When You Receive a Dataset, How Should You Scrutinize Its Representativeness and Validity from a Sampling Perspective?

I. Introduction: Why Does the Sampling Theory You Learned Seem to "Disappear" When You Need It?

Anyone who has taken an econometrics methods course has likely studied sampling—simple random sampling, stratified sampling, cluster sampling, systematic sampling, and PPS sampling, memorizing the formulas and characteristics of each. For exams, you know stratified sampling reduces variance and how to calculate the design effect for cluster sampling.

But when it comes to real-world application, this knowledge seems to vanish.

When you open a CFPS or CGSS dataset, when you read a paper stating "this study uses data from the 2018 XX Survey," or when you collect 500 questionnaires yourself—do you instinctively scrutinize the data through the lens of sampling theory?

Most people don't. People care more about "how to define the dependent and independent variables" and "whether the regression results are significant." As for where the data came from and whether the sample can represent the population under study—honestly, few people seriously investigate this.

This is precisely the problem this article aims to solve: to help you transform sampling theory from "knowledge forgotten after the exam" into "a practical toolkit for scrutinizing any dataset you encounter."

Core message: The upper bound of data quality is determined at the moment the data is collected. All subsequent econometric techniques merely approximate the truth as closely as possible given existing conditions—they cannot conjure information out of thin air.

---

II. Scrutinizing Data Provenance: Five Core Questions You Must Ask

When you face a dataset—whether self-collected or downloaded from a public database—the following five questions constitute the basic framework for scrutinizing its representativeness and validity. Each question you ask eliminates one risk that could undermine your entire study.

Question 1: What Is the Target Population? Does the Sampling Frame Cover It?

This is the most fundamental question. If an error occurs here, all subsequent analysis rests on a false premise.

Target Population: The group you truly intend to study. For example, "China's working-age population aged 18–60."

Sampling Frame: The list or scope from which you can actually draw your sample. For example, "the national database of mobile phone number prefixes."

Key point: The sampling frame ≠ the target population. The gap between them is called coverage bias, and this is the first diagnostic you should perform after obtaining data.

Examples that should immediately raise red flags:

• You want to study "political attitudes of Chinese adults," but the data comes from an online questionnaire. The sampling frame is "netizens who can access the internet and are willing to fill out surveys"—many elderly, rural residents, and low-income individuals are systematically excluded. Your conclusions do not represent "Chinese people," only "that subset of netizens willing to complete surveys."
• You want to study "innovation behavior of Chinese firms," but the data consists of A-share listed companies. How many firms does the sampling frame cover? China has tens of millions of firms, yet only a few thousand are listed. Small and medium-sized enterprises—which may innovate in entirely different ways—are completely outside your field of vision.
• You want to study "human risk preferences," using data from a psychology subject pool consisting of undergraduates enrolled in introductory psychology at one university. The sampling frame is students from a single department, while the target population is "humanity." How large is this gap? Nature and Science have debated it for decades.

Your action: After obtaining data, the first step is to ask—what is the sampling frame of this data? What is the target population I intend to study? Along which dimensions, and for which groups, are the gaps systematic? If your conclusions might not hold for the excluded groups, does this direction make the effect stronger or weaker?

Question 2: What Sampling Method Was Used? Is It Truly "Random"?

The term "random sampling" is arguably as overused as "significant."

Strictly defined random sampling: Each individual in the population has a known and nonzero probability of being selected. The emphasis is on "known"—you know the probability of each individual being chosen. Precisely because you know this probability, you can use it to construct sampling weights, recovering unbiased estimates of the population from the sample.

Several common scenarios of "fake randomness" that should immediately raise alarm:

• Street-intercept surveys: It appears random to stop passersby, but in reality, who gets stopped depends on the interviewer's subjective judgment (more likely to approach friendly-looking people) and who happens to pass that street at that time. Each individual's probability of selection is completely unknown—this is not random sampling.
• Online questionnaires shared via social media: Respondents are all volunteers. "Voluntary" and "random" are precisely antonyms—those who are not interested have a zero probability of being selected.
• "We randomly selected three classes"—but was the "selection" process truly a random draw, or was it "picking a few convenient classes"? If the latter, systematic differences between classes (different teachers, different class times, different student backgrounds) will be confounded with your treatment effects.

