inference statistics

Inference statistics is the branch of statistics that uses a sample of data to make probability-based statements about the population from which the sample was drawn, and it is the entire mathematical engine that turns the thousands of measurements collected on a semiconductor fab floor into confident engineering decisions. Whereas descriptive statistics summarizes what is in the sample itself, inference statistics extrapolates beyond the sample to the larger truth that the engineer actually cares about, whether that is the true mean of a process, the proportion of defective dies, the difference between two process conditions, or the effect of a parameter change on yield. The fundamental problem of inference is that the sample is never the whole truth: a sample of fifty wafers cannot reveal the exact behavior of the millions of wafers the process will eventually produce, and yet the engineer must act on the sample every day. Inference statistics solves this problem by quantifying exactly how much the sample can and cannot tell, producing not just a point answer but an honest statement of uncertainty that allows a decision to be made with a known probability of error. This document develops the ideas of point estimation, confidence intervals, hypothesis testing, and the model-based methods that form the core of statistical inference, and it shows how each applies to the decisions that run a modern semiconductor factory. **The logic of inference rests on the relationship between three distinct objects: the population, the sample, and the statistic.** The population is the complete set of all units that the engineer wishes to understand, such as every wafer the process has produced or will produce; the sample is the subset actually measured, such as the fifty wafers selected for metrology; and a statistic is any number computed from the sample, such as the sample mean, the sample variance, or the observed defect rate. The central act of inference is to use a statistic, which is known and computable, to learn about a parameter, which is an unknown but fixed characteristic of the population, such as the true mean film thickness or the true proportion of defective dies. The bridge between the statistic and the parameter is the sampling distribution, which describes how the statistic would vary if the sampling were repeated many times, and it is this sampling distribution that carries the uncertainty that the engineer must quantify. **A statistic is itself a random variable, because a different random sample would produce a different value of the statistic, and the distribution of the statistic across all possible samples is called its sampling distribution.** The sampling distribution is the key to everything in inference, because it tells the engineer how much the sample mean, for example, is likely to wander around the true population mean. If the population has mean $\mu$ and variance $\sigma^2$, then the sample mean $\bar{X}$ based on a sample of size $n$ has an expected value of $\mu$ and a variance of $\sigma^2/n$, so that the standard deviation of the sample mean, called the standard error, is $\sigma/\sqrt{n}$. The standard error is the single most important number in inference, because it quantifies how much uncertainty remains after the sample has been taken, and it shrinks as the sample size grows, which is precisely why larger samples support more confident conclusions. Every confidence interval, every hypothesis test, and every estimate of an effect is ultimately an expression of this standard error. **The field of inference statistics rests on two complementary modes of reasoning, and the distinction between them is the intellectual foundation of the discipline.** The first mode, estimation, asks how big the unknown parameter is and how precisely it is known, and it produces a point estimate together with a measure of uncertainty such as a confidence interval. The second mode, testing, asks whether the data are consistent with a particular hypothesis, and it produces a decision about whether an effect is real or merely the result of random variation. Estimation and testing answer different questions: estimation answers how large, while testing answers whether, and both are needed in practice. A process engineer estimating a deposition rate uses estimation to report the rate and its uncertainty, and uses testing to decide whether the rate has drifted from a target specification. The two modes are deeply connected, because a confidence interval and a hypothesis test are two ways of looking at the same body of evidence, and a decision can be read off either one. **The theoretical foundation of inference was laid over more than a century by a sequence of remarkable mathematicians and statisticians, and their names mark the milestones of the discipline.