With So Many Data Classifications, Do We Need to Learn Them All? What Are They For?

1. Introduction: Isn't Data Just Numbers?

For beginners in econometrics, the first wave of shock often comes from the classification of data.

The instructor fluently lists terms at the podium: quantitative data, qualitative data, cross-sectional data, time series, panel data, experimental data, observational data, balanced panel, unbalanced panel...

You sit below, with only one question in mind: What is the point of all these classifications? Aren't they all just numbers? Can't we just throw them into a regression?

No, you can't. And "throwing them into a regression" is precisely the most common source of errors for beginners.

The type of data determines which models you can use, what questions you can answer, and what conclusions you can draw. Choosing the wrong model is not a matter of precision—it means the entire analytical framework is invalid.

This article follows three main classification threads to help you build a "classification map" of data, so that when you encounter a new dataset in the future, you can immediately judge: What kind of data is this? What is it suitable for? What pitfalls are there?

Core message: Data classification is not a taxonomic game—each classification dimension corresponds to a choice of analytical strategy.

---

2. First Classification Thread: By Variable Measurement Scale—Quantitative, Qualitative, and Ordinal

This is the most basic classification and the starting point of all data analysis.

2.1 Definitions and Examples of the Three Types

Quantitative Data: The values of the variable are themselves numbers, and the differences between numbers have substantive meaning.

• Continuous: wage (yuan), GDP (100 million yuan), height (cm), temperature (°C)
• Discrete: number of children, number of patents, years since firm establishment

Core feature of quantitative data: You can perform arithmetic on it. 100 yuan is twice 50 yuan—this operation is meaningful.

Qualitative Data: The values of the variable are categories; numbers are merely labels and cannot be used in arithmetic. In statistics, this is usually called categorical data.

• Gender (male/female—coded as 0/1, but 0+1 or 0×1 is meaningless)
• Marital status (unmarried/married/divorced/widowed)
• Industry classification (manufacturing/services/agriculture/...)
• Region (eastern/central/western/northeastern)

Core feature of qualitative data: You can only compare "whether they are the same," not "which is larger than which."

Ordinal Data: Falls between the two. Values have an order and can be compared in magnitude, but the differences between values have no numerical meaning.

• Education level (primary < junior high < high school < college < graduate school—but "college minus high school" does not equal "graduate school minus college")
• Satisfaction ratings (very dissatisfied/dissatisfied/neutral/satisfied/very satisfied—the gap between "satisfied" and "very satisfied" is not equal to the gap between "dissatisfied" and "neutral")
• You can rate a performance on a 1–5 scale in a competition

Core feature of ordinal data: You can rank, but you cannot compute differences. Treating an ordinal variable directly as a continuous variable in a regression is one of the most common mistakes beginners make.

2.2 Why Distinguish Them?—Three Practical Reasons

Reason 1: It determines which model you use.

Dependent variable type	Applicable basic model
Continuous quantitative	OLS (ordinary least squares)
0/1 binary (qualitative)	Logit / Probit (binary choice model)
Multicategory (qualitative)	Multinomial Logit
Ordinal	Ordered Logit / Probit
Count (discrete quantitative)	Poisson regression / negative binomial regression

If you run OLS on a 0/1 variable, your predicted values may be -0.3 or 1.2, which is logically nonsensical. In undergraduate econometrics courses, "why OLS cannot be used when Y is a binary variable" is usually one of the first topics covered.

Reason 2: It determines how you interpret coefficients.

Consider the same question, "When X increases by 1 unit, how does Y change?":

• Continuous dependent variable + OLS: $\beta_1$ = when X increases by 1 unit, Y changes by $\beta_1$ units on average. Intuition is clear.
• Logit model: The coefficient itself cannot be interpreted directly. You need to compute the marginal effect or the odds ratio—this is where beginners often get confused.
• Ordered Logit model: The coefficient tells you "how much the log-odds of moving into a higher category change when X increases by 1 unit"—the interpretation hurdle is even higher.

Reason 3: It determines whether you can perform certain operations.

• Taking the mean of a qualitative variable? Meaningless. (What is the "average gender"?)
• Taking the standard deviation of an ordinal variable? Computable mathematically, but the meaning is unclear. (What is a "standard deviation of satisfaction of 0.8"?)
• Treating an ordinal variable as continuous in a regression? This is common practice in many settings (e.g., using years of schooling to represent education level), but you must be aware of what you are doing and what you are assuming—you are effectively assuming that "the jump from primary to junior high" and "the jump from high school to college" have equal effects. If this assumption fails, your estimates may be wrong.