Your action: For large public datasets (CFPS, CGSS, etc.), read the technical documentation—these surveys provide detailed sampling design descriptions. For self-collected data, honestly describe the recruitment method. If it is not random sampling, do not call it "random." Furthermore, even if the data comes from a rigorous complex multi-stage sampling design (such as CFPS), you cannot simply use the default standard errors from the reg command—it assumes simple random sampling, whereas your data has clustering and stratification structures, requiring adjusted standard errors (e.g., clustered standard errors).

Question 3: Does the Sample Structure Match the Population Structure?

Even if the sampling frame perfectly covers the target population and the sampling method is strictly random, sampling error still exists—draw another 1,000 people and the sample structure will fluctuate.

What you need to answer is: How large is the fluctuation? On certain key dimensions, has the sample systematically deviated from the population?

The procedure is straightforward:

1. Select several key benchmark variables—gender, age group, education level, urban-rural distribution, region.
2. Find the true population distribution of these variables from authoritative sources (census bulletins, statistical yearbooks).
3. Create a "sample vs. population" comparison table.

A simulated example—suppose you analyze household income using data from a survey:

Dimension	Sample Distribution	National Population	Direction of Deviation
Urban population share	72%	64% (2020 census)	Urban overrepresented
College degree or above	35%	15% (2020 census)	Highly educated overrepresented
Age 60+ share	8%	19% (2020 census)	Elderly severely underrepresented

If you find the sample overrepresents urban and highly educated individuals—and these two groups happen to have systematically higher incomes—then the "average income" you compute from the sample will be significantly higher than the true national level.

But note: You do not need the sample to perfectly match the population on every dimension (impossible with finite samples). What you truly need to judge is—are the dimensions of deviation systematically correlated with your core variables of interest?

If your research question concerns "returns to education," a slight skew in the sample's age distribution may not matter much; but if the sample's education distribution severely deviates from the population (too many highly educated individuals), your estimated returns to education will be problematic—because the sample lacks sufficient low-education individuals for valid comparison.

Question 4: Who Did Not Respond?—Nonresponse Bias

This issue is severely underestimated in empirical research.

You sampled 1,000 people, but only 600 ultimately answered your questions. What about the other 400?

If "nonrespondents" and "respondents" are identical on all relevant characteristics—then nonresponse merely reduces the sample size without affecting the unbiasedness of your estimates.

But reality is almost never like this.

• Income surveys: Both the very rich and the very poor are less willing to respond than middle-income earners. The former cite privacy concerns; the latter feel alienated from and distrustful of "official surveys."
• Health surveys: The least healthy may refuse to participate because "there's no hope anyway," or they may be more willing because "someone finally cares about me"—the direction of bias is uncertain, but bias certainly exists.
• Firm surveys: The worst-performing firms often do not return questionnaires. The "average firm profit margin" you observe is biased toward the better-performing half.

Your action: If the data includes response rate information, report it. If the data has a tracking design (such as prior-wave information in panel data), you can compare "this-wave refusers" and "this-wave respondents" on observable characteristics from the previous wave—this at least gives you a clue about the direction of bias. If nothing is available, at minimum state clearly in the discussion section: "The nonresponse mechanism is unknown. If nonresponse correlates with characteristic X, our estimates may be biased in direction Y."

Question 5: Does the Data Have Sampling Weights? Are You Using Them Correctly?

Many large surveys (CFPS, CGSS, CHFS, U.S. CPS, etc.) provide sampling weights. Yet a large number of users, upon obtaining the data, immediately proceed to descriptive statistics and regression analysis, completely ignoring the weight variables.

Why do weights exist? Because complex sampling designs are not simple "equal probability of selection for everyone." Certain groups (e.g., ethnic minorities, specific provinces) are sampled at higher probabilities (to ensure adequate subsample sizes for subgroup analysis), different strata may use different sampling ratios, and post-stratification adjustments are made based on known population characteristics to correct for nonresponse—all of these operations are condensed into the "sampling weight" variable.

To weight or not to weight? Two principles:

• For descriptive statistics—estimating means, proportions, totals—weights must be used. Without them, your estimates are biased.
• For regression analysis—there is scholarly debate. If the model is correctly specified (correctly characterizing the data-generating process), weighted and unweighted coefficient estimates should be consistent in large samples. The safest practice is to report both weighted and unweighted results as a robustness check.

---

III. Five Recurring Types of Problems in Empirical Data

The five questions above constitute the "physical examination" after obtaining data. The five types of problems below are the "common diseases" that recur in empirical research. Open any issue of an economics journal, and at least one or two articles are grappling with one of them.

