Undergraduate MATH201 ⭐ 3 Credits Semester 3

Probability and Statistics for Scientists

Probability theory and statistical inference with emphasis on applications in the natural sciences. Topics include descriptive statistics, probability distributions, hypothesis testing, regression, and experimental design. Develops quantitative skills for designing and analyzing chemistry and biology experiments.

📅 Studied by Skipper on 2026-10-01

📖 Core Concepts

Claude Opus 5.5

Core Concepts: Probability and Statistics for Scientists

1. What Is This Subject About?

Every measurement you made in CHEM101L and CHEM102L, such as a titration endpoint, a mass on an analytical balance, or an absorbance reading, was really one draw from a distribution of possible values. Run the titration again and you get 24.37 mL instead of 24.41 mL. Neither number is "the answer." Each is a sample. Statistics is the discipline that takes uncertainty, which you might be tempted to treat as an annoyance to minimize, and turns it into a quantity you can model, measure, and reason about.

The subject has two halves that run in opposite directions:

  • Probability runs forward, from model to data. If a coin is fair, what fraction of 100 flips will be heads? If a coupling reaction succeeds 99% of the time, what is the chance that all 50 couplings in a synthesis succeed? Here you know the mechanism and predict the outcome.
  • Statistical inference runs backward, from data to model. I flipped a coin 100 times and got 61 heads; is it fair? My peptide came out 58% pure; what was my per-step coupling efficiency? Here you know the outcome and want to learn about the mechanism.

The backward direction is harder, and it is where most of the conceptual traps are. Its central difficulty is that many different "true states of the world" could have produced the same data. Inference does not give certainty. It gives calibrated statements about how compatible the data are with various hypotheses, and about how often a given procedure would mislead you if you used it repeatedly.

"All models are wrong, but some are useful." [Box, 1979]

George Box, a co-author of one of your textbooks, made this point repeatedly. A normal distribution is never exactly correct for real data, and neither is a straight-line calibration curve or an assumption of independent errors. The skill this course builds is knowing when a wrong model is still useful, and how to check whether it has stopped being useful (residual analysis, which comes later, is the main tool for this).

A useful analogy is to treat statistics as the thermodynamics of data. In PHYS201 and CHEM102 you saw that you cannot predict the trajectory of a single gas molecule, yet pressure, temperature, and entropy are sharply predictable because they are averages over huge numbers of molecules. Statistics uses the same move. Individual observations are unpredictable, but aggregates like means, proportions, and variances behave lawfully. The Central Limit Theorem, discussed below, is the mathematical reason this works, and the reason it works in both settings is closely related.

The two textbooks emphasize complementary things. Devore [Devore, 2016] builds the mathematical machinery carefully: distributions, estimators, test statistics. Box, Hunter, and Hunter [Box, Hunter & Hunter, 2005] focus on the experimenter's question of how to design data collection so that the analysis is informative in the first place. A well-designed experiment analyzed with simple methods usually beats a poorly designed one analyzed with sophisticated methods.


2. Describing Data: Center, Spread, and Shape

Before any inference, you need to summarize what you have. Suppose you have n observations x1, …, xn.

Measures of center

Sample mean:     x̄ = (1/n) Σ xᵢ

Sample median:   the middle value of the sorted data
                 (average of the two middle values if n is even)

The mean behaves like a center of mass. If you place unit masses at each data point on a number line, the mean is the balance point. You know from PHYS201 that a single mass far from the pivot has a lot of leverage, and the same is true here: the mean is sensitive to outliers. The median counts positions rather than weighing them, so it is resistant. One wildly wrong HPLC injection can drag a mean substantially while leaving the median nearly unchanged [Devore, 2016].

Measures of spread

Sample variance:            s² = Σ (xᵢ − x̄)² / (n − 1)

Sample standard deviation:  s = √s²

Interquartile range:        IQR = Q₃ − Q₁
                            (spread of the middle 50% of the data)

Why divide by n − 1 and not n? Students often accept this as a convention, but there is a real reason. The deviations (xi − x̄) are constrained to sum to zero, because that is what it means for x̄ to be the mean. Once you know n − 1 of the deviations, the last one is determined. Only n − 1 pieces of independent information about spread remain, which is what "n − 1 degrees of freedom" means.

There is also a more concrete reason. The sample mean x̄ is by construction the point that minimizes Σ(xi − c)². Measuring deviations from x̄ rather than from the unknown true mean μ therefore systematically underestimates the spread. Dividing by n − 1 corrects this bias exactly, so that E[s²] = σ². Degrees of freedom will reappear throughout the course, in the t, χ², and F distributions, and in each case they count independent pieces of information left over after you have estimated something.

Shape

Histograms and boxplots show shape: symmetric or skewed, one peak or two, gaps, outliers. A bimodal histogram of melting points from a recrystallization lab, for instance, might reveal that you have two polymorphs, or two students using differently calibrated apparatus. Always plot your data first. Summary statistics can hide structure that is obvious in a picture. Anscombe's quartet is the classic demonstration: four datasets with nearly identical means, variances, correlations, and regression lines, but completely different shapes.


3. Probability: Axioms, Conditioning, and Bayes' Theorem

The axioms

Probability rests on three axioms, usually credited to Kolmogorov. For a sample space S and events A, B:

1.  P(A) ≥ 0                                  for every event A
2.  P(S) = 1
3.  P(A ∪ B ∪ …) = P(A) + P(B) + …            for mutually exclusive events

Everything else follows from these three, including the complement rule P(A′) = 1 − P(A) and the general addition rule P(A ∪ B) = P(A) + P(B) − P(A ∩ B) [Devore, 2016]. It is worth noticing how little is assumed. The axioms do not say what probability means. It could be a long-run frequency or a degree of belief. They only say how probabilities must combine to be consistent.

Conditional probability and independence

P(A | B) = P(A ∩ B) / P(B)        provided P(B) > 0

A and B are independent  ⟺  P(A ∩ B) = P(A)·P(B)  ⟺  P(A | B) = P(A)

Conditioning means shrinking the universe. Once you know B occurred, B becomes your new sample space, and you ask what fraction of it is also A. You already used this in BIOL102. When you computed the probability that a phenotypically dominant F₂ offspring of an Aa × Aa cross is a heterozygote, you conditioned on the dominant phenotype. The answer was 2/3, not 1/2, because the homozygous recessives had been removed from consideration.

Bayes' theorem

                 P(B | A) · P(A)
P(A | B)  =  ─────────────────────────────────────
              P(B | A)·P(A) + P(B | A′)·P(A′)

Bayes' theorem tells you how to reverse a conditional probability. This matters because the conditional you can measure is often the opposite of the one you care about. A worked example shows why the reversal is so often counterintuitive.

Worked example. A diagnostic test has 99% sensitivity, so P(+ | disease) = 0.99. It has 95% specificity, so P(− | no disease) = 0.95. The condition has 1% prevalence. You test positive. What is the probability you have the disease?

                       (0.99)(0.01)                0.0099
P(D | +) = ──────────────────────────────── = ──────── ≈ 0.167
             (0.99)(0.01) + (0.05)(0.99)        0.0594

The answer is about 17%, not 99%. Imagine 10,000 people. About 100 have the disease, and 99 of them test positive. Of the 9,900 healthy people, 5% (495) also test positive. The false positives outnumber the true positives five to one, simply because healthy people are so much more common. This base-rate effect is one of the most important ideas in the course. It returns in disguised form when you interpret p-values (Section 8) and when proteomics software reports peptide identifications (Section 13).


4. Random Variables and Discrete Distributions

A random variable is a function that assigns a number to each outcome of a random experiment, for example "number of heads in 10 flips" or "number of failed couplings in a 30-residue synthesis." A discrete random variable has a probability mass function (pmf) p(x) = P(X = x). Two summaries are central:

Expected value:  E[X] = μ = Σ x · p(x)

Variance:        V(X) = σ² = E[(X − μ)²] = E[X²] − (E[X])²

The expected value is again a center of mass, this time of the probability distribution. It need not be a value X can actually take: the expected number of heads in 3 flips is 1.5.

Two properties are used constantly:

  • Linearity of expectation: E[aX + bY] = aE[X] + bE[Y], always, even when X and Y are dependent.
  • Variance of independent sums: V(X + Y) = V(X) + V(Y) only if X and Y are independent (more precisely, uncorrelated). Also V(aX) = a²V(X).

These two rules underlie the error propagation you did informally in your chemistry labs.

The binomial distribution

Suppose you run n independent trials, each with the same success probability p, and let X count the successes.

P(X = x) = C(n, x) · pˣ · (1 − p)ⁿ⁻ˣ,      x = 0, 1, …, n

E[X] = np          V(X) = np(1 − p)

The binomial coefficient C(n, x) counts the arrangements. The rest is the probability of any one specific arrangement. Mendelian genetics is binomial: the number of affected children among four, when each has a 1/4 chance, follows Binomial(4, 1/4).

The geometric distribution

Now let X count the number of trials up to and including the first success.

P(X = x) = (1 − p)ˣ⁻¹ · p,      x = 1, 2, 3, …

E[X] = 1/p

The mean has an intuitive form. If a reaction works 20% of the time, you expect to need about 5 attempts. Checking that the probabilities sum to 1 uses the geometric series from MATH102.

The Poisson distribution

The Poisson distribution counts rare events occurring at a constant average rate, such as radioactive decays per minute, mutations per kilobase, or photons hitting a detector.

P(X = x) = e^(−λ) · λˣ / x!,      x = 0, 1, 2, …

E[X] = λ          V(X) = λ

Two features are worth slowing down for.

  • It sums to 1 because of a Taylor series. From MATH102 you know that eλ = Σ λx/x!. The factor e−λ is exactly the normalizing constant that makes the probabilities sum to 1. The Poisson pmf is the Taylor expansion of the exponential function, read term by term as probabilities.
  • It is the limit of the binomial. If n → ∞ and p → 0 with np = λ held fixed, Binomial(n, p) → Poisson(λ). Many independent opportunities, each very unlikely, give a Poisson count.

The equality of mean and variance is a useful diagnostic. If your count data have variance much larger than the mean ("overdispersion"), the events are probably not independent or the rate is not constant. In biology this often signals clustering or heterogeneity.


