AP Statistics Unit 9: Inference for Regression — Free Review Games.

H0: beta = 0 tests whether there is a significant linear relationship.

The t-statistic divides the sample slope by its standard error.

df = n - 2 in simple linear regression.

0.003 < 0.05 means we reject H0 and conclude a significant linear relationship exists.

LINE conditions: Linearity, Independence, Normal residuals, Equal variance.

The CI for beta uses the sample slope, the t* critical value, and SE of the slope.

Since 0 is not in the interval, the slope is significantly different from 0.

Inference assumes residuals are approximately normal with mean 0.

Residuals should show roughly the same spread at all levels of x.

SE(b) quantifies how much the sample slope would vary across different random samples.

A fan shape shows increasing variance, violating the equal variance condition.

With r=0.7 and n=30, the relationship is fairly strong and the sample moderate, so likely significant.

With large samples, the CLT helps ensure the t-test is approximately valid.

t = b / SE(b) = 0.60 / 0.15 = 4.0.

df = n - 2 = 22 - 2 = 20.

The null hypothesis for a slope test claims there is no linear relationship between the explanatory and response variables, which is written as \(H_0: \beta_1 = 0\). The choice "\(H_0: \beta_0 = 0\)" is wrong because \(\beta_0\) refers to the population intercept, not the slope being tested. Always identify which population parameter (slope \(\beta_1\)) the hypothesis test targets before writing \(H_0\) and \(H_a\).

\(\beta_1\) denotes the unknown population slope parameter that we estimate using sample data. The choice "\(b_1\)" is incorrect because that is the sample statistic (the estimated slope from the regression line), not the population parameter. Distinguishing parameters (Greek letters) from statistics (Roman letters) is essential in all inference procedures.

The sample slope \(b_1\) is a point estimate for the unknown population slope \(\beta_1\), calculated from the least-squares regression line fit to sample data. The distractor "The correlation coefficient \(r\)" is wrong because \(r\) measures the strength and direction of a linear relationship, not the rate of change estimated by the slope. Remember that every inference procedure begins with a sample statistic estimating a corresponding population parameter.

The Random condition ensures that the data collection method (random sampling or random assignment) allows for valid generalization to the population, which is necessary before performing inference. The choice "Equal variance condition" is incorrect because that condition concerns the spread of residuals being consistent across x-values, not how the data were collected. Checking the Random condition first protects against bias in any inference procedure.

The Linear condition requires that the actual relationship between the explanatory and response variables be linear, which is checked using a scatterplot and residual plot showing no curved pattern. The distractor "The sample size is at least 30" is wrong because sample size relates to the Normal/Large Counts condition, not linearity. A residual plot with no discernible pattern supports the assumption that a linear model is appropriate.

A confidence interval for the slope uses the sample slope \(b_1\) plus or minus a critical \(t^*\) value times the standard error of the slope, since we use the t-distribution when the population standard deviation is unknown. The choice "\(b_1 \pm z^* \cdot SE(b_1)\)" is incorrect because slope inference relies on the t-distribution, not the standard normal (z) distribution. Always match the correct critical value type (t vs z) to the inference procedure being used.

\(SE(b_1)\) estimates how much the sample slope would vary from sample to sample if we repeatedly took samples from the same population, forming the sampling distribution of \(b_1\). The distractor "The average of the residuals" is wrong because the average of residuals is always zero in least-squares regression and has nothing to do with slope variability. Understanding standard error as sampling variability is key to interpreting all confidence intervals and test statistics.

One of the LINER conditions requires that residuals be approximately normally distributed, especially important for small samples, so that the sampling distribution of \(b_1\) is approximately normal. The choice "The explanatory variable is categorical" is incorrect because linear regression inference requires a quantitative explanatory variable, not a categorical one. Checking residual normality via a histogram or normal probability plot is a standard step before performing slope inference.

The t-statistic is calculated as \(t = \frac{b_1 - 0}{SE(b_1)} = \frac{3.2}{0.8} = 4.0\), following the standard formula for slope inference. The distractor "\(3.2\)" is wrong because that is simply the slope value itself, not divided by its standard error. Always divide the statistic by its standard error to obtain a t-statistic, never report the raw estimate as the test statistic.