In short: The measurement scale of a variable is the first threshold for choosing a model and interpreting results. If you go through the wrong door, everything that follows will be off track.

---

3. Second Classification Thread: By Spatiotemporal Structure—Cross-Sectional, Time Series, Panel, and Pooled Cross-Sectional

This classification thread is directly related to "how the data were collected," and it determines which models you can use and what types of questions you can primarily answer.

3.1 Definitions of the Four Types

Cross-Sectional Data: Data obtained by observing multiple individuals at the same point in time (or approximately the same point in time).

• Typical example: The 2020 Seventh National Population Census of China—surveying the entire population of the country in the same year.
• In economics: A single wave of the China Family Panel Studies (CFPS), annual reports of listed companies in a given year, or all returned samples from a particular survey are all cross-sectional data.

Key reminder—"same point in time" is often an approximation: In practice, the "same point in time" for cross-sectional data is usually an approximation that ignores micro-level time differences. For example, household interviews in a given CFPS wave may run from January of that year through March of the following year, spanning more than a dozen months. In analysis, however, we typically treat these data as if they were collected at "the same point in time"—because differences of a dozen months are negligible relative to the questions we study (e.g., the long-term effect of education on income).

But note: if a major event occurred during the survey period (e.g., a survey during the pandemic spanning the pre- and post-lockdown periods), this "approximation" can become problematic. Different respondents interviewed at different times may reflect vastly different external environments—yet you assume they are at "the same point in time" when modeling.

Time-Series Data: Repeated observations of the same subject at multiple points in time.

• Typical examples: China's annual GDP from 1990 to 2023, the daily closing price of a stock, the monthly CPI index of a city.
• Core feature: There is only one observation subject (or one aggregated economy), but it is observed over a long period.

Panel Data: The intersection of cross-sectional and time-series data—tracking multiple individuals at multiple points in time. Also called longitudinal data.

• Typical examples: CFPS tracks the same households with follow-up visits every two years, forming a structure of "multiple households × multiple survey waves." The Chinese Industrial Enterprises Database tracks the same firms over multiple years and is also panel data.
• Core feature: The same individuals are observed repeatedly, and you capture each person's trajectory of change over time.

Pooled Cross-Sectional Data: The type most easily confused with panel data. It also involves "multiple time periods × multiple individuals," but—each wave draws a different sample, and the same individuals are not tracked.

• Typical example: The Chinese General Social Survey (CGSS) is conducted annually, but each year draws different respondents. You cannot track "the same person" from 2015 to 2020.
• Key difference from panel data: Panel data tracks the same set of individuals, while pooled cross-sections draw different individuals each time.

3.2 A Table to Distinguish the Four Structures

Data type	Individual dimension	Time dimension	Tracks the same sample?	Typical example
Cross-sectional	Multiple	One point (approximate)	Not applicable	A year's census
Time series	One	Multiple points	Not applicable	GDP over years
Panel data	Multiple	Multiple points	Yes	CFPS tracking survey
Pooled cross-section	Multiple	Multiple points	No, independent draws each time	CGSS annual survey

3.3 Why Distinguish Them?—Each Type Answers Different Questions

Cross-sectional data primarily answers: differences between individuals.

You find that "people with more education earn higher wages"—this compares different people in a cross-section. But you cannot answer "how much would a person's wage increase if they studied one more year"—because you do not observe changes within the same person. You only see that "highly educated people earn more," which could be because education raises productivity, or because more capable people tend to pursue more education in the first place—this cannot be distinguished in a cross-section.

Time-series data primarily answers: how a variable evolves over time.

Trend analysis of GDP growth rates, inflation rates, and unemployment rates—these are classic applications of time series. But the limitation of time series is: you cannot use a single subject's time series to study heterogeneity across individuals. You see China's overall GDP growth, but not "why Guangdong grows faster than the Northeast"—because the data have already been aggregated.

Panel data can answer the most questions—this is why it has become the dominant data structure in modern empirical research.