5. Continuous Distributions and the Central Limit Theorem

For continuous random variables, probability is spread over intervals and described by a probability density function f(x):

P(a ≤ X ≤ b) = ∫ₐᵇ f(x) dx          ∫₋∞^∞ f(x) dx = 1

E[X] = ∫ x f(x) dx

This is where MATH101 pays off. The density f(x) is not a probability; it can exceed 1. The probability of any single exact value is zero. Only areas under the curve are probabilities, and density is "probability per unit x," just as mass density is mass per unit volume.

Uniform

The uniform distribution assigns f(x) = 1/(b − a) on [a, b], with mean (a + b)/2 and variance (b − a)²/12. Rounding error from a digital readout is roughly uniform over ± half the last digit.

Exponential, and a connection you already know

f(x) = λ e^(−λx),   x ≥ 0          E[X] = 1/λ          median = ln 2 / λ

Memoryless property:   P(X > s + t | X > s) = P(X > t)

This should look familiar. In CHEM102 you studied first-order kinetics, [A] = [A]₀e−kt, and in PHYS201 or CHEM101 radioactive decay. Here is the statistical meaning of those curves. Each individual molecule or nucleus has an exponentially distributed lifetime with rate parameter k. The macroscopic decay curve is the survival function P(X > t) = e−kt, averaged over an enormous population. Then:

  • The half-life t½ = ln 2 / k is the median lifetime. Half the molecules react before t½ and half after.
  • The mean lifetime 1/k is longer than the half-life, because the distribution is right-skewed with a long tail of persistent molecules.
  • Memorylessness is the statistical statement that a first-order reaction has no "aging." A molecule that has survived an hour is no more likely to react in the next second than a freshly made one. That is exactly what a unimolecular process with constant rate constant implies.

The exponential distribution also describes waiting times between Poisson events. If counts per unit time are Poisson(λ), the gaps between events are Exponential(λ). These are two descriptions of the same underlying process.

The normal (Gaussian) distribution

f(x) = [1 / (σ√(2π))] · exp[ −(x − μ)² / (2σ²) ]

Standardization:   Z = (X − μ) / σ   ~  N(0, 1)

Empirical rule:    ≈68% within μ ± σ
                   ≈95% within μ ± 2σ   (more precisely, ±1.96σ)
                   ≈99.7% within μ ± 3σ

The bell curve has no closed-form antiderivative, which is why you use Z-tables or software. Standardizing converts any normal into the standard normal, so a single table works for all of them.

The Central Limit Theorem

Here is the result that makes most of the course work.

Let X₁, X₂, …, Xₙ be independent and identically distributed
with mean μ and finite variance σ².  Then for large n:

    X̄  ≈  N(μ, σ²/n)

equivalently,

    (X̄ − μ) / (σ/√n)  →  N(0, 1)   as n → ∞

The theorem holds regardless of the shape of the original distribution, provided its variance is finite. The data can be skewed, lumpy, or discrete, and their average still becomes approximately bell-shaped. Roll a single die and the outcomes are uniform. Average 30 dice and the average is very nearly normal.

The CLT explains why the Gaussian appears so often in nature. Any quantity that results from adding up many small, independent contributions tends toward normality. Measurement error in an analytical balance is the sum of tiny effects: air currents, vibration, thermal drift, electronic noise. That is why errors are often approximately normal, and why the Maxwell–Boltzmann velocity components you met in PHYS201 are Gaussian.

Equally important is the √n in the denominator. The spread of the average shrinks like 1/√n. To halve your uncertainty you need four times as many measurements. To reduce it tenfold you need a hundred times as many. This square-root law governs the economics of every experiment you will design.


6. Sampling Distributions: Standard Error, t, χ², and F

A statistic (x̄, s², a regression slope) is computed from a random sample, so it is itself a random variable with its own distribution, called the sampling distribution. This is the conceptual pivot of the course. You stop asking only "what are my data like?" and start asking "if I repeated this whole experiment many times, how would my summary vary?"

Standard deviation (SD):   σ or s       — spread of individual observations
Standard error (SE):       σ/√n or s/√n — spread of the sample mean across repeated experiments

The SD describes the population and does not shrink with more data. The SE describes your estimate's precision and does shrink.

Distributions derived from the normal

When the data come from a normal population, three important distributions follow. All of them can be built from standard normals Z:

Chi-square:  χ²_ν  =  Z₁² + Z₂² + … + Z_ν²
             (n − 1)s² / σ²  ~  χ²_{n−1}

Student's t: t_ν   =  Z / √(χ²_ν / ν)
             (X̄ − μ) / (s/√n)  ~  t_{n−1}

F:           F_{ν₁,ν₂}  =  (χ²_{ν₁}/ν₁) / (χ²_{ν₂}/ν₂)
             s₁²/s₂² (when σ₁ = σ₂)  ~  F_{n₁−1, n₂−1}

Why the t-distribution exists. If you knew σ, then (X̄ − μ)/(σ/√n) would be exactly standard normal. In practice you substitute the estimate s, which is itself noisy. Sometimes s happens to be too small, which makes the ratio too large. The t-distribution accounts for this extra uncertainty by having heavier tails than the normal. As n grows, s becomes a reliable estimate of σ and tν → N(0, 1). With 3 replicates, the honest uncertainty multiplier is noticeably larger than 1.96: for a 95% interval, t0.025,2 ≈ 4.30. This is why triplicate measurements give such wide confidence intervals.

W. S. Gosset developed this distribution while working at the Guinness brewery, where small samples were unavoidable, and published it under the pseudonym "Student" in 1908. The t-distribution exists because real experimental work forced the issue.


7. Estimation and Confidence Intervals

A point estimator is a rule for guessing a parameter from data, such as x̄ for μ or s² for σ². Good estimators are unbiased (correct on average) and have small variance. When the two goals conflict, a common criterion is mean squared error:

MSE = Variance + Bias²

Devore also introduces maximum likelihood estimation [Devore, 2016]. You choose the parameter value that makes the observed data most probable. This general principle produces most of the standard estimators, and it reappears in regression.

A point estimate alone hides its uncertainty, so we report an interval:

100(1 − α)% CI for μ (σ unknown, normal population or large n):

    x̄  ±  t_{α/2, n−1} · s/√n

General form:   estimate  ±  (critical value) × (standard error)

What "95% confidence" actually means. This is one of the most frequently misunderstood ideas in science, so it is worth stating precisely. The true μ is a fixed, unknown constant. It does not vary. The interval is what varies, because it is computed from random data. "95% confidence" is a statement about the procedure. If you repeated the experiment many times and built an interval each time, about 95% of those intervals would contain μ.

For any single computed interval, say [12.3, 14.1], μ is either inside it or not. Under the frequentist framework of Devore, you cannot say "there is a 95% probability μ lies in [12.3, 14.1]." A helpful analogy is a ring toss by a player who rings the peg 95% of the time. Once a particular ring has landed, it either encircles the peg or it does not. The 95% describes the player's skill, not that particular throw.


8. The Logic of Hypothesis Testing

Hypothesis testing is proof by contradiction made probabilistic.

  1. State a null hypothesis H₀, usually "no effect" or "no difference" (e.g., μ = μ₀), and an alternative Hₐ (μ ≠ μ₀, μ > μ₀, or μ < μ₀).
  2. Assume H₀ is true and work out the distribution of a test statistic under that assumption.
  3. Compute the statistic from your data.
  4. Ask how surprising that value would be if H₀ were true. This is the p-value.
  5. If the data would be very surprising under H₀ (p ≤ α), reject H₀.
p-value = P( test statistic at least as extreme as observed | H₀ true )

The p-value is computed assuming H₀ is true. It is P(data this extreme | H₀), not P(H₀ | data). Recall the Bayes example from Section 3: these two conditionals can differ enormously, and converting one into the other requires a prior probability, which plays the role of the base rate.

Two kinds of error

                        H₀ true               H₀ false
                   ┌────────────────────┬────────────────────┐
  Reject H₀        │  Type I error (α)  │  Correct (power)   │
                   │  "false positive"  │  1 − β             │
                   ├────────────────────┼────────────────────┤
  Fail to reject   │  Correct           │  Type II error (β) │
                   │  1 − α             │  "false negative"  │
                   └────────────────────┴────────────────────┘

The courtroom analogy is standard because it fits well. H₀ is "innocent." Conviction requires evidence beyond reasonable doubt, which corresponds to small α. A verdict of "not guilty" does not mean the jury found the defendant innocent; it means the evidence was insufficient. In the same way, failing to reject H₀ is not evidence that H₀ is true. It may simply mean your experiment was too small to detect the effect.

Power, 1 − β, is the probability of correctly detecting a real effect. It increases with:

  • a larger true effect size,
  • a larger sample size n,
  • a smaller noise level σ,
  • a larger α (trading more false positives for fewer false negatives).

An underpowered study is wasteful in two ways. It usually misses real effects, and when it does reach significance, it tends to overestimate the effect size, because only the lucky overestimates cleared the threshold.


9. t-Tests: One-Sample, Two-Sample, and Paired

One-sample:     t = (x̄ − μ₀) / (s/√n),              df = n − 1

Two-sample      t = (x̄₁ − x̄₂) / √(s₁²/n₁ + s₂²/n₂)
(Welch):        df from the Welch–Satterthwaite approximation

Paired:         dᵢ = x₁ᵢ − x₂ᵢ
                t = d̄ / (s_d/√n),                    df = n − 1

Every t-statistic has the same structure: signal ÷ noise, or (observed difference) ÷ (standard error of that difference). The statistic is large when the difference is big relative to how much it would wobble by chance alone.

Welch's version does not assume equal variances. Devore recommends it as the default over the pooled-variance test, because it loses little when variances are equal and protects you when they are not [Devore, 2016].

Pairing: the first lesson in experimental design

Suppose you test whether a new resin improves crude peptide purity. You synthesize 8 different sequences on both the old and new resin. Peptide-to-peptide variation is large, since some sequences are simply harder to make than others. A two-sample test treats that variation as noise, and the noise can drown the resin effect. A paired test takes the difference within each sequence, which cancels the sequence-difficulty effect entirely. Mathematically:

V(X₁ − X₂) = V(X₁) + V(X₂) − 2·Cov(X₁, X₂)

When paired measurements are positively correlated, which they are when they share the same peptide, the covariance term subtracts out much of the variance. Box, Hunter, and Hunter's well-known boys'-shoes example makes this point. Two sole materials were each worn on one foot of each boy. Analyzed as independent samples, the difference was not detectable. Analyzed as pairs, it was clear [Box, Hunter & Hunter, 2005]. Pairing is the simplest case of blocking (Section 12).