Type 1: Convenience Samples—Data Convenience Affects Conclusions in Unknowable Ways

Typical manifestations:

• Studying "the impact of corporate social responsibility on performance" using CSI 300 constituent stocks—not because "A-shares are most representative," but because "this data is easiest to download."
• Studying "determinants of social trust" using questionnaires from 500 students at one university—justified by "convenience of distribution."
• Behavioral economics experiments whose subjects are all students from the same department and cohort at the researcher's own institution.

Core problem: What is most troubling about convenience samples is not "I know it is biased," but I do not know the direction or magnitude of the bias. CSI 300 firms are not only large but also highly visible to capital markets; their motivations for disclosing social responsibility information are entirely different from those of a small or medium-sized enterprise. To what extent do conclusions drawn from such firms hold for the population of "firms"? You cannot answer—because your sample simply does not include those firms for which "data was inconvenient to obtain."

Coping strategies:

• Treat the "convenience sample" itself as the upper bound of the scope of your conclusions. Say only "we find among CSI 300 constituent stocks…," not "Chinese firms…."
• If possible, find a public database that overlaps with your sample on key characteristics for external validation.

Type 2: Coverage Bias—It Is Not That the Sample Is Biased, but That the Entire Sampling Frame Misses Key Groups

Typical manifestations:

• Using mobile phone number prefixes as the sampling frame—missing those who do not use mobile phones (typically the poorest and the oldest).
• Using Dianping review data to study "restaurant service quality"—the sampling frame is "people willing to write online reviews," who are motivated to write by extreme satisfaction or extreme dissatisfaction. The experiences of ordinary diners are not recorded.
• Using LinkedIn resumes to analyze "occupational mobility"—the sampling frame is "white-collar professionals with LinkedIn profiles"; blue-collar workers do not exist in this data.

Core problem: Coverage bias is more fundamental than "biased samples"—sample bias can be partially adjusted through weights, but groups never covered by the sampling frame can never be recovered through weighting. Any pattern you observe in your analysis holds only for the population within the sampling frame.

Coping strategies:

• Explicitly define the "study population" as the group covered by the sampling frame, rather than the target population.
• Discuss the relevance of excluded groups to your research question. If external aggregate statistics about the excluded groups are available (e.g., census data), devote at least a brief paragraph to describing "what the excluded individuals look like."

Type 3: Survivorship Bias—What You See Happens to Be What Was Not Eliminated

This is the most widely recognized data trap, yet it remains ubiquitous in empirical research.

Three classic scenarios:

• Studying "the operating performance of family firms"—you extract family firms that survived until 2020 from a database, but those that went bankrupt or exited before then are invisible to you. The conclusion that "family firms have decent average profitability" is an optimistic illusion.
• Studying "factors behind entrepreneurial success"—you interview 50 successful entrepreneurs and distill common traits (risk-taking, having read a certain book, waking up early, etc.). But 5,000 failures may have read the same book; they simply never made it onto your interview list.
• Using a fund database to study "fund performance persistence"—funds that performed so poorly they were liquidated are no longer in the database. The fund sample you observe has upward-biased overall performance.

Core problem: The insidious nature of survivorship bias is that—you can only observe the world "after survival," and the elimination process itself is nonrandom. Individuals with certain characteristics are systematically "deleted" from the observable world, and you cannot reconstruct a complete picture of those characteristics from the remaining sample.

Coping strategies:

• Optimal solution: Use databases containing complete entry and exit records (e.g., business registration databases that include delisted firms).
• Second-best solution: Analyze in the discussion section—if the elimination process makes "a certain type of individual" more likely to disappear, is your estimate biased upward or downward? Provide a directional judgment.
• Robustness check: Run regressions separately on subgroups with different "survival durations" within the sample, and examine whether the effect changes systematically with survival time.

Type 4: Self-Selection—People Do Not Randomly Appear Where They Are

Classic example: You want to evaluate "the impact of vocational training on wages." You obtain data containing both participants in government training programs and non-participants. Directly comparing wages—you find that training participants actually earn less.

Can you conclude that "training is harmful"? No. Because those who participate in training are precisely those who already face difficulties in the labor market. Participation in training is not random; it is self-selected.