A residual plot shows the residuals plotted against the explanatory variable or predicted values, and a random scatter with no pattern supports the Linear condition. The choice "A boxplot of the response variable" is incorrect because a boxplot only shows the distribution and center of one variable, not the relationship between two variables. Residual plots are a core diagnostic tool students should use to assess linearity, equal variance, and outliers together.

Degrees of freedom for slope inference equal \(n - 2\), so with \(n = 15\) data points, \(df = 13\), and the appropriate \(t^*\) value comes from the t-distribution table at 90% confidence with 13 df. The choice "\(t^*\) with 14 degrees of freedom" is wrong because it subtracts only 1 from \(n\), forgetting that regression estimates two parameters (slope and intercept), which uses up two degrees of freedom. Always subtract 2 for degrees of freedom in simple linear regression inference, unlike the \(n-1\) used for single-mean inference.

Since the confidence interval \((-1.2, 3.4)\) contains 0, a slope of zero (no relationship) is a plausible value, so we fail to reject \(H_0: \beta_1 = 0\) at the 5% significance level. The distractor "No conclusion can be made without the p-value" is incorrect because a confidence interval and a two-sided hypothesis test at the same confidence/significance level always give consistent conclusions, making the p-value unnecessary here. Recognize that whenever a confidence interval for the slope includes 0, the corresponding two-sided test is not significant.

The t-statistic is \(t = \frac{2.1}{0.7} = 3.0\), and degrees of freedom for slope inference equal \(n - 2 = 23\). The choice "\(t = 3.0\), \(df = 25\)" is incorrect because it fails to subtract 2 from the sample size for the two estimated regression parameters. Always compute both the t-statistic and correct df together, since an accurate p-value depends on using the right degrees of freedom.

Measuring the same students repeatedly creates dependent (correlated) observations over time, violating the Independence condition, which assumes that individual observations don't influence each other. The choice "Randomly sampling 50 unrelated households for income and spending data" is incorrect as a violation because random sampling of unrelated units typically satisfies independence. Watch for repeated measures or clustered data as common independence violations in regression scenarios.

As sample size increases, the standard error of the slope typically decreases because more data reduces sampling variability, which narrows the confidence interval. The distractor "The interval tends to become wider because \(t^*\) increases" is wrong because \(t^*\) actually decreases (approaching the z-value) as degrees of freedom increase with larger sample sizes. Larger samples generally produce more precise slope estimates, an important principle for evaluating study design.

Since the researcher expects a positive association between study time and test scores, the correct one-sided alternative is \(H_a: \beta_1 > 0\), paired with the null of no relationship, \(H_0: \beta_1 = 0\). The choice "\(H_0: \beta_1 = 0\), \(H_a: \beta_1 \neq 0\)" is incorrect here because that represents a two-sided test, not the directional hypothesis implied by "higher test scores." Match the wording of the research question (increase, decrease, or just "different") to a one-sided or two-sided alternative hypothesis.

Since the p-value of 0.08 is greater than \(\alpha = 0.05\), we fail to reject \(H_0\), meaning the data do not provide statistically significant evidence of a linear relationship. The distractor "Reject \(H_0\); there is sufficient evidence of a linear relationship" is incorrect because rejection only occurs when the p-value is less than or equal to alpha. Always compare the p-value directly to the chosen significance level before stating a conclusion.

The correct long-run frequency interpretation is that 95% of confidence intervals constructed this way, across repeated samples, would contain the true population slope \(\beta_1\). The choice "There is a 95% probability that the true slope falls in this specific interval" is a common misconception because the true slope is a fixed number, not a random variable, so it either is or isn't in this one interval. Confidence level describes the reliability of the method over many samples, not the probability for one specific interval.

A curved pattern in the residual plot indicates that a linear model is not appropriate, so the Linear condition is violated and a transformation or nonlinear model should be considered before performing inference. The distractor "Increase the sample size to eliminate the curve" is wrong because a larger sample size does not fix a fundamentally curved relationship; it may even make the lack of fit more apparent. Always check the residual plot for curvature as a diagnostic that linear regression inference conditions are met before trusting the t-test results.