10. Chi-Square Tests and ANOVA

Chi-square tests for categorical data

χ² = Σ (Observed − Expected)² / Expected

Goodness of fit:   df = k − 1 − (number of parameters estimated from data)
Independence
(r × c table):     df = (r − 1)(c − 1)
Requirement:       expected counts generally ≥ 5 in each cell

You may have used this in BIOL102 to test whether F₂ counts fit a 9:3:3:1 ratio. The statistic measures how far the observed counts are from what H₀ predicts, scaled by the expected count. Dividing by the expected count makes sense because for Poisson-like counts the variance is about equal to the mean, so each term is roughly a squared standardized deviation. This is why the sum follows a χ² distribution.

One historical note: R. A. Fisher's re-analysis of Mendel's data suggested the fit to Mendelian ratios was too good, with χ² values improbably small. Data that agree too closely with theory can also be statistically suspicious.

One-way ANOVA

To compare the means of k groups (for example, coupling yields with four different activating reagents), you might think of running all the pairwise t-tests. The problem is that with 4 groups there are 6 comparisons, and the chance of at least one false positive rises far above α. ANOVA addresses this with a single test of H₀: μ₁ = μ₂ = … = μk.

SST  =  SSTr  +  SSE
(total)  (between groups)  (within groups)

MSTr = SSTr / (k − 1)          MSE = SSE / (N − k)

F = MSTr / MSE   ~   F_{k−1, N−k}   under H₀

Although the method is called analysis of variance, it tests means. The idea is to estimate the noise variance σ² in two ways:

  • MSE uses only within-group scatter. It estimates σ² whether or not H₀ is true.
  • MSTr uses the scatter of the group means. It estimates σ² only if H₀ is true. If the true means differ, MSTr is inflated.

F ≈ 1 is therefore consistent with H₀, and F ≫ 1 is evidence against it. The partition SST = SSTr + SSE is the Pythagorean theorem in n-dimensional space: the deviation vector splits into orthogonal "between" and "within" components. Two-way ANOVA extends the partition to two factors plus their interaction. An interaction means the effect of one factor depends on the level of the other, for example when one reagent is best at room temperature but a different one is best at 50 °C.

Multiple comparisons

A significant F says only that some means differ. To find which ones without inflating false positives:

  • Bonferroni: test each of m comparisons at level α/m. It is simple, always valid, and conservative.
  • Tukey's HSD: designed specifically for all pairwise comparisons, using the studentized range distribution. It is less conservative than Bonferroni for this purpose.

The family-wise error rate makes the danger concrete. For m independent tests at α = 0.05, P(at least one false positive) = 1 − 0.95m. For m = 20 that is about 0.64.


11. Regression: Modeling Relationships

Simple linear regression

Model:   Yᵢ = β₀ + β₁xᵢ + εᵢ,     εᵢ ~ independent N(0, σ²)

Least squares estimates:
    β̂₁ = Sxy / Sxx = Σ(xᵢ − x̄)(yᵢ − ȳ) / Σ(xᵢ − x̄)²
    β̂₀ = ȳ − β̂₁x̄

Coefficient of determination:
    R² = 1 − SSE/SST   (fraction of variation in y explained by the model)

Your Beer–Lambert calibration curves in CHEM101L and CHEM102L were regressions. Least squares minimizes the sum of squared vertical residuals. There is a deeper justification than convenience: if the errors are normal, the least-squares line is also the maximum likelihood line. Choosing squared error amounts to assuming Gaussian noise.

In matrix form, which generalizes to many predictors:

y = Xβ + ε           β̂ = (XᵀX)⁻¹ Xᵀy

"Linear" means linear in the parameters, not in x. The model y = β₀ + β₁x + β₂x² is a linear regression model, and so is y = β₀ + β₁ log x. A model such as Michaelis–Menten, v = Vmax[S]/(Km + [S]), is genuinely nonlinear in its parameters and requires iterative nonlinear least squares.

The historical workaround of linearizing (for example, the Lineweaver–Burk double-reciprocal plot) distorts the error structure. Taking 1/v enormously magnifies the noise in the smallest rates, so the least reliable points end up with the most leverage. Fitting the nonlinear model directly is better practice.

Residual analysis

The assumptions (linearity, independence, constant variance, normality) are checked by plotting residuals ei = yi − ŷi:

  • Residuals vs. fitted values should look like a structureless horizontal band.
  • Curvature indicates a missing term or the wrong functional form.
  • A funnel shape indicates non-constant variance (heteroscedasticity). This is common in analytical chemistry, where error grows with concentration, and it suggests weighted regression or a transformation.
  • A normal probability (Q–Q) plot checks the normality of the errors.

A high R² does not mean the model is correct. A calibration curve with curvature can have R² = 0.995 and still give systematically biased concentrations at the ends of the range. Only the residual plot reveals this.


12. Experimental Design: Randomization, Blocking, Replication

Box, Hunter, and Hunter argue that the most important statistical decisions are made before any data are collected [Box, Hunter & Hunter, 2005]. Three principles organize good design.

  1. Replication. Repeat independent runs to estimate the noise σ and to reduce the standard error. Be careful about pseudoreplication. Injecting the same sample into the HPLC three times measures instrument noise, not the batch-to-batch variability of your synthesis. Only truly independent experimental units count as replicates.
  2. Blocking. Group runs into homogeneous sets (same day, same batch of reagent, same peptide sequence) and compare treatments within blocks, so that known nuisance variation cancels out. This generalizes the paired test.
  3. Randomization. Assign treatments to runs at random. This protects you against nuisance factors you did not think of or cannot measure, such as instrument drift, room temperature, or reagent aging, by making them unlikely to line up with your treatments. Randomization is also what justifies the probability calculations in the analysis.
"Block what you can and randomize what you cannot." [Box, Hunter & Hunter, 2005]

Factorial designs are a major theme of BHH. The intuitive approach of changing one factor at a time is inefficient and blind to interactions. A 2k factorial design varies all k factors simultaneously at two levels each. With 3 factors (temperature, equivalents of reagent, time) you need 8 runs. Every run contributes to the estimate of every main effect, and interactions come for free. Fractional factorials screen many factors with even fewer runs.

Sample size calculation combines everything above. For a two-sided, two-sample comparison of means with common σ, detecting a difference Δ with power 1 − β requires approximately:

n per group  ≈  2 · [ (z_{α/2} + z_β) · σ / Δ ]²

Example: α = 0.05 (z = 1.96), power 0.80 (z = 0.84), Δ = σ
    n ≈ 2 · (2.80)² ≈ 15.7  →  16 per group

Note that n scales with (σ/Δ)². Detecting an effect half as large requires four times as many samples, which is the CLT's square-root law appearing again.

Nonparametric alternatives

When normality is doubtful and samples are small, rank-based tests replace the data with their ranks. The Mann–Whitney (Wilcoxon rank-sum) test is the nonparametric counterpart of the two-sample t-test. The Kruskal–Wallis test is the counterpart of one-way ANOVA. Ranks are immune to outliers, just as the median is, at the cost of a modest loss of power when the data really are normal.


13. Connections to Peptide Chemistry

You will meet each idea below again in peptide synthesis, analysis, and biology. In each case statistics is part of the chemistry itself, not just a tool applied afterward.

Solid-phase peptide synthesis is a binomial process

In SPPS, a peptide of n residues requires roughly n − 1 sequential couplings. If each succeeds independently with probability p, the fraction of chains that are full length is:

P(full-length) = pⁿ

p = 0.99,  n = 50:   0.99⁵⁰  ≈ e^(−0.503) ≈ 0.605   (≈ 60% full-length)
p = 0.995, n = 50:   0.995⁵⁰ ≈ e^(−0.251) ≈ 0.778   (≈ 78% full-length)

Number of deletion events per chain ~ Binomial(n, 1 − p);
mean deletions at p = 0.99, n = 50:  50 × 0.01 = 0.5

A half-percent improvement per step yields nearly 18 percentage points more product. This compounding is why long peptides are so hard to make, why coupling efficiency is monitored so closely, and why longer proteins are often assembled by ligating shorter peptides. The same arithmetic applies to cumulative racemization risk across many residues.

Mass spectrometry isotope patterns are binomial and Poisson

Natural carbon is about 1.07% 13C. For a peptide with n carbons, the number of 13C atoms is approximately Binomial(n, 0.0107), well approximated by Poisson(λ = 0.0107n).

Fraction at monoisotopic mass (all ¹²C):  (0.9893)ⁿ

n = 100 carbons:  (0.9893)¹⁰⁰ ≈ e^(−1.076) ≈ 0.34

Ratio of M+1 to M peak:  P(1)/P(0) ≈ λ ≈ 0.0107n

As a peptide grows, the monoisotopic peak shrinks relative to its neighbors. Once λ exceeds 1, which happens at roughly 93 carbons from carbon alone (fewer once 15N, 2H, and other heavy isotopes are included), the M+1 peak exceeds the monoisotopic peak. For large proteins the monoisotopic peak is often barely visible. The pattern of isotope peaks you will interpret in mass spectra is a probability distribution plotted directly by the instrument.

Proteomics and the base-rate problem

When a mass spectrometer identifies peptides by matching spectra against a database, thousands of matches are scored simultaneously. Without correction for multiple testing, false identifications pile up, as with the family-wise error calculation in Section 10. The field therefore controls the false discovery rate, often estimated with "target–decoy" strategies that search reversed or scrambled sequences. The underlying logic is the Bayes calculation from Section 3: a good score is not enough when true matches are rare among all possible candidates.

Dose–response and binding curves

Measuring the potency of a bioactive peptide (IC50, EC50, Kd) means fitting sigmoidal models such as the four-parameter logistic. These are nonlinear regressions that need iterative fitting, residual checking, and confidence intervals on the fitted parameters. Comparing two peptide analogs' potencies is a hypothesis test on those parameters.

Optimizing synthesis and purification

Coupling conditions (reagent, equivalents, temperature, time, solvent) and HPLC gradients (slope, temperature, ion-pairing agent) involve several factors that interact. This is the setting factorial and response-surface designs from BHH were built for, and they are standard in process chemistry. Purity and yield comparisons across conditions call for ANOVA, and comparing the same sequences under two protocols calls for paired designs.

Statistical mechanics of conformations

Looking further ahead, peptide conformational ensembles, helix–coil transitions, and folding are described by Boltzmann distributions over states. The populations you will compute are probability distributions, and the averaged observables measured by NMR and CD spectroscopy are expected values over those ensembles.