** Carl Friedrich Gauss developed the method of least squares in the early nineteenth century, giving inference its first powerful estimator for linear problems. Jacob Bernoulli provided the law of large numbers that guarantees a sample proportion converges to the population proportion, while Pierre-Simon Laplace connected the binomial to the normal distribution and championed the central limit theorem. William Sealy Gosset, writing under the pseudonym Student, derived the t-distribution in 1908 to handle small samples where the population variance is unknown, and the t-test that bears his name is still the workhorse of comparing two groups. Ronald Fisher built the modern theory of maximum likelihood, analysis of variance, and experimental design, while Jerzy Neyman and Egon Pearson formalized hypothesis testing as a decision problem with control of error rates. Karl Pearson contributed the chi-square test and the correlation coefficient, and William Shewhart, though an engineer rather than a mathematician, founded statistical process control at Bell Labs. This parade of names shows that inference is not a single discovery but a layered body of technique built over generations. **The central limit theorem is the single most important result in all of inference, because it guarantees that the sampling distribution of the sample mean is approximately normal regardless of the shape of the underlying population, provided the sample size is large enough.** Formally, the theorem states that for a random sample of size $n$ from a population with finite mean $\mu$ and variance $\sigma^2$, the standardized sample mean $(\bar{X} - \mu)/(\sigma/\sqrt{n})$ converges in distribution to the standard normal as $n$ grows. The beauty of the theorem is that it does not require the population to be normal: even if the underlying measurements are strongly skewed or discrete, the average of many independent observations becomes bell-shaped. This is why the normal distribution appears everywhere in engineering, why the sample mean can be treated as normal even when the raw data are not, and why the same normal-based formulas can be applied to averages from a wide range of processes. The practical warning is that the theorem is an asymptotic result, so it works well for large samples but can fail badly for small samples from very skewed populations, which is precisely the situation that calls for the t-distribution or for bootstrap methods. **The workhorse method of estimation is the method of maximum likelihood, developed by Ronald Fisher, and it produces estimators with excellent theoretical properties.** The likelihood function $L(\theta)$ is the probability of observing the actual data as a function of the unknown parameter $\theta$, and the maximum likelihood estimator is the value of $\theta$ that makes the observed data most probable. For a normal population the maximum likelihood estimator of the mean is the sample mean, and for a binomial proportion it is the sample proportion, so the method reproduces the intuitive estimators while also giving a principled way to handle far more complex models. Maximum likelihood estimators are consistent, meaning they converge to the true parameter as the sample grows, asymptotically unbiased, and asymptotically efficient, meaning they achieve the smallest possible variance among a wide class of estimators. The method is also the foundation of likelihood-based inference, including likelihood ratio tests and the Wald and score tests, which provide general tools for testing hypotheses in complex models. In a fab, maximum likelihood is the engine behind fitting distribution models to failure-time data, fitting regression models to process outputs, and estimating the parameters of yield models from observed die counts. **A good estimator must satisfy several criteria, and the most important are unbiasedness, consistency, and efficiency, which together ensure that the estimator is accurate, stable, and as precise as possible.** An estimator is unbiased if its expected value equals the true parameter, so that on average it does not systematically over- or under-estimate; the sample mean is unbiased for the population mean, while the sample variance with $n-1$ in the denominator is unbiased for the population variance, which is why the $n-1$ correction appears. An estimator is consistent if it converges to the true parameter as the sample size grows, and an estimator is efficient if it has the smallest variance among all unbiased estimators, a property captured by the Cramer-Rao lower bound. The choice among estimators involves trade-offs between bias and variance, and the bias-variance decomposition of the mean-squared error formalizes this tension: mean-squared error equals variance plus the square of the bias. This decomposition is the organizing principle of all model fitting, because it shows that a slightly biased estimator can sometimes be better than an unbiased one if it is much less variable, which is exactly the insight behind ridge regression and other regularized estimators used in modern data analysis. **A confidence interval converts a point estimate and its standard error into a range that is likely to contain the true parameter, and it is the most honest way to report what a sample has learned.** A 95% confidence interval for a normal population mean with known variance is $\bar{X} \pm 1.96 \cdot \sigma/\sqrt{n}$, where $1.96$ is the standard normal quantile that leaves 2.5% in each tail. The interpretation of the interval is subtle but crucial: a 95% confidence interval means that if the sampling procedure were repeated many times, then 95% of the intervals so constructed would contain the true parameter. It does not mean that there is a 95% probability that the true parameter lies in the specific interval just computed, because the parameter is fixed and the interval is the random object, and this distinction is the source of endless confusion in practice. The width of a confidence interval is proportional to the standard error and therefore to $1/\sqrt{n}$, so that quadrupling the sample size halves the width of the interval, which is the direct quantitative link between sample size and precision that drives metrology budget decisions. **When the population variance is unknown, which is almost always the case in practice, the normal distribution must be replaced by the t-distribution, and this is the source of the t-statistic and the t-test.