Data Quality Issues - Business Statistics

1. A researcher is analyzing a large survey dataset with 100 variables and 1000 respondents. Missing values are scattered sparsely throughout the dataset and are credibly assumed to be Missing Completely At Random (MCAR). The researcher's primary goal is to compute a full correlation matrix for all 100 variables. What is the key trade-off when choosing between pairwise and listwise deletion in this context?
   Answer: Pairwise deletion uses more of the available data for each correlation, but may result in a correlation matrix that is not positive semi-definite. Listwise deletion would remove any respondent with even one missing value, which could be a very large portion of the sample given 100 variables. Pairwise deletion calculates each correlation using all rows with valid data for just that pair of variables, thus maximizing sample size for each calculation. However, a known mathematical problem with this method is that the resulting correlation matrix, assembled from pairs with different sample sizes, may not be positive semi-definite, which can cause problems in subsequent analyses like factor analysis or regression.
2. A company wants to measure employee satisfaction with a new software system. A manager sends out a survey with the following question: 'Considering the significant productivity improvements and enhanced user interface of the new software, how would you rate your satisfaction with this essential upgrade?' This question is a poor measurement instrument primarily due to what type of issue?
   Answer: It contains leading and loaded language, which introduces a systematic bias into the responses. When evaluating survey questions in business statistics, you need to identify sources of bias that can compromise data quality. Poor question design can introduce systematic errors that skew results in predictable directions. This survey question is problematic because it contains leading and loaded language that pushes respondents toward positive answers. The question assumes the software has "significant productivity improvements" and an "enhanced user interface," calling it an "essential upgrade" before even asking for the employee's opinion. This language primes respondents to give favorable ratings regardless of their actual experience, creating systematic bias where responses are artificially inflated. Option A is incorrect because while the question mentions both productivity and interface, it's asking for overall satisfaction with the software system, not separate ratings for each feature. This isn't truly double-barreled. Option B is wrong because the question doesn't show us the response scale - we can't determine if it's ambiguous from the question text alone. The reliability issue stems from the biased wording, not scale ambiguity. Option C misses the mark because the language used ("productivity," "interface," "upgrade") is standard business terminology that employees would reasonably understand, so this wouldn't cause random measurement error. Watch for survey questions that contain assumptions, superlatives ("significant," "enhanced"), or value judgments ("essential") before asking for opinions. These are red flags for leading questions that introduce systematic bias. Always look for neutral, objective wording that doesn't influence the respondent's answer.
3. A supply chain analyst notices that delivery time data contains several values that seem implausible: 15 orders show delivery times of exactly 0 days, 8 orders show negative delivery times (ranging from -2 to -7 days), and 23 orders show delivery times exceeding 200 days when the maximum service level agreement is 30 days. The dataset contains 5,000 total orders. What is the most appropriate initial approach to address these data quality issues?
   Answer: Investigate the data generation process to understand whether these values represent coding errors, system glitches, or legitimate business scenarios before deciding on treatment Understanding the data generation process is crucial before making corrections. The apparent errors might have systematic causes (e.g., timezone issues, data entry protocols, or legitimate scenarios like order cancellations) that affect treatment decisions. Choice A assumes the values are errors without investigation. Choice C imputes without understanding the underlying issues. Choice D makes arbitrary corrections that could introduce new biases without understanding why these values occurred.
4. Based on the data collection issues described in the passage, which data quality problem requires the most immediate attention before conducting statistical analysis?
   Answer: The '99' values from technical issues, because these represent clear measurement errors that will severely distort any analysis if not corrected first The '99' values represent unambiguous data errors that must be addressed before any analysis, as they will completely distort descriptive statistics and relationships. Choice A addresses a real concern but missing data can be handled with appropriate methods after fixing obvious errors. Choice C identifies a problem but response bias is harder to detect and correct than the clear '99' errors. Choice D is overly broad and doesn't prioritize the most immediately fixable issue.
5. A management consulting firm assesses employee productivity by having consultants observe employees and rate their 'focus level' on a 1-10 scale. Analysis reveals that employees observed by a specific senior consultant received systematically lower focus scores than those observed by other consultants. This discrepancy most strongly suggests which data quality issue?
   Answer: Systematic measurement error, specifically observer bias, related to the senior consultant. Observer bias occurs when the characteristics or expectations of the person collecting the data influence the measurements. Here, the scores are systematically lower for one specific observer. This indicates a bias, not just random variation between raters (which would be random error / low reliability). The pattern is consistent and tied to a single source, making it a systematic error.
