Variables, data types and measurement scales

Start with what a row and a value mean

A variable is a characteristic recorded or considered for observation units. The units might be people, manufactured parts, river samples, households, events or repeated measurements of one object. Before calculating, specify what counts as one observation, which characteristic is being recorded and when or under what procedure it is observed. A table with one row per person differs from one with one row per person per visit. Treating repeated visits as independent people changes the scientific question even if every column contains valid-looking numbers.

For a collection of water samples, one column might identify the sampling site, another record an ordered clarity category, another give a count of visible particles and another give a measured volume with units. Each value needs a definition. Does particle count refer to the whole container or a specified subsample? Does volume include the container’s empty space? Which procedure assigns the clarity category? A short codebook that answers such questions is part of the mathematical information, not an optional decoration added after analysis.

An algebraic variable can stand for an unknown or changing quantity in a formula; a data variable records a characteristic across observations. A random variable additionally belongs to a probability model, which will be developed in later lessons. A column of observed values does not by itself establish a probability law, independence or a sampling design. This article needs proportions, basic algebra and units, not calculus or the preceding differential-equation branch of the map.

Separate the target characteristic from its recorded representation. A person’s height, the rounded height printed by an instrument and a category such as ‘above the doorway limit’ are three related variables with different information. A number without its reference can also be ambiguous: 12 may mean 12 kilograms, category code 12 or the twelfth position in an ordering. Naming the observation unit, target, procedure, unit or category reference, and recording convention prevents an operation from acquiring a meaning it never had.

A number can be a label rather than an amount

Categorical variables distinguish classes, while quantitative variables describe numerical amounts or magnitudes under a stated interpretation. A route identifier 12, a material class encoded 3 and a student identifier stored as an integer can all be categorical. Adding their codes does not normally combine routes, materials or students in a meaningful physical or statistical amount. Conversely, storing a measured mass as the text ‘2.40 kg’ does not turn the underlying characteristic into a category. Scientific meaning and machine storage type are different layers.

Some variables have both a category structure and an intentional numerical function. For a nominal outcome A or B, define I_A=1 when the outcome is A and 0 otherwise. The average of this indicator is the observed proportion in category A. That interpretation follows from the explicit definition of the function, not from a claim that arbitrary category codes have an ordinary arithmetic average. Changing labels from A, B to red, blue preserves which observations belong to the designated category if the codebook preserves the correspondence.

The same recorded numeral can describe different kinds of variable. A value 4 might be a count of defects, an ordered satisfaction category, a team code or a temperature in degrees Celsius. Their allowed comparisons differ. Four defects is twice the count of two defects when the counting rules and exposure are comparable. Category 4 need not be twice category 2. A temperature ratio formed from Celsius numerals depends on an arbitrary origin and needs a different analysis. The numeral alone cannot choose the operation.

The classification depends on the question and operational definition. A postal code is a label even if it happens to sort geographically in a useful way; an analyst can derive a distance or region variable from external information, but that is additional structure. A test score defined as the number of correctly answered items is a quantitative count of successes on that test. Treating it as an equal-interval measure of an underlying ability requires a further measurement model. Keep the directly defined variable and the inferred construct distinct.

Discrete does not mean integer, and continuous does not mean infinitely printed digits

A discrete numerical description has a finite or countable set of allowed values. Counts 0, 1, 2, … are familiar examples, but integers are not required. A score restricted to 0, 1/2, 1, 3/2, 2 is discrete. Even an allowed set such as {1,1/2,1/3,…} is countable despite its shrinking gaps. Inspect the set defined by the procedure, not whether a spreadsheet displays a decimal point. Category codes also form a finite set, but their discreteness does not make their labels quantitative amounts.

A continuous quantity is commonly modeled as varying through an interval without a prescribed minimum increment. An idealized length or elapsed time can be treated this way even when the instrument reports only a finite number of digits. The observed record then belongs to a grid produced by a recording rule. The underlying target and its recorded version have different allowed-value descriptions. Neither the idealization nor the instrument should be mistaken for an exhaustive assertion about physical reality at every possible scale.

