Correlation & Association - Business Analytics

1. Which summary is most appropriate for this bivariate relationship?
   Answer: Approximately $49\%$ of the variation in satisfaction is associated with a linear relationship with delivery time. When you see a Pearson correlation coefficient (r), a key follow-up tool is the coefficient of determination ($r^2$), which tells you how much of the variation in one variable is linearly explained by the other. Here, $r = -0.70$, so $r^2 = (-0.70)^2 = 0.49$, meaning 49% of the variability in satisfaction scores is associated with a linear relationship with delivery time. That makes D the correct and precise interpretation. Choice A is tempting but commits two serious errors: it replaces "variation explained" with "dissatisfaction caused," introducing causation language that correlation never supports, and it uses 70% (the raw r) instead of 49% (the proper $r^2$). Correlation describes association, not causation. Choice B misreads $r^2$ as a frequency count — the idea that 49% of observations fall below average under certain conditions. The coefficient of determination says nothing about how individual data points are distributed relative to their means; that's a different statistical concept entirely. Choice C invents a rule that doesn't exist: a negative correlation does not reduce the coefficient of determination or leave a portion "unpredictable." Squaring r always produces a non-negative value regardless of direction, so the sign of the correlation is irrelevant to $r^2$. Study tip: Always square r to get $r^2$, and frame your interpretation as "__% of the variation in Y is associated with (or explained by) a linear relationship with X." Watch for distractors that sneak in causal language ("caused by") or misuse the raw correlation value instead of $r^2$.
2. Which statement best describes what the company can infer from this association?
   Answer: The course may improve performance, but prior motivation or ability could influence both enrollment and later ratings. Whenever you see a question linking an observed association to a causal conclusion, your first instinct should be to ask: could something else explain this pattern? This is the core of confounding variable reasoning, one of the most tested concepts in business analytics. Here, employees who choose to enroll in the analytics course likely differ from non-enrollees before the course even begins — they may be more motivated, more ambitious, or already more skilled. These pre-existing traits could independently drive both the decision to enroll and higher performance ratings later. That's exactly what B captures: the course may help, but you can't isolate its effect because self-selection creates a confound. Voluntary enrollment means the two groups aren't comparable from the start, so the correlation alone doesn't establish causation. A is wrong because the timing of measurement — after enrollment rather than before — doesn't eliminate confounding. The post-enrollment design would only establish causation in a controlled experiment, not an observational study like this one. C goes too far in the opposite direction; voluntary enrollment doesn't make the data useless, it just limits causal inference. You can still generate hypotheses and observe patterns. D is a distractor built on a false measurement rule — there's no requirement that two variables share the same numerical scale to produce a valid correlation. A reliable strategy: when you see correlation data from a self-selected group, always ask "who opted in, and why?" If enrollment wasn't random, a confounding variable is almost certainly lurking.
3. Which interpretation is most appropriate?
   Answer: A meaningful nonlinear association may exist even though Pearson correlation detects little overall linear association. Whenever you see a question involving Pearson correlation, remember that it measures only linear association — the degree to which two variables move together along a straight line. It is completely blind to curved or U-shaped relationships, which is exactly the trap this question sets. The passage describes a classic inverted-U (curvilinear) pattern: response rises with moderate email frequency and falls at both extremes. When a relationship bends like this, the positive slope on the left and the negative slope on the right cancel each other out mathematically, producing a Pearson $r \approx 0$ — even though the association is actually quite strong and systematic. Choice C correctly identifies this: a near-zero Pearson correlation does not mean no association exists; it means no linear association was detected. A nonlinear relationship can be very meaningful and still invisible to Pearson's formula. Choice A is wrong because it treats $r \approx 0$ as evidence that the variables are completely unrelated. That conclusion is only valid if you've already confirmed the relationship is linear — which the passage clearly contradicts. Choice B goes further by making a causal claim, but correlation (or its absence) never implies causation, and more importantly, the near-zero $r$ here reflects the shape of the relationship, not the absence of one. Choice D fabricates a conclusion the data don't support — a strong negative correlation would require response to decline consistently across all frequency levels, not just at high values. Your study tip: always ask whether a relationship could be nonlinear before interpreting a low correlation as "no relationship." Scatter plots are your best diagnostic tool here.