6. In analyzing its sales data, a company's business intelligence team encounters three distinct issues: 1. One sales entry is 100 times larger than any other, which they attribute to a misplaced decimal point. 2. For 10% of entries, the 'customer region' field is blank because a legacy data entry system often crashed when used by new sales associates. 3. The 'sales amount' is self-reported by salespersons who informally round to the nearest hundred dollars. Which sequence correctly classifies these three data quality issues?
   Answer: Outlier, followed by Missing At Random (MAR) data, followed by measurement error. 1. The extremely large sales entry is an outlier, likely due to a data entry error. 2. The missing 'customer region' data is dependent on an observable characteristic (the newness of the sales associate), making it Missing At Random (MAR). It's not MCAR because it's not completely random. 3. The practice of rounding introduces a discrepancy between the true sales amount and the recorded amount, which is a form of measurement error (specifically, reduced precision).
7. A human resources analyst studying employee turnover finds that exit interview data is missing for 40% of departing employees. Among employees who completed exit interviews, 85% cite 'career growth' as their primary reason for leaving, while industry benchmarks suggest this reason typically accounts for only 35% of departures. What data quality issue is most likely affecting the analysis?
   Answer: Selection bias combined with response bias, where employees who complete interviews may differ systematically from those who don't, and may provide socially desirable responses When analyzing data quality issues, you need to identify whether problems stem from how data is collected, who provides it, or how it's measured. This question presents a classic scenario where multiple biases compound to create misleading results. The key insight is recognizing that 40% missing data combined with results dramatically different from industry benchmarks (85% vs 35%) suggests multiple systematic problems. Answer D correctly identifies this as selection bias (certain types of employees are more likely to complete exit interviews) combined with response bias (those who do participate may give socially acceptable answers rather than truthful ones). Employees citing "career growth" sounds more professional than admitting they're fleeing bad management or toxic culture. Answer A focuses only on measurement error from leading questions, but doesn't address the substantial missing data problem that's likely creating the bigger distortion. Answer B mentions sampling bias but incorrectly attributes it to demographic representation rather than the fundamental difference between employees who do versus don't complete interviews. Answer C suggests random measurement error from inconsistent administration, but the consistently high percentage citing career growth indicates systematic rather than random error. The 40% missing data rate is the crucial clue here. When combined with results that deviate so significantly from industry norms, this strongly suggests that the employees completing interviews are fundamentally different from those who don't participate, and they're providing responses that paint their departure in the best possible light. Study tip: When you see large amounts of missing data paired with unexpected results, always consider whether the missing data is random or if certain groups systematically opt out.
8. A market researcher notices that income data in a consumer survey shows suspicious patterns: exactly 47% of respondents reported incomes ending in '000' (like $50,000, $75,000), and the income distribution shows unusual spikes at round numbers. Additionally, 12% of high-income respondents (>$150,000) refused to answer the income question, while only 3% of lower-income respondents had missing income data. What analytical strategy best addresses these data quality issues?
   Answer: Use income ranges instead of exact values and apply multiple imputation for missing data, treating missingness as Missing Not at Random (MNAR) When you encounter data quality issues in survey research, you need to recognize different types of measurement problems and missing data patterns to choose appropriate analytical strategies. The scenario presents two key issues: heaping (respondents rounding to convenient numbers) and systematic missingness (high earners are more likely to skip income questions). The 47% reporting round numbers indicates substantial heaping, making exact income values unreliable. More critically, the differential missing rates (12% vs 3%) suggest the missingness depends on the unobserved income values themselves—high earners systematically avoid disclosure regardless of other observed characteristics. Answer A correctly addresses both problems. Income ranges reduce heaping bias by acknowledging that many responses aren't precise anyway. Multiple imputation with Missing Not at Random (MNAR) assumptions is essential because the probability of missing data depends on the actual (unobserved) income level, not just observed respondent characteristics. Answer B fails because treating this as Missing at Random (MAR) assumes missingness only depends on observed variables, ignoring the clear systematic pattern where income level itself drives non-response. Answer C discards too much valuable data—removing nearly half your sample (round numbers) and all high-income non-responders creates severe bias and reduces power. Answer D's smoothing techniques artificially manipulate the underlying distribution, while hot-deck imputation doesn't account for the systematic missingness pattern. Remember: when missing data rates vary dramatically across the variable of interest itself, always suspect MNAR and choose methods that explicitly model this dependency.
9. An analyst is screening a dataset of 1,000 transaction values for outliers. The data distribution is symmetric but has heavier tails than a normal distribution (leptokurtic). The analyst considers using either the Z-score method (flagging values where $|Z| > 3$) or the IQR method (flagging values outside $Q_1 - 1.5 \times IQR$ and $Q_3 + 1.5 \times IQR$). Which statement accurately compares the expected outcomes?