** The t-statistic for a single sample is $t = (\bar{X} - \mu_0)/(s/\sqrt{n})$, where $s$ is the sample standard deviation, and it follows a t-distribution with $n-1$ degrees of freedom. The t-distribution was derived by William Sealy Gosset, writing as Student, and it is slightly wider than the normal distribution with heavier tails, reflecting the extra uncertainty that comes from having to estimate the population variance from the sample. As the sample size grows, the t-distribution converges to the normal, so the two give nearly identical results for large samples and differ only when samples are small. The degrees of freedom, equal to $n-1$ for a single sample, count the independent pieces of information available after estimating the mean, and they determine how much heavier the tails of the t-distribution are. Every confidence interval and every test that uses the sample standard deviation in place of the population standard deviation is a t-based procedure. **The two-sample t-test is the standard tool for comparing the means of two groups, and it is used constantly in semiconductor engineering to decide whether two process conditions truly differ.** When comparing the means of two independent samples, the t-statistic is the difference of the sample means divided by the standard error of that difference, which is the pooled or separate estimate of the within-group variability. The test decides whether the observed difference is large relative to the sampling variability, and if the difference is large enough the test rejects the null hypothesis that the two means are equal. In a fab the two-sample t-test compares the film thickness from two chambers, the defect rate from two cleaning recipes, or the performance of two process windows, and it is the quantitative answer to the recurring question of whether a change made a real difference. The test assumes that the observations are independent and approximately normal within each group, and the Welch variant of the test relaxes the assumption of equal variances, making it the default choice in modern statistical software. The two-sample t-test is the bridge from the abstract idea of sampling distributions to the concrete decisions of process improvement. **The formal machinery of hypothesis testing was established by Jerzy Neyman and Egon Pearson, and it frames the testing problem as a decision between a null hypothesis and an alternative hypothesis with controlled error rates.** The null hypothesis $H_0$ is the default assumption that no effect exists, such as the claim that a process mean equals its target, while the alternative hypothesis $H_a$ is the claim that an effect exists, such as the claim that the mean has drifted away from target. The test computes a test statistic from the sample, compares it to a reference distribution under the null hypothesis, and rejects $H_0$ if the statistic is sufficiently extreme. The logic is analogous to a criminal trial: the null is presumed true until the evidence against it is strong enough, and the burden of proof falls on the data. Neyman-Pearson theory formalizes this by fixing the probability of the two types of error and choosing the test procedure to control them, which turns hypothesis testing from an informal judgment into a rigorous decision procedure with known operating characteristics. **The two kinds of error in hypothesis testing are named Type I and Type II, and every test is a balancing act between them.** A Type I error, with probability $\alpha$, is rejecting the null hypothesis when it is actually true, which in a fab context is a false alarm that declares a process out of control when it is in fact fine. A Type II error, with probability $\beta$, is failing to reject the null when the alternative is actually true, which is a missed signal that fails to detect a real drift or a real defect problem. The significance level $\alpha$ is the probability of a Type I error that the engineer is willing to tolerate, conventionally set to 0.05 or 0.01, and the power of a test is the probability $1 - \beta$ of correctly detecting an effect when one exists. The engineer cannot make both errors arbitrarily small with a fixed sample size, because reducing $\alpha$ makes the test less sensitive and raises $\beta$, and the only way to reduce both simultaneously is to increase the sample size. A power analysis, performed before an experiment, computes the sample size needed to achieve a desired power for a given effect size and variability, and it is the disciplined answer to the question of how many wafers to run. **The p value is the most widely used and most widely misunderstood number in all of statistics, and it deserves a precise statement of its meaning.