For a grid spacing δ>0, a nearest-grid recorder can be written schematically as Y=Q_δ(X). Once tie handling and the origin of the grid are specified, this is a defined map from underlying values to recorded values. Many different X values produce the same Y. Printing extra zeros afterward does not recover the lost information. Recording 1.73 m as 1.730000 m changes the display, not the measurement resolution or the original uncertainty information.

Some data require a mixed description rather than a forced two-way label. A device can report an exact ‘no event’ state and otherwise a positive continuously modeled amount; a censored measurement can combine a category flag with a numerical limit. Later probability models distinguish point masses, continuous components and mixtures. Here the practical requirement is to preserve that layered information in the codebook. Treating every zero as an ordinary measured amount or every decimal as a fully resolved continuum value can erase the distinction before analysis starts.

Nominal data support equality and counting without arbitrary code arithmetic

A nominal classification distinguishes categories without supplying an intrinsic order or numerical spacing. Material type, a designated treatment label and an observation-site code can be nominal variables. A valid relabeling is a one-to-one correspondence of categories: it changes names while preserving which observations match. Counts and proportions then refer to the same groups. Sorting the labels alphabetically may help display a table, but it does not discover a scientific ranking.

For observations A, A, B, codes 1, 1, 2 have arithmetic mean 4/3. Relabel the same categories with codes 9, 9, 2 and the code mean becomes 20/3. Nothing about the category memberships changed. This demonstrates that the raw code average does not describe an intrinsic average category. The modal category is still A, with count 2, and the category frequencies remain 2/3 and 1/3. The result follows from the permitted change of labels, rather than a vague rule that numbers appearing near categories must never be calculated.

Define the A-indicator explicitly: I_A=(1,1,0) for these three observations. Its sum is 2 and its mean is 2/3, exactly the sample proportion in A. Indicators for all categories add to 1 for each observation when the categories are mutually exclusive and exhaustive. Those assumptions need checking: multi-select responses, overlapping diagnoses or an unrecorded category can violate them. A count table should retain missing or inapplicable cases separately when its denominator requires that distinction.

Nominal outcomes can support sophisticated statistical models, comparisons of proportions and prediction of category probabilities. Such work requires a model, design and purpose, not an arithmetic ordering of arbitrary labels. Coding A as 0 and B as 1 can be convenient when the event corresponding to 1 is explicit. Swapping that event changes the interpretation to its complement. The codebook must preserve the event definition so a numerical summary remains a statement about the intended category.

Order does not establish equal gaps

An ordinal variable supplies an ordering of categories or levels without, by that fact alone, assigning equal distances between them. Low, medium and high may be ordered, while the change from low to medium need not match the change from medium to high. A permitted recoding preserves the order, so a strictly increasing function of the assigned codes can represent the same ordered information. In metrology, an ordinal quantity is established by a specified procedure with an ordering; it does not automatically acquire the ordinary addition and subtraction rules of interval or ratio quantities.

For ordered categories coded 1, 2, 3, replace the codes by 1, 2, 100. The middle ordered observation remains the middle category, but the arithmetic mean changes from 2 to 103/3. A statement such as ‘the typical observation is the middle category’ can survive the recoding when based on that order statistic. A statement that the top category lies exactly one equal-sized step above the middle does not. Counting observations below, at or above a specified category also uses the ordering without inventing a metric gap.

Median conventions need care. For an odd sample size, the middle order statistic commutes with a strictly increasing recoding: applying the recoding to that observed category gives the new middle category. For an even size, reporting the central ordered pair or an appropriate median category/interval respects the available information. Averaging two central numeric codes need not commute with an arbitrary increasing recoding and may produce a value that is not an allowed category. An unqualified claim that every numerical median convention is automatically ordinally meaningful is too broad.

Ordered response scales can be modeled using latent variables, threshold models or substantively justified scores. These approaches add assumptions about distances or probabilities; they should be explained and evaluated rather than smuggled in through labels 1 to 5. A sum of several item codes is a defined score, but interpreting it as an interval measure of a psychological or educational construct requires more than knowing that each item is ordered. Scale classification informs the question; it does not alone settle the validity of a particular model or estimator.