14. Common Misconceptions to Watch Out For

  1. "The p-value is the probability that H₀ is true." It is P(data this extreme | H₀), not P(H₀ | data). Getting from one to the other requires Bayes' theorem and a prior. A p-value of 0.03 can correspond to a high probability that H₀ is true when the prior plausibility of the effect is low.
  2. "Failing to reject H₀ proves there is no effect." Absence of evidence is not evidence of absence. A non-significant result from an underpowered study is uninformative. If you want to claim "no meaningful difference," report a confidence interval and show it excludes differences you would care about, or use equivalence testing.
  3. "A 95% CI has a 95% probability of containing μ." In the frequentist framework, 95% is the long-run success rate of the procedure, not a probability statement about one realized interval.
  4. "The CLT makes my data normal." The CLT concerns the distribution of the sample mean (and other sums), not the raw data. Skewed data stay skewed no matter how many you collect. Also, the "n ≥ 30" rule of thumb is not universal: heavily skewed distributions may need much larger n.
  5. Confusing SD and SE. Error bars showing SE look smaller than those showing SD, and they mean different things. SD describes variability among individuals. SE describes precision of the mean. Always state which one you are plotting.
  6. "Statistically significant means important." With a large enough n, a trivially small effect becomes significant. A 0.1% purity improvement can be "highly significant" and still chemically irrelevant. Report effect sizes and intervals, not just p-values.
  7. "High R² means a good model." R² can be high for a model with systematic misfit, and low for a correct model of noisy data. Look at residual plots. Never extrapolate a regression beyond the range of the data used to fit it.
  8. "Correlation implies causation." Observational associations can come from confounders. Randomized experiments are the main tool for supporting causal claims, which is why randomization is central in Box, Hunter, and Hunter.
  9. Pseudoreplication. Technical replicates (repeat injections, repeat readings) are not independent experimental replicates. Treating them as such shrinks the SE artificially and inflates false positives.
  10. Ignoring multiple comparisons. Testing 20 hypotheses at α = 0.05 and reporting only the "hit" is a reliable way to publish noise. Plan comparisons in advance, and correct with Bonferroni, Tukey, or FDR methods.
  11. "Variances always add." V(X + Y) = V(X) + V(Y) requires independence (zero covariance). Correlated errors, such as two measurements sharing the same miscalibrated pipette, violate this rule and can make your error propagation badly wrong.
  12. "Linear regression can only fit straight lines." "Linear" refers to the parameters. Polynomials and log terms are fine. Conversely, linearizing a truly nonlinear model (as in Lineweaver–Burk) distorts the error structure, so fit the nonlinear model directly.

A Closing Thought

The intellectual core of this course is a change in what you treat as the object of study. In chemistry so far, the objects were molecules and reactions. In statistics, the object is the experiment itself, considered as one draw from the set of all experiments you could have run. A confidence interval, a p-value, or a power calculation is a statement about how your procedure would behave across that whole set of possible experiments. Once that shift becomes natural, the formulas follow from it, and you will start designing experiments by asking first what answer each possible outcome could give you.

References

  • [Devore, 2016] Devore, J. L. Probability and Statistics for Engineering and the Sciences, 9th ed. Boston: Cengage Learning, 2016.
  • [Box, Hunter & Hunter, 2005] Box, G. E. P., Hunter, J. S., & Hunter, W. G. Statistics for Experimenters: Design, Innovation, and Discovery, 2nd ed. Hoboken, NJ: Wiley, 2005.
  • [Box, 1979] Box, G. E. P. "Robustness in the strategy of scientific model building." In R. L. Launer & G. N. Wilkinson (Eds.), Robustness in Statistics. New York: Academic Press, 1979.

⚙️ Mechanisms & Applications

GPT-6 Astra

Mechanisms & Applications: Learning to Reason from Experiments

Skipper, in chemistry we often ask, “What mechanism could have produced this result?” Statistics adds a second question: “How much would this result change if we repeated the experiment?” That question connects molecular mechanisms to reliable scientific knowledge. A measured difference in peptide activity might reflect a genuine structural effect, assay noise, plate position, cell-culture history, or a combination of these. Statistics helps separate those possibilities—but only when the design and the mathematical model respect how the data were generated.

The central workflow is therefore not “collect numbers, then choose a test.” It is define the scientific target → design the comparison → model the variation → estimate the effect → quantify uncertainty → challenge the model. Devore develops the probability machinery behind this workflow; Box, Hunter, and Hunter emphasize that its greatest value comes from improving experiments, not merely analyzing completed ones. [1, 2]

1. From Laboratory Measurements to a Model of Variation

A useful starting model for repeated measurements is:

Observed measurement = systematic component + random variation

Yᵢ = μ + εᵢ

E(εᵢ) = 0
Var(εᵢ) = σ²

This does not say that every source of error is random or normally distributed. A miscalibrated balance can introduce bias; consecutive chromatographic injections can have correlated errors; fluorescence variability may increase with signal. The model is a hypothesis about the measurement process, not a law of nature.

  1. Identify the observational and experimental units. A measurement is an observation. The experimental unit is the entity independently assigned a treatment: a patient, animal, culture flask, or manufacturing batch.
  2. Plot the observations. Histograms reveal distribution shape; scatterplots reveal relationships; plots against run order can expose drift. For small samples, show individual observations rather than relying only on a histogram.
  3. Describe location and spread. Use the mean and standard deviation when they answer the scientific question; use the median and quartiles when robustness to extreme observations matters.
  4. Investigate unusual values. An outlier may be a transcription error, an instrument fault, or the most scientifically important observation. It is not automatically disposable.
Sample mean:
x̄ = (1/n) Σᵢ xᵢ

Sample variance:
s² = [1/(n − 1)] Σᵢ (xᵢ − x̄)²

Sample standard deviation:
s = √s²

Interquartile range:
IQR = Q₃ − Q₁

The denominator n − 1 has a structural explanation. Once the sample mean is fitted, the deviations must sum to zero; only n − 1 deviations can vary freely. For independent observations with a common mean and finite variance, this correction makes s² unbiased for the population variance. It does not make s exactly unbiased for the population standard deviation. [1]

Peptide-industry example: Suppose three independently synthesized peptide batches are each injected into an HPLC instrument four times. There are twelve injections, but only three independent batches for inference about batch-to-batch synthesis performance. Repeated injections estimate analytical precision; independent batches estimate manufacturing reproducibility. Treating all twelve injections as independent batch replicates is pseudoreplication.

Connection to laboratory chemistry: High precision is not high accuracy. Repeating a biased assay can make the mean extremely precise around the wrong value. Reference standards, blanks, calibration checks, and independent methods address problems that a small standard error cannot fix.

2. Conditional Probability: Moving from Mechanisms to Evidence

Probability begins with a coherent accounting system: probabilities are nonnegative, the total probability is one, and probabilities of mutually exclusive events add. Conditional probability changes the reference population. Instead of asking how often an event occurs among all cases, we ask how often it occurs among cases satisfying another condition. [1]

Conditional probability:
P(A | B) = P(A ∩ B) / P(B),    provided P(B) > 0

Multiplication rule:
P(A ∩ B) = P(A | B)P(B)

Independence:
P(A ∩ B) = P(A)P(B)

Bayes' theorem:
P(A | B) =
P(B | A)P(A) /
[P(B | A)P(A) + P(B | not A)P(not A)]

Medical application: interpreting a positive diagnostic test. Assume disease prevalence is 1%, sensitivity is 95%, and specificity is 95%.

  1. Among 10,000 representative people, about 100 have the disease.
  2. Of those 100, about 95 test positive.
  3. Of the 9,900 without disease, about 495 also test positive.
  4. Thus, among positive results, only 95 out of 590 represent disease.
P(disease | positive)
= (0.95 × 0.01) / [(0.95 × 0.01) + (0.05 × 0.99)]
≈ 0.161

The surprising result is that a positive test implies only about a 16% probability of disease in this population. Sensitivity is not the probability of disease after a positive test. The same logic applies when screening thousands of peptide candidates: if genuinely active compounds are rare, apparently strong assay performance can still produce many false leads.

Historical context: Thomas Bayes’s essay, published posthumously in 1763 through Richard Price, addressed a particular inverse-probability problem. Pierre-Simon Laplace subsequently developed inverse probability much more broadly. The important conceptual reversal was from predicting observations given a model to learning about a model from observations. [3, 4]

3. Choosing a Distribution by Understanding Its Generating Process

A distribution should be chosen because its assumptions resemble the experiment—not merely because its curve resembles a histogram. The expected value is a probability-weighted average, and variance measures expected squared deviation from that average. [1]

Discrete:
E(X) = Σₓ xP(X = x)

Continuous, with density f:
E(X) = ∫ x f(x) dx

Variance:
Var(X) = E[(X − E(X))²]
       = E(X²) − [E(X)]²
  • Binomial: Count successes in a fixed number of independent trials having the same success probability. Example: the number of independently sampled vials failing a specified test.
  • Geometric: Count independent trials until the first success, with a constant success probability. Example: an idealized series of independent synthesis attempts until one meets specification. Here the count includes the successful trial.
  • Poisson: Count events in a specified exposure under a constant-rate, independent-event model. Example: background detector counts or idealized contamination events per sampled volume.
  • Uniform: Every equal-length interval inside a bounded range has equal probability. This describes a randomization device or a deliberately uniform sampling scheme—not ignorance in general.
  • Exponential: Model a waiting time with a constant instantaneous event rate. Example: a molecular lifetime under a single first-order decay mechanism.
  • Normal: Model continuous variation arising approximately from many small additive contributions, provided dependence and unusually dominant contributions do not invalidate that approximation.
Binomial: X ~ Binomial(n, p)
P(X = k) = C(n,k)pᵏ(1 − p)ⁿ⁻ᵏ
E(X) = np;  Var(X) = np(1 − p)

Geometric: T = 1, 2, ...
P(T = k) = (1 − p)ᵏ⁻¹p
E(T) = 1/p;  Var(T) = (1 − p)/p²

Poisson: X ~ Poisson(λ)
P(X = k) = e⁻λ λᵏ/k!
E(X) = Var(X) = λ

Uniform on [a,b]:
f(x) = 1/(b − a), for a ≤ x ≤ b
E(X) = (a + b)/2;  Var(X) = (b − a)²/12

Exponential with rate k:
f(t) = k e⁻ᵏᵗ, for t ≥ 0
P(T > t) = e⁻ᵏᵗ
E(T) = 1/k;  Var(T) = 1/k²

Normal: X ~ N(μ, σ²)
f(x) = exp[−(x − μ)²/(2σ²)] / [σ√(2π)]

Connection to CHEM102 kinetics: For first-order disappearance, concentration satisfies dC/dt = −kC, giving C(t)/C(0) = e−kt. Under a constant-hazard molecular model, this fraction is also the probability that a molecule survives beyond time t. A deterministic concentration curve and a stochastic molecular-lifetime distribution describe the same mechanism at different scales.