** The p value is the probability of obtaining a test statistic at least as extreme as the one actually observed, computed under the assumption that the null hypothesis is true. It is a measure of how surprising the data are under the null, not a probability that the null is true, and it is not a probability that the alternative is true, nor is it the probability that the result is due to chance. A small p value means that the data would be unusual if the null were true, which casts doubt on the null and suggests rejecting it; a p value of 0.03 means that, if the null were true, there is a 3% chance of seeing data this extreme or more. The decision rule is to reject the null when the p value is less than the significance level $\alpha$, and the convention is to call p values below 0.05 statistically significant, though this threshold is arbitrary and is now widely criticized. The p value depends on the sample size, so a large sample can produce a tiny p value for an effect that is trivially small, which is why the p value must always be accompanied by an estimate of the effect size. **The idea of statistical significance must be sharply distinguished from the idea of practical importance, and this distinction is one of the most important lessons of inference for an engineer.** Statistical significance is a statement about whether an effect is distinguishable from zero given the sampling variability, while practical importance is a statement about whether the effect is large enough to matter for the business or the process. With a sufficiently large sample, almost any difference, however small, becomes statistically significant, because the standard error shrinks and the test gains power to detect even trivial effects. Conversely, a genuinely important effect may fail to reach significance in a small sample simply because there is not enough data to detect it. The correct practice is to report the estimated effect size and its confidence interval alongside the p value, so that the reader can judge both whether the effect is real and whether it matters. An engineer who reports only a p value, without the size of the effect and its uncertainty, is reporting at most half of the information that the data contain. **The chi-square test, developed by Karl Pearson, tests hypotheses about categorical data and is the tool for questions about counts and proportions.** The chi-square goodness-of-fit test compares observed counts in categories to the counts expected under a hypothesized distribution, and the chi-square test of independence tests whether two categorical variables are associated in a contingency table. The test statistic is a sum of squared deviations of observed from expected counts, each normalized by the expected count, and it follows a chi-square distribution with degrees of freedom determined by the number of categories or the dimensions of the table. In a fab the chi-square test assesses whether the number of defects in different zones of a wafer differs from a uniform expectation, whether the distribution of failure modes has changed after a process change, or whether two attributes such as wafer zone and defect type are related. The test requires that the expected counts in each cell are not too small, typically at least five, and when they are smaller Fisher's exact test provides an alternative. The chi-square test is the bridge from the normal-based world of means to the categorical world of counts and proportions. **Analysis of variance, universally abbreviated as ANOVA, generalizes the comparison of two means to the comparison of many groups at once, and it was developed by Ronald Fisher for agricultural experiments before becoming a staple of industrial analysis.** The fundamental idea of ANOVA is to partition the total variability in the data into a component due to the differences between groups and a component due to the variation within groups, and then to compare these components with an F-test. If the between-group variation is large relative to the within-group variation, the groups are judged to have different means, and the F-statistic, the ratio of these two variance estimates, follows an F-distribution under the null hypothesis that all group means are equal. ANOVA is the correct method when comparing three or more process conditions, because doing many pairwise t-tests inflates the chance of a false positive, whereas ANOVA tests all groups simultaneously while controlling the error rate. In a fab ANOVA compares the film thickness across several chambers, the yield across multiple lots, or the performance of several process recipes, and when it finds a significant difference, follow-up comparisons identify which specific groups differ. ANOVA is also the gateway to experimental design, because the same decomposition of variance is the engine behind designed experiments and the identification of which factors matter. **The F-test used in ANOVA is an example of a broader family of tests based on the ratio of variances, and the F-distribution that underlies it was named by George Snedecor in honor of Ronald Fisher.** The F-statistic is the ratio of two independent chi-square variables each divided by its degrees of freedom, and its distribution depends on two degrees-of-freedom parameters, one for the numerator and one for the denominator. In ANOVA the F-test asks whether the mean square between groups, which measures how different the groups are, is large relative to the mean square within groups, which measures the background noise. The F-test is also used to compare the variances of two populations, where it tests whether one process is more variable than another, which is directly relevant to judging whether a new process is more consistent than an existing one. The F-test is sensitive to the normality assumption, so it is often paired with tests such as Bartlett's test or Levene's test for equality of variances that are more robust. The F-distribution is the last of the major reference distributions, completing the trio of normal, t, and chi-square that together supply the null distributions for nearly all classical tests. **Regression analysis is the inference method for questions about relationships between a response and one or more explanatory variables, and it is one of the most powerful and most used tools in engineering.