Test a claim by changing an allowed representation

The classic nominal, ordinal, interval and ratio classification is useful when expressed through transformations that preserve the intended structure. Nominal labels allow bijective relabeling. Ordinal codes allow strictly increasing recoding. An interval representation allows y=ax+b with a>0: the unit and origin may change while equal differences remain comparable. A ratio representation allows y=ax with a>0: a meaningful zero is retained while the unit changes. These are descriptions of represented information, with the application’s assumptions stated, not a claim that every real variable fits one simple class without qualification.

For an interval change y=ax+b, the difference between two values becomes y₂−y₁=a(x₂−x₁). Its sign is preserved, equal differences remain equal, and ratios of nonzero differences are unchanged. A raw ratio y₂/y₁ generally changes because of b. For a ratio change y=ax, x₂/x₁ is preserved whenever the denominator is nonzero. These calculations explain why a change of origin and a change of unit have different consequences. The condition a>0 preserves the chosen direction of the ordering.

A meaningful summary need not keep the identical numerical value after changing units. If a mean length is 2 m, the same mean is 200 cm. The required property is consistent transformation of the statement or statistic, often called equivariance. An invariant comparison may be a ratio or an order relation, while an equivariant value may change by the representation’s rule. Demanding that every useful statistic be numerically unchanged would incorrectly reject ordinary unit conversion.

Stevens’s historical scale framework organizes these distinctions, but it is not a universal menu that automatically approves or bans statistical tests. Independence, distributional assumptions, measurement validity, sampling design and the target of inference remain separate. A continuous ratio variable need not be normally distributed; an ordinal outcome can be analyzed with a suitable ordinal model; nominal outcomes can be represented with explicit indicators. Ask what the calculation means, which transformations preserve that meaning, and what additional model makes an inferential claim justified.

Temperature makes the role of an origin visible

For Celsius numerical value t_C and thermodynamic temperature numerical value T_K, T_K=t_C+273.15. Fahrenheit numerical values satisfy t_F=(9/5)t_C+32. Thus 20 °C corresponds to 293.15 K and 68 °F, while 30 °C corresponds to 303.15 K and 86 °F. These are representations of the same two temperature states. Their Celsius-number ratio is 3/2, their kelvin-number ratio is 6063/5863, and their Fahrenheit-number ratio is 43/34. An arbitrary-origin numerical ratio is not representation independent.

The difference between these states is 10 Celsius degrees, 10 kelvins or 18 Fahrenheit degrees. Temperature intervals have the same numerical size in kelvins and Celsius degrees, while Fahrenheit intervals rescale by 9/5. Two Celsius intervals of 10 and 20 degrees have ratio 1/2; after conversion to Fahrenheit intervals, 18/36 is still 1/2. Ratios of differences behave differently from ratios of the original readings because the offset cancels before division.

The SI kelvin is defined using a fixed value of the Boltzmann constant; the former water triple-point definition is historical. The Celsius offset of 273.15 is exact in the relation to kelvins. A ratio of positive thermodynamic temperatures has a specific physical interpretation within that quantity scale. It must not be paraphrased as a universal claim of ‘twice as hot’, twice the thermal energy of arbitrary objects or twice the perceived warmth. Heat content and human sensation depend on additional physical and physiological information.

A mean temperature can transform consistently under these affine changes even though a raw Celsius ratio does not. The mean of 20 °C and 30 °C is 25 °C, corresponding to 298.15 K and 77 °F, which are also the means of the converted readings. Whether this arithmetic mean answers a particular physical question is another matter: pooling temperatures to predict an equilibrium can require masses, heat capacities and a model of energy exchange. Representation consistency is necessary for interpretation, not a substitute for the application’s physics.

A meaningful zero allows some ratios, with their denominators checked

For a ratio quantity, changing the unit rescales numerical values by a common positive factor while preserving the intended zero. Masses of 2 kg and 3 kg become 2000 g and 3000 g, and the second-to-first ratio remains 3/2. The same argument works for lengths or durations under a specified definition. Ratios compare quantities of the same kind with a nonzero denominator; mixing a mass and a temperature does not become dimensionless merely because both are recorded as numbers.

