Mathematical Statistics Data Analysis Chapter 4 Solutions

Getting Through Chapter 4 of Mathematical Statistics and Data Analysis

Chapter 4 in John P. Howe's book covers estimation theory, hypothesis testing, and confidence intervals. It is where the course goes from probability theory into actual statistical inference. Most students treat it like a collection of formulas. It is not. The material rewards people who spend time understanding what each estimator is actually doing rather than memorizing which formula applies to which problem type. The core topics break down into several areas. Point estimation comes first, and that means learning about method of moments estimators, maximum likelihood estimators, and the properties that separate good estimators from bad ones. Bias, efficiency, and consistency are not optional vocabulary. You need all three terms when the exam asks you to compare two estimators for the same parameter. Interval estimation follows, which introduces confidence intervals for means, variances, and proportions under different distributional assumptions. Then hypothesis testing covers the likelihood ratio framework, Neyman-Pearson lemma, uniformly most powerful tests, and p-value interpretation. The last section usually handles goodness-of-fit tests with the chi-squared statistic and sometimes an introduction to nonparametric methods. I worked through a version of this chapter while tutoring several undergraduates last semester. One student kept losing points on a problem involving the Fisher information matrix for a shifted exponential distribution. The distribution itself is straightforward, but the support depends on the parameter, which violates the regularity conditions that most textbook examples quietly assume. She tried applying the standard Cramér-Rao lower bound formula blindly and got a result that was mathematically inconsistent with the problem constraints. The fix was recognizing that the MLE for the shift parameter in a shifted exponential is actually the minimum order statistic, which has a completely different variance structure. The Fisher information approach simply does not apply there. That single realization cleared up about half her confusion on the problem set.

Maximum Likelihood Estimation: What Actually Matters

Maximum likelihood estimation is the workhorse of Chapter 4, and the procedure is mechanically simple. Write the likelihood function, take the log, differentiate, set equal to zero, solve. The mechanical part takes ten minutes. The understanding part takes longer. A lot of students write down the likelihood and immediately start differentiating without checking whether the likelihood is actually well-defined over the parameter space. That skips a critical step. For discrete distributions like the geometric or negative binomial, the MLE usually follows the same pattern as continuous cases, but boundary issues appear more frequently. When your parameter appears in the support of the distribution, like the uniform distribution on [0, theta], differentiation will not give you the answer. The likelihood is a step function in theta, and the maximum occurs at the boundary, specifically at the largest observed value. I see this mistake repeatedly. Students differentiate the log-likelihood, get zero, and conclude the MLE does not exist or that they made an algebra error. Neither is true. The MLE exists and equals the maximum observation. Another thing textbooks underplay is the difference between the score equation having a solution and that solution actually being a maximum. The second derivative test matters, and so does checking the behavior at the boundaries of the parameter space. In practice, most exam problems are constructed so the critical point is the global maximum, but not all of them are. A few problems in the end-of-chapter section set traps this way.

Method of Moments: Don't Dismiss It

Method of moments estimators get dismissed as inferior to MLEs, and for large samples they usually are. But method of moments still shows up on exams, and sometimes it is the only tractable approach. The procedure is mechanical. Compute the theoretical moments in terms of the unknown parameters, compute the sample moments, set them equal, and solve the resulting system of equations. For a distribution with two parameters, you need the first and second moments. For three parameters, you need up to the third moment. The real question is when to use which method. MLEs are generally more efficient when the model is correctly specified. Method of moments estimators are more robust to model misspecification because they only depend on the first few moments rather than the full likelihood structure. If you are in a course where the grading is straightforward calculation-based, know both methods cold. The exam will almost certainly ask you to derive both and compare their properties.

Hypothesis Testing: The Neyman-Pearson Framework

The Neyman-Pearson lemma is the most important theoretical result in this chapter for constructing optimal tests. It states that for testing a simple null hypothesis against a simple alternative hypothesis, the likelihood ratio test is the uniformly most powerful test at any given significance level. The proof is not difficult, but understanding when the lemma applies and when it does not apply is where students struggle. The lemma requires both hypotheses to be simple. Once you introduce composite hypotheses, where the parameter can take any value in a range, you cannot directly apply the lemma. That is when you move toward uniformly most powerful tests, which rely on the monotone likelihood ratio property. Distributions with MLR, including the exponential family, admit UMP tests for one-sided hypotheses. This is a fact worth memorizing because it saves time on exams. One pitfall I want to highlight involves the interpretation of p-values. A p-value is not the probability that the null hypothesis is true. It is the probability of observing data as extreme or more extreme than what you actually observed, assuming the null hypothesis is true. Confusing these two concepts leads to serious errors in reasoning, and professors love to test that confusion on written exams. A p-value of 0.03 does not mean there is a 3 percent chance the null is correct. It means that if the null were correct, data this extreme would occur roughly 3 percent of the time. The distinction matters for things like Bayesian updating and for understanding the replication crisis in applied statistics.