The exponential model is memoryless: surviving for an hour does not change the distribution of the remaining lifetime. This is appropriate for some molecular processes but often inappropriate for aging equipment or disease progression.

Model check: Under a Poisson model, mean and variance are equal. Much greater observed variance can signal clustering, heterogeneous rates, or an omitted batch effect. “Overdispersion” is often a clue to biology, not just an inconvenience in calculation.

4. Sampling Distributions: Why Averaging Works—and When It Does Not

A sampling distribution describes how a statistic changes across hypothetical repetitions of the whole sampling process. The distribution of individual peptide concentrations is not the same object as the distribution of their sample mean.

For independent, identically distributed observations
with mean μ and finite variance σ²:

E(X̄) = μ
Var(X̄) = σ²/n
SD(X̄) = σ/√n

Estimated standard error:
SE(x̄) = s/√n

Central Limit Theorem:
(X̄ − μ)/(σ/√n) approaches N(0,1) as n increases.

The Central Limit Theorem does not say that the raw observations become normal. It says that their standardized average becomes approximately normal under suitable conditions. There is no universal rule that n = 30 is sufficient: strongly skewed or heavy-tailed data may require much larger samples. [1]

SD versus SE: Standard deviation describes variation among individual observations. Standard error describes uncertainty in an estimated quantity. Reporting a small SE does not establish that a manufacturing process produces uniform vials.

The square-root law is both beautiful and expensive: halving the standard error requires four times as many independent observations. Dependence makes the situation worse:

Without assuming independence:

Var(X̄) =
[Σᵢ Var(Xᵢ) + 2Σᵢ<ⱼ Cov(Xᵢ, Xⱼ)] / n²

Positive correlations among measurements from the same culture, patient, or batch reduce the information gained by adding more measurements within that cluster. This is why biological replication and technical replication are not interchangeable.

5. Confidence Intervals and the Logic of the t-Distribution

When the population standard deviation is unknown, replacing it with a sample estimate introduces additional uncertainty. For independent normal observations, the resulting standardized mean follows a t-distribution rather than a standard normal distribution.

T = (X̄ − μ)/(S/√n) ~ t with n − 1 degrees of freedom

Two-sided 100(1 − α)% confidence interval for μ:
x̄ ± t₁₋α/₂,ₙ₋₁ × s/√n
  1. Specify the parameter: for example, mean peptide recovery under a defined purification protocol.
  2. Calculate the point estimate and its standard error.
  3. Choose a confidence level before interpreting the results.
  4. Use the relevant sampling distribution to obtain the multiplier.
  5. Interpret the interval together with the design and model assumptions.

A frequentist 95% confidence procedure produces intervals that contain the fixed true parameter in 95% of repeated applications under its assumptions. It does not mean that 95% of individual observations lie inside the interval, nor does it directly assign a 95% posterior probability to the parameter lying inside the particular observed interval. [1]

Paired example: Ten independent peptide batches are each split between purification methods A and B. Let the within-batch difference be B minus A. Suppose the mean difference is 2.4 percentage points and its standard deviation is 1.9 percentage points.

n = 10;  d̄ = 2.4;  s_d = 1.9

SE(d̄) = 1.9/√10 ≈ 0.601
t₀.₉₇₅,₉ ≈ 2.262

95% CI = 2.4 ± 2.262(0.601)
       ≈ [1.04, 3.76] percentage points

Pairing works by subtracting away shared batch characteristics. Algebraically, Var(B − A) = Var(B) + Var(A) − 2Cov(A,B). When the paired measurements are positively correlated, the difference can be much less variable than an unpaired comparison.

Historical context: William Sealy Gosset developed small-sample methods while working at Guinness and published “The Probable Error of a Mean” under the name “Student” in 1908. The industrial problem was practical: how can limited experimental material support dependable conclusions? [5]

6. Hypothesis Testing: Measuring Conflict, Not Proving a Claim

A hypothesis test asks whether the observations are unusually incompatible with a specified null model. The null is not automatically “what science believes”; it is a reference model used to calibrate evidence.

  1. State the estimand and hypotheses. For example, H0: μB − μA = 0 versus a two-sided alternative.
  2. Choose the analysis and error criterion. Specify the significance level, endpoints, and any multiplicity adjustment.
  3. Compute a test statistic. Usually this is an observed discrepancy divided by its standard error.
  4. Calculate the p-value. Under the null model, find the probability of a statistic at least as extreme as observed, according to the selected test.
  5. Report the effect and uncertainty. A p-value without an effect estimate is scientifically incomplete.
One-sample t statistic:
t = (x̄ − μ₀)/(s/√n)

Paired t statistic:
t = d̄/(s_d/√n)

Welch two-sample t statistic:
t = (x̄₁ − x̄₂ − Δ₀) / √(s₁²/n₁ + s₂²/n₂)

Welch approximate degrees of freedom:
ν = (s₁²/n₁ + s₂²/n₂)² /
    [(s₁²/n₁)²/(n₁ − 1) + (s₂²/n₂)²/(n₂ − 1)]

Welch’s test does not require equal population variances and is generally preferable to pooling variances without justification. Small-sample exact one-sample and paired t inference requires normal observations or normal paired differences, respectively; larger-sample approximations can be useful but remain sensitive to severe outliers and dependence. [1]

  • Type I error: Rejecting a true null. A valid level-α procedure controls its probability at no more than α under the null assumptions.
  • Type II error: Failing to reject the null at a specified alternative effect.
  • Power: The probability of rejecting the null at that specified alternative; power = 1 − β.

Critical interpretation: A p-value is not the probability that the null is true, and it is not the probability that the result was “caused by chance.” Failing to detect a difference also does not establish equivalence. A pharma study seeking to show that two formulations are sufficiently similar needs prespecified equivalence margins and an appropriate equivalence analysis.

Fisher emphasized significance tests as tools for learning from experiments. Neyman and Pearson developed a complementary framework focused on decision rules, error rates, and power. Modern practice often combines these traditions; understanding their different goals prevents mechanical interpretation of a threshold such as 0.05. [6, 7]

7. Count Data: Testing Genetic Ratios and Clinical Associations

Chi-square methods compare observed counts with counts predicted under a null model.

Pearson chi-square statistic:
χ² = Σcategories (Observed − Expected)² / Expected

Goodness-of-fit, with fully specified category probabilities:
Expectedᵢ = n pᵢ
Degrees of freedom = number of categories − 1

Independence in an r × c contingency table:
Expectedᵢⱼ = (row totalᵢ × column totalⱼ)/(grand total)
Degrees of freedom = (r − 1)(c − 1)

Genetics application: In BIOL102, a dihybrid cross under specific Mendelian assumptions predicts a 9:3:3:1 phenotype ratio. For 160 offspring, the expected counts are 90, 30, 30, and 10. A goodness-of-fit test evaluates disagreement with that complete model—not merely with the arithmetic ratio. Linkage, differential survival, phenotype misclassification, or nonindependent sampling can all matter.

Clinical application: A contingency table can test association between treatment group and occurrence of an adverse event. The test does not quantify clinical importance; risk differences, risk ratios, or other suitable effect measures should accompany it.

The usual chi-square reference distribution is an approximation requiring adequate expected counts. Sparse tables may need exact or simulation-based methods. Estimating model parameters from the same data can reduce goodness-of-fit degrees of freedom, with the details depending on the fitting procedure. Pearson’s 1900 work helped make goodness-of-fit a general quantitative scientific question. [1, 8]

8. ANOVA and Factorial Experiments: Separating Sources of Variation

Analysis of variance is best understood as a comparison of two estimates of noise. Under the null hypothesis that all treatment means are equal, both between-group and within-group variation estimate the same error variance. Real treatment differences inflate the between-group component.

One-way model:
Yᵢⱼ = μ + τᵢ + εᵢⱼ

SS_between = Σᵢ nᵢ(x̄ᵢ − x̄_grand)²
SS_within  = ΣᵢΣⱼ(xᵢⱼ − x̄ᵢ)²

F = [SS_between/(g − 1)] / [SS_within/(N − g)]

Under the classical null model:
F ~ F with (g − 1, N − g) degrees of freedom

Here g is the number of groups and N the total number of independent observations. Classical ANOVA assumes independent normal errors with a common variance. The underlying distributional link is:

If Z₁, ..., Zν are independent standard normals:
Σ Zᵢ² ~ χ²ν

If U ~ χ²a and V ~ χ²b are independent:
(U/a)/(V/b) ~ F(a,b)

A significant omnibus ANOVA indicates that the means are not all equal; it does not identify which means differ. Tukey procedures target all pairwise mean comparisons, while Bonferroni controls familywise error for m selected tests by using a per-test threshold of α/m. Bonferroni does not require independence among those tests. [1]

Two-way peptide-synthesis experiment: Cross two coupling reagents with three temperatures and independently replicate every combination.

Two-way fixed-effects model:
Yᵢⱼₖ = μ + αᵢ + βⱼ + (αβ)ᵢⱼ + εᵢⱼₖ

αᵢ       = reagent effect
βⱼ       = temperature effect
(αβ)ᵢⱼ   = interaction

An interaction means that the effect of one factor depends on the other. If reagent A performs best at low temperature but reagent B performs best at high temperature, reporting a single “best reagent” averaged across temperatures can conceal the chemistry.

Why factorial experiments matter: Changing one factor at a time can miss interactions entirely. Crossing factors reveals the response surface more efficiently, and independent replication supplies information about experimental error. In an unrestricted two-way interaction model, one observation per cell leaves no residual degrees of freedom for estimating pure error. [2]

9. Regression: Connecting Quantitative Mechanisms to Measurements

Regression models the mean response as a function of explanatory variables while separately describing residual variation. In ordinary least squares, fitting means choosing coefficients that minimize squared residuals.