Since the p-value of 0.002 is less than \(\alpha = 0.01\), we reject \(H_0: \beta_1 = 0\) and conclude there is significant evidence of a linear relationship between \(x\) and \(y\). The distractor "The slope is not different from 1.2" is incorrect because 1.2 is the standard error value in the output, not a hypothesized comparison value for the slope. Reading computer output correctly means identifying which number is the estimate, which is the standard error, and which is the resulting p-value.

The t-distribution is used because the standard error of the slope is estimated using the sample standard deviation of residuals rather than a known population parameter \(\sigma\), introducing extra variability accounted for by the t-distribution. The distractor "Because the sample size is always small in regression studies" is wrong because the t-distribution is used regardless of sample size whenever \(\sigma\) is estimated from data, not only for small samples. Whenever a standard error is estimated rather than known exactly, the t-distribution (not z) is the appropriate inference tool.

With a small sample size like \(n=10\), even a fairly strong correlation can produce a standard error large enough, combined with few degrees of freedom, that the t-statistic does not clear the critical threshold for significance—though in this specific case \(r=0.9\) with \(n=10\) is actually significant, illustrating why sample size must always be checked rather than assumed. The distractor "Strong correlation guarantees rejection regardless of \(n\)" is wrong because significance depends jointly on both the strength of correlation and the sample size through the degrees of freedom. Students should recognize that both correlation strength and sample size jointly determine statistical significance in slope tests.

Degrees of freedom equal \(n - 2 = 10\), and since a 90% confidence level uses a smaller critical value \(t^*\) than a 95% confidence level, the resulting interval is narrower. The distractor "10 degrees of freedom; wider than the 95% CI" is incorrect on the width comparison because lower confidence levels always produce narrower intervals, not wider ones, all else equal. Remember the tradeoff: higher confidence requires a wider interval, while lower confidence yields a narrower, less certain interval.

An outlier that lies far from the regression line vertically (a large residual) but near the center of the x-values in leverage terms increases the scatter of residuals, inflating \(SE(b_1)\), without necessarily pulling the slope estimate much because it lacks high leverage. The distractor "The outlier was located exactly at \(\bar{x}\), giving it maximum leverage" is wrong because a point at \(\bar{x}\) actually has minimal leverage on the slope, not maximum, since leverage increases with distance from \(\bar{x}\). Distinguish between outliers affecting residual scatter (and thus standard error) versus high-leverage points that can dramatically shift the slope estimate itself.

Because the sample slope is positive, the one-sided test (Analysis B) concentrates all the p-value area in the upper tail, making its p-value exactly half of the two-sided test's p-value (Analysis A), which splits the area between both tails. The distractor "Analysis A's p-value is exactly half of Analysis B's p-value" reverses the correct relationship, since the two-sided test always has the larger (not smaller) p-value when the observed slope is in the direction of the alternative. Recall that one-sided tests have smaller p-values than two-sided tests when the sample result agrees with the direction of the one-sided alternative.

A statistically significant slope only demonstrates a linear association between \(x\) and \(y\); causation can only be established through a well-designed randomized experiment, not through observational regression data from 40 cities. The distractor "The claim is correct because the interval does not contain 0" is wrong because ruling out a zero slope confirms significance of the association, but says nothing about the mechanism of causation. Always distinguish statistical significance of a relationship from proof of causation, especially with observational data.

With four times as many observations, the standard error of the slope typically shrinks (roughly by a factor related to \(\sqrt{n}\)), producing a narrower confidence interval even with the same point estimate. The distractor "Identical, since the point estimate is the same" is incorrect because the width of a confidence interval depends on the standard error and degrees of freedom, not solely on the point estimate. Larger sample sizes generally increase the precision of slope estimates, narrowing confidence intervals accordingly.

Violating independence undermines the very formula used to calculate the standard error and p-value, because that formula assumes each observation contributes independent information; this problem persists and can worsen regardless of sample size. The distractor "Normality alone, since large samples always fix non-normal residuals" is only partly true because the Central Limit Theorem helps with non-normality for large \(n\), but does nothing to correct violations of independence. Independence is the condition least forgiving of large sample sizes, making it the most critical condition to verify through study design.

