the question is what themes appear, how often, and in whom, not what the lived experience is in depth.

Define the corpus and what stayed out. Choose the unit: article, sentence, sequence, image. Build the code grid from the question and from a trial reading, with definitions and examples. Two analysts code a part, agreement is calculated, and the grid is clarified. Then code the rest. Report the categories, the frequencies if counting makes sense, the differences between sources, and quotes that show how the code was applied.

A code without a definition cannot be repeated. The frequency of a theme is not its importance.

Thematic analysis

It identifies themes of meaning that cross the material and interprets them, not only counts them. A theme is not a repeated word. It is a pattern of meaning that says something about the question. It is the usual method for interviews when you want a map of meanings, not a formal theory.

Transcribe and read the whole material. Code relevant passages, close to the participants’ language. Group, split, and test the codes on the whole corpus. Keep a theme if it is supported in several places, if it can be named precisely, and if it is not only a reformulation of the question. Write quotes for it and cases that do not fit. Say what the researcher chose to put at the centre.

Themes produced from the guide questions, without work on the material, repeat the guide. An isolated quote does not support a theme.

Frame analysis

It shows how a problem is framed: what is presented as fact, who is the cause, who is the victim, and what solution becomes reasonable. Two texts can speak about the same event and demand opposite actions because they frame it differently.

Collect comparable texts: the same problem, different speakers or periods. Extract the statements about the definition of the problem, the cause, the moral evaluation, and the remedy. Group them into frames. Note which actor has a voice and which explanation is missing. Compare the frequency and placement of frames if the corpus allows it, but the interpretation remains on the structure of the frame.

Calling any adjective a frame empties the method. The frame is demonstrated on statements, not on the overall impression.

Discourse analysis and critical discourse analysis

Discourse analysis looks at how language constructs a version of reality and what actions that version makes possible. The critical variant links these versions to power, institutions, and inequality: who can speak, what is taken as given, which positions are excluded.

Choose texts in their context: who speaks, to whom, in which institution, with what stake. Analyse the vocabulary, the metaphors, the presuppositions, and the way groups are constructed. Link the text to practices that support it or that it justifies. In the critical variant, make explicit the position from which the critique is made and bring the evidence, not only the evaluation.

Critical does not mean the text is read to confirm an already chosen political conclusion. The institutional context is not decoration.

Narrative analysis

It follows stories as a form of meaning: where it starts, what breaks, what the character chooses, how it ends, and what evaluation the teller makes. It is appropriate when people organise their experience as narrative, not as a list of opinions.

Identify the narratives in the interview or the text. Decompose them into scenes, characters, turning points, and moral evaluation. Compare versions of the same event and what function the story has: justification, identity, request. Keep the teller’s language in the analysis of the sequence, not only the extracted theme.

Not every answer is a narrative. Without sequence and evaluation, the method has no object.

Phenomenological analysis and IPA (Interpretative Phenomenological Analysis)

It looks for the structure of a lived experience, as the person understands it, before translating it into a ready-made theory. IPA, interpretative phenomenological analysis, works on few cases and keeps the difference between cases, not only the common theme.

Interview in depth people who have lived the experience, with open questions. Analyse each case separately: units of meaning, then a structure of that person’s experience. Only after that compare. Mark the researcher’s interpretation as distinct from the participant’s formulation. Look for what is specific to the experience, not sociodemographic variables.

A large sample, coded quickly on common themes, is no longer this method. Experience is not deduced from theory and then illustrated with a quote.

Hermeneutic analysis

It interprets a text, a practice, or a symbol in the context that gives it meaning. The part is understood through the whole, and the whole is corrected through the parts. It is the method of dense texts: legal, religious, historical, literary, institutional.

Establish what kind of text it is and in which tradition it stands. Propose a reading of the whole. Check it on the passages that seem to contradict it. Expose the alternative reading as well, with the reason it is weaker. Use historical context to limit anachronistic readings, not as decoration.

An interpretation that cannot say what evidence would disconfirm it is not yet an analysis.

Conversation analysis

It analyses speech as coordinated action: how turns are taken, how a misunderstanding is repaired, what a pause does, which response is preferred. The object is the sequence, not the theme discussed.

Record natural interactions when possible. The transcription includes overlaps, pauses, stress, and interruptions. Take a sequence and show how one turn makes the next relevant. Compare several occurrences of the same type of sequence. Support the claim on the detail of the transcript, not on a summary.

An interview transcribed only on words, without the organisation of turns, does not allow this analysis.

Semiotic analysis

It studies signs and the relations between them: what they denote, what they connote, and in which system of oppositions they acquire meaning. It applies to images, brands, rituals, spaces, and texts in which meaning sits in contrast, not only in explicit content.

Isolate the sign and its context of appearance. Describe the expression and the content. Reconstruct the oppositions in the system: what other sign could have stood there. Distinguish denotation and the connotations supported by use. Check the reading on several appearances or on recipients, not on a single chosen image.

Declaring a symbol without the system in which it functions is free association, not semiotics.

Ethnographic analysis

It describes practices, norms, and local meanings through prolonged presence, observation, and conversation. The question is what people do and what that means in their world, not what they declare on a form.

Enter the site, observe repeatedly, and write notes the same day, separating what was seen from interpretation. Participate as far as the site allows. Link interviews to observed scenes. Look for exceptions to the declared norm. Check the interpretation through long stay and through new situations. On exit, say what access existed and what could not be seen.

A few visits and quotes are not ethnography. The researcher’s presence changes the scene and must be treated as data.

Documentary analysis

It uses documents produced independently of the study: files, regulations, correspondence, reports, archives. It is necessary when the practice under study has left written traces or when the actors can no longer be asked.

Establish who produced the document, for whom, in what procedure, and what interest they had to include or omit. Extract the relevant content with a reference to place. Compare versions and series. Describe the gaps in the archive. Corroborate the document with another source when the stake is a fact, not the text itself.

Reading an official report as a neutral description of reality ignores the purpose for which it was written.