Simple linear regression:
Yᵢ = β₀ + β₁xᵢ + εᵢ

Multiple linear regression:
Yᵢ = β₀ + β₁xᵢ₁ + ... + βₚxᵢₚ + εᵢ

Residual:
eᵢ = yᵢ − ŷᵢ

Ordinary least-squares objective:
Minimize SSE = Σᵢ eᵢ²

For an ordinary least-squares fit with an intercept:
R² = 1 − SSE/SST
SST = Σᵢ(yᵢ − ȳ)²
  1. Choose predictors from the scientific question. Include plausible batch, concentration, temperature, or interaction terms where justified.
  2. Fit the model. Estimate the coefficients and their uncertainty.
  3. Plot residuals against fitted values, predictors, and run order. Curvature suggests a missing mean-response feature; a funnel suggests unequal variance; temporal structure suggests dependence or drift.
  4. Assess influential observations and predictor redundancy. Strongly correlated predictors can make individual coefficients unstable.
  5. Evaluate prediction beyond the fitted data where possible. High in-sample R² is not proof of causality, a correct mechanism, or reliable extrapolation.

Analytical application: Beer–Lambert calibration relates absorbance to concentration. Ordinary least squares assumes the predictor values are effectively known and response errors have constant variance. Uncertain standard concentrations or concentration-dependent noise may require a different analysis. Weighted least squares uses weights proportional to inverse error variance when that variance structure can be justified.

Biochemical application: enzyme kinetics.

Michaelis–Menten mean-response model:
v(S) = Vmax S/(Km + S)

Additive-error observation model:
vᵢ = Vmax Sᵢ/(Km + Sᵢ) + εᵢ

Nonlinear least-squares objective:
Minimize Σᵢ[vᵢ − Vmax Sᵢ/(Km + Sᵢ)]²

Taking reciprocals produces a straight-line equation, but it changes the error structure and can magnify noise at low velocities. Fitting the original nonlinear equation with an appropriate error model is often preferable. “Linear regression” means linear in its coefficients: a polynomial in concentration can still be a linear regression model.

Design insight: If every substrate concentration is far below Km, then v ≈ (Vmax/Km)S. The data primarily identify the ratio, not the two parameters separately. Measurements near and above Km are needed to reveal curvature and saturation. A sophisticated fitting algorithm cannot recover information the experiment never collected. [1, 2]

10. Experimental Design: Creating Information Rather Than Just Data

Randomization, blocking, and replication solve different problems:

  • Randomization prevents systematic allocation based on known or unknown characteristics and supplies a basis for randomization-based inference. It balances nuisance factors in expectation, not necessarily in every small experiment.
  • Blocking compares treatments within relatively homogeneous groups, such as donor, synthesis batch, assay day, or plate. It removes predictable nuisance variation from the treatment comparison.
  • Replication provides independent opportunities to observe treatment responses and estimate variability.
  • Blinding reduces measurement, handling, and interpretation biases arising from knowledge of treatment identity.

Concrete assay workflow: To compare two peptide formulations, include both formulations on every plate, randomize their positions within each plate, and include controls that can detect assay failure. Use independent biological preparations across plates or explicitly account for shared preparations. If all A samples are measured on Monday and all B samples on Tuesday, formulation and day are confounded; statistical software cannot separate their effects from those data alone. [2]

Sample-size planning: First define the smallest effect worth detecting, then estimate variability, choose error rates, and account for the actual design.

Approximate sample size per group for two independent groups,
equal allocation, common SD σ, and a two-sided normal-approximation test:

n ≈ 2(z₁₋α/₂ + z₁₋β)² σ²/δ²

δ = effect size to detect
α = Type I error probability
1 − β = target power

For α = 0.05, 80% power, σ = 10 units, and δ = 5 units:

n ≈ 2(1.96 + 0.84)²(10²)/(5²)
  ≈ 62.7

Round upward: approximately 63 independent units per group.

This is a planning approximation, not a universal answer. A t-based calculation, attrition, clustering, multiple endpoints, or unequal allocation may increase the requirement. Pairing or blocking may reduce it when they genuinely reduce relevant variance.

Fisher’s work at Rothamsted made randomization and analysis of variance central to agricultural experimentation. The same design logic transfers directly to fermentation, cell culture, peptide synthesis, and clinical trials: nuisance variation must be managed in the experiment, not merely wished away afterward. [2, 6]

11. Nonparametric Methods: Changing What We Ask of the Data

Rank methods are useful when numerical distances are unreliable, observations are ordinal, or a location-based normal model is unsuitable. They trade some numerical information for different assumptions.

  1. Pool observations across independent groups.
  2. Replace values with ranks, using suitable handling of ties.
  3. Measure how strongly group membership is associated with rank.
  4. Compare that statistic with its null distribution.
  • Mann–Whitney: Compares two independent groups through relative ordering. Its statistic is related to the probability that an observation from one group exceeds one from the other, with half-weight for ties.
  • Kruskal–Wallis: Extends a rank-based comparison to several independent groups.

These are not assumption-free tests, and they are not automatically tests of medians. A simple median-shift interpretation requires additional assumptions, such as distributions with the same shape differing only in location. Unequal shapes and spreads can affect conclusions. For paired observations, Mann–Whitney is inappropriate because it discards the pairing. [1]

Biochemical application: An ordinal histopathology score may be better suited to rank-based analysis than to a model assuming equal numerical spacing between scores. Conversely, an assay that reports all concentrations below its detection limit as the same value creates a censoring problem; merely ranking those values does not recover the missing information.

Skipper’s Working Principle

The unit of independence, the scientific estimand, and the mechanism of variation come before the test. A hundred technical measurements cannot replace independent biological replication. A small p-value cannot turn a negligible effect into a useful medicine. A high R² cannot validate an incorrect kinetic mechanism. Statistics becomes most powerful when it is used with chemistry and biology to decide what experiment would distinguish competing explanations.

Sources and Historical References

  1. Devore, J. L. Probability and Statistics for Engineering and the Sciences. 9th ed. Cengage Learning, 2016. Principal source for the distributional results, estimation, testing, ANOVA, regression, and nonparametric methods presented here.
  2. Box, G. E. P., Hunter, J. S., and Hunter, W. G. Statistics for Experimenters: Design, Innovation, and Discovery. 2nd ed. Wiley, 2005. Principal source for experimental strategy, blocking, randomization, factorial designs, interactions, and iterative model improvement.
  3. Bayes, T. “An Essay towards Solving a Problem in the Doctrine of Chances.” Communicated by Richard Price. Philosophical Transactions of the Royal Society of London 53 (1763): 370–418. doi:10.1098/rstl.1763.0053.
  4. Laplace, P.-S. Théorie analytique des probabilités. Paris: Courcier, 1812.
  5. Student [W. S. Gosset]. “The Probable Error of a Mean.” Biometrika 6, no. 1 (1908): 1–25. doi:10.1093/biomet/6.1.1.
  6. Fisher, R. A. Statistical Methods for Research Workers. Oliver and Boyd, 1925; and The Design of Experiments. Oliver and Boyd, 1935.
  7. Neyman, J., and Pearson, E. S. “On the Problem of the Most Efficient Tests of Statistical Hypotheses.” Philosophical Transactions of the Royal Society of London, Series A 231 (1933): 289–337. doi:10.1098/rsta.1933.0009.
  8. Pearson, K. “On the Criterion That a Given System of Deviations from the Probable in the Case of a Correlated System of Variables Is Such That It Can Be Reasonably Supposed to Have Arisen from Random Sampling.” Philosophical Magazine, series 5, 50, no. 302 (1900): 157–175. doi:10.1080/14786440009463897.

🪬 Skipper's Reflection

Gemini 3.1 Pro

1. What surprised me most about this material

Coming off the heels of Physics I and Calculus II, my brain had become deeply wired for determinism. If I know the initial velocity of a projectile, or the rate constant of a chemical reaction, I can integrate and find the exact state of the system at time t. The most surprising thing about MATH201 was realizing that randomness has its own strict geometry.

The Central Limit Theorem (CLT) is, without exaggeration, one of the most beautiful and shocking concepts I have encountered so far. The idea that you can take a population with a completely chaotic, asymmetrical, non-normal distribution, and yet, if you take large enough samples, the means of those samples will perfectly organize themselves into a pristine Gaussian bell curve... it feels like magic. As Devore writes, the CLT is the bridge between the unpredictable individual and the predictable aggregate. It made me realize that statistics isn’t the study of guessing; it’s the macroscopic mechanics of uncertainty. It reminded me of thermodynamics in CHEM102—we can't track a single gas molecule, but the aggregate behaves with perfect mathematical predictability.

2. The conceptual breakthrough moments — when things clicked

My biggest breakthrough came with Hypothesis Testing and the dreaded p-value. Before this course, I assumed a p-value of 0.05 meant "there is a 5% chance the null hypothesis is true." Devore ruthlessly corrected this misconception. The click happened when I tied it back to conditional probability: the p-value is $P(\text{Data} \mid H_0)$, not $P(H_0 \mid \text{Data})$. It is the probability of observing data this extreme assuming the null hypothesis is entirely true. If that probability is tiny, we reject the null. This realization completely changed how I read scientific literature.

A second massive "aha!" moment happened with ANOVA (Analysis of Variance). Initially, I was terribly confused by the name. Why are we analyzing variance to figure out if means are different? Then I visualized it. Box, Hunter, and Hunter explain this brilliantly: ANOVA is just a signal-to-noise ratio. You compare the variance between the groups (the signal of your treatment) to the variance within the groups (the background noise). If the "between" variance is significantly larger than the "within" variance, your treatment actually did something. Suddenly, the F-distribution wasn't just a formula; it was a magnifying glass for finding truth in noisy data.

3. How this changes how I think about peptides

This course radically upgraded my ability to evaluate the peptide industry, bridging my biochemistry knowledge with clinical and manufacturing realities.

Take Tirzepatide and Retatrutide. When I read the clinical trial data for these multi-receptor agonists, I used to just look at the average weight loss percentages. Now, I immediately look for the standard error, the confidence intervals, and the multiple comparison corrections. Because Retatrutide is a triple agonist (GLP-1, GIP, Glucagon), clinical trials must compare multiple dosage groups against a placebo and potentially against single-agonist baselines. I now understand why they use tools like Tukey's or Bonferroni's corrections in their ANOVA—if you run enough t-tests, you'll eventually find a false positive (Type I error). The statistics protect the biology from our own wishful thinking.