4. Which statement accurately summarizes both the observed association and its limitation?
   Answer: Offer recipients had a $4$-percentage-point higher purchase rate, but representative targeting may confound the estimated effect of the offer. Whenever you see a question combining a statistical comparison with a study design flaw, you need to evaluate two things separately: the math of the observed difference, and what that difference actually means given how the data were collected. Start with the numbers. The purchase rate among offer recipients was $12\%$ versus $8\%$ for non-recipients. The absolute difference is $12\% - 8\% = 4$ percentage points. That's the correct way to express the observed gap — straightforward subtraction. Now for the design issue: because sales representatives chose who received offers (not a random process), the two groups likely differ in ways beyond just receiving the offer. Representatives probably targeted customers who already seemed more likely to buy — a classic confound. This means the $4$-percentage-point gap may reflect pre-existing differences, not the true effect of the offer itself. Answer A captures both points precisely, making it correct. Answer B makes two errors: $50$ percentage points is wildly wrong (that would require one group at $58\%$ and the other at $8\%$, or similar), and nonrandom assignment doesn't merely affect precision — it introduces bias, a fundamentally different problem. Answer C misquotes the lift as $4\%$ relative (the relative increase is actually $\frac{4}{8} = 50\%$), and fatally claims nonrandom assignment is sufficient for causal inference — the opposite of the truth. Answer D correctly states the $12\%$ figure, but that's the raw rate for one group, not the "lift." More importantly, nonrandom assignment doesn't make comparisons meaningless — it just limits causal interpretation. Study tip: Always distinguish absolute difference, relative difference, and raw rates — exam questions frequently mix them up to test whether you're paying attention to units and framing.
5. Which explanation is most defensible based only on this information?
   Answer: Restricting the range of income can reduce observed correlation even if the broader population has a meaningful association. Whenever you see a question comparing correlations across different sample sizes or ranges, think about range restriction — a core concept in statistics that explains why a relationship can look weaker in a narrow slice of data than it does in the full population. Here's the key idea: correlation measures how consistently two variables move together across the variation present in your sample. When you restrict your sample to a narrow band of income ($\$95{,}000$ to $\$100{,}000$), you dramatically reduce the spread of the independent variable. With almost no variation in income to work with, the statistical algorithm has little signal to detect, and the correlation naturally compresses toward zero — even if the true underlying relationship is strong. This is exactly what happened here: the full-sample correlation of $0.58$ collapses to $0.12$ simply because the income range is too tight to reveal it. Answer B correctly identifies this mechanism. Answer A is wrong because a valid relationship does not need to produce identical correlations in every subgroup — that's not how statistics works. Subgroup correlations legitimately vary, especially under range restriction. Answer C makes an unjustified causal claim: a near-zero correlation in one narrow income band does not prove that spending and income are independent above $\$95{,}000$ — the restricted range is a far simpler explanation. Answer D invents a causal reversal (spending causes income) that the data never support; correlation, in any direction, cannot establish causation. Your study tip: whenever a correlation shrinks dramatically in a subgroup, always ask "Is the range of the key variable restricted here?" before drawing any conclusions about independence or causality.
6. What is the correlation between the two transformed variables?
   Answer: $0.62$, because reversing the discount scale changes the sign while rescaling profit leaves correlation unchanged. Whenever you see a question about correlation and variable transformations, your first instinct should be to recall how Pearson correlation responds to linear changes. The key rule: multiplying or adding a constant to a variable changes the sign of the correlation only if the multiplier is negative; it never affects the magnitude. Here, two transformations occur. First, discount percentage is replaced by $(100 - \text{discount\%})$. This is equivalent to multiplying the original variable by $-1$ (plus a constant, which is irrelevant). That negative multiplier flips the sign of the correlation from $-0.62$ to $+0.62$. Second, gross profit is divided by 1,000 to convert dollars to thousands. Dividing by a positive constant leaves the sign unchanged and, crucially, does not affect the magnitude of the correlation at all. The result is $+0.62$, confirming B is correct. A is wrong because it claims sign cannot change — it can, whenever a variable is multiplied by a negative constant. The sign does flip here. C and D both claim that rescaling profit to thousands shrinks the correlation coefficient numerically (e.g., dividing by 1,000 to get $0.0062$). This is a classic trap: correlation is unitless and dimensionless — it is already a standardized measure, so multiplying or dividing either variable by any positive constant has zero effect on its value. D compounds that error by also flipping the sign incorrectly. Study tip: Remember that Pearson correlation is invariant to any positive linear rescaling. Only a negative multiplier changes the sign. Units never change the magnitude.
7. How does the fifth store affect the Pearson correlation between advertising and sales?
   Answer: It produces a negative correlation because the fifth store is a high-leverage point whose outcome contradicts the original trend. When a question involves how a new data point changes correlation, your first instinct should be to think about leverage — how far that point sits from the center of the existing data, and whether its outcome aligns with or contradicts the established trend. Here, the first four stores form a perfect line: advertising $(1,2,3,4)$ maps exactly to sales $(2,4,6,8)$, giving a correlation of $r = 1$. The fifth store spends $10$ units on advertising — far above the others — but produces only $1$ unit of sales, the lowest in the dataset. This is a classic high-leverage outlier: it sits at an extreme position on the horizontal axis and pulls hard against the existing upward trend. When you compute the new correlation, that single contradictory point is powerful enough to drag $r$ into negative territory, making A the correct answer. Choice B is tempting but wrong — it assumes a single point can only weaken, never reverse, a correlation. In reality, a sufficiently extreme observation can absolutely flip the sign, especially when it contradicts the trend as dramatically as this one does. Choice C incorrectly assumes the perfect relationship among the first four stores is immune to new data; correlation is recalculated across all observations every time you add one. Choice D contains a grain of truth — extreme horizontal values do distort correlations — but "always eliminates" (drives to zero) is too strong and inaccurate; the actual direction depends on where the outlier's outcome falls, not just its position. As a study tip: watch for outliers that combine extreme $x$-values with outcome values that contradict the trend — these are the scenarios most likely to reverse, not merely weaken, a correlation.