Because panel data contain both cross-individual differences and within-individual temporal variation, it can do far more than the previous two types:

• Controlling for unit fixed effects: Each individual serves as its own control. You can control for all time-invariant omitted variables—such as ability, personality, family background, and cultural traditions. This is the most powerful weapon of panel data.
• Studying dynamic adjustment processes: How long does it take from shock to response? Is the wage response to policy changes immediate or gradual?
• Separating age effects, cohort effects, and period effects—this is nearly impossible in pure cross-sections or pure time series.

Pooled cross-sectional data lies between cross-sections and panels. It adds a "time dimension" beyond a pure cross-section, allowing you to compare aggregate characteristics across periods (e.g., "comparing income distributions between 2010 and 2020"); but it cannot achieve the "within-individual tracking" of panel data—because you draw different people each time. Nevertheless, pooled cross-sections have their own unique advantages: you can include time dummy variables to control for macro trends, or in policy evaluation, use "different individuals at different time points" to construct a control group.

3.4 Points That Are Easy to Overlook

(1) The "micro-level time differences" problem in cross-sections.

As noted above, the "same point in time" in cross-sectional data is usually an approximation. In your data description, you should note the specific time span of the survey and consider in robustness checks: if I use only data from the first three months, do the results still hold?

(2) The "attrition" problem in panel data.

Panel data track the same people, but people move, refuse to participate, or die. If you find that "those who dropped out" and "those who remained" differ systematically on key characteristics (e.g., low-income individuals are more likely to drop out), then your estimates may suffer from selection bias.

(3) Pooled cross-sections cannot substitute for panels.

A common mistake among beginners is to take CGSS 2015 and CGSS 2020 data and run a fixed effects model as if it were panel data. But CGSS draws different people each time—you have not tracked individual changes, so a fixed effects model is simply out of the question. Pooled cross-sections can only be used as "cross-sections with a time dimension," not as panels.

(4) The "non-stationarity" problem in time series.

If two variables both trend upward over time (e.g., GDP and telephone ownership are both growing), their regression may appear highly significant—but this is only spurious correlation driven by time trends, not a true causal relationship. In time-series analysis, you must first conduct stationarity tests, which are completely unnecessary in cross-sectional analysis.

---

4. Third Classification Thread: By Data Generation Process—Experimental Data vs. Observational Data

This classification thread directly determines whether you can—and how easily you can—conduct causal inference.

4.1 Definitions and Differences

Experimental Data: The researcher actively manipulates a treatment, then randomly assigns subjects to treatment and control groups, and records outcomes under controlled conditions.

• Core elements: Manipulation + randomization + control
• Typical sources: Medical RCTs, psychology laboratory experiments, field experiments in development economics, A/B testing

Observational Data: The researcher does not intervene in the data generation process but merely collects and records facts that have already occurred or occur naturally.

• Core feature: You observe things after they have happened
• Typical sources: Survey data, administrative records, annual reports of listed companies, satellite remote sensing data, social media data

4.2 Why Is This Distinction Crucial?

Consider an example:

You want to study whether "attending college raises income."

• If you had experimental data: You randomly draw a group of people, randomly assign half to attend college and half not to, then compare their wages after graduation. Because of randomization, the two groups are comparable in ability, family background, and motivation—the wage difference has only one explanation: whether they attended college.
• If you only have observational data: You have data on people who have already finished college and are working. Those who attended college differ systematically from those who did not in ability, family background, and academic motivation—they would likely earn higher wages anyway, regardless of college attendance. The wage difference you observe is not entirely the causal effect of education.

This logical distinction is the starting point of the entire field of causal inference.

4.3 A Checklist of Core Problems with Observational Data

Core problem 1: Selection bias—"People do not appear randomly in the positions they occupy."

People in high-income regions are healthier—is it because high income improves health, or because healthier people are more likely to earn high incomes? People who attend training programs earn more—is it because training is effective, or because motivated people both sign up for training and would earn high wages anyway?

In observational data, you must always ask yourself: Why are these individuals in their current state? Does this "why" contain systematic factors that also affect the outcome?

Core problem 2: Reverse causality—"The correlation you see may run in the opposite direction from what you assume."

You find that "areas with more police have higher crime rates"—do police cause crime, or do high-crime areas deploy more police? You find that "time spent on social media is positively correlated with depression"—does social media cause depression, or do depressed people tend to spend more time scrolling on their phones?

In cross-sectional observational data, reverse causality and forward causality look identical—you have the same correlation coefficient, and both stories are plausible.