** In simple linear regression the model is $Y = \beta_0 + \beta_1 X + \epsilon$, where $Y$ is the response, $X$ is the predictor, $\beta_0$ and $\beta_1$ are the intercept and slope to be estimated, and $\epsilon$ is the random error. The coefficients are estimated by least squares, the method that minimizes the sum of squared residuals and that was introduced by Gauss, and the inference problem is to test whether the slope $\beta_1$ is different from zero, which is a test of whether the predictor actually affects the response. The regression output provides estimates of the coefficients, their standard errors, t-tests for each coefficient, and an R-squared that measures the fraction of variability explained by the model. In a fab regression relates process variables such as deposition temperature or pressure to the resulting film properties, relates metrology readings to physical parameters, and provides the response surfaces used in process optimization. Regression inference is the statistical arm of process modeling, connecting the collected data to predictive equations. **The assumptions that underlie classical inference must be checked before the results can be trusted, and a responsible analysis always inspects the residuals and the model diagnostics.** The assumptions of the t-test and ANOVA include independence of the observations, approximate normality of the errors, and equality of variances across groups, while regression additionally assumes linearity and that the errors are homoscedastic, meaning they have constant variance. Violations of these assumptions can invalidate the p values and confidence intervals, so the engineer checks them with residual plots, normal probability plots, and tests such as Levene's test or the Shapiro-Wilk test for normality. When the assumptions fail, the remedy is either to transform the data, such as taking logarithms to stabilize variance, or to use a robust or nonparametric method that does not rely on the assumptions. Nonparametric methods such as the Mann-Whitney U test, the Wilcoxon signed-rank test, and the Kruskal-Wallis test provide distribution-free alternatives that replace the mean with the median and the normal-based tests with rank-based tests. The discipline of checking assumptions is what separates inference that is trustworthy from inference that merely looks rigorous. **The bootstrap is a resampling method, introduced by Bradley Efron in 1979, that computes standard errors and confidence intervals without assuming a particular distribution, and it is the modern answer to the problem of unknown sampling distributions.** The idea of the bootstrap is to treat the observed sample as a stand-in for the population and to draw many new samples from it with replacement, computing the statistic of interest on each resample to build an empirical sampling distribution. From this bootstrap distribution the engineer reads the standard error, constructs percentile-based confidence intervals, and assesses the stability of almost any estimator, no matter how complicated, without needing an analytic formula for its sampling distribution. The bootstrap is especially valuable for statistics whose sampling distributions are hard to derive, such as medians, quantiles, correlations, and complex model parameters, and it works by the same large-sample logic as the central limit theorem but with the distribution replaced by the empirical distribution of the data. In a fab the bootstrap estimates the uncertainty of yield estimates, the confidence intervals of Cpk indices, and the variability of parameters in fitted models, and it is the default tool when the textbook formulas are not available. **The methods of classical inference can also be contrasted with the Bayesian approach, and the contrast clarifies the philosophical and practical stakes of each framework.** Classical, or frequentist, inference treats the parameter as a fixed unknown and quantifies uncertainty through the long-run behavior of the procedure, such as the confidence level of an interval, whereas Bayesian inference treats the parameter as a random variable and combines a prior with the likelihood to produce a posterior distribution. The two frameworks answer different questions: the frequentist confidence interval is a property of the procedure repeated many times, while the Bayesian credible interval is a direct statement about the parameter given the data and the prior. The choice between them depends on the context, with the frequentist framework dominating regulated and high-stakes industrial settings where error rates must be controlled, and the Bayesian framework offering advantages when prior knowledge is available or when data are scarce. In the series of keywords built for this domain, the probability stats and statistics basics keywords set up the distributions and descriptive tools, while the bayesian statistics keyword develops the Bayesian machinery in full, and inference statistics stands alongside them as the frequentist counterpart. The engineer who masters both frameworks can choose the appropriate tool for each problem and translate between them fluently. **The design of experiments, abbreviated as DOE, is the discipline of deciding how to collect data so that the inference will be efficient and the conclusions valid, and it is the practical interface between inference and engineering action.