Grounded theory

It builds concepts and relations between them starting from data, through repeated comparison. It does not apply a ready-made theory to quotes. It stops when new cases no longer change the central categories.

Code the first materials openly. Compare categories between cases. Choose the next sample to force the category: a case where it should fail, an extreme case. Write memos about relations, not only codes. Gradually integrate around a central category. Stop at an argued saturation, not at a round number of interviews.

Coding an already collected set, without letting the analysis change the sample, is only partly this method.

Case study

An in-depth analysis of a bounded case: a decision, an organisation, a patient, a locality, a project. The case is chosen because it allows the mechanism to be seen, not because it is the average. The conclusion transfers to similar cases on the mechanism, not statistically.

Define the case and what is outside it. Use several sources that can contradict each other. Build the chronology. Test the explanation on the facts of the case, including the uncomfortable ones. Say on what dimensions another case would have to resemble for the transfer to be reasonable.

An example that illustrates an already written conclusion is not a case study. Unspecified limits of the case make the analysis unfalsifiable.

Qualitative comparative analysis (QCA)

It compares configurations of conditions on a medium number of cases and looks for the combinations linked to an outcome. It sits between the case study and statistics on large samples. The result is a combination of conditions, not an average effect.

Choose the conditions from theory, few relative to the number of cases. Code presence and outcome with explicit rules. Build the truth table. Keep the consistent configurations and simplify only with declared assumptions about unobserved combinations. Check the solution by returning to the cases that contradict it.

Conditions added after seeing who has the outcome empty the method. Contradictory cases are not hidden in the simplified solution.

Historical comparative analysis

It explains large outcomes, such as reforms, revolutions, or institutional trajectories, through the comparison of historical sequences. It looks for the point at which similar cases diverged and the conditions that made the divergence possible.

Choose positive and negative cases, not only the successes. Reconstruct the sequence from sources, with dates. Compare the conditions at the point of divergence. Test alternative explanations on the same sequence. Say what would have been expected in each case if the explanation is correct.

Choosing only the cases that confirm a thesis, or reading the past exclusively through the outcome, produces an apparent cause.

Contextual analysis

It places the phenomenon in the environment without which its meaning or effect cannot be understood: institutions, moment, place, relations, rules. It is not a background chapter. It is the analysis that says what in the context enters the explanation.

List the elements of context that can change the meaning of the observation. Bring evidence for those that actually differ between cases or moments. Show the link between the element and the phenomenon, not only coexistence. Remove what remains decoration.

A context described at length, but unlinked to the result, explains nothing.

Reflexive analysis

It examines how the analyst’s position, access, and decisions shaped what could be seen and what was put at the centre. It is not a confession. It is a check on one’s own analyses.

Note the relation to the site or the participants, the methodological choices, what access was missing, and what interpretations were rejected. Show what would change in the conclusion if an important choice had been different. Reflexivity is linked to a concrete point of the analysis, not to the analyst’s identity in general.

A paragraph about positionality, without effect on a conclusion, does not fulfil the role of the method.

Critical analysis

It evaluates the validity, the limits, the interests, and the power effects of an argument, a measure, or a practice. It asks what is supported, what is assumed, and who is affected by the fact that that version passes as obvious.

Reconstruct the claim in its strongest form, not in its weak version. Check the evidence and the inference. Identify the presuppositions and the excluded alternatives. Say what follows in practice if the claim is accepted. The critique stops where the evidence ends.

Rejecting an argument for its origin, without examining its content, is not critical analysis.

Cultural analysis

It interprets values, norms, and symbols that make a practice intelligible for a group. It differs from the mere description of customs by reconstructing the code through which members judge what is appropriate.

Collect practices, objects, stated rules, and sanctioned deviations. Reconstruct the local contrast between appropriate and inappropriate. Check the reading on new situations or with members of the group. Avoid immediate translation into the analyst’s categories.

A norm declared in an interview, contradicted by practice and still kept as culture, is an incomplete description.

Historical analysis

It reconstructs an evolution and its past context from period sources. It serves when the question is how we got here and what was possible then, not only what the state is now.

Date the sources, compare versions, and classify what is a contemporary document, what is a later recollection, and what is interpretation. Build the sequence from dates, not from the outcome. State the gaps in the archive. Limit the conclusion to what the sources can support.

Explaining every event through the known result makes the causes seem necessary after they have happened.

Strategic, economic, and decisional

These methods exist to support a choice under constraints. Their failure mode is almost always the same: a label without evidence, or an optimisation of a criterion that is not the one the decision actually turns on.

SWOT (Strengths, Weaknesses, Opportunities, Threats)

A summary of the internal factors the organisation can influence, strengths and weaknesses, and the external factors, opportunities and threats. It is a way to order a strategic discussion, not a proof. It becomes an analysis only if each factor has evidence and a consequence.

Collect facts for each quadrant: shares, costs, capacity, rules, moves by rivals. Remove adjectives that do not change a decision. Cross the quadrants: which strength can use an opportunity, which weakness exposes to a threat. Turn them into a few options, each with what would have to be true for it to work.

Long, unsourced lists, produced in a workshop, give the illusion of a strategy. SWOT does not choose by itself.

PESTEL (Political, Economic, Social, Technological, Ecological, Legal)

It scans the political, economic, social, technological, ecological, and legal environment to see which forces outside the organisation can change the conditions of the game. It does not describe the world in general. It selects what can move the objective in the chosen horizon.

For each dimension, note the state, the direction, the horizon, and the mechanism by which it would reach the organisation. Keep the forces with a concrete effect. Link them to scenarios or to SWOT; do not leave them as a list. Write the sources and the uncertainty next to the factor.

Six exhaustive lists, without effect on the decision, are not an environmental analysis.

Porter’s five forces

It evaluates the pressure on margin in an industry: existing rivalry, the threat of new entrants, substitutes, the power of suppliers, and the power of customers. It says how hard it is to keep a profit, not how good a firm is.

Define the industry from the perspective of substitution at the customer. For each force, bring data: concentration, switching costs, barriers to entry, differentiation. Draw the conclusion about which force presses the margin and which position can weaken it. Redo if the industry is delimited differently.