Confidence Intervals: Beyond the Standard Formula

Confidence intervals for the population mean follow the familiar t-interval when the variance is unknown and the data are approximately normal. The formula is standard, but the conditions under which it remains valid are where students lose points. The t-distribution result is exact only for normal data. For non-normal data, the interval is approximately correct by the central limit theorem, and the approximation improves with sample size. A sample of twenty from a heavily skewed distribution will give you a confidence interval that does not actually have the nominal coverage probability. This is worth testing empirically with simulation before you trust the formula on a real dataset. For the difference between two means, the pooled variance approach assumes equal population variances. If that assumption does not hold, the unpooled Welch-Satterthwaite approach is safer. The degrees of freedom adjustment is messy by hand but straightforward in practice. Most courses accept the unpooled version as the default recommendation now because the equal variance assumption is rarely justified in applied work. Confidence intervals for variances use the chi-squared distribution, but that result is sensitive to non-normality in a way that the t-interval for means is not. Even moderate skewness can destroy the coverage accuracy of a chi-squared-based interval for the variance. If your data are not approximately normal and you need an interval for the variance, a bootstrap approach is more reliable than the textbook formula. I learned this the hard way when a colleague used the chi-squared interval on a dataset with moderate right skew and reported a confidence interval that missed the true variance in nearly half of repeated samples instead of the expected five percent.

Chi-Squared Goodness-of-Fit: Common Errors

The chi-squared goodness-of-fit test compares observed frequencies to expected frequencies under a hypothesized distribution. The test statistic follows a chi-squared distribution asymptotically, and the degrees of freedom equal the number of categories minus one minus the number of estimated parameters. The "-1" accounts for the constraint that probabilities sum to one, and the subtraction for estimated parameters accounts for the loss of degrees of freedom when you estimate parameters from the data rather than specifying them a priori. A frequent mistake is forgetting to subtract the number of estimated parameters from the degrees of freedom. If you estimate the mean and variance of a normal distribution from the data before performing the goodness-of-fit test, you lose two degrees of freedom. Skipping this adjustment inflates the test statistic's critical value and can lead to incorrect rejections. Another common error is having expected cell counts below five in too many categories. The chi-squared approximation breaks down when expected frequencies are small, and the recommended workaround is to combine adjacent categories until each expected count is at least five. This reduces the resolution of the test but restores validity.

Working Through the Problems Efficiently

If you are looking for Mathematical Statistics Data Analysis Chapter 4 Solutions to check your work, the most useful approach is to attempt each problem independently first and then compare your derivation steps, not just the final answer. Most online solution manuals give the final result but skip the algebraic manipulation, and that is exactly where the learning happens. A solution that shows every step from the likelihood function through the differentiation to the final expression is worth far more than a bare answer key. When checking your work against a solution, pay attention to whether the estimator or test they present is unique or if multiple valid approaches exist. Some problems in this chapter admit both a method of moments and a maximum likelihood solution, and both can be correct depending on the instructions. The exam will usually specify which method to use, but being able to derive both gives you flexibility if the problem wording is ambiguous. The chapter also tends to include problems involving the relationship between sufficient statistics and completeness. Recognizing a sufficient statistic early in a problem can simplify the entire solution. The factorization theorem is the tool you reach for, and it applies directly. Once you identify the sufficient statistic, you can often reduce a multi-dimensional problem to a one-dimensional one, which makes the rest of the derivation much faster. This shortcut does not always work, but when it does, it cuts the work in half.

Limitations of This Chapter's Approach

The main weakness of this chapter in Howe's book is that it presents estimation and testing within a frequentist framework without extensive discussion of Bayesian alternatives. That is fine for an introductory mathematical statistics course, but it leaves a gap if you plan to apply these methods in research. Bayesian credible intervals, for instance, do not require the normality assumptions that make confidence intervals fragile, and they handle small sample variance estimation more gracefully. If you find the frequentist coverage properties unsatisfactory for your application, the Bayesian route is the natural next step. Another practical limitation is that the textbook problems assume clean, well-behaved data. Real data rarely cooperate. Missing values, outliers, and censored observations all appear in actual applications but are barely addressed in this chapter. If you are using this material for a project or research, you will need supplementary resources for those scenarios. The theoretical foundation here is solid, but the bridge to messy real-world data requires additional learning. The computational methods section is also thin. Modern statistical practice relies heavily on bootstrap and simulation-based inference, and while the bootstrap is mentioned, it is not developed in depth. For confidence intervals where the analytic form is unknown or intractable, the percentile bootstrap is often the simplest viable alternative, and it is worth learning alongside the theoretical material in this chapter.