For simple linear regression, \(R^2\) can be expressed in terms of the t-statistic for the slope via \(R^2 = \frac{t^2}{t^2 + df}\), showing that a stronger relationship (higher \(R^2\)) corresponds to a larger magnitude t-statistic given fixed df. The distractor "\(R^2\) has no mathematical connection to the slope's t-statistic" is wrong because in simple linear regression there is exactly one explanatory variable, so \(R^2\) and the slope's significance test are directly linked through this formula. Understanding that \(R^2\), correlation \(r\), and the slope's t-statistic are all mathematically connected helps clarify why strong linear fits tend to produce significant slope tests.

The correct contextual interpretation describes the plausible range for the true population slope \(\beta_1\), meaning for each one-unit (here $1000) increase in advertising spending, true mean sales are estimated to increase between 0.10 and 0.90 units, with 99% confidence in the method. The distractor "There is a 99% chance that the sample slope falls between 0.10 and 0.90" is wrong because the sample slope \(b_1\) is already a known, fixed calculated value from this data, not a random quantity to be captured by the interval. Confidence interval interpretations for slopes must always be phrased in the context of the original variables and units, referencing the true population slope, not the sample statistic.

Greater scatter of the data points around the regression line means larger residuals, which increases the estimated standard deviation of residuals \(s\) and thus increases \(SE(b_1)\). The distractor "A wider spread of x-values in the sample" is incorrect because a wider spread of x-values actually decreases \(SE(b_1)\), since more spread-out x-values provide more leverage to precisely estimate the slope. Standard error of the slope depends on both the scatter of residuals and the spread of the explanatory variable, so students must consider both factors.

A p-value represents the probability, assuming the null hypothesis is true (no true linear relationship), of obtaining a sample slope as extreme as or more extreme than what was observed. The distractor "There is a 0.001 probability that there is no linear relationship" misinterprets the p-value as a probability about the hypothesis itself, which is not how frequentist p-values work. Always phrase p-value interpretations in terms of the probability of the observed (or more extreme) data given that the null hypothesis is true, never as a probability about the hypothesis.

The scatterplot offers an initial overall look at the relationship, helping to identify potential nonlinearity, outliers, or clusters before modeling, while the residual plot removes the linear trend and makes deviations from linearity or equal variance far easier to detect. The distractor "The scatterplot checks independence while the residual plot checks linearity" mismatches the actual purposes, since neither plot type is the primary tool for verifying independence, which instead depends on the data collection design. Using both plots together provides complementary information: an overall pattern check and a magnified diagnostic of leftover variation.

The population slope \(\beta_1\) is a fixed, unknown constant, so it either lies within this specific computed interval or it does not; the 95% confidence level instead describes the long-run success rate of the interval-construction method across many samples. The distractor "The interval should not include negative values" is wrong because a legitimate confidence interval can certainly include zero or negative values, especially when the true relationship might be weak or absent. This common misinterpretation of confidence intervals is one of the most frequently tested conceptual points in AP Statistics.

The t-statistic is calculated as \(t = \frac{-2.4}{0.9} \approx -2.67\), and the negative sign reflects the direction of the sample slope, indicating that as x increases, y tends to decrease. The distractor "\(t \approx -0.375\)" is wrong because it inverts the formula (dividing \(SE(b_1)\) by \(b_1\) instead of the reverse). The sign of the t-statistic always matches the sign of the sample slope and conveys the direction of the estimated association, though it does not by itself indicate whether the test is valid.

The regression model assumes that for each value of x, the responses vary normally around the true regression line, which translates to the residuals (deviations from the fitted line) being approximately normally distributed, not the raw unconditional distribution of \(y\). The distractor "Because checking residuals is easier computationally but conceptually irrelevant" is wrong because the choice to examine residuals is conceptually essential, not just a computational convenience, since it directly reflects the model's normality assumption. Recognizing that regression normality assumptions apply to the residuals (conditional on x), not the marginal distribution of y, is a key conceptual distinction in AP Statistics.

Points with unusually large residuals but typical (non-extreme) x-values have low leverage, meaning they add scatter (noise) to the model that inflates \(SE(b_1)\) without substantially shifting the slope estimate itself, consistent with the described results. The distractor "They must have had extreme x-values far from \(\bar{x}\)" is contradicted by the question, which explicitly states the removed points had "typical x-values," meaning they lacked the leverage needed to pull the slope in one direction. Differentiating between outliers that affect standard error versus high-leverage points that affect the slope estimate itself is essential for advanced regression diagnostics.