An industry defined too broadly mixes different games. Forces are evaluated with facts, not with adjectives.

Value chain analysis

It decomposes the activities through which the offer passes, from inputs to after-sales service, and the support activities. It shows where cost is added and where something is added for which the customer pays.

List the primary and support activities as they happen, not after a generic scheme. Attach the cost, the time, and the difference from the alternative. Mark the links that can be redesigned, automated, or removed. Keep the links between stages, because an economy in one place can create cost in another.

A list of departments is not a value chain. Without cost or without a differentiator, the decomposition does not help the decision.

BCG matrix (Boston Consulting Group)

It classifies activities by market growth and relative share against the largest rival. The labels — question mark, star, cash cow, and dog — are a short way to talk about cash need and position. They are not a strategy.

Calculate share against the leader and market growth under the same definition. Place the activities, with size proportional to revenue or profit. Link the placement to the allocation of resources and to what would have to change for an activity to move to another quadrant. Check whether share actually brings advantage in that industry.

Share is not automatically advantage. A small activity in an artificially defined market appears as a leader.

Stakeholder analysis

It identifies who is affected by a decision and who can stop it, change it, or make it hold. It serves the planning of involvement, not an organigram of opinions.

List the parties from the real effects of the project, not only from the formal ones. For each, evaluate interest, power, position, and what would change it. Group by role in the decision. Establish what information each receives and where agreement is needed, not only information. Redo the map when the project changes who is affected.

A map made once, without those who can block implementation, fails exactly at them.

Benchmarking

It compares a performance with an internal or external benchmark to see the gap and the practices that explain it. The benchmark must do comparable work. Otherwise the difference cannot be copied.

Choose few indicators, with the same definition and the same period. Normalise for scale. Measure the gap. Then decompose: what the benchmark does differently in process, not only what number it has. Separate what can be transferred from what depends on a context you do not have.

Comparing an indicator defined differently, or copying a practice without its conditions, produces false targets.

Gap analysis

It measures the distance between the current state and a desired state, on common criteria. It is the beginning of a plan, not the plan. Without the cause that maintains the gap, the measure does not say what to change.

Describe the target in observable terms. Measure the present state with the same criteria. Quantify the gap on each criterion. Identify the causes that maintain it and which of them is the constraint. Link to actions that close a particular gap, with a later check.

A target formulated as an adjective cannot be compared with the current state.

Competitor and positioning analysis

It shows who is actually competed with and on which axes the offer is chosen. The competitor is the customer’s alternative, including not buying or solving it another way. Positioning is the relative place on axes that matter to the decision.

Reconstruct the alternatives from purchases or from interviews, not from the internal list of rivals. Choose two axes supported by buying criteria. Place the offers on data of price, performance, or measured perception. Interpret gaps only if there are customers for that combination.

Axes chosen because they differentiate one’s own offer favourably do not describe the customer’s choice.

Balanced scorecard

It translates strategy into a few objectives and indicators on finance, customers, processes, and learning capacity, with links between them. The idea is that the financial indicator comes too late to be the only instrument of management.

Write the strategic objectives, then the indicator, the target, the data source, and the owner for each. Check the chain: the process indicator should move the customer indicator, and that the financial result. Remove measures that do not change a decision. Revise when the strategy changes; do not keep the board out of inertia.

A board with dozens of unlinked indicators is a reporting system, not a scorecard.

Cost-benefit analysis

It compares the costs and benefits of an option in the same unit, usually discounted money, against the alternative. It can include external effects if the purpose is public. The result is the net value, not a list of advantages.

Define the alternative. List the flows over the whole duration, including maintenance costs and costs avoided. Monetise non-monetary benefits only with an explicit rule, or leave them outside the sum, but named. Discount. Calculate the net value and test it at the discount rate, at the investment cost, and at the principal benefit.

Benefits counted twice, or an artificially weak alternative, inflate the net value.

Cost-effectiveness analysis

It compares incremental cost with a common physical outcome: case avoided, kilometre, pupil reaching a threshold. It is used when the benefit cannot be monetised credibly, but the outcome is the same across options.

Measure the extra cost against the alternative and the extra effect, in the same unit. Report the ratio and the magnitudes it comes from. Eliminate options that cost more and produce less. If the outcomes differ qualitatively, the method no longer suffices and one moves to cost-utility or to multi-criteria analysis.

Average cost, instead of incremental cost, changes the ranking when the alternatives have different bases.

Cost-utility analysis

It reports incremental cost against a measure of utility, frequently the quality-adjusted life year. It allows comparison of interventions with different medical outcomes. It requires a threshold at which the utility purchased is considered acceptable.

Estimate the utility gain against the alternative, with the source of the utility values. Estimate the incremental cost. Report the ratio and the sensitivity to utilities and to horizon. State the threshold used and what remains outside it, such as effects on relatives if they were not included.

Utilities taken from another context, without checking that they describe the states being compared, move the ratio.

Break-even analysis

It finds the volume at which revenues cover costs, at a given price and cost structure. It is a survival threshold, not a measure of value. It says how much you must sell so as not to lose, not how much it is good to sell.

Separate fixed cost from cost that grows with volume. Calculate the margin per unit. Divide fixed cost by margin. Check whether price and variable cost remain constant in the region of the threshold and whether there are fixed costs in steps. Report the threshold in units and in time.

A variable cost treated as fixed, or a price that falls as volume grows, moves the threshold.

Profitability analysis

It judges an investment by the cash flows it adds. Net present value sums the discounted flows. The internal rate is the rate at which net value is zero. Payback period says how long until capital is recovered, but ignores what follows.

Write the incremental flows against not investing, including working capital and residual value. Choose the discount rate by the risk of the flows. Calculate the net value. Show at what combination of price, cost, and duration the value passes through zero. Compare mutually exclusive projects on net value, not on rate.

Flows that include non-incremental benefits, or a rate chosen so that the result is positive, do not evaluate the investment.