Zero needs an operational meaning. Zero counted defects means no events meeting the defined defect rule in the inspected exposure. Zero elapsed duration is the absence of elapsed time between the specified endpoints. A balance reading after taring refers to mass relative to the declared tare procedure; its metadata still matters. An instrument reading of zero can also mean below detection, a rounded small value or a missing-value code. Those recording states must not be confused with a known exact zero of the target quantity.

A ratio statement also needs a relevant denominator and comparison context. Twice the number of particles in twice the sampled volume does not imply twice the particle concentration. A count per volume has its own units and is a derived variable. A ratio of 0 to a positive amount is defined; a ratio with denominator 0 is not. Negative recorded values can arise from a difference variable, calibration correction or error and need interpretation rather than an automatic declaration that every field with a physical unit is a nonnegative ratio measurement.

Absolute scale structure does not make an application model automatic. A mass ratio is meaningful under common units, but it does not imply that two materials have the same density, value or energy content in that ratio. A duration ratio says nothing by itself about productivity or causal effects. Preserve what the quantity represents and which operation created a derived rate. This allows unit-aware algebra to support a scientific comparison without turning a valid numerical ratio into a broader unsupported claim.

A summary should transform consistently with its meaning

For readings x₁,…,xₙ and an affine change yᵢ=axᵢ+b with a>0, the arithmetic mean satisfies ȳ=a x̄+b. Subtracting gives yᵢ−ȳ=a(xᵢ−x̄). Consequently a variance computed with the same denominator in the two representations scales by a², and the standard deviation scales by a. The origin shift disappears from differences. These are algebraic transformation facts; they do not require the observations to be independent or normally distributed. Such assumptions enter later inferential uses, not this identity.

For nominal labels, a mode refers to the corresponding category with greatest count; a bijective relabeling changes its name consistently. For ordinal observations, an odd-sample middle order statistic follows an increasing recoding. The arithmetic mean does not generally do so under arbitrary increasing transformations. An interval mean is affine-equivariant, while a raw ratio of two interval means is generally not invariant to shifting the origin. Asking which property should survive is more precise than asking whether a summary is simply ‘allowed’ or ‘forbidden’ by a four-word scale classification.

The coefficient of variation illustrates the role of a zero. For an original mean x̄≠0 and a transformed mean a x̄+b≠0, CV=s/x̄ changes under y=ax+b, a>0, to a s/(a x̄+b). Positive rescaling with b=0 preserves the ratio, but an arbitrary shift does not. Reporting a Celsius CV as if it were an origin-independent relative temperature variability therefore lacks that interpretation. Additional complications arise for zero or negative means and for whether relative variability is substantively useful. Having a computable formula is not enough to justify its scientific meaning.

A nonlinear transformation can intentionally create a new variable and a new question. Its mean need not be the transformation of the original mean. For example the arithmetic mean of 2 and 8 is 5, while averaging their logarithms and transforming back gives the geometric mean 4. Neither is a mistake when the target and assumptions are clear. They summarize different structures. The later lessons on averages and models will develop these choices; the present task is to retain the transformation and interpretation in the record.

A display step is not a complete uncertainty statement

Instrument resolution concerns the smallest change in the measured quantity that causes a perceptible change in the corresponding indication. Display resolution instead concerns the smallest difference between displayed indications that can be meaningfully distinguished. A chosen recording or rounding increment is another convention; it should not be substituted for instrument resolution or for measurement uncertainty. Suppose positive lengths are rounded to the nearest centimetre, with exact half ties rounded upward. A record of 1.73 m then corresponds, for the rounding rule alone, to underlying input values in [1.725,1.735) m. The lower endpoint rounds to 1.73 and the upper endpoint to 1.74. Different tie rules can change endpoint inclusion, so an exact interval claim must specify that convention rather than rely only on a printed decimal place.

Relative to a stipulated reference value of 1.734 m, the signed difference of the recorded value and reference is 1.730−1.734=−0.004 m. The sign indicates that the record is smaller. Metrology distinguishes measurement error, a measured value minus a reference value, from measurement uncertainty, which characterizes dispersion attributed to the measurand using available information. A real reference can itself have uncertainty. One cannot generally infer a known error relative to an unknown true value just because an uncertainty number has been supplied.