8. Which additional analysis would best assess whether the reported association mainly reflects the shared upward trend?
   Answer: Compare month-to-month changes or remove the time trends before recalculating the association between the series. Whenever you see a high correlation between two variables measured over time, your first instinct should be to ask: are these truly related, or do they just both happen to trend upward together? This is the classic spurious correlation problem driven by a shared time trend. When two series both climb steadily over months or years — here, due to new store openings and rising prices — their correlation will naturally be high even if advertising has no real effect on sales. The $r = 0.91$ may be capturing nothing more than "both numbers got bigger over time." The remedy is to isolate the relationship after removing that shared trend. You can do this by analyzing month-to-month changes (first differences) or by detrending both series and correlating the residuals. If the association weakens substantially, the original correlation was mostly trend-driven. This is exactly what C proposes, making it the correct answer. A is wrong because changing units (dollars vs. another denomination) has no effect on correlation — it is scale-invariant by construction. You'd get the same $r$ regardless. B is wrong because squaring the correlation to get $r^2$ only removes information about direction; it does nothing to address whether a shared trend is inflating the magnitude of the association. D is wrong and statistically destructive — deliberately removing observations where both variables increased would bias your sample and tell you nothing meaningful about the trend problem. Study tip: On any question pairing two time-series with a suspiciously high correlation, immediately flag the possibility of a spurious trend relationship and look for answers that involve differencing, detrending, or controlling for time.
9. Which conclusion is best supported by these results?
   Answer: Campaign X has the higher purchase rate within each segment, but campaign Y has the higher aggregate rate because campaign assignment differs by segment. Whenever you see a question presenting group-level data broken into subgroups, you should immediately think about Simpson's Paradox — the phenomenon where a trend that appears in combined data reverses or disappears when the data is separated by subgroup. Let's calculate the rates. In the high-propensity segment, Campaign X converts $9/10 = 90\%$ versus Campaign Y's $80/100 = 80\%$ — X wins. In the low-propensity segment, Campaign X converts $20/100 = 20\%$ versus Campaign Y's $1/10 = 10\%$ — X wins again. So Campaign X outperforms Y within every segment. Yet in aggregate, Campaign X totals $9 + 20 = 29$ purchases out of $110$ customers ($\approx 26\%$), while Campaign Y totals $80 + 1 = 81$ purchases out of $110$ customers ($\approx 74\%$). Campaign Y appears dominant overall. Why? Because Campaign Y was disproportionately assigned to the high-propensity segment, which inflates its aggregate rate. Answer A correctly captures this: X wins each segment, but Y wins in aggregate due to unequal segment assignment. Answer B is wrong because Campaign Y does not have the higher rate within each segment — X does in both. Answer C reverses the truth: X does not have the higher aggregate rate; Y does. Answer D is wrong on two counts — performance is not identical after adjustment, and the distractor fabricates a "same total customers" justification that doesn't resolve the paradox. When you see subgroup data, always calculate rates at both levels. If a campaign looks better overall but worse in every subgroup, you're looking at Simpson's Paradox — and the segmented analysis is the more honest comparison.
10. What can the analyst conclude about the correlation after pooling the two regions?
    Answer: It cannot be determined from the two correlations because differences between regional means also affect the pooled correlation. Whenever you see a question about combining groups with different means, think about Simpson's Paradox and the broader phenomenon of ecological aggregation — the idea that pooling subgroups can dramatically shift a statistic beyond what any individual group shows. Here's why D is correct: When two regions have different average levels of both variables (spending and service calls), pooling them introduces between-group variation that wasn't present in either within-group analysis. Imagine plotting both regions together — the cluster of high-spending, low-call customers from one region sits far from the cluster of low-spending, high-call customers from the other. That separation between clusters creates its own trend line in the combined data, which can pull the pooled correlation higher, lower, or even in the opposite direction compared to either region alone. Because you don't know the exact means and standard deviations of each region, you simply cannot calculate the pooled correlation from the two regional correlations alone. A is wrong because equal sample sizes don't neutralize the effect of differing means. Equal n only means each region contributes the same number of data points — it says nothing about where those points sit in the variable space. B is wrong because more variation doesn't guarantee a predictable directional change in correlation, which is already a standardized, scale-free measure. C is the trickiest distractor: two positive within-group correlations can actually produce a negative pooled correlation if the group means are arranged the right way — that's precisely Simpson's Paradox. Study tip: Any time a question mentions subgroups with different means being combined, immediately ask yourself whether between-group differences could be distorting the pooled statistic.

Start this quiz