Marginal analysis

It follows the effect of an additional unit of production, price, or effort, not the average of everything done so far. The correct decision is made at the margin: continue as long as the benefit of the next unit covers its cost.

Estimate the cost and benefit of the next unit, not the historical average cost. Include congestion or discount if they appear at higher volume. Find the point at which marginal benefit equals marginal cost. Use the average only as description, not as a rule of extension.

A small average cost can coexist with a large marginal cost. Extending by the average passes the useful point.

DuPont and financial ratios

DuPont decomposes the return on capital into margin, asset turnover, and leverage. Liquidity, solvency, and efficiency ratios say whether the firm covers its obligations and how quickly it turns assets into sales. They are diagnostic instruments, not a complete valuation.

Calculate on series and on stable definitions. Decompose the change in profitability into the three parts. Check whether a good margin is supported by debt or by a turnover that cannot be repeated. Compare with firms that have a similar model, not with a heterogeneous industry average.

A good ratio in an exceptional year, or a comparison with firms that keep accounts differently, leads to a false diagnosis.

Fundamental analysis, in markets

It values an asset by its capacity to produce cash, by its assets, and by the price paid for them. It answers what it is worth relative to price, not where the price will move this week.

Reconstruct revenues and margins over a cycle, not over an atypical quarter. Estimate the flows a holder can receive. Discount with a rate linked to risk. Compare with the price and say what assumption is already in the price. Give an interval, because small changes in the rate move the value.

A target precise to two decimals, on unreliable assumptions, claims a precision the method does not have.

Technical analysis, in markets

It studies price, volume, and sometimes positioning, to identify regimes and entry or exit rules. It does not measure the value of the firm and does not explain why an asset produces cash. It can be tested only as a rule defined in advance.

Write the rule before seeing the test period: what signal, what horizon, what cost. Apply it on a later period, with transaction costs and with points at which the rule could not have been executed. Report the distribution of results, not a chosen figure. If the rule changes after the test, redo the test on new data.

A figure drawn after the move is known is not a test. Without costs, almost any rule looks better.

Risk analysis

It identifies what can go wrong, how probable it is, how serious, and what can be done. Risk is an event with consequences, not a feeling. The analysis serves prioritisation and the choice between avoidance, reduction, transfer, and acceptance.

List the events within the scope of the decision, not all imaginable misfortunes. For each, estimate probability and impact, with the source. Rank. For the large ones, choose a treatment and an owner. Follow whether the treatment reduced the risk, not whether it was ticked. Mark risks that occur together.

A coloured matrix without estimates and without treatment is a poster. Treating risks as independent underestimates them when they come as a package.

FMEA (Failure Mode and Effects Analysis)

It walks through a process or a product and, for each way it can fail, estimates severity, how often it appears, and how hard it is to detect before it reaches the customer. Priority goes to the combination that is serious and hard to see.

Describe the function of each step. List the failure modes, the effect, and the cause. Score severity, occurrence, and detection with defined scales. Multiply or judge together, by the chosen scale. Redesign the high scores first. After the measure, rescore, to see whether detection or occurrence actually fell.

Scores put from memory, without scales, cannot be compared between teams. A low score on a catastrophic failure does not remove it from attention.

Fault tree analysis

It puts an undesired event at the top and decomposes it into the combinations of causes that produce it, linked by “and” and “or” gates. It shows the minimal cut sets: the smallest set of faults that leads to failure. It is used in technical systems and in security.

Define the top event without ambiguity. Build the tree down to basic faults that can be estimated. Find the minimal combinations. If there are rates, calculate the probability of the top and the contribution of each cut. Check the tree with someone who knows the system, for gates put wrongly.

An “or” gate put in place of “and” changes the risk completely. Common causes across several branches must be linked, otherwise the risk looks smaller.

Feasibility analysis

It checks whether a project can be done, before optimising it. It covers technical, financial, legal, operational, and, when it matters, demand. The conclusion is one of: it can, it can if certain conditions are met, it cannot.

Write the conditions that, if missing, stop the project. For each, bring evidence: demonstrated technology, available capital, authorisation, operating capacity, observed demand. Estimate the resources and the time. Formulate unfulfilled conditions as thresholds, not as hopes. Stop the study at the first hard blockage if there is no way around it.

A study written to justify an already taken decision does not test feasibility.

Multi-criteria analysis (MCDA)

It chooses between options judged on criteria that do not reduce to a single number: cost, time, risk, quality, equity. AHP (Analytic Hierarchy Process) derives weights from pairwise comparisons. TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) chooses the option closest to an ideal solution. Other MCDA (Multi-Criteria Decision Analysis) normalise and aggregate differently. The method does not remove judgement from the decision. It makes it visible.

Choose criteria without overlap. Assign weights explicitly, by those who bear the decision. Score each option on the same scale, with data. Aggregate by the chosen rule. Redo with reasonably different weights and see whether the ranking holds. Apply eliminatory criteria before weighting.

Weights chosen after seeing the ranking justify it, they do not produce it. Correlated criteria count the same thing twice.

Conjoint analysis

It decomposes preference for an offer into the utility of its parts: price, brand, features. The respondent chooses between combinations, does not declare how important an attribute seems. From the choices, one estimates what would change the share if the offer changes.

Choose the attributes and levels that actually exist in the decision. Build combinations a person can compare. Collect choices. Estimate the utilities. Simulate shares at new offers and see sensitivity to price. Check on held-out choices whether the model predicts the choice.

Too many attributes tire the respondent and produce noisy utilities. A share simulation is not a market share if the sample does not buy.

Segmentation

It divides customers or beneficiaries into groups that are similar inside and different between them on a dimension useful to action. The segment must be large enough, stable, and addressable differently. Otherwise it is a label.

Choose variables linked to need or behaviour, not only demographic if they do not change the offer. Group. Describe the groups by what they do and what they value. Check size, stability on another sample, and whether the message or the product would differ. Merge segments that require the same action.

Demographic segments that buy the same way do not justify a different strategy.

Funnel analysis

It measures how many pass from one stage to another, from attention to purchase or from eligible to outcome. It shows where volume is lost. It does not say why it is lost, until it is linked to a cause in that stage.