The rounding cell bounds only the error from that idealized quantization stage. Calibration, sensor response, environment, repeatability, sampling location and the definition of the measurand can contribute additional uncertainty. Assigning a uniform distribution within a rounding cell is an extra model, not a consequence of the interval endpoints. Under that explicitly adopted model, the rounding component’s standard deviation is δ/√12; without the model, the cell alone does not provide that probabilistic interpretation or a confidence level.

More displayed digits, a smaller standard deviation of repeated readings and a smaller total uncertainty are different claims. A miscalibrated device can repeat almost identical readings while remaining systematically displaced from a reference. Repeated measurements may also share environmental or calibration effects. Keep the resolution, rounding convention, reference information, repeated-measurement conditions and uncertainty assessment separately documented. This preserves the information needed for later error propagation and statistical modeling without presenting an instrument grid as the whole measurement process.

Preserve states that a numerical column can easily hide

A codebook should specify the variable’s name, observation unit, target meaning, allowed values, units or category labels, ordering if any, recording procedure and version. It should distinguish actual values from flags for missing, not applicable, refused, invalid or not yet observed. A sentinel such as −999 is a storage convention, not a measured value. Including it in an arithmetic mean can produce a result dominated by the code rather than the observations. Validating a column requires its schema and meaning together.

Censoring is not ordinary missingness. A ‘below 0.2 mg/L’ record contains an upper-limit statement under the detection procedure; it is not an exact concentration of zero or necessarily a value of 0.1 mg/L. Similarly, a device that records every value above its range as ‘greater than 100’ supplies a lower bound rather than an exact reading of 100. Replacing such records by fixed numbers adds an imputation rule or model. Preserve the original flag and threshold even if a later analysis introduces a justified replacement.

A categorical ‘not applicable’ state can be scientifically different from ‘unknown’. For example an item about equipment-maintenance time may be inapplicable when no equipment is present, whereas a blank response from an equipment owner is missing. Collapsing both into zero minutes asserts a fact that was never observed. Denominators for proportions and summaries should follow the intended population of applicable observations and the analysis rule, with exclusions or missingness explicitly reported. Selection and nonresponse will be developed further in the sampling lesson.

Machine types remain useful safeguards. A string identifier can preserve leading zeros; a category dictionary can prevent accidental new spellings; unit-aware fields can reject adding kilograms to metres. But a schema cannot alone verify that a label is scientifically valid, a calibration is correct or a missingness mechanism is harmless. Keep raw records and documented transformations so an interpretation can be traced back to its source. Correcting metadata should not silently rewrite the underlying historical observations.

A transformation changes the question unless it is just a representation change

A unit conversion preserves a quantity while changing its representation. A nonlinear transformation generally defines a new variable with a new interpretation. For a positive quantity x and a stated positive reference x₀ of the same kind, z=ln(x/x₀) has a dimensionless logarithm argument. If x and x₀ are both converted from centimetres to metres, their ratio and hence z remain unchanged. Taking a logarithm of an isolated dimensional numerical value without documenting the reference hides a unit-dependent choice.

The identity ln(x₂/x₀)−ln(x₁/x₀)=ln(x₂/x₁) turns a positive ratio into a difference on the log scale. Multiplying all x values by a factor becomes an additive shift when the reference is held fixed. Changing the reference also shifts the log values in a known way. These are explicit algebraic relationships, not proof that the transformed observations are normally distributed or have a linear relationship with another variable. The transformation’s usefulness depends on the scientific and statistical model.

Zeros and negative values require attention. The real logarithm is undefined at zero and for negative inputs. Adding an arbitrary constant before logging defines another transformation whose meaning depends on that constant and its units. A true zero, a below-detection flag and a missing code should not be made indistinguishable merely to make software return finite numbers. A model or a different representation may be needed; the original records should remain recoverable.

Averaging on a transformed scale also defines a different summary. For positive values xᵢ relative to x₀, exp((1/n)Σln(xᵢ/x₀))x₀ is the geometric mean. For 2 kg and 8 kg it is 4 kg, while the arithmetic mean is 5 kg. Both transform properly under a common unit rescaling, but they answer different averaging questions. Choosing between them requires a rationale about additive amounts, multiplicative factors, a distribution model or another substantive target, rather than a preference for a smoother plot.

