# Chapter 14 · Complete draft exercise solutions

**Constructed teaching examples. Technical and author review pending.**

1. $\bar x=3$, $\bar y=4$, $S_{xx}=10$, $S_{xy}=9$, $b=0.9$, $a=1.3$.
2. Fitted values: 2.2, 3.1, 4.0, 4.9, 5.8. Residuals: −0.2, −0.1, 1.0, −0.9, 0.2. Their sum is zero, as is $\sum x_ie_i$ up to rounding.
3. $SSE=1.9$, $SST=10$, $R^2=1-1.9/10=0.81$. This compares in-sample squared errors with a mean-only reference; it is not a causal share or independent predictive score.
4. With $u=x-3$, the fitted line is $\hat y=4+0.9u$. The predictions remain the same because this is only a change of origin.
5. $0.9$ points per hour is $0.9/60=0.015$ points per minute. If zero hours is still zero minutes and only the predictor unit changes, the intercept remains 1.3.
6. Fifty sessions lies far beyond the observed one-to-five range. The algebra gives a number, but the linear mechanism and measurement regime may fail there. Mark extrapolation and seek new data or substantive mechanism.
7. A curve may mean the conditional mean is nonlinear or that omitted groups have different trends. Inspect raw values, time, group and measurement definitions; compare a prespecified alternative model without deleting points solely to improve fit.
8. Cases within one instrument may share calibration, environment and drift. Five instruments, not fifty readings, may dominate independent instrument-level information. Use a design-aware uncertainty calculation.
9. At 10% baseline, odds are 1/9; doubled odds are 2/9 and probability 2/11 = 18.18%. At 50% baseline, doubled odds of 2 give probability 2/3 = 66.67%. Absolute changes are 8.18 and 16.67 points.
10. Logistic probabilities are 0.5, 0.8 and 0.2 at $z=0,\log4,-\log4$. A fixed log-odds increment changes probability most near the curve's middle.
11. Perfect separation can send unpenalised logistic coefficients toward extreme magnitudes. Check sample size, feature timing and leakage, uncertainty, penalisation and out-of-sample calibration before interpreting an odds ratio.
12. One defensible report says: “For five constructed observations with predictor values one to five and outcomes 2, 3, 5, 4, 6, the ordinary least-squares line is $\hat y=1.3+0.9x$ with in-sample $R^2=0.81$. Residuals include +1.0 and −0.9, and the endpoint values help determine slope. The data are invented and too small for a population or intervention claim; do not use this line to allocate practice time or predict at 50 sessions.”
13. The simple linear-projection slope is $0.2+1.5(0.4)=0.8$ points per practice hour. The extra 0.6 arises from the stipulated association between practice and preparation. The formula is a constructed projection; a causal practice effect would additionally require a credible intervention comparison and control of relevant common causes.
14. Below $c$ the hinge term is zero and the slope is $b_1$. Above $c$ it changes one-for-one with $X$, giving slope $b_1+b_2$. If the analyst searched over cutoffs, report the search, evaluate the selected model on independent data and avoid treating a naive interval as though the cutoff were fixed in advance.
15. The line gives $1.3+0.9(6)=6.7$. With $s^2=1.9/3$, $S_{xx}=10$ and $\bar x=3$, the conditional-mean standard error is $\sqrt{(1.9/3)\{1/5+9/10\}}\approx0.8347$. This calculation assumes the line continues; unmodeled curvature beyond five is not represented by it.