Define stages without overlap and a rule by which a case enters a stage. Count on the same cohort; do not mix entries from different periods. Calculate passage and abandonment. Compare segments. The stage with the largest avoidable loss becomes the object of diagnosis.

Rates calculated on stocks from different periods, or stages that count the same person twice, move the place of the loss.

Attribution analysis

It tries to allocate a result among several points of contact or causes that preceded the result. It can use a rule, such as last contact or equal split, or a model on data. It is not an experiment unless exposure was separated from the customer’s choice.

Declare the rule or the model. Reconstruct the sequence of contacts. Check the overlap of channels and the cases without measured contact. Compare several rules to see how much the credit moves. If cause is required, look for a test in which a channel is turned on or off, not only a rule model.

The last point before purchase receives credit even when it only collected an already made decision.

RFM (Recency, Frequency, Monetary value)

It classifies customers by how recently they bought, how often, and how much they spent. It is a simple behavioural segmentation, useful when there are no other good data and when the action is retention or reactivation.

Calculate the three indicators on a declared window. Divide into classes with stable limits. Combine into a score or a grid. Describe the segments by value and by risk of not returning. Differentiate the action: the recent and valuable do not receive the same treatment as those absent for a long time. Check later whether the action changed recency or value.

A score without a time window mixes active customers with already lost ones. RFM does not say why they buy.

Demand and elasticity

It estimates how the quantity demanded changes when price, income, or the price of a substitute changes. Elasticity is the ratio between the percentage change in quantity and the percentage change in the factor. It serves pricing, taxes, and capacity.

Use observed variations in which the factor moved for a reason that is not demand itself, or price experiments. Control for season and promotions. Estimate the elasticity and the interval. Say on what price range it is valid. Estimate cross-elasticity against real substitutes.

A relation between price and quantity in observed data mixes demand with supply. Without a variation that moves one of them, elasticity is not identified.

Input-output analysis

It follows how the production of one branch becomes input for others. It allows estimation of the direct and the indirect effect of an additional demand, through the chain of suppliers. It is an analysis of economic structure, not of a firm.

Build or take a table of flows between branches, for an economy and a year. Calculate the technical coefficients. Propagate an additional demand through these coefficients. Report the direct and indirect effect and the assumption that the proportions between inputs remain fixed. State the year of the table.

Fixed proportions do not hold when a large demand changes the technology or when branches import differently from the table.

Equilibrium and game theory

Equilibrium looks for the state in which the actors’ plans fit: no one wants to change the decision alone. Game theory analyses decisions in which the payoff depends on what the other does. They serve competition, negotiation, and policies in which the other party’s reaction is essential.

Specify the players, the actions, the information, and the payoffs. Look for the equilibria. Check whether there are several and which is plausible in context. Say what changes if the information or the order of moves is different. Compare the prediction with observed behaviour when data exist.

Invented payoffs produce invented equilibria. An equilibrium that exists mathematically may not be the one played.

Delphi

It gathers the judgement of a group of knowledgeable people in anonymous rounds, with feedback between rounds. Anonymity reduces the effect of status. It does not measure facts. It produces a collective judgement and a map of disagreement.

Formulate precise questions, with unit and horizon. Choose people for different knowledge, not for agreement. Send the estimates back in aggregate, possibly with the extreme arguments. Ask for revisions. Stop when the answers stabilise or when the disagreement is explicit and argued. Report the dispersion, not only the median.

A group that thinks the same from the start gains nothing from rounds. Consensus does not make the judgement true.

Systemic and relational

These methods refuse the isolated piece. Their common risk is the map that includes everything and explains nothing.

Systemic analysis

It treats a whole through limits, components, inputs, outputs, and loops, not through isolated pieces. A local intervention can move the problem elsewhere or amplify it through feedback. The analysis looks for these links before the solution.

Draw the limit of the system and what remains environment. Put the elements, the flows, and the delays. Look for loops that amplify and loops that stabilise. Identify where a local measure changes the signal received by another part. Test, at least mentally and then with data, what happens if one intervenes at a point.

A map with all possible elements, without loops and without a limit, is not yet systemic.

Structural analysis

It describes the stable relations between elements: hierarchy, dependence, opposition, belonging. The question is how the system is linked, not what function each part fulfils. Structure limits which behaviours are possible.

List the elements and the type of link, with an inclusion rule. Build the scheme. Check on cases whether the structure predicts who depends on whom or what blocks together. Separate occasional relations from stable ones.

A formal organigram, contradicted by the real dependencies, describes the rule, not the structure that functions.

Functional analysis

It asks what role a component fulfils for the system it sits in. A practice can have a declared function and an observed one. The analysis separates them and sees what happens if the component is missing.

Identify the declared function. Observe what changes in the system when the component is missing or fails. Note the unannounced functions, including those that maintain an arrangement. Avoid the explanation that it exists because it is useful, without proof that usefulness explains its presence.

Attributing to a practice any distant effect as function makes the explanation untestable.

Network analysis

It measures nodes and links: who is central, who links groups, where the communities are, and along which path information or a shock can circulate. The node can be a person, an organisation, an article, or a place. The link must be defined, otherwise the graph has no meaning.

Define the node, the link, and the time window. Say what a missing link means: absence or non-knowledge. Calculate the indicators required by the question, not all of them. Check communities under different methods. Test a very central link or node by removing it, to see whether the description depends on it.

Centrality is not influence. A network collected from a single source favours the nodes visible in that source.

Flow and value-stream analysis

It follows the circulation of material, information, or work and the times between steps. It shows how much of the duration is processing and how much is waiting. It is the usual instrument when a process looks busy but the result comes out late.

Walk the real flow, not the procedure. Measure for each step the processing time, the waiting, and the stock. Draw on a single piece or request, from entry to exit. Mark the rechecks and the loops. The principal constraint becomes the object of intervention, then remeasure.

Times declared by those responsible, without observation, underestimate the waiting. Optimising a step that is not the constraint does not shorten the flow.