Some values have geometry beyond a straight ordered line

Directions measured around a circle wrap around: 0° and 360° represent the same direction. Bearings 359° and 1° are close, although their ordinary numerical difference is 358°. Their arithmetic code average is 180°, pointing the opposite way from the neighborhood occupied by both observations. The problem is the geometry of the represented variable, not insufficient decimal precision. Choosing a different angular origin can move the apparent numerical discontinuity without changing the directions themselves.

As an optional geometric connection, represent direction θ by the unit vector (cosθ,sinθ). Average these vectors and, if the resulting vector is nonzero, use its direction as a circular mean. For equal weights at 359° and 1°, the sine components cancel and the cosine components are positive, so the mean direction is 0°. For antipodal directions such as 90° and 270°, the mean vector is zero and this mean direction is undefined. The method therefore has an explicit nonzero-result condition rather than a universal numerical answer for every dataset.

Other scientific variables have several components. A position vector, a velocity and a collection of separate measurements require their component meanings, coordinate conventions and units. A list of numbers is not automatically a set of exchangeable scalar observations. Changing coordinates can transform all components together, and combining quantities with different units in an ordinary Euclidean distance requires reference scales or another justified metric. These issues extend the earlier algebraic lesson that a representation carries structure, not just values.

The four classic scale classes are useful for scalar comparisons, but they should not force a circular direction, a graph label or a structured object into an inappropriate straight-line ordering. One can derive scalar variables, such as an angular separation under a stated convention or a speed from a velocity vector. Doing so is a defined mathematical operation with its own loss of information and interpretation. Preserve the original structured record when a later analysis uses such a derived value.

A data type does not certify a statistical model

Knowing that a variable is a count does not establish a Poisson model, independence or a common exposure. Knowing that a measured value is continuously modeled does not establish a normal distribution or rule out bounds, mixtures or censoring. Knowing that a response is ordinal does not identify the spacing between its categories. A measurement description states what was recorded and which comparisons are meaningful; a statistical model adds a structure for variation, relationships and inference that must be justified for the task.

An ordinal threshold model is one example of an added structure. It imagines an unobserved variable Z and cut points such that the recorded category indicates the interval containing Z. The observed ordering alone does not identify the distribution, origin or scale of Z. A particular model supplies assumptions and identification conventions. Its estimated coefficients are conditional on those choices, not facts obtained just by replacing category labels with integers. Similar care applies to scoring rules that combine several ordered items.

For a nominal outcome, a model can predict category probabilities through explicitly defined indicator variables. Those probabilities have numerical meaning even though the labels do not have a meaningful arithmetic average. For a ratio quantity, a regression or a test can still be inappropriate if the design, dependence, functional form or error assumptions fail. Neither the most detailed measurement scale nor an impressive software output supplies all the missing model conditions.

Transformations can be diagnostic tools rather than automatic licenses. If a conclusion changes under a harmless relabeling or unit conversion in a way that its claimed meaning should not, inspect the calculation and interpretation. If a deliberately nonlinear transformation changes the target, document that change and justify it. Questions about robustness, uncertainty and model fit then become explicit parts of the analysis. This is more informative than applying a rigid list of permitted statistics without considering what they estimate.

A useful dataset retains meaning through the whole calculation

To design a small study of packaged material, define one observation unit, such as one selected package at a specified inspection. Keep the package identifier as a nominal label, the material category with its controlled vocabulary, the count of defined defects with the inspected exposure, the mass with its unit and procedure, and any ordered quality rating with its category descriptions. A count of inspections, a count of packages and a count of defects are different denominators. Their meanings should remain visible when constructing summaries.

Document transformations as operations: converting grams to kilograms, forming a defect rate per inspected area, defining an indicator for a target category, rounding a displayed value or creating a log-ratio variable. Record missing and censored states separately, retain original values, and specify which observations enter each calculation. The observation design also matters: repeated packages from one production run can share conditions, and a convenience sample may not represent the intended population. The next lesson takes up that sampling question directly.

The connections to earlier mathematics are concrete. Equality classes explain nominal labels; order relations explain ordinal information; affine functions explain interval conversions; proportions and dimensional analysis explain ratios; functions explain quantization and derived variables; geometry explains directions and vectors. These connections provide reasons for a calculation instead of a memorized vocabulary list. They also show where information is lost or additional assumptions are introduced.