** The core principles of experimental design, established by Ronald Fisher, are randomization, replication, and blocking, which together ensure that the data are collected without bias and with enough precision to detect real effects. A factorial design varies several factors simultaneously in a structured pattern, so that the main effects and interactions of the factors can be estimated with far fewer runs than testing each factor one at a time. The results of a designed experiment are analyzed with ANOVA and regression, linking the design directly to the inference methods of this document, and the design determines the power of the analysis to detect effects of a given size. In a fab, designed experiments optimize process windows, characterize the effect of temperature, pressure, and flow on film properties, and identify the few factors that dominate yield, and the sample-size and power calculations of the design are the same ones developed earlier in this document. The full treatment of DOE as a keyword in its own right belongs to a separate topic, but its logic is inseparable from the inference that interprets its results. **The practice of inference in a fab follows a disciplined sequence of decisions, and the sequence is best stated as a checklist that an engineer can apply to any measurement problem.** First, define the population of interest and the parameter to be estimated or tested, whether it is a mean, a proportion, a difference, or a relationship. Second, collect a random sample of adequate size, using a power analysis to justify the number of units measured. Third, compute the relevant statistic, its standard error, and its confidence interval, and check the assumptions of the method against the data. Fourth, if a decision is required, perform the appropriate test, report the p value, and interpret it in light of the estimated effect size and its practical importance. Fifth, document the result with both the estimate and its uncertainty, so that the conclusion can be evaluated and reproduced. Each step has its own pitfalls, and the checklist is a guard against the most common errors, from sampling that is not random to reporting a p value without its effect size. **A comparison of the major inference tools makes their roles concrete, and the following table organizes them by the question they answer, the type of data they use, and the typical engineering decision they support.** The table condenses the methods developed throughout this document into a single reference, so that an engineer facing a new question can locate the right tool quickly. The choice among tools is driven first by the type of data, whether continuous or categorical, and second by the structure of the question, whether it involves one group, two groups, many groups, or a relationship between variables. | Inference tool | Question answered | Data type | Typical engineering decision | |---|---|---|---| | One-sample t-test | Is the mean different from a target? | continuous, one group | is a chamber mean on target? | | Two-sample t-test | Do two group means differ? | continuous, two groups | did a recipe change move the mean? | | Paired t-test | Do paired before/after measurements differ? | continuous, paired | did a clean reduce defect counts? | | ANOVA + F-test | Do several group means differ? | continuous, many groups | do multiple chambers run alike? | | Chi-square test | Are counts or proportions independent? | categorical | is defect type related to wafer zone? | | Correlation | Is there a linear association? | continuous, two vars | do two metrology tools track each other? | | Simple regression | Does X predict Y and by how much? | continuous | does temperature drive thickness? | | Bootstrap | Standard error and CI for any statistic | any | uncertainty of Cpk or yield estimate | | Mann-Whitney U | Do two distributions differ (median)? | continuous, nonnormal | robust two-group comparison | | Power analysis | What sample size do I need? | design | how many wafers to run the study? | **The choice of a statistical test is governed by a clear decision tree, and the following flowchart routes an inference question to the correct method based on the type of data and the structure of the problem.** The first question is whether the response is continuous or categorical; the second is how many groups are being compared; and the third is whether the data meet the assumptions of the normal-based methods. Working through these questions selects the appropriate test, and each branch leads to the methods whose formulas and interpretations were developed in this document. ```flowchart A([Inference question]) --> B{Response type?} B -- categorical --> C[Chi-square / proportions test] B -- continuous --> D{How many groups?} D -- one group --> E{Compare to target?} E -- yes --> F[t-test for a single mean] E -- no --> G[Confidence interval for mean] D -- two groups --> H{Independent or paired?} H -- independent --> I[Two-sample t-test / Welch] H -- paired --> J[Paired t-test] D -- three or more --> K[ANOVA + F-test] K --> L{Significant?} L -- yes --> M[Multiple comparisons] L -- no --> N[Conclude groups similar] D -- relationship --> O{Linear model?} O -- yes --> P[Regression + t-test on slope] O -- no --> Q[Correlation / rank methods] ``` **The calculation of the sample size needed for a desired level of precision or power is one of the most practical skills in inference, and it illustrates the interplay between the quantities developed throughout this document.