Cybernetic analysis

It studies control through feedback: there is a target, a sensor, a comparison, and a correction. The question is whether the loop sees the deviation in time and whether the correction does not amplify it. It applies to organisations, to markets, and to technical systems.

Identify the target, what is actually measured, and the delay between deviation and correction. See whether the signal is distorted. Describe what the correction does and whether it produces oscillation. Mark the absence of a sensor or an ambiguous target as a lack of control, not as a problem of execution.

An indicator followed without an action linked to it is not a loop. An ignored delay makes the correction arrive when the problem has reversed.

Life-cycle analysis

It sums the impacts of a product or service from extraction to waste, against a functional unit. It allows comparison between options only if they deliver the same service and if the limits are the same.

Define the functional unit and the limits. Inventory the materials and the energy. Translate into impact categories with declared factors. Report the contribution of the stages. Do sensitivity on limits, on lifetime, and on the energy source. Compare only systems with the same function.

Limits that remove the stage unfavourable to an option make the comparison useless. An unmeasured impact is not zero.

Process and bottleneck analysis

It finds the stage that limits the rhythm of the whole process. The capacity of the process is the capacity of that stage, not the average of the stages. Unblocking a stage that is not the constraint does not increase output.

Measure the capacity of each stage in the same unit and in the same period. Identify the smallest. Check whether the queue collects before it. Intervene there and remeasure, because the constraint moves. Occupied time of people in other parts is not read as lack of capacity.

High utilisation of an equipment does not prove that it is the constraint. It can produce stock that the next stage cannot take.

By domain

Domain methods inherit the disciplines above and add the constraints of a field. The label does not exempt them from the same standards of comparison, measurement, and limit.

Legal analysis

It establishes which rule applies to facts and what consequence it produces. It works with norms, qualifications, and case law. It does not replace the proof of facts. It takes them as to be established and shows what follows if they are qualified one way or another.

Identify the applicable source and its rank. Qualify the facts in the categories of the norm. Look for precedents on similar facts and state the differences. Build the argument for the chosen qualification and the one for the alternative qualification. Limit the conclusion to the law and the facts used.

Jumping from a general norm to a conclusion, without qualifying the facts, skips exactly the step of the analysis.

Clinical and epidemiological analysis

Clinical combines the probability of a condition before the test with the result of the test, to say how plausible the condition is in a patient. Epidemiological links exposures to health events in a population and must separate association from confounding and from selection.

In a patient, start from the frequency in that group, apply the characteristics of the test, and arrive at the probability after the test. In a population, define the case, the exposure, and the time window. Choose the comparison. Control known confounders. Report the association with an interval and discuss what might remain uncontrolled. Judge a screening also by how many false positives it produces.

A precise test applied to a very rare condition can have many false positives. Association in a population is not individual diagnosis.

Economic and public policy analysis

It evaluates how agents behave under incentives and what effect a measure has on groups, including beyond the direct beneficiary. A policy is judged against the alternative, with the reactions it provokes.

Specify the mechanism: what changes in price, rule, or income, and who modifies behaviour. Define the group. Build the counterfactual. Estimate the effect on beneficiaries and on those who pay or are displaced. Separate the average effect from the effect on the most exposed. State the cost and the behavioural assumptions.

The effect on beneficiaries, without those displaced and without cost, is not a policy evaluation.

Environmental and sustainability analysis

It evaluates pressure on resources and on the environment and whether an activity can continue without crossing explicitly chosen thresholds. Sustainability without a threshold and without a physical indicator remains an adjective.

Choose physical indicators linked to the question: emissions, water, soil, biodiversity, waste. Measure the pressure and, if it exists, the state. Compare with a declared threshold. Show the uncertainty and what remains outside the indicators. Separate the trend from an atypical year.

An effort indicator, such as the sum spent on environment, does not measure pressure. Undeclared thresholds make any conclusion possible.

Spatial analysis and GIS (Geographic Information System)

It links a phenomenon to place, distance, neighbourhood, or barrier. The question is whether place organises the phenomenon, not only where to put pins on a map. The map is the result, not the analysis.

Geocode with a rule for failed addresses. Choose the scale: point, neighbourhood, region. Test whether the clustering is greater than under a random allocation with the same density. Check whether the pattern holds at another scale. Support a causal conclusion with another design, not with neighbourhood alone.

Aggregation on large zones can create or hide clusters. A map without a denominator, only with counts, follows the population, not the risk.

Gender analysis

It examines differences of access, role, pay, or outcome linked to gender and how much of them remains after other relevant factors. It can include norms and the division of labour, not only a disaggregated table.

Disaggregate the indicator. Check whether the difference holds after occupation, working time, age, or other factors that actually belong to the question. Describe the candidate mechanism and what evidence supports it. State the limits of the categories used. The conclusion does not jump from the gap to a single cause.

A raw gap, without disaggregation, mixes composition with treatment. Adjustment on a variable that is itself the result of inequality can hide the effect.

Ethical analysis

It evaluates whether an action respects principles that can enter into conflict: not to harm, to respect autonomy, to distribute burdens, to tell the truth. It does not give a score. It makes the conflict and the ground of the choice visible.

Identify those affected, including those without a voice in the decision. Formulate the principles in tension and the real options. Describe the consequences of each. Justify the choice and what moral cost remains. If there is a threshold that is not negotiated, say it before the weighing.

Calling a preference a principle, without saying what would disconfirm it, does not structure the decision.

Educational analysis

It analyses learning, an assessment, or a curriculum. The question is what should be learned, what task demonstrates it, and what the result says about teaching, not only about the pupil.

Link the objective to the task through which it is observed. Check whether the test asks exactly that objective or a collateral skill. Interpret results on items, not only on the mean. Separate what depends on access to teaching from what depends on the test. Compare groups only if the test and the conditions are common.

A mean score without item analysis does not say what was not learned. A test unaligned with the objective measures something else.