Three typical directions of self-selection:

• Positive selection: More capable individuals are more likely to attend college, participate in training, and move to cities with more opportunities—at least part of the observed "higher income for college graduates" reflects ability effects rather than education effects.
• Negative selection: The most disadvantaged groups are more likely to receive subsidies and participate in poverty-alleviation programs—the observed "lower income among program participants" does not mean the program is harmful, but rather that the poor are more likely to participate.
• Polarization: Those who care deeply about an issue or are strongly opposed to it are both more likely to respond to your survey—extreme opinions are amplified in the sample.

Coping strategies:

• If you are designing your own data collection, record basic information (age, gender, geographic location, etc.) for all invitees at the recruitment stage, even if they ultimately decline to participate. You can then compare "accepters" and "refusers" on observable dimensions for systematic differences.
• If analyzing secondary data, candidly analyze in the discussion section the direction and magnitude of bias that self-selection may induce.

Type 5: Rare Events and Small Samples—Large-Sample Theory Cannot Help You

Typical manifestations:

• Studying "the economic causes of war"—there have only been a few dozen international wars in history; the sample size is inherently limited.
• Studying "characteristics of unicorn firms"—there are only a few hundred unicorns globally, yet you want to add five control variables.
• Studying "bank loan defaults"—defaults are ultimately a minority event.

Core problem: Most of the econometric methods you learned (asymptotic normality of OLS, asymptotic properties of MLE, t-tests and F-tests) are built on large-sample theory. When the sample size is only a few dozen—or more critically, when the number of "event occurrence" observations in your dependent variable is only a dozen or so—these asymptotic theories no longer provide reliable guidance. A single leverage point can flip the sign of an entire regression coefficient.

Coping strategies:

• Choose methods specifically designed for rare events or small samples: King & Zeng's rare-events Logit, Firth's penalized likelihood estimation, exact logistic regression.
• Reduce model complexity: with only a few dozen observations, add control variables one at a time rather than running a fully saturated model.
• Be candid: wide confidence intervals are themselves information—they tell readers "we are actually very uncertain about the magnitude of the effect."

---

IV. How to Discuss Limitations and Remediate Different Data Problems?

Identifying problems is only the first step. What truly distinguishes good empirical research from average research is—how you handle and discuss these problems.

4.1 Three Levels of Discussing Limitations

Lowest level—generic apology: "This study has a limited sample size and may suffer from sampling bias."

This style of writing says almost nothing. Reviewers and readers cannot determine: (1) In which specific direction does the bias lie? (2) How severe is the bias? (3) Is the bias sufficient to overturn your core conclusions?

Intermediate level—specific diagnosis: "The sample's urban population share is 72%, higher than the national level of 64%. Since urban residents have higher average incomes, our estimated national average income may be upward-biased by approximately X%."

This identifies the specific dimension and rough direction but lacks further quantitative analysis.

Higher level—quantitative assessment: "The sample overrepresents the urban population. We applied sampling weights for correction; the estimated average income changed from XX yuan (unweighted) to YY yuan (weighted), a decline of approximately Z%. This change does not affect our core conclusion regarding returns to education (see weighted regression results in Appendix Table A). If nonresponding urban residents differ from respondents on unobservable dimensions, our estimates may still be upward-biased—but the magnitude is not expected to exceed [specific range], and the qualitative conclusions remain unaffected."

This is what effective limitation discussion looks like—it not only identifies the problem but also quantifies its potential impact and uses evidence (rather than rhetoric) to argue that core conclusions remain credible despite the problem.

4.2 Remediation Strategy Toolkit

Problem Type	Available Remediation Methods	Applicable Prerequisites
Sample structure deviates from population	Post-stratification weighting (Raking/Calibration)	Requires knowledge of the true population distribution on certain dimensions
Nonresponse bias	Heckman two-stage selection model	Requires an exclusion variable that affects response but not the core outcome directly (such variables are difficult to find—this is the main limitation)
Selective entry/exit	Inverse probability weighting (IPW)	Requires observable predictors
Small-sample inference	Bootstrap standard errors or exact tests	Generally requires sample size ≥ 30
Rare events	Rare-events Logit (King & Zeng), Firth penalized likelihood	Logit/Probit framework
Missing data	Multiple imputation	Missingness mechanism must be at least missing at random (MAR)

Important reminder: All of the above statistical remediation methods rest on additional assumptions, and these assumptions themselves typically cannot be verified with data. The Heckman model assumes the existence of an exclusion variable; weighting methods assume ignorability; imputation methods assume missing at random. Using these methods does not mean "the problem is solved"—it means—"If the reader is willing to accept specific assumptions, then under those assumptions, the bias can be controlled within a certain range."