On the manufacturing and quality control side, for peptides like BPC-157 and TB-500, Box, Hunter, and Hunter's focus on Experimental Design (DOE) blew my mind. When optimizing the solid-phase synthesis of these peptides, you don't just change one variable (like temperature or solvent) at a time. That ignores interaction effects! By using factorial design and blocking, you can test temperature, solvent, and coupling time simultaneously, isolating the main effects and how they interact. Furthermore, tracking batch purity isn't just about passing a test; it’s about establishing control charts using normal distributions to detect when a synthesizer is drifting out of calibration before a batch actually fails.

4. Questions this course raises that I want to investigate further

First, Bayes' theorem makes so much intuitive sense for updating our beliefs based on prior knowledge. Why, then, is the frequentist framework (null hypotheses, p-values) the absolute gold standard for FDA clinical trials? I want to investigate Bayesian clinical trial design and see if it offers a better way to test peptide efficacy, especially in early-phase trials with smaller sample sizes.

Second, we covered linear and multiple linear regression, which are elegant, but I know from CHEM201 and BIOL101 that biological systems rarely scale linearly forever. Receptors saturate. How do we rigorously apply nonlinear regression to model Michaelis-Menten kinetics or receptor binding affinities for peptides like Retatrutide? I need to understand how residual analysis works when the underlying model is a curve, not a line.

Finally, the concept of "Power" (1 - Type II error) made me wonder about the hundreds of anecdotal "biohacker" reports on peptides. Since their sample sizes (n=1) are virtually powerless, the probability of them falsely accepting a null effect—or falsely attributing a placebo effect to the peptide—is astronomically high. How can we crowdsource n=1 data into a statistically powered aggregate?

5. Rating: Confidence levels after studying

Most Confident: Continuous distributions (Uniform, Exponential, Normal) and the Central Limit Theorem. Thanks to MATH101 and MATH102, calculating probabilities by integrating the area under a probability density function (PDF) felt like second nature. I also feel highly confident in standard one-sample and two-sample t-tests and interpreting confidence intervals.

Medium Confident: ANOVA and Multiple Regression. I completely grasp the concept of ANOVA as a signal-to-noise ratio, and R² as explained variance. However, setting up a Two-Way ANOVA by hand, specifically calculating the interaction sum of squares, is algebraically tedious, and I feel I'd rely heavily on software to execute it perfectly.

Least Confident: Nonparametric tests (Mann-Whitney, Kruskal-Wallis) and Sample Size/Power calculations. Nonparametric tests feel like a bit of a "black box" right now—I understand we use them when data isn't normally distributed by ranking the data, but the underlying mechanics of how the test statistics are derived don't feel intuitive yet. As for Power calculations, balancing effect size, alpha, and beta feels circular; I need to program a few simulations in Python to watch how these variables push and pull against each other before it truly solidifies in my mind.

🗺️ Visual Aids & Key Takeaways

Claude Opus 5.5

Diagram 1: The Statistical Inference Pipeline

╔══════════════════════════════════════════════════════════════════════════════════╗
║              THE STATISTICAL INFERENCE PIPELINE                                ║
║         from raw reality to defensible conclusions                             ║
╚══════════════════════════════════════════════════════════════════════════════════╝

      ┌─────────────────────────────────────────────────────────────────┐
      │                     POPULATION                                  │
      │   the true, unknowable distribution of all possible values      │
      │   parameters: μ (true mean),  σ² (true variance),  p (true     │
      │   proportion) — these exist but we can NEVER measure them all   │
      └────────────────────────────┬────────────────────────────────────┘
                                   │
                          random sampling
                     (goal: every member equally likely)
                                   │
                                   ↓
      ┌─────────────────────────────────────────────────────────────────┐
      │                   SAMPLE  (size n)                              │
      │   x₁, x₂, … , xₙ  — the data we actually have                 │
      │   statistics: x̄ (sample mean),  s² (sample variance),  p̂      │
      │   these are RANDOM VARIABLES — they vary from sample to sample  │
      └────────────────────────────┬────────────────────────────────────┘
                                   │
                    two parallel paths of inference
                    ┌──────────────┴──────────────┐
                    ↓                             ↓
      ┌─────────────────────────┐   ┌──────────────────────────────┐
      │   POINT ESTIMATION      │   │   INTERVAL ESTIMATION        │
      │                         │   │                              │
      │   x̄  estimates  μ       │   │   x̄ ± t*(s/√n)             │
      │   s²  estimates  σ²     │   │   → confidence interval      │
      │   p̂  estimates  p       │   │   captures μ with stated     │
      │                         │   │   probability (e.g. 95%)     │
      │   single best guess,    │   │   honest about uncertainty   │
      │   no uncertainty shown  │   │                              │
      └─────────────────────────┘   └──────────────────────────────┘
                    │                             │
                    └──────────────┬──────────────┘
                                   ↓
      ┌─────────────────────────────────────────────────────────────────┐
      │               HYPOTHESIS TESTING                                │
      │                                                                 │
      │   H₀ (null):  e.g. μ = μ₀   vs.   Hₐ (alternative): μ ≠ μ₀  │
      │                                                                 │
      │   compute test statistic  →  find p-value  →  decide           │
      │                                                                 │
      │   p-value = P(data this extreme | H₀ is TRUE)                  │
      │   small p → data are surprising under H₀ → reject H₀           │
      │                                                                 │
      │   ┌───────────────────────────────────────────────────────┐    │
      │   │  ERROR TYPES                                          │    │
      │   │  Type I  (α): reject H₀ when H₀ is true  — false +   │    │
      │   │  Type II (β): fail to reject when H₀ is false — miss  │    │
      │   │  Power = 1 − β : probability of catching a real effect │    │
      │   └───────────────────────────────────────────────────────┘    │
      └─────────────────────────────────────────────────────────────────┘

The full path from population to conclusion, showing how point estimates, confidence intervals, and hypothesis tests are three different answers to the same underlying question: what can we learn about the population from our sample?

Diagram 2: Family of Probability Distributions and When to Use Each

╔══════════════════════════════════════════════════════════════════════════════════╗
║              PROBABILITY DISTRIBUTIONS: A DECISION MAP                         ║
╚══════════════════════════════════════════════════════════════════════════════════╝

  What kind of outcome are you modeling?
             │
      ┌──────┴──────┐
      ↓             ↓
  DISCRETE        CONTINUOUS
  (counts,        (measurements,
   successes)      lengths, times)
      │                   │
  ┌───┴────────────┐   ┌──┴─────────────────────────────┐
  │                │   │                                │
  ↓                ↓   ↓                                ↓
BINARY           COUNT  BOUNDED or      UNBOUNDED
outcomes         data   uniform         (real-valued)
  │                │       │                │
  ↓                ↓       ↓                ↓
┌──────────────┐ ┌────────────────┐ ┌────────────┐ ┌───────────────────┐
│  BINOMIAL    │ │    POISSON     │ │  UNIFORM   │ │  NORMAL (Gauss)   │
│              │ │                │ │            │ │                   │
│ n trials,    │ │ events per     │ │ every value│ │ symmetric, bell   │
│ each P(S)=p  │ │ interval; λ   │ │ in [a,b]   │ │ shaped            │
│              │ │ = avg rate     │ │ equally    │ │ μ, σ fully define │
│ E[X] = np    │ │                │ │ likely     │ │ it                │
│ Var = np(1-p)│ │ E[X] = λ      │ │            │ │                   │
│              │ │ Var  = λ      │ │ E[X]=(a+b)/2│ │ E[X] = μ         │
│ "how many    │ │               │ │            │ │ Var  = σ²        │
│  successes   │ │ "how many     │ │ "random    │ │                   │
│  in n tries?"│ │  mutations in │ │  number    │ │ "measurement      │
│              │ │  a gene?"     │ │  generator"│ │  errors, heights, │
│ limit: when  │ │               │ │            │ │  many natural     │
│ n→∞, p→0,   │ │ approx.       │ │            │ │  phenomena"       │
│ np=λ fixed   │ │ Binomial when │ └────────────┘ │                   │
│ ⇒ Poisson   │ │ n large, p sm.│               └───────────────────┘
└──────────────┘ └────────────────┘
                                                            │
                                                   ┌────────┴────────┐
                                                   ↓                 ↓
                                            EXPONENTIAL         t / χ² / F
                                                                (sampling
                                         time between          distributions)
                                         Poisson events
                                                                used for
                                         E[X] = 1/λ            inference from
                                         Var  = 1/λ²           small samples;
                                         memoryless!           arise naturally
                                                               from Normal data

  ─────────────────────────────────────────────────────────────────────────────
  KEY RELATIONSHIPS
  Binomial(n,p)  ──n large, p small──→  Poisson(λ=np)
  Any distribution ─ average of n──→  Normal   [Central Limit Theorem]
  Normal(0,1)² summed ──────────────→  Chi-square(k)
  ─────────────────────────────────────────────────────────────────────────────

A taxonomy of the major probability distributions, showing when each applies, its key parameters, and the mathematical relationships that connect them — especially the Central Limit Theorem as the unifying bridge.

Diagram 3: The Central Limit Theorem — Why It Is Astonishing

╔══════════════════════════════════════════════════════════════════════════════════╗
║         THE CENTRAL LIMIT THEOREM: ANY SHAPE IN, BELL CURVE OUT               ║
╚══════════════════════════════════════════════════════════════════════════════════╝

  Start with ANY population distribution — it does not have to be Normal:

  SKEWED POPULATION          BIMODAL POPULATION         UNIFORM POPULATION
  (e.g. reaction times)      (e.g. two species mixed)   (e.g. random integers)

       ▐█                         ▐█   █▌                  ▐████████▌
       ▐██                        ▐██ ██▌                  ▐████████▌
       ▐███▌                      ▐████▌                   ▐████████▌
       ▐████▌                     ▐████▌                   ▐████████▌
  ─────────────────         ─────────────────         ─────────────────
  mean μ,  variance σ²      mean μ,  variance σ²      mean μ,  variance σ²

         │                          │                          │
         │  draw random             │  draw random             │  draw random
         │  samples of size n       │  samples of size n       │  samples of size n
         │  compute x̄ each time    │  compute x̄ each time    │  compute x̄ each time
         ↓                          ↓                          ↓

  ╔═════════════════════════════════════════════════════════════════════════════╗
  ║  SAMPLING DISTRIBUTION OF  x̄   (after many, many samples)                ║
  ║                                                                            ║
  ║                              ▐███▌                                        ║
  ║                           ▐███████▌                                       ║
  ║                        ▐███████████▌                                      ║
  ║                     ▐███████████████▌                                     ║
  ║               ──────────────────────────────                              ║
  ║                                                                            ║
  ║  Shape:  Normal  N(μ,  σ²/n)   regardless of original population shape   ║
  ║  Mean of x̄ :  μ       (same as population)                               ║
  ║  Std dev of x̄ :  σ/√n  (shrinks as n grows — this is standard error)     ║
  ║  Works well for n ≥ 30 in most practical cases                            ║
  ╚═════════════════════════════════════════════════════════════════════════════╝

  WHY THIS MATTERS FOR CHEMISTRY AND BIOLOGY:
  ┌───────────────────────────────────────────────────────────────────────────┐
  │  • Your assay measurements almost certainly aren't normally distributed  │
  │    at the individual level — but your SAMPLE MEANS are, approximately.   │
  │  • This is why t-tests work even on non-normal data (with large enough n)│
  │  • This is why averaging replicates is so powerful — σ/√n shrinks error  │
  │    by factor √n: triple your replicates ⇒ error drops by factor ~1.73   │
  └───────────────────────────────────────────────────────────────────────────┘

  STANDARD ERROR  =  σ/√n    (NOT the same as standard deviation!)
       │
       └──→ measures precision of the MEAN estimate, not spread of raw data

The Central Limit Theorem applied to three differently shaped populations, showing that the sampling distribution of the mean converges to Normal with standard deviation σ/√n — the single most important theorem justifying most of classical statistics.