Linguistic analysis

Lexical looks at units and meanings. Syntactic, at the structure of the utterance. Semantic, at composed meaning. Pragmatic, at what the utterance does in context. Corpus analysis looks for patterns in large collections, with frequencies and contexts. Each has its unit and they are not mixed without saying at what level one is working.

Define the unit and the corpus. Annotate by a scheme with examples. In corpus work, calculate relative frequencies and read the context, not only the number. In pragmatics, reconstruct the situation of utterance. Limit the result to the declared corpus.

The frequency of a word is not its meaning. A structure described on chosen examples is not yet a pattern.

Bibliometric and scientometric analysis

It measures scientific production and citation: volume, collaboration, journals, citation networks. It says how the literature visible in a database is organised, not which work is true.

Declare the database, the period, and the query. Clean duplicates and name variants. Calculate the required indicators and normalise by domain and by age of publication when comparing. Interpret networks after saying what a link means. State the limits of the database: language, document type, domain.

Citation count, uncorrected by domain and year, also measures the size of the community, not only influence.

Reliability and requirements analysis

Reliability estimates how and when a system fails under given conditions. Requirements analysis establishes what the system must do, in statements that can be verified. One without the other leaves either a system that does not do what it should, or requirements that cannot be tested.

In reliability, collect times to failure, stress conditions, and failure mode, then estimate the rate and which component dominates it. In requirements, write each need as a testable statement, follow it to verification, and separate the requirement from the already preferred solution. Pass changes through the same chain.

A requirement formulated as technology can no longer be verified except circularly. A failure rate without the conditions of use does not transfer.

Security and threat analysis

It identifies what must be protected, who or what could affect it, what vulnerability makes that possible, and what measure reduces the risk. A threat without a vulnerability, or a vulnerability without a consequence, does not have the same priority.

Inventory the assets by the consequence of loss. Describe the credible scenarios, not all imaginable ones. Link each to an existing vulnerability. Evaluate probability and impact. Judge measures by how much they reduce risk relative to cost and by what happens if the measure fails.

A list of ticked controls, without a scenario, does not say what remains uncovered.

Laboratory and signal methods

These measure a physical phenomenon. They are not the same class as SWOT or regression. Their discipline is calibration, blank, interference, and the limit of detection.

Gravimetric and volumetric analysis

Gravimetry determines a quantity from the mass of an isolated product. Volumetry, or titrimetry, determines it from the volume of reagent consumed up to the equivalence point. They are classical methods, used when the reaction is known and interferences can be removed.

Prepare the sample so that the species of interest is separated. Dry and weigh to constant mass, or titrate to the detected point. Run a blank. Calculate the concentration from stoichiometry. Repeat and report the scatter. Test known interferences on a standard.

An incomplete precipitation or a wrongly seen equivalence point propagates directly into the concentration.

Instrumental analysis

It links a signal produced by an instrument to the concentration of a species: absorption or emission of light, electrochemical signal, signal after chromatographic separation. It is the usual method when concentrations are small or when species must be separated before measurement.

Calibrate with standards that cover the range of the samples. Check the blank and the interferences. Measure the sample in the zone where the signal responds proportionally. Report the concentration, the repeatability, and the limit of detection. Dilute a sample above the calibration range; do not extrapolate.

A signal without calibration in the sample matrix can be shifted by other components.

Spectral analysis and Fourier

It decomposes a signal into the frequencies of which it is formed. The Fourier transform does this for sampled signals. It serves vibrations, sound, electrical signals, and any series in which the periodic matters more than the shape in time.

Sample often enough relative to the frequencies sought. Choose the window and the length. Apply the transform. Identify the components and separate the noise. Check whether a frequency is real or an artefact of sampling and of the window. Link the interpretation to the physical source of the component.

Sampling too rarely folds high frequencies over low ones and creates components that do not exist.

Thermal analysis

It follows mass or heat flow while the sample is heated in a controlled way. TGA (Thermogravimetric Analysis) sees the loss of mass. DSC (Differential Scanning Calorimetry) sees the heat absorbed or released at a transformation. They are used to identify decompositions, transitions, and stability.

Heat the sample by a declared programme, with a declared gas. Record the signal against a reference. Attribute the steps to transformations by temperature and by comparison with known materials. Repeat, because sample mass and heating rate move the steps.

A step without attribution remains a signal. Different conditions make the temperatures hard to compare.

Image analysis

It decomposes an image into measurable regions, shapes, textures, or objects. In control and in medicine it serves detection and measurement. In social research it can describe the composition of a visual corpus, but then the coding rules must be as explicit as for text.

Correct the acquisition: scale, lighting, distortion. Segment by a reproducible rule. Measure the required descriptors. Validate on images labelled by a person, with agreement between evaluators. Report on what type of image the rule fails.

A model that works on the training images and has not been checked on another instrument or another lighting is not validated.

Elemental and isotopic analysis

Elemental analysis determines which elements and in what proportion are in a sample. Isotopic analysis determines the ratio between the isotopes of an element. The ratio can say the source or the process, if the reference system is known.

Prepare the sample without contamination and without selective loss. Measure against standards. Report the ratio in the declared reference system, with the uncertainty. Interpret source only if the possible sources have different ratios and if later processes have not mixed them.

Contamination or an unsuitable standard moves the ratio. A ratio compatible with several sources does not identify one.

Further methods that complete the map

These are not leftovers. They are the methods most often missing from catalogues, and the ones that most often decide whether a claim survives contact with an alternative explanation.

Counterfactual analysis

It estimates what would have happened to the same units without the intervention. The effect is the difference between the observed outcome and this counterfactual. Without a credible comparison, the before-after difference mixes the intervention with the rest of the changes.

Build the comparison: random allocation, a similar group that did not receive the intervention, an interrupted series, or a threshold. Show why the groups would have gone the same way in the absence of the intervention. Estimate the difference and the interval. State the assumptions that, if they fall, cancel the conclusion.

A group that chose the intervention because it was already different is not a counterfactual. Its difference is not the effect.

Mechanism analysis

It shows the path through which an effect is produced, not only that the two ends are linked. It is necessary when the decision depends on the link: if the link is missing, the intervention has no way to work.

