In the four weeks that began with week 50 of 2016, Wyoming recorded 62 deaths from cancer, a rate of 138.8 per 100,000 person-years. Over the rest of the six years from 2014 to 2019 its rate was 152.9. California's four-week rate, from 4,703 deaths, was 156.4, against 152.1 for the rest of the six years. Both rates are unbiased, but Wyoming's rests on 62 deaths and California's on 4,703, so Wyoming's is far more at the mercy of chance. The 51 rates (the 50 states and the District of Columbia) estimated from the same window are thus not equally trustworthy. Should each area's rate be estimated from its own deaths alone, or can the other 50 areas help estimate Wyoming's, and by how much?
A hierarchical model answers by treating the 51 true rates as draws from a common distribution, whose center and spread are themselves estimated from all the data. Each area's estimate then becomes a weighted average of its own rate and the center, with a weight on the center that is large when the area's own data are thin and small when they are plentiful. This is partial pooling: between estimating every area separately (no pooling) and using one rate for all (complete pooling). For Wyoming in that window the weight on the center is 0.329, and its estimate moves from 138.8 to 157.2; California's weight is 0.007, and its estimate hardly moves.
First we set out the normal hierarchical model and derive its posterior means, its shrinkage factors and how much they reduce the error when the model is right. Then we estimate the common distribution from the data (empirical Bayes) and connect the result to Tweedie's formula and the winner's curse of Selective Inference. Next we fit the gamma-Poisson version to the state death rates and measure, over 30 seeded windows, when shrinkage helps. Then we shrink hypertension rates in 36 NHANES cells toward one mean and toward age-group means. Finally we prove the James-Stein theorem: shrinkage beats separate estimates in total squared error for every set of true means, with no prior at all, as soon as there are three or more. The article assumes Priors, Posteriors and Conjugate Families for the conjugate posteriors and Sampling Distributions from the Normal for chi-squared distributions of projections.
Data. The death counts are the National Center for Health Statistics (NCHS) weekly counts of deaths from malignant neoplasms by jurisdiction of occurrence, 2014 to 2019 (313 weeks), as used in Overdispersion in Count Data; New York and New York City are added together, and the populations are the Census Bureau's Vintage 2019 estimates for July 1 of each year, so a week contributes the year's population times $7/365.25$ person-years. NCHS suppresses weekly counts from 1 to 9; Alaska has 3 such weeks and Wyoming 8. No window used below contains one, and in the six-year rates each is counted as 5 deaths, which changes Wyoming's six-year rate by at most 0.6% whichever value from 1 to 9 is used. No count of 9 or fewer is shown. Rates are crude: they are not adjusted for age, so they differ between areas partly because populations differ in age (from 106.0 per 100,000 in Utah to 254.6 in West Virginia, with 184.6 for all 51 areas together). The deaths are counted where they occurred and the populations where people live, a small mismatch for areas where many people die in a neighboring state. The hypertension data are the 5,182 NHANES 2017 to 2018 adults of Comparing Two Proportions. Every number is computed by tools/experiments/statistics/hierarchical-models-and-shrinkage.py.
Three ways to estimate many rates
Suppose $N$ groups each have a quantity $\theta_i$ to estimate, and an estimate $y_i$ from their own data. Gelman et al. (2013, §5.4) set out the choices.
- No pooling: estimate $\theta_i$ by $y_i$. Each estimate is unbiased, but those from small groups are noisy.
- Complete pooling: estimate every $\theta_i$ by one pooled value. The noise is small, but real differences between groups are erased.
- Partial pooling: estimate $\theta_i$ by a weighted average $\lambda_iy_i + (1 - \lambda_i)\bar y$ with $\lambda_i$ between 0 and 1.
A classical approach chooses between the first two with a test of equal means; Gelman et al. note that "we are not forced to choose between complete pooling and none at all." They show that the three choices are the posterior means under three priors: independent flat priors give no pooling, a single common value gives complete pooling, and independent normal priors with a common mean and variance give partial pooling, which contains the other two as its limits (infinite and zero prior variance).
On the state rates the two extremes are far apart. In the window of the opening, the total squared error of the 51 raw four-week rates against the six-year rates is 8,751 (in squared deaths per 100,000 person-years), and that of the pooled rate of the window, 187.9, given to every area, is 40,798. The partial pooling below gives 7,382. The question is how to choose the weights.
The normal hierarchical model
Here is the model in one sentence: each group's estimate is its true value plus normal noise of known variance, and the true values are themselves drawn from one normal distribution.