Before accepting an output, ask four questions: what was observed, what structure was retained, what operation was applied, and what claim is now being made? A correctly computed mean of identifiers still lacks an average-identifier interpretation. A correctly converted temperature ratio can still answer the wrong physical question. Conversely, an explicitly defined category proportion or model-based ordinal analysis can be entirely meaningful. Good data reasoning keeps the scientific target, representation, calculation and inference connected without treating any one of them as a substitute for the rest.

Worked examples

Try it yourself

1. A table has one row per household per month. Is the number of rows necessarily the number of distinct households, and are the rows automatically independent?

   Show the answer and reasoning

   Neither follows; households can contribute repeated rows, and independence needs a design or model.

   The observation unit includes both household and month. An identifier and time field preserve the repeated structure, which can affect weighting and uncertainty even when each value has valid units.

2. A material type is stored as integer code 12. Does its machine type make it a quantitative amount?

   Show the answer and reasoning

   No; it can be a nominal label.

   The codebook determines what equality, order and arithmetic mean. Storing a label as an integer supplies no intrinsic size or spacing, just as storing a mass as text does not erase its quantitative meaning.

3. A score is restricted to 0, 1/2, 1, 3/2, 2. Is it continuous because some values have decimal representations?

   Show the answer and reasoning

   No. It has five allowed values and is discrete.

   Discrete numerical values can be nonintegers. Allowed values and the measurement procedure matter; a decimal point is only a recording convention.

4. Does recording an idealized continuously varying length to the nearest millimetre make the underlying target a millimetre-grid quantity?

   Show the answer and reasoning

   No; the recorded variable is quantized while the target can remain continuously modeled.

   Many underlying inputs map to the same record. Preserve both the target definition and recording rule, and do not treat appended digits as recovered information.

5. For category observations A, B, A, A, define I_A=1 for A and 0 otherwise. What is its mean, and what has not been assumed?

   Show the answer and reasoning

   3/4, the observed A proportion; independence and population representativeness have not been assumed.

   The indicator is an intentionally defined numerical membership function. Its average is meaningful even though an arithmetic mean of arbitrary nominal category codes would lack an intrinsic category interpretation.

6. Three ordered categories use codes 1, 2, 3. Does their order establish that the upper-to-middle gap equals the middle-to-lower gap?

   Show the answer and reasoning

   No; equal spacing is an additional assumption.

   The increasing recoding 1, 2, 100 preserves order and changes the gaps. Ordered comparisons and cumulative counts survive, while arbitrary gap arithmetic does not automatically do so.

7. Which statement is justified under every strictly increasing recoding: an odd-sample middle order statistic follows the recoding, or the arithmetic average of two central numeric codes always does?

   Show the answer and reasoning

   The odd-sample middle order statistic follows the recoding; the averaged two-code convention does not generally do so.

   Increasing functions preserve ranks but need not preserve arithmetic midpoints. For even ordinal samples, retain the central ordered pair or state a justified median convention.

8. Under y=3x+7, what happens to x₂−x₁ and to a ratio of two nonzero differences?

   Show the answer and reasoning

   Each difference is multiplied by 3, and the ratio of differences is unchanged.

   The offset cancels in subtraction and the common positive factor cancels in a ratio. A raw ratio of the original readings generally does not have that shift invariance.

9. Is 30 °C ‘one and a half times the temperature’ of 20 °C in an origin-independent numerical-ratio sense?

   Show the answer and reasoning

   No. The Celsius numerical ratio changes when the same states are represented in kelvins or Fahrenheit degrees.

   Temperature intervals convert consistently, and positive thermodynamic-temperature ratios have their own physical interpretation. Neither licenses an unqualified claim about perceived warmth or arbitrary heat content.

10. Masses 4 kg and 10 kg become 4000 g and 10000 g. Which ratio is preserved, and what denominator condition is required?

    Show the answer and reasoning

    The second-to-first ratio 5/2 is preserved, with a nonzero first mass.

    A common positive scaling cancels in the ratio. A zero denominator would make that ratio undefined; a zero instrument indication also needs its recording interpretation checked.