   Answer: The Z-score method may flag fewer outliers because the heavy tails will inflate the standard deviation used in its calculation. The Z-score is calculated as $(x - \mu) / \sigma$. In a distribution with heavy tails, the extreme values will inflate the calculated standard deviation ($\sigma$). This larger denominator will result in smaller Z-scores for all points, including the extreme ones. Consequently, fewer points may cross the $|Z| > 3$ threshold. The IQR method, based on quartiles, is robust to extreme values and will not be similarly affected, making it more likely to flag outliers in a heavy-tailed distribution.
10. A financial analyst is comparing the quarterly returns of two portfolios. Portfolio A's returns are approximately normally distributed. Portfolio B's returns have a similar mean but are subject to occasional extreme positive and negative events, creating a distribution with fat tails. To accurately compare the central tendency and risk of the two portfolios, which pair of statistics would be most appropriate and robust?
    Answer: Median for central tendency and Interquartile Range (IQR) for risk. For Portfolio B, the presence of extreme events (outliers) makes the mean and standard deviation unreliable. The mean is sensitive to extreme values, and the standard deviation is even more so. Robust statistics are needed. The median is a robust measure of central tendency, and the interquartile range (IQR) is a robust measure of dispersion. For a fair comparison, the same set of statistics should be used for both portfolios. Therefore, median and IQR are the most appropriate choices.
11. A marketing analyst discovers that 15% of survey responses about customer satisfaction are missing, with the missing pattern showing that customers who purchased premium products are 3 times more likely to have missing satisfaction scores compared to standard product customers. Additionally, 8% of recorded satisfaction scores are exactly 5.0 (on a 1-7 scale), which preliminary investigation suggests may represent default values entered by data entry personnel. What combination of data quality issues is most likely present?
    Answer: Missing Not at Random (MNAR) and systematic measurement error from default value substitution The missing data is MNAR because the probability of missingness depends on the unobserved satisfaction values themselves (premium customers likely have different satisfaction patterns). The 5.0 values represent systematic measurement error from default value substitution by data entry staff. Choice A is wrong because MCAR would show no relationship between missingness and customer type. Choice B is wrong because MAR would require missingness to depend on observed variables, not the satisfaction values themselves. Choice D is wrong because the 5.0 pattern indicates systematic, not random, measurement error.
12. A call center wants to assess agent performance using an AI algorithm that scores transcribed calls for 'empathy'. To test the algorithm, the manager feeds the same 100 call recordings through the system on five different occasions. The resulting 'empathy' scores for each individual call vary significantly across the five runs. This variability points to a significant problem with which aspect of data quality?
    Answer: Low reliability, as the measurement tool does not produce consistent results under identical conditions. When evaluating data quality in business statistics, you need to distinguish between reliability (consistency) and validity (accuracy). This question tests your ability to diagnose measurement problems based on the symptoms described. The key clue here is that the same 100 recordings produced significantly different empathy scores across five identical runs. This inconsistency under identical conditions is the textbook definition of poor reliability. A reliable measurement tool should produce nearly identical results when measuring the same thing repeatedly. Think of a scale that shows different weights each time you step on it within minutes - that's a reliability problem. Answer A correctly identifies this as a reliability issue because the algorithm fails to produce consistent results under identical conditions. Answer B describes a validity problem, which would involve the algorithm consistently measuring something other than what we intend (empathy). However, the question doesn't suggest the AI is measuring the wrong thing - just that it's measuring inconsistently. Answer C focuses on sampling bias, but the problem isn't about whether the 100 calls represent all calls well. The issue is that the same calls produce different scores each run. Answer D suggests systematic error, which would show up as consistent over- or under-scoring across all measurements. But systematic errors are typically consistent, not variable like described here. Remember: Reliability = consistency, Validity = accuracy. When you see measurement results varying dramatically under identical conditions, think reliability first.
13. A scatter plot of employee performance score versus age shows a weak, positive, linear association for employees aged 20-60. The dataset also includes one 65-year-old senior executive with a very high performance score. This data point is located far from the other points but in the same general direction of the trend. How will the inclusion of this single outlier most likely affect the Pearson correlation coefficient (r)?
    Answer: It will substantially increase the correlation coefficient, making the linear relationship appear stronger than it is for the bulk of the data. This type of outlier is an 'influential point'. Because it lies far from the mean of the predictor variable (age) and reinforces the existing trend (high age, high performance), it will 'pull' the regression line towards it and increase the magnitude of the Pearson correlation coefficient. It creates a stronger statistical relationship than what is representative of the rest of the employees.
14. An analyst is studying the effect of a new training program by comparing the performance scores of employees who completed the training with those who did not. The training was optional. The analyst finds that 20% of the scores are missing from the 'trained' group, while only 5% are missing from the 'untrained' group. It is likely that lower-performing employees in the 'trained' group were less inclined to report their scores. What is the most significant risk in this analysis?