Core problem 3: Omitted variable bias—"Factors you did not measure may be driving everything."

Ice cream sales and drowning deaths are highly positively correlated—does eating ice cream cause drowning? No, it is because of hot weather—hot weather → ice cream consumption, hot weather → swimming → drowning. If you do not measure "temperature," you will mistakenly interpret the correlation between ice cream and drowning as causal.

Core problem 4: Measurement error—"What you measure is not what you want."

Using "self-reported height" instead of "actual height"—people tend to report slightly higher values. Using "firm-reported profits" instead of "true profits"—firms with tax avoidance motives will underreport.

When a key independent variable contains measurement error, the classic result is: OLS estimates are attenuated toward zero (attenuation bias)—you underestimate the true effect. The most frightening aspect is: many empirical studies inevitably contain measurement error due to data limitations, yet researchers may not realize how it affects their conclusions.

4.4 Is Experimental Data Always Perfect?

No. Experimental data also have their own problems:

• External validity: Can findings from the laboratory be generalized to the real world?
• Hawthorne effect: Subjects change their behavior when they know they are being observed.
• Compliance problems: Random assignment is done, but some people do not follow the assignment (noncompliance)—in this case, instrumental variables are needed to estimate the "local average treatment effect (LATE)" for compliers.
• Ethical constraints: Not all questions can be answered with experiments. You cannot randomly assign one group to smoke and another not to, then wait 20 years to see who gets lung cancer.

The core advantage of experimental data lies in internal validity (you can be confident that X caused Y), but external validity (how far the conclusions generalize) requires additional argumentation. The core weakness of observational data lies in internal validity (it is difficult to cleanly identify causal relationships), but its advantage lies in external validity—you are studying what is actually happening in the real world.

---

5. Additional Important Classifications—Four Concept Pairs You Cannot Miss

Beyond the three main threads above, several other data classification concepts recur frequently in applied research and deserve their own section.

5.1 By Level of Analysis: Micro Data vs. Macro Data

Micro Data: The unit of observation is an individual—a person, household, firm, or school. Each record represents a specific micro-level decision-making unit.

Macro Data: The unit of observation is an aggregated economy or region—a country, province, or industry. Each record represents a statistic at an aggregate level.

Why does this matter?

There is a famous trap between these two levels of data: the ecological fallacy—using macro-level correlations to infer micro-level individual behavior.

A classic example: At the provincial level, you find that "provinces with a higher proportion of migrant workers have higher wages for local residents." But this does not mean that "a local resident who has more contact with migrant workers will see his own wage rise"—the provincial correlation may reflect "migrants tend to move to provinces with better economies and higher wages," not an individual-level causal relationship.

There is also a trap in the opposite direction: the atomistic fallacy—using micro-level individual behavior to directly infer macro-level phenomena. Individual thrift is a virtue at the individual level, but at the macro level it can lead to the "paradox of thrift"—if everyone saves, aggregate demand collapses.

5.2 By Panel Completeness: Balanced Panel vs. Unbalanced Panel

Balanced Panel: Every individual has observations in every period. Like a complete rectangle—N individuals × T periods, with no gaps.

Unbalanced Panel: Some individuals are missing in some periods. Firms have no data before they are established, no data after they go bankrupt, and respondents may refuse to participate in a given year—all of these create an unbalanced panel.

Why does this matter?

• Unbalanced panels are more common—data in the real world are almost always unbalanced.
• But imbalance can introduce selection bias: if "missing observations" are not random—for example, poorly performing firms disappear early from the panel (bankruptcy)—then analyzing only surviving firms will overestimate average performance (survivorship bias).
• Most software implementations of panel data models (e.g., Stata's xtreg) automatically handle unbalanced panels, but they do not automatically handle the selection problems that cause the imbalance—this requires your own judgment and argumentation.

5.3 By Data Source: Primary Data vs. Secondary Data

Primary Data: Data collected by the researcher themselves, using instruments they designed. You design the questionnaire, you recruit subjects, you run the experiment—the data are "yours."

Secondary Data: Data collected by others that you use. You use CFPS, CGSS, listed company annual reports, or satellite data—the data are "borrowed."

Why does this matter?