- Sampling level. $y_i \mid \theta_i$ is normal with mean $\theta_i$ and known variance $\sigma_i^2$, independently for $i = 1, \ldots, N$.
- Population level. $\theta_1, \ldots, \theta_N$ are independent draws from the normal distribution with mean $\mu$ and variance $\tau^2$.
The hyperparameters $\mu$ and $\tau^2$ describe the population of groups. Define the shrinkage factor
$$B_i = \frac{\sigma_i^2}{\sigma_i^2 + \tau^2},$$
the share of the total variance of $y_i$ that is noise.
Theorem 1 (posterior and marginal distributions). Given $\mu$ and $\tau^2 > 0$:
- the posterior distribution of $\theta_i$ given all the data is normal with mean and variance $$E[\theta_i \mid \mathbf{y}] = (1 - B_i)\,y_i + B_i\,\mu, \qquad \operatorname{Var}(\theta_i \mid \mathbf{y}) = (1 - B_i)\,\sigma_i^2,$$ and the $\theta_i$ are independent given $\mathbf{y}$;
- marginally, the $y_i$ are independent and normal with mean $\mu$ and variance $\sigma_i^2 + \tau^2$.
Proof. (1) The pairs $(\theta_i, y_i)$ are independent across groups, so the posterior factors and $\theta_i$ depends on $y_i$ alone. For one group, the prior on $\theta_i$ is normal with mean $\mu$ and variance $\tau^2$ and the likelihood is that of one normal observation with variance $\sigma_i^2$. By the normal-normal update of Priors, Posteriors and Conjugate Families, the posterior precision is the sum of the precisions, $1/\tau^2 + 1/\sigma_i^2 = (\sigma_i^2 + \tau^2)/(\sigma_i^2\tau^2)$, so the posterior variance is $\sigma_i^2\tau^2/(\sigma_i^2 + \tau^2) = (1 - B_i)\sigma_i^2$. The posterior mean is the precision-weighted average of $y_i$ and $\mu$: $$\frac{y_i/\sigma_i^2 + \mu/\tau^2}{1/\sigma_i^2 + 1/\tau^2} = \frac{\tau^2}{\sigma_i^2 + \tau^2}\,y_i + \frac{\sigma_i^2}{\sigma_i^2 + \tau^2}\,\mu = (1 - B_i)y_i + B_i\mu.$$ (2) Write $y_i = \theta_i + (y_i - \theta_i)$, where $y_i - \theta_i$ is normal with mean 0 and variance $\sigma_i^2$ whatever $\theta_i$ is, hence independent of $\theta_i$. A sum of independent normal variables is normal with the means and variances added, as in Corollary 1 of Bayes Factors and Posterior Model Probabilities; independence across groups is inherited from the pairs. $\square$
Each estimate is pulled toward $\mu$ by the fraction $B_i$. A group with a small sampling variance ($\sigma_i^2 \ll \tau^2$) keeps nearly its own estimate; a group whose noise dominates ($\sigma_i^2 \gg \tau^2$) is pulled nearly to $\mu$. The amount of pooling is decided by comparing the noise in each group with the spread between groups.
How much does it help? Over the joint distribution of $\theta_i$ and $y_i$ (the Bayes risk), the expected squared error of the posterior mean is the posterior variance averaged over the data, $(1 - B_i)\sigma_i^2$, against $\sigma_i^2$ for $y_i$. That holds because $E[(\theta_i - E[\theta_i \mid \mathbf{y}])^2] = E[\operatorname{Var}(\theta_i \mid \mathbf{y})]$, and here the posterior variance does not depend on the data. The posterior mean is also the best of all estimates in this sense, since the conditional mean minimizes expected squared error (Conditional Expectation and the Laws of Total Expectation and Variance). The gain is the factor $1 - B_i$: largest for the noisiest groups.
Example (a pulled estimate can be worse). Shrinkage is not a promise for each group. In the window of the opening, Alaska's raw rate was 146.1, its six-year rate 123.3, and the estimate below moves it to 159.7, further from the six-year rate. The model's claim is about the average squared error over the groups (or over the population of groups), and a group whose true rate is far from the center can be made worse. Efron and Hastie (2016, ch. 7) describe shrinkage as introducing deliberate biases to improve overall performance, "at a possible danger to individual estimates".