    Answer: The data are MNAR, and listwise deletion will likely cause the average score of the 'trained' group to be artificially inflated. The missingness of scores is related to the scores themselves (lower-performers not reporting), which is the definition of Missing Not At Random (MNAR). If the analyst uses listwise deletion, they will disproportionately remove the lower scores from the 'trained' group. This will artificially inflate the average score of the remaining members of the 'trained' group, potentially making the training program look far more effective than it actually was.
15. An analyst is preparing a dataset of employee salaries for a report. The salary distribution is known to be right-skewed. About 5% of the salary values are missing and assumed to be Missing Completely At Random (MCAR). To prepare for calculating the departmental average salary, which method for handling the missing data is most appropriate?
    Answer: Median imputation, because the median is a robust measure that is less affected by the skewness of the salary distribution. In a right-skewed distribution, the mean is pulled upward by high-value outliers. Using mean imputation would involve filling the missing values with this inflated average, which would not accurately represent the central tendency and could bias subsequent calculations. The median is robust to skew and outliers. Imputing with the median fills missing values with a more representative value of the center of the distribution, preserving the overall shape and providing a better basis for calculating a meaningful average.
16. A city's economic development office wants to measure the 'financial health' of its small business community. Lacking direct access to revenue data, they decide to use the 'number of social media posts per week by businesses' as a proxy variable. What is the most significant potential data quality issue with using this proxy?
    Answer: A potential lack of construct validity, as social media activity may not accurately reflect a business's actual financial performance. Construct validity is the degree to which a variable measures the theoretical construct it is intended to measure. Here, the construct is 'financial health'. While there might be some correlation, social media activity is not a direct or necessarily accurate reflection of revenue, profit, or stability. A struggling business might increase posting out of desperation, while a highly successful one might not need to. This mismatch between the proxy and the construct is the most fundamental issue.
17. A company surveys employees, asking them to rate their job satisfaction and also whether they are actively seeking a new job. The company finds that employees who report lower job satisfaction are significantly more likely to leave the 'actively seeking' question blank. How would an analyst classify this pattern of missing data?
    Answer: Missing At Random (MAR), as the probability of missingness depends on another observed variable in the dataset. This scenario is a classic example of Missing At Random (MAR). The data is not missing completely at random, because the likelihood of a missing value depends on another variable. However, it is not necessarily MNAR because the missingness depends on an observed variable (job satisfaction), not the unobserved value of the 'actively seeking' variable itself. With MAR, the reason for missingness is captured by other information in the dataset, which can potentially be used in more advanced imputation methods.
18. A company uses a new employee engagement survey. When administered to the same group of employees one month apart, the survey yields highly consistent scores for each employee. However, the survey scores show no significant correlation with employee turnover rates, a key indicator the survey was intended to predict. Which statement best describes the quality of the survey data?
    Answer: The survey has high reliability but low validity. Reliability refers to the consistency of a measurement. Since the survey gives consistent scores over time, it has high reliability. Validity refers to whether the instrument measures what it is intended to measure. Since the scores do not correlate with a relevant outcome (turnover), its predictive validity is low. Therefore, the survey is consistent but does not appear to be measuring the construct of 'engagement' in a way that is useful for prediction.
19. A retail analyst identifies an outlier in weekly sales data for a specific store. The sales for that week were five times the average. Before deciding how to handle this data point, the analyst investigates the context. Which of the following discoveries would most strongly justify keeping the data point in the analysis without modification?
    Answer: The store is located in a city that hosted a major national sporting event during that specific week. The core task when dealing with an outlier is to determine if it is a legitimate data point representing a real, albeit unusual, event or if it is an error. A major sporting event provides a valid, real-world explanation for a massive surge in sales. In this case, the outlier represents a true event and should be kept to have an accurate understanding of sales performance, although its influence might be noted or handled with robust statistical methods.
20. A marketing firm is building a predictive model for customer churn. The dataset has missing values for the 'customer_tenure' variable, which is a key predictor. An analyst uses mean imputation to fill these missing values before training the model. What is the most likely consequence of this action on the final model?
    Answer: The predictive power of 'customer_tenure' will be artificially weakened, potentially reducing the model's overall accuracy. Mean imputation replaces all missing values with a single constant (the mean). This artificially reduces the natural variation in the 'customer_tenure' variable. By concentrating many data points at the mean, it weakens the observed relationship between tenure and churn, as the model sees many instances with the same tenure value but different churn outcomes. This dilution of the relationship reduces the variable's predictive power.

Start this quiz