Using \(df = n-1\) instead of the correct \(df = n-2\) means selecting a slightly smaller \(t^*\) critical value (since more degrees of freedom yield a smaller \(t^*\)), producing a marginally narrower interval than is technically correct. The distractor "The interval becomes centered on the wrong value" is wrong because the center of the confidence interval is always the sample slope \(b_1\), which is unaffected by the degrees of freedom used to find \(t^*\). Small errors in degrees of freedom typically cause only minor changes in interval width, but using the correct df = n - 2 remains essential for exam accuracy.

When testing against a nonzero hypothesized slope value, the general t-statistic formula subtracts the hypothesized value (here, 2) from the sample slope before dividing by the standard error, giving \(t = \frac{b_1 - 2}{SE(b_1)}\). The distractor "\(t = \frac{b_1 - 0}{SE(b_1)}\)" is the formula for testing against a hypothesized slope of 0, which does not match this scenario testing against a hypothesized value of 2. Always substitute the specific hypothesized parameter value into the general t-statistic formula, rather than defaulting to zero.

With very large samples, \(SE(b_1)\) becomes extremely small, so even a slope that is close to zero in practical terms can still be many standard errors away from 0, resulting in a small p-value and statistical significance despite limited real-world importance. The distractor "Statistical significance requires practical significance by definition" is a common misconception, since statistical and practical significance are distinct concepts that do not automatically align, particularly with very large datasets. AP Statistics emphasizes that students should always consider both statistical significance and practical significance (effect size) when interpreting regression results.

A negative linear effect corresponds to a slope less than zero, so the correct one-sided alternative hypothesis is \(H_a: \beta_1 < 0\), reflecting the researcher's directional claim. The distractor "\(H_a: \beta_1 \neq 0\)" represents a two-sided alternative, which does not capture the specific direction (negative) stated in the research question. Carefully translate directional research claims like "negative effect" or "positive relationship" into the matching one-sided alternative hypothesis.

Since the p-value of 0.045 is less than \(\alpha = 0.05\), we reject the null hypothesis of no linear relationship, concluding there is statistically significant evidence of an association between the variables. The distractor "Accept \(H_0\) as true" is incorrect because statistical hypothesis tests never allow us to "accept" or prove a null hypothesis true; we can only fail to reject it or reject it. Always use precise statistical language—reject or fail to reject—rather than saying a hypothesis is proven or accepted.

LINER is the common mnemonic AP Statistics students use to remember the five conditions: Linear, Independent, Normal, Equal variance, and Random, all of which should be checked before performing regression inference. The distractor "NICER" is incorrect because it is not a standard mnemonic taught for these particular conditions in this order. Memorizing the LINER checklist helps ensure no condition is overlooked when justifying the use of a t-distribution for slope inference.

The critical value \(t^*\) is determined jointly by the chosen confidence level (such as 90%, 95%, or 99%) and the degrees of freedom, which for slope inference equal \(n-2\). The distractor "Only the confidence level" is incomplete because \(t^*\) also changes with degrees of freedom, unlike \(z^*\) values which depend only on confidence level. Always look up \(t^*\) using both the confidence level and the correct degrees of freedom from a t-table.

There is no requirement that the sample size be a multiple of 10; this is not part of the LINER conditions used to justify regression inference. The choice "Residuals are approximately normal at each x" is a genuine required condition, part of the Normal condition, and thus is not the correct answer to this question asking for the condition that is NOT required. Be careful to distinguish real statistical conditions from arbitrary numeric rules that sound plausible but are not part of the actual checklist.

Since the entire interval \((0.8, 3.2)\) lies above zero, this provides strong evidence that the true population slope \(\beta_1\) is positive, meaning taller individuals tend to weigh more on average. The distractor "We are confident the true slope is exactly 2.0" is wrong because a confidence interval provides a plausible range of values, not a single exact point estimate for the true parameter. When an entire confidence interval falls on one side of zero, that strongly implies statistical significance in the direction indicated by the interval.

Because the data are observational rather than from a randomized experiment, we can only conclude there is a significant positive association between coffee consumption and heart rate, not that one variable causes the other. The distractor "Coffee consumption causes an increase in heart rate" wrongly asserts causation, which requires controlled experimental design to rule out confounding variables. Statistical significance from observational regression data speaks only to association, never automatically to causation.