• Primary data: You can measure precisely the variables you want. The disadvantage is the extremely high cost, sample sizes are usually limited, and external validity requires additional argumentation.
• Secondary data: Large sample sizes, broad coverage, low cost, and much of it is publicly available. But the variable definitions are not yours to set—if you want to study "social capital," the database has no variable directly called "social capital," so you must approximate it with proxy variables such as "whether one participates in community activities" or "neighborhood trust." This operation of "using available variables to approximate the concept you want" runs through virtually all empirical research using secondary data.
• Evaluation criterion: When reading an empirical paper that uses secondary data, you should pay special attention to how the authors argue that "this proxy variable truly represents the concept I am referring to." This is not pedantry—it is the core of persuasiveness in secondary data research.

5.4 By Data Structure: Long Format vs. Wide Format

This is a classification at the data-processing level, but every empirical researcher must deal with it.

Wide Format: Each individual occupies one row, with variables from different periods laid out horizontally.

ID   income_2018   income_2019   income_2020
1    50000         52000         55000
2    38000         40000         .

Long Format: Each individual–period combination occupies one row.

ID   year   income
1    2018   50000
1    2019   52000
1    2020   55000
2    2018   38000
2    2019   40000

Why does this matter?

• Most econometric software (Stata, R, Python's statsmodels) regression commands require long format (panel format) by default.
• Converting from wide to long format (reshape long / pivot_longer) is one of the most frequent operations in data cleaning.
• Human reading habits favor wide format, while statistical software favors long format—you need to be proficient at converting between the two. Although this is not a theoretical issue, the time spent on wide-to-long conversion may exceed the time spent understanding the essence of the model—this is the daily reality of empirical research.

---

6. Summary: Classification Map + Practical Wisdom

Four Questions to Ask Yourself When You Get a New Dataset:

Question 1: What measurement scale does each variable belong to? (quantitative/qualitative/ordinal)
      → Determines which models I can use and how I interpret coefficients

Question 2: What is the spatiotemporal structure of the data? (cross-sectional/time series/panel/pooled cross-section)
      → Determines what types of questions I can answer and whether I can control for unobservable factors

Question 3: How were the data generated? (experimental/observational)
      → Determines how far I am from causal inference and how much effort I need to rule out alternative explanations

Question 4: What implicit limitations does the data have? (level, missing patterns, source, format)
      → Determines which potential biases I need to remain alert to

Three Takeaways

These classifications do not need to be memorized, but they need to be "recognized."

The next time you open the data description section of a paper and see "we use a balanced panel of 2,345 firms from 2010 to 2020," you should immediately realize: balanced panel → possible survivorship bias; firm data → observational data → causal identification requires persuasiveness; panel structure → fixed effects can be used to control for time-invariant omitted variables.

The value of classification lies not in classification itself, but in what it tells you about what you can and cannot do.

The moment you know your data are "observational cross-sectional data with a qualitative dependent variable," you have already ruled out 80% of inapplicable methods, you know you must confront the challenge of selection bias, and you know you need Logit rather than OLS—this is the meaning of classification. It is not about labeling data, but about helping you see the terrain clearly before you begin operations.

The ultimate payoff of learning data classification is not how many classification terms you remember—but that when you later see a research finding with "R² = 0.98," you can instinctively ask:

"What type of data is this? Does the model match the method? Is that high R² because the model is genuinely good, or because two variables in a time series happen to lie on the same trend line?"

This instinct is what data classification knowledge looks like once it has entered your bloodstream.

---

7. Presentation Suggestions for Bilibili/WeChat Official Account

• Bilibili video: For the four classification threads (measurement scale, spatiotemporal structure, data generation process, and other supplements), use an animated metaphor of "four roads leading to the same city," with specific case visuals for each thread (e.g., "quantitative vs. qualitative" using a scale and a label book as visual associations).
• WeChat official account: There are many tables (4 comparison tables in total), so consider making them into a series of infographics with a unified color scheme; for the code block (wide-to-long format conversion), use a handwriting-style presentation to reduce "code anxiety."
• Recommended titles:
  • Main title: "With So Many Data Classifications, Do We Need to Learn Them All?"
  • Alternative title: "The First Thing to Do with New Data Is Not to Run a Regression—But to Figure Out What Kind of Data It Is"
• Memorable quotes (suitable for cover cards or closing frames):

  "The type of data determines which models you can use, what questions you can answer, and what conclusions you can draw. Choosing the wrong model is not a matter of precision—it means the entire analytical framework is invalid."