11. For y=ax+b with a>0, how do the mean, variance and standard deviation transform when the variance convention is kept the same?

    Show the answer and reasoning

    Mean becomes a x̄+b, variance is multiplied by a², and standard deviation by a.

    Subtracting the new mean gives yᵢ−ȳ=a(xᵢ−x̄). The result is an algebraic transformation identity, not a claim that the data are independent, normal or physically summarized adequately by an unweighted mean.

12. Why is s/x̄ generally not an origin-independent coefficient of relative variability for an interval scale?

    Show the answer and reasoning

    An origin shift changes it to a s/(a x̄+b), which generally differs from s/x̄.

    A pure positive rescaling preserves CV, but an arbitrary interval origin shift does not. A nonzero mean and a scientifically justified relative-variability interpretation are also needed.

13. Does a nearest-centimetre display by itself establish a uniform rounding distribution, a 95% confidence interval or the total measurement uncertainty?

    Show the answer and reasoning

    No.

    The increment and tie rule can define a quantization cell, but a probability law within it is an added assumption. Calibration, repeatability, environment and reference information can contribute additional uncertainty.

14. A recorded length is 2.50 m and a stipulated reference is 2.503 m. What is the signed recorded-minus-reference difference?

    Show the answer and reasoning

    −0.003 m, or −3 mm.

    The negative sign means the record is below that reference. The calculation does not imply that all real measurement errors are known, or that an actual reference has no uncertainty.

15. Should ‘missing’, ‘not applicable’, ‘below detection’ and a measured zero automatically share the same numerical code for arithmetic?

    Show the answer and reasoning

    No; they represent distinct information states.

    A common zero code can invent an exact value, change denominators and lose censoring information. Preserve flags and thresholds; any imputation or analysis exclusion requires its own stated rule.

16. Why write ln(x/x₀) for positive same-kind quantities with a positive reference rather than logging an unexplained dimensional number?

    Show the answer and reasoning

    The ratio is dimensionless and remains unchanged when both values and reference undergo the same unit conversion.

    The reference states what the transformed variable means. Zero and negative inputs are outside the real logarithm’s domain, and a nonlinear transformation changes the summary question rather than merely adding display digits.

17. Why does the vector-based circular mean direction fail for equal-weight directions 0° and 180°?

    Show the answer and reasoning

    Their unit vectors average to zero, which has no direction.

    A geometry-aware method still has applicability conditions. It must report the undefined direction rather than assign a misleading default angle.

18. A variable is continuous and on a ratio scale. Does that alone justify a normal model, independence, a causal comparison or a particular statistical test?

    Show the answer and reasoning

    No; those require their own design, substantive and statistical assumptions.

    Measurement structure describes recorded information and meaningful transformations. It does not certify a distribution, sampling mechanism, functional relationship or causal identification strategy.

What to carry forward

• Define the observation unit, target variable, allowed values, procedure and recording convention before calculating.
• A numerical storage code can be a label, and a text record can represent a quantitative value.
• Discrete variables can have noninteger values; a rounded record and a continuously modeled underlying target are different variables.
• Nominal labels support equality and corresponding counts; explicit indicators supply meaningful category proportions.
• Ordinal information supplies order without equal spacing, and median conventions need their correct transformation qualifications.
• Interval representations allow positive affine changes; ratio representations preserve an intended zero under positive rescaling.
• Meaningful values can transform equivariantly rather than remain numerically unchanged after a unit conversion.
• Celsius and Fahrenheit origins invalidate raw arbitrary-origin temperature ratios; differences and affine means transform consistently.
• Ratios need compatible quantities, a relevant denominator and a stated interpretation of zero.
• Resolution, rounding, signed reference error and uncertainty are distinct parts of a measurement description.
• Missing, inapplicable, censored and true-zero records must not silently become the same exact numerical value.
• Logarithms require dimensionless references and valid domains; averaging after transformation can define a different target.
• Circular and multicomponent variables require their geometry and metadata, beyond a forced ordinary scalar ordering.
• A scale label does not approve a statistical model: design, dependence, distribution and scientific purpose remain separate.
• Keep original records, transformation rules and codebook history so each result can be traced to what was actually observed.