** For estimating a mean with a desired margin of error $E$ at a given confidence level, the required sample size is $n = (z_{\alpha/2} \cdot \sigma / E)^2$, which shows that the sample size grows with the square of the ratio of the standard deviation to the margin of error. For a two-sample test with a target power, the required sample size depends on the effect size, the variability, and the significance level, and it can be solved from the power formula. The message is that precision and power both scale like the square of the ratio of the signal to the noise, so that halving the desired margin of error requires quadrupling the sample size, while reducing the variability of the process is often a more economical path to precision than measuring more wafers. Every metrology budget in a fab is ultimately an answer to this sample-size question, and the formulas convert the engineering goal into a concrete count of measurements. **The historical and conceptual arc of inference statistics is complete when the reader sees how each method answers a fundamental human question about uncertainty, and the names of the field are the landmarks of that arc.** Gauss gave least squares, Bernoulli gave the law of large numbers, Laplace championed the central limit theorem, Gosset gave the t-distribution for small samples, Fisher gave maximum likelihood, ANOVA, and experimental design, Neyman and Pearson gave the decision-theoretic framework of hypothesis testing, and Efron gave the bootstrap for the computer age. The arc shows that inference is cumulative, with each generation building on the last, and that the modern engineer stands on a body of technique refined over two centuries. The central thread running through all of it is the standard error, the quantification of sampling uncertainty, and the recognition that every estimate is incomplete without its uncertainty and every decision is incomplete without its error rate. Mastering inference statistics means internalizing this thread: the sample is never the whole truth, but with the right methods it is enough to act on with confidence. **The connection between inference statistics and the other keywords in the series is direct and worth making explicit, because the series has been designed so that each topic builds on the previous ones.** The probability stats keyword supplies the probability distributions and expectations that define the sampling distributions, the statistics basics keyword supplies the descriptive statistics and the concept of estimation that inference refines, the bayesian statistics keyword offers the alternative framework that contrasts with the frequentist methods here, and the stochastic processes keyword supplies the temporal models into which these inference tools are applied. Inference statistics, in turn, is the discipline that ties data to decisions across all of them, providing the confidence intervals, the hypothesis tests, and the model-based estimates that turn raw measurements into engineering action. The engineer who moves through the series from probability through inference acquires a complete toolkit: the language of probability, the discipline of statistics, the power of Bayesian reasoning, the modeling of random processes, and the machinery of inference that unifies them into decisions. **The practice of inference in a real fab is illustrated by a concrete example that ties the methods together, and the example of deciding whether a new etch recipe reduces defect density shows how every tool in this document contributes.** The engineer defines the population as all wafers the new recipe would produce, collects a random sample of wafers from the new and old recipes, and computes the mean defect count and its standard error for each. A two-sample t-test, or its Welch variant, tests whether the observed difference is real, while a confidence interval for the difference reports its size and precision, and a power analysis confirms that the sample was large enough to detect an effect of practical interest. The assumptions are checked by inspecting the defect-count distributions, and if they are heavily skewed the engineer falls back to a nonparametric test or a bootstrap interval. The final report states the estimated reduction, its confidence interval, and the p value, so that the decision to adopt the recipe is made with both the evidence and its uncertainty in view. This single example demonstrates that inference is not a collection of isolated formulas but a coherent procedure for turning data into decisions. **The closing lens for inference statistics is that it is the discipline of making decisions under uncertainty, and the reader should approach the subject through that lens.** The value of the subject is not the individual equation for the t-statistic or the confidence interval, but the recognition that every estimate is incomplete without its uncertainty and every decision is incomplete without its error rate. With this lens the engineer sees the p value not as a magical verdict but as a measure of surprise under a null hypothesis, sees the confidence interval not as a probability statement but as a property of a procedure, and sees the standard error as the honest currency of all sampling. The mastery of inference statistics is the mastery of acting on partial information with known, quantified risk, which is precisely what a semiconductor engineer must do every day on a noisy fab floor. Read inference statistics through a decision-and-uncertainty lens rather than a formula-recitation lens.