Diagram 4: ANOVA Conceptual Structure — Partitioning Variance

╔══════════════════════════════════════════════════════════════════════════════════╗
║         ONE-WAY ANOVA: WHERE DOES THE VARIATION COME FROM?                    ║
╚══════════════════════════════════════════════════════════════════════════════════╝

  Suppose you measure enzyme activity under 3 different pH conditions (groups).
  You have n observations total, spread across k = 3 groups.

  TOTAL OBSERVED VARIATION IN THE DATA
  ─────────────────────────────────────
  SSTotal  =  Σᵢⱼ (xᵢⱼ − x̄grand)²     ← every observation vs. grand mean

         is partitioned EXACTLY into two parts:

  ┌──────────────────────────────┐     ┌──────────────────────────────────────┐
  │   SS BETWEEN groups          │     │   SS WITHIN groups                   │
  │   (SS Treatment)             │  +  │   (SS Error / SS Residual)           │
  │                              │     │                                      │
  │  Σ nᵢ(x̄ᵢ − x̄grand)²       │     │  Σᵢⱼ (xᵢⱼ − x̄ᵢ)²                  │
  │                              │     │                                      │
  │  variation DUE TO the        │     │  variation WITHIN each group,        │
  │  different treatments        │     │  purely from random noise            │
  │                              │     │                                      │
  │  df = k − 1                  │     │  df = n − k                          │
  └──────────────────────────────┘     └──────────────────────────────────────┘
              │                                         │
              ↓                                         ↓
       MS Between = SS_B / (k−1)            MS Within = SS_W / (n−k)
              │                                         │
              └────────────────┬────────────────────────┘
                               ↓
                         F  =  MS_Between / MS_Within

  ┌─────────────────────────────────────────────────────────────────────────────┐
  │  Intuition: F asks "is the spread BETWEEN group means much bigger than     │
  │  the spread we'd expect just from within-group noise?"                     │
  │                                                                             │
  │  If H₀ is true (all group means equal): F ≈ 1  (both MS estimate σ²)     │
  │  If treatments have real effect:        F >> 1 (MS_Between inflated)      │
  │                                                                             │
  │  F follows an F-distribution with df₁ = k−1, df₂ = n−k under H₀         │
  └─────────────────────────────────────────────────────────────────────────────┘

  AFTER ANOVA: if F is significant, WHERE are the differences?
  Use multiple comparison corrections to avoid inflating Type I error:

  BONFERRONI:  divide α by number of comparisons — simple, conservative
  TUKEY HSD:   uses studentized range distribution — preferred for all pairwise
  ─────────────────────────────────────────────────────────────────────────────
  WARNING: Running 20 t-tests at α=0.05 gives ~64% chance of at least one
           false positive. ANOVA + Tukey controls this. Box, Hunter & Hunter
           stress this as central to proper experimental design. (p. 147)
  ─────────────────────────────────────────────────────────────────────────────

The variance-partitioning logic of one-way ANOVA, showing how SS_Total splits into between-group and within-group components, how the F-statistic is constructed from those components, and why post-hoc corrections are necessary after a significant result.

Key Takeaways

  1. The p-value is NOT the probability that H₀ is true. It is P(observing data this extreme or more extreme | H₀ is true). This is a conditional probability in the frequentist direction: it assumes H₀ and asks how surprising the data are. Confusing it with P(H₀ | data) is called the "prosecutor's fallacy" and requires Bayes' theorem to correct. Devore (Ch. 9) is explicit on this point.
  2. Standard error (SE = s/√n) and standard deviation (s) measure entirely different things. Standard deviation measures the spread of individual observations; standard error measures how precisely the sample mean estimates the population mean. SE shrinks with larger n; s stabilizes. In a lab report, reporting SE when you mean SD (or vice versa) is a substantive error, not a cosmetic one.
  3. The Central Limit Theorem states that the sampling distribution of x̄ approaches N(μ, σ²/n) as n → ∞, regardless of the population's shape. In practice, n ≥ 30 is often sufficient unless the population is extremely skewed. This theorem is the reason t-tests and z-tests work on non-normal data at moderate sample sizes.
  4. A 95% confidence interval does NOT mean "there is a 95% probability that μ lies in this interval." μ is a fixed (though unknown) constant — it either is or isn't in the interval. The correct interpretation is: if this procedure is repeated many times, 95% of the resulting intervals will contain μ. The probability is in the procedure, not in the parameter.
  5. Type I error (α) and Type II error (β) are inversely coupled for fixed sample size. Lowering α (making the test more stringent) raises β (making it harder to detect a real effect). Power = 1 − β is the probability of correctly detecting a real effect. Increasing n is the only way to simultaneously reduce both error types, which is why sample size calculation before an experiment is not optional — it is an ethical and scientific obligation (Box, Hunter & Hunter, Ch. 3).
  6. Bayes' theorem, P(A|B) = P(B|A)·P(A) / P(B), can dramatically change your interpretation of a test result depending on the prior probability P(A). A diagnostic test with 99% sensitivity and 99% specificity applied to a disease with 0.1% prevalence yields a positive predictive value of only ~9%: most positives are false positives. This is the base-rate neglect problem and it is directly connected to the difference between p-values and posterior probabilities.
  7. The Poisson distribution is the limiting case of Binomial(n, p) as n → ∞ and p → 0 with λ = np held fixed. It has the unique and remarkable property that its mean equals its variance (both equal λ). In biology, it describes rare events: spontaneous mutations per genome per generation, radioactive decay counts, receptor-ligand binding events at low concentrations. Overdispersion (variance > mean) signals the data are not truly Poisson.
  8. One-way ANOVA partitions total sum of squares (SS_Total) exactly into SS_Between and SS_Within, and the F-statistic is the ratio of their mean squares. Under H₀ (all group means equal), both MS_Between and MS_Within are unbiased estimators of σ², so F ≈ 1. A large F indicates treatment effects. ANOVA tests whether any means differ; it does not tell you which ones. Post-hoc tests (Tukey HSD for all pairwise comparisons, Bonferroni for pre-planned comparisons) are required for that.
  9. In simple linear regression, R² is the fraction of total variance in y explained by the linear relationship with x. R² = SS_Reg / SS_Total = 1 − SS_Residual / SS_Total. A high R² does not mean the model is appropriate — always inspect residual plots. Residuals should be randomly scattered around zero with no pattern; any systematic curve in the residuals means the linear model is wrong, regardless of R².
  10. The exponential distribution is memoryless: P(X > s + t | X > s) = P(X > t) for all s, t ≥ 0. This means a component that has already lasted s hours has exactly the same remaining lifetime distribution as a brand-new component. The exponential is the only continuous distribution with this property. It is the waiting-time distribution for Poisson processes and appears constantly in pharmacokinetics (drug clearance) and radioactive decay.
  11. The t-distribution with ν degrees of freedom has heavier tails than the Normal. When ν is small (small samples), critical values are substantially larger than the Normal z* — t with 5 df and α = 0.05 (two-tailed) gives t* = 2.571 vs. z* = 1.960. As ν → ∞, the t-distribution converges to N(0,1). This is why you must use t-tables (not z) when σ is unknown and n is small, as is routine in chemistry and biology experiments.
  12. Randomization, replication, and blocking are the three foundational principles of experimental design (Box, Hunter & Hunter, Ch. 3–4). Randomization protects against unknown confounders by distributing them evenly across treatments on average. Replication provides an estimate of pure experimental error. Blocking removes known sources of nuisance variation (e.g., day-to-day drift, batch effects) from the error term, increasing power without increasing n.
  13. The chi-square goodness-of-fit test statistic χ² = Σ (Observed − Expected)² / Expected follows a χ² distribution with k − 1 degrees of freedom (where k is the number of categories), provided expected counts are ≥ 5 in each cell. The test of independence in a contingency table uses the same statistic with df = (rows − 1)(columns − 1). From your genetics background: this is the test Mendel's ratios are evaluated against.
  14. Nonparametric tests (Mann-Whitney U, Kruskal-Wallis) replace raw data values with their ranks, making them robust to non-normality and outliers. Mann-Whitney is the rank-based analogue of the two-sample t-test; Kruskal-Wallis is the rank-based analogue of one-way ANOVA. The cost is lower power than parametric tests when the normality assumption actually holds, so they are not automatically preferable — they are appropriate when sample sizes are small and distributions are clearly non-normal.
  15. Two-way ANOVA with an interaction term tests three things simultaneously: main effect of Factor A, main effect of Factor B, and the A × B interaction — whether the effect of A depends on the level of B. From Box, Hunter & Hunter's perspective, detecting interactions is often the most scientifically important result of a factorial experiment, because additive models (no interaction) are the exception rather than the rule in chemistry and biology. A significant interaction term means the main effects cannot be interpreted in isolation.

References

  1. Jay L. Devore, Probability and Statistics for Engineering and the Sciences, 9th edition. Cengage Learning, Boston, 2016. ISBN 978-1-305-25180-9.
  2. George E. P. Box, J. Stuart Hunter, and William G. Hunter, Statistics for Experimenters: Design, Innovation, and Discovery, 2nd edition. Wiley-Interscience, Hoboken, NJ, 2005. ISBN 978-0-471-71813-0.