Write the links before measurement. Measure whether the intervention moves the link and whether the link moves the outcome, in the correct order. Look for cases in which the link was blocked and see whether the effect decreases. A link that does not move eliminates that path, even if the overall effect exists.

A variable correlated with both ends, measured at the same moment, does not prove the path.

Process tracing

It tests a mechanism in a case, on the traces the mechanism should have left. It does not look for themes. It looks for clues that would be improbable if the mechanism is false and probable if it is real.

Write the mechanism and the expected clues before searching in the material. Classify each clue by how strongly it tests the mechanism. Look also for the clues that should be absent if the mechanism is false. Say which test was passed, which failed, and what remains undecided.

Quotes chosen because they fit, without clues that could have disconfirmed, are not process tracing.

Sentiment analysis and topic modelling

Sentiment classifies the tone of a text, usually positive, negative, or neutral, sometimes on aspects. Topic modelling looks for groups of words that appear together and treats them as latent themes. Both scale on large corpora and both require human control.

Clean the text and declare the language and the domain. For sentiment, choose a model suited to the language and check on a sample labelled by people, with the typical errors. For topic modelling, choose the number of themes, read the words and the representative documents, and keep only the intelligible themes. Do not publish an automatic theme without this reading.

Irony, negation, and domain terms break sentiment. A statistically coherent theme can be an artefact of frequent words.

Jobs-to-be-done

It describes what a person is trying to achieve in a situation, including the criterion by which they judge success. The product is hired for that job and can be fired in favour of an unexpected alternative. It differs from demographic segmentation.

Reconstruct real situations of use: what triggered the search, what alternative was fired, what mattered in the choice, and what would have made the choice fail. Formulate the job independently of the current product. Check on other cases whether the same job appears. Judge the offer by how well it closes the job, not by features.

A persona with age and hobbies, without the situation of use, is not this analysis.

Post-mortem and after-action review

It reconstructs an incident after it has been closed, in order to change the conditions that allowed it. The object is the system: missed signals, pressures, tools, local decisions that were rational and that together produced the failure. It is not a search for a culprit.

Establish the chronology from traces, not from the most confident memory. Identify the points at which information existed and did not change the decision. Describe the conditions that made that decision reasonable at the time. Choose measures that change the condition, with a way to check later that they were applied.

A conclusion that stops at one person’s error, without the conditions that made the error probable, will repeat.

Opportunity and addressable market

It estimates the demand that can be reached with the existing channel, price, and capacity, not the theoretical market of everyone who might have a need. The difference between the broad market and the addressable one is the object of the analysis.

Start from the population that has the need. Apply in turn the real filters: access to channel, ability to pay, regulation, area served, alternative already used. Each filter has a source. Check the result with observed demand, a pilot, or orders, not only with a percent applied to a large market.

A small percent of a huge market, without filters, produces an opportunity that does not exist.

Substitution and cannibalisation

It measures how much of the demand for a new offer comes from rivals’ alternatives and how much from one’s own offers. Substitution against rivals can be desired. Cannibalisation of one’s own volume is a cost that must be subtracted from the apparent revenue.

Follow customers before and after launch: who came from outside, who switched an own product, who added. Separate additional volume from moved volume. Calculate margin on both, because moved volume can have a different margin. Check whether the move would have happened anyway.

The growth of sales of the new product, without what was lost on the old product, overestimates the effect of the launch.

Seasonality analysis

It separates in a series what repeats in cycles, usually annual, from trend and from remainder. Without this separation, a weak month can be read as decline, and a seasonally good month as the success of a measure.

Estimate the pattern that repeats over enough years. Remove it from the series. Check whether the shape of the season is stable or has moved. Make decisions and comparisons on the adjusted series, and state the seasonal target explicitly. Do not treat an incomplete year as a year.

Adjustment on too few years confuses a unique event with a season. The season removed from a series must not be read as its disappearance from operations.

Equity and disparity analysis

It measures differences of access or outcome between groups and how much of them remains after relevant factors. It is broader than gender analysis. It does not by itself establish the cause and does not say which remedy is justified. It says where the gap is and how sensitive it is to adjustment.

Disaggregate the indicator on the groups in the question. Check small numbers. Adjust with the factors that belong to the question, knowing that some can themselves be the result of the disparity. Report the raw gap and the adjusted one. Say what could not be separated.

Adjusting until the gap disappears does not prove that the gap does not exist. Groups that are too small do not support a precise comparison.

Validation analysis

It checks whether an instrument measures what it claims. Content validity concerns coverage of the construct. Construct validity concerns relations with other measures. Criterion validity concerns prediction of an external outcome. Reliability concerns how stable the measurement is on repetition or between evaluators.

Link the items to the construct, with what remains uncovered. Test the expected relations and those that should be absent. Compare with an external criterion if one exists. Repeat the measurement or calculate agreement between evaluators. Do not use a new instrument for a decision before these steps, at least on a trial sample.

High internal consistency does not prove that the desired construct is being measured. It can mean that the items repeat the same phrasing.

How to choose

Start from the question, not from the method.

If the question is descriptive, a causal model is surplus. If cause is required, a correlation, a SWOT (Strengths, Weaknesses, Opportunities, Threats), or a map does not suffice: there must be a comparison that says what would have happened otherwise, or a test that eliminates the candidate cause. If the data are texts and the cases are few, stay with qualitative methods, with process tracing, or with QCA. If the decision depends on a threshold, sensitivity is part of the analysis, not of the annex. If an instrument is new, validation precedes the conclusion.

The practical test is simple. Before the work starts, write the question in one sentence, the comparison or the corpus that can answer it, and the result that would make you abandon the claim. If those three cannot be written, the method has not been chosen. A label has.

Analysis is not a performance of seeing through things. It is a constraint on what can be said, and on what must be left unsaid until the comparison, the measurement, or the disconfirming case is in hand. The gap between a method’s name and the work it actually did is where most public arguments fail. Closing that gap is not a technical preference. It is the condition under which a conclusion can be acted on without pretending the uncertainty was never there.