4.3 The Most Powerful Move: Robustness Checks

When you cannot perfectly remediate a data problem through statistical methods, the most powerful argumentative strategy is: demonstrate that even if the data has this problem, your core conclusions remain unaffected.

Several of the most commonly used robustness strategies:

• Subsample analysis: If you worry that a few extreme observations drive all your results, delete them and rerun. Results unchanged? Reviewers breathe a sigh of relief.
• Altering variable definitions: If conclusions hold when the definition of "unemployment" changes, your conclusions are insensitive to measurement choices.
• Weighted vs. unweighted dual reporting: If the core coefficients have consistent signs and similar significance, the sampling design is unlikely to drive your conclusions.
• Bounding analysis: Assume the worst case—nonrespondents' characteristics equal those of the most extreme individuals in your sample—how wide is the estimated effect interval? Under the worst case, does the sign flip?
• Placebo tests: If your conclusion is "a certain policy increased employment in a certain region," then in years before the policy was implemented ("pseudo-treatment time") or in unaffected regions ("pseudo-control group"), you should not observe an "effect." If you do, your results may reflect an inherent pattern in the data rather than the policy's causal effect.

The soul of robustness checks is not "I ran many tests," but "I can articulate: if my data has problems on certain dimensions, to what extent do my conclusions still hold?"

---

V. Summary: A Scrutiny Checklist You Can Post on Your Monitor

Before you next open Stata/R/Python to run regressions, run your data through the following seven questions:

Question	Corresponding Risk	Your Action
① Which sampling frame does the data come from? Where is the gap from the target population?	Coverage bias	Define the study population; discuss characteristics of excluded groups
② What sampling method was used? Is it truly random?	Unknown selection probabilities	Honestly describe the sampling method; adjust standard errors for complex sampling designs
③ Does the sample structure match the population structure?	Systematic deviation on certain dimensions	Create a "sample-population structure comparison table"; assess whether deviations correlate with core variables
④ What is the nonresponse rate? Who did not answer?	Self-selection bias	Compare observable characteristics of respondents/nonrespondents; discuss direction of bias
⑤ Does the data include sampling weights? Have I used them?	Biased descriptive statistics	Descriptive statistics must be weighted; report both weighted and unweighted regressions
⑥ Is there survivorship bias?	Sample contains only "non-eliminated" individuals	Check whether the database includes exiters; discuss the directional impact of the elimination process on estimates
⑦ Is the sample size sufficient for key subgroups?	Unreliable inference under small samples	Reduce model complexity; switch to exact tests or Bootstrap

---

A final reminder:

Those who do empirical work easily fixate on "how to produce significant results." But from a data quality perspective, here is a fact you may not want to face but must:

A "methodologically fully armed" regression—complete with instrumental variables, difference-in-differences, regression discontinuity, and five sets of robustness checks—if based on a sample with severely incomplete coverage, a nonresponse rate exceeding 50%, and a completely unknown nonresponse mechanism, is merely meticulously sculpting a castle built on sand.

Conversely, data with rigorous sampling design, complete weight information, traceable response rates, and clear sample-population structure comparisons—even if analyzed with the simplest OLS—may yield conclusions far more credible than the former.

The data source determines your starting point. Methodological sophistication can only help you approach the upper bound of that starting point—it cannot help you exceed it.

---

VI. Content Segmentation Recommendations: Suitable for a 3-Part Series

This article covers substantial ground—from five scrutiny questions to five typical data problems, and then to the remediation strategy toolkit. If produced as a single video (or a single WeChat article), the information density would be too high, and audiences might lose focus midway. We recommend splitting it into a 3-part series, each focusing on one core task, independent yet sequentially connected.

---

Part 1: Who Can Your Data Represent?—Three Core Questions from a Sampling Perspective

Core theme: The first round of scrutiny after obtaining data—starting from the fundamental question of "who can this data represent?"