With small samples, there are too few observations for the Central Limit Theorem to smooth out non-normality in the residuals, so if the residuals are strongly skewed or have outliers, the sampling distribution of \(b_1\) may not be well approximated by a t-distribution. The distractor "The Normal condition only matters for large samples" reverses the actual relationship, since large samples are more robust to violations of normality, not small ones. The smaller the sample size, the more critical it becomes to verify normality of residuals using a graph like a normal probability plot or histogram.

The confidence interval is calculated as \(150 \pm 2.02(20) = 150 \pm 40.4\), giving an interval of approximately \((109.6, 190.4)\). The distractor "\((130, 170)\)" is incorrect because it uses a margin of error of only 20, which is the standard error itself rather than \(t^*\) multiplied by the standard error. Always multiply the standard error by the correct critical value \(t^*\) to obtain the margin of error before constructing the interval.

The slope's confidence interval captures the plausible values for the average change in the response variable per one-unit increase in the explanatory variable, while the intercept's confidence interval reflects plausible values of the response when \(x=0\), a value that is often outside the observed data range and thus may lack real-world meaning. The distractor "The intercept's CI is always more useful for prediction than the slope's CI" is incorrect because the slope, not the intercept, is generally the primary quantity of interest for describing the relationship between variables. Students should be cautious interpreting the intercept when \(x=0\) represents an extrapolation beyond the range of observed data.

The coefficient ('Coef') row for the explanatory variable in typical regression output gives the estimated slope \(b_1\), which serves as the numerator (minus the hypothesized value, usually 0) in the t-statistic formula. The distractor "The 'Constant' row" is incorrect because that row reports the y-intercept \(b_0\), a separate parameter unrelated to the slope's significance test. When reading computer output, always locate the correct row corresponding to the explanatory variable, not the constant term, when testing the slope.

Since the sample slope \(b_1 = 4.2\) is positive and matches the direction of the one-sided alternative \(H_a: \beta_1 > 0\), the one-sided p-value is exactly half of the two-sided p-value, giving \(0.032 / 2 = 0.016\). The distractor "\(0.064\)" incorrectly doubles the two-sided p-value instead of halving it, reversing the correct relationship between one-sided and two-sided tests. When the observed effect matches the direction of a one-sided alternative, always halve the two-sided p-value to get the correct one-sided p-value.

A residual plot with a fan or cone shape, where residuals grow larger as x increases, directly indicates that the variance of the residuals is not constant across values of x, violating the Equal Variance condition. The distractor "A residual plot shows residuals randomly scattered around zero with consistent spread across all x-values" actually describes what a residual plot should look like when the Equal Variance condition is satisfied, not violated. Fan-shaped or funnel-shaped residual plots are the classic visual signal of heteroscedasticity that AP Statistics exams frequently test.

Even when the true relationship is curved, a linear regression fit can still detect an overall positive or negative trend, producing a statistically significant slope, but this linear model may poorly capture the actual pattern and lead to inaccurate predictions, especially near the extremes. The distractor "The p-value being small means the Linear condition is automatically satisfied" is incorrect because statistical significance of the slope is a separate issue from whether the Linear condition (checked via residual plots) is actually met. A small p-value does not validate model conditions; conditions must be checked independently using graphical diagnostics like residual plots.

Sample B's t-statistic is \(5/1.0 = 5.0\) compared to Sample A's \(5/2.5 = 2.0\), and combined with more degrees of freedom (48 vs 8), Sample B produces a much smaller p-value and thus stronger evidence against the null hypothesis. The distractor "Sample A, because a smaller sample size always yields more significant results" is incorrect because smaller samples generally have larger standard errors and fewer degrees of freedom, both of which tend to reduce statistical significance, not increase it. Comparing t-statistics (not just point estimates) is essential when judging the strength of evidence across different studies.

Log transformations compress larger values proportionally more than smaller ones, which can reduce the increasing spread of residuals across x-values, helping to satisfy the Equal Variance condition when data show this fanning pattern. The distractor "Log transformations always fix violations of the Independence condition" is wrong because independence relates to how data were collected, not to the mathematical scale of the response variable, so no transformation can fix a design-based independence violation. Transformations like logarithms are common tools to address non-constant variance or nonlinearity, but they cannot resolve violations related to study design such as independence.