Content covered:

• Target population vs. sampling frame (coverage bias)
• Whether the sampling method is truly random sampling (three common scenarios of fake randomness)
• Comparison of sample structure with population structure (constructing and interpreting the "sample-population comparison table")

Design rationale: These three questions constitute the "first line of defense" in scrutinizing data provenance. They answer the same overarching question—"To which population can the conclusions from this data be generalized?" This question is the foundation of all subsequent discussion and must logically be addressed first. Moreover, these three questions have a natural progression: you first need to know who your desired population is (population → sampling frame), then examine whether the sampling method is valid (sampling frame → sample), and finally test whether the sample truly represents the population (sample → population comparison).

Recommended length: Video 12–15 minutes; WeChat article 3,000–4,000 characters.

Ending hook (to guide viewers to the next part): "However, even if the sampling method is sound, your data contains deeper hidden problems—some people did not answer your questionnaire, and some firms and individuals have already 'disappeared' from your data. How do you detect and address these issues? We'll discuss that next time."

---

Part 2: The "Silent" and the "Vanished" in Your Data—Nonresponse, Survivorship Bias, and Self-Selection

Core theme: Three types of systematically formed data biases—not problems of sampling design, but the voids left in data by "those who did not respond" and "those who were already eliminated."

Content covered:

• Question 4: Nonresponse bias (Who did not answer? How do they differ from those who did?)
• Type 3: Survivorship bias (What you see happens to be what was not eliminated)
• Type 4: Self-selection (People do not randomly appear where they are)
• Coping strategies for all three biases

Design rationale: The essence of these three problems is the same—systematic, rather than random, differences between observed and omitted individuals in the data. Presenting them together enables the audience/readers to realize that data problems are not just "poor sampling"; often, "some people simply will not appear in your data, and their absence follows a pattern." This cognitive shift is the key step from "memorizing sampling formulas" to "truly understanding data."

Recommended length: Video 15–18 minutes; WeChat article 3,500–4,500 characters.

Ending hook (to guide viewers to the next part): "Now you know how to diagnose the 'illnesses' of data. But diagnosis alone is not enough—once problems are identified, how do you remediate them? How do you write a convincing limitations discussion in your paper? Next time, we'll discuss remediation strategies."

---

Part 3: What to Do After Identifying Problems?—Limitations Discussion, Remediation Strategies, and Robustness Checks

Core theme: From "diagnosis" to "treatment"—how to translate data problems into compelling academic discussion.

Content covered:

• Question 5: Correct use of sampling weights (descriptive statistics must be weighted; report both weighted and unweighted regressions)
• Type 5: Rare events and small samples (and corresponding specialized methods)
• Three levels of limitations discussion (generic apology → specific diagnosis → quantitative assessment)
• Remediation strategy toolkit (post-stratification weighting, Heckman model, IPW, multiple imputation, etc.) and their applicable prerequisites
• The most powerful move: robustness check strategies (subsample analysis, altering variable definitions, bounding analysis, placebo tests)
• Final scrutiny checklist (summary table of seven questions)

Design rationale: This part concludes the series and serves two functions. First, it consolidates the "pending" solutions from the previous two parts (such as sampling weights and robustness checks). Second, it provides the audience with a "ready-to-use" checklist—after watching all three parts, they can quickly scrutinize the representativeness and validity of any dataset using the seven questions. Placing the checklist at the end of the final part serves both as a summary and as a "toolkit"—audiences can screenshot and save it.

Recommended length: Video 18–20 minutes; WeChat article 4,000–5,000 characters.

---

Overview of the Three-Part Structure

Part	Theme	Core Questions	Audience Takeaway
Part 1	Who can your data represent?	Population vs. sampling frame, whether sampling is random, whether sample structure matches population	Can quickly judge the "provenance" of data using three questions
Part 2	The silent and the vanished in data	Nonresponse, survivorship bias, self-selection	Can identify "invisible biases" in data
Part 3	What to do after identifying problems?	Weight usage, remediation strategies, limitations discussion writing, robustness checks	Can move from diagnosis to treatment, writing convincing limitations paragraphs

---

If Only One Part Can Be Produced (Alternative Plan)

If constraints allow only one long video or one long article, we recommend allocating content as follows:

• Five scrutiny questions (30% of space): Focus on Question 1 (coverage bias) and Question 3 (sample vs. population comparison)—these are the most easily overlooked yet most impactful.
• Five typical problems (40% of space): Focus on survivorship bias and self-selection—they are the most common in empirical research and have the most insidious effects on conclusions.
• Remediation strategies and robustness (30% of space): Focus on sampling weights and the core logic of robustness checks, without methodological elaboration.