Empirical Bayes: estimating the population of groups
Theorem 1 needs $\mu$ and $\tau^2$. They describe the population of groups, and the $N$ groups are a sample from it, so the data can estimate them. Here is the idea: by Theorem 1(2), the estimates $y_i$ are a sample from a normal distribution with mean $\mu$ and variances $\sigma_i^2 + \tau^2$, so $\mu$ and $\tau^2$ can be estimated by maximum likelihood on that marginal distribution, and then plugged into Theorem 1(1).
Proposition 1 (estimating the hyperparameters). The marginal log-likelihood is
$$\ell(\mu, \tau^2) = -\frac12\sum_{i=1}^N\Big[\ln(\sigma_i^2 + \tau^2) + \frac{(y_i - \mu)^2}{\sigma_i^2 + \tau^2}\Big] + \text{const}.$$
For fixed $\tau^2$ it is maximized by the precision-weighted mean $\hat\mu(\tau^2) = \sum_i w_iy_i/\sum_i w_i$ with $w_i = 1/(\sigma_i^2 + \tau^2)$.
Proof. The expression is the logarithm of the product of the normal densities of Theorem 1(2). In $\mu$ it is a downward parabola, $-\frac12\sum_i w_i(y_i - \mu)^2$ plus terms free of $\mu$; its derivative $\sum_i w_i(y_i - \mu)$ vanishes at the weighted mean. $\square$
Substituting $\hat\mu(\tau^2)$ leaves a function of $\tau^2$ alone, which is maximized numerically over $\tau^2 \ge 0$ (the maximum can be at $\tau^2 = 0$, which is complete pooling). Plugging the estimates into Theorem 1 gives the empirical Bayes estimates
$$\hat\theta_i = (1 - \hat B_i)\,y_i + \hat B_i\,\hat\mu, \qquad \hat B_i = \frac{\sigma_i^2}{\sigma_i^2 + \hat\tau^2}.$$
Robbins (1956) introduced the term, as Efron and Hastie (2016, §6.5) report, and the approach was justified at first by frequentist arguments (Efron and Hastie, 2016, §7.5). Gelman et al. (2013, §5.1) prefer to avoid the name, because it "misleadingly suggests" that a full Bayesian analysis is not empirical. They regard empirical Bayes as an approximation to the full hierarchical analysis, which puts a prior on $(\mu, \tau)$ and averages over its posterior instead of plugging in a point estimate. Empirical Bayes ignores the uncertainty in $\hat\tau^2$, which matters most when $N$ is small or $\hat\tau^2$ is near 0. The full analysis is usually computed by sampling, as in the Gibbs sampler of Markov Chain Monte Carlo.
Equal variances: an unbiased shrinkage factor. When every $\sigma_i^2 = 1$, set $A = \tau^2$, so that $B = 1/(1 + A)$. Efron and Hastie (2016, §7.1) estimate $\mu$ by $\bar y$ and the shrinkage factor by $(N - 3)/S$, with $S = \sum_i(y_i - \bar y)^2$, and note that it is unbiased when $N > 3$. (They write $B$ for the weight on the data, $A/(A + 1)$, the complement of ours.)
Lemma 1 (the mean of a reciprocal chi-squared variable). If $X$ has the chi-squared distribution with $\nu > 2$ degrees of freedom, $E[1/X] = 1/(\nu - 2)$.
Proof. The chi-squared density is $x^{\nu/2 - 1}e^{-x/2}/(2^{\nu/2}\Gamma(\nu/2))$ for $x > 0$ (The Chi-Squared and t Distributions). Then $$E[1/X] = \int_0^\infty \frac{x^{\nu/2 - 2}e^{-x/2}}{2^{\nu/2}\Gamma(\nu/2)}\,dx = \frac{2^{\nu/2 - 1}\,\Gamma(\nu/2 - 1)}{2^{\nu/2}\,\Gamma(\nu/2)} = \frac{1}{2(\nu/2 - 1)} = \frac{1}{\nu - 2},$$ using the gamma integral $\int_0^\infty x^{c - 1}e^{-x/2}\,dx = 2^c\Gamma(c)$ with $c = \nu/2 - 1 > 0$, and $\Gamma(c + 1) = c\,\Gamma(c)$. For $\nu \le 2$ the integral diverges at 0. $\square$
Proposition 2 ($(N - 3)/S$ is unbiased). If $y_1, \ldots, y_N$ are independent normal with mean $\mu$ and variance $1 + A$, and $N \ge 4$, then $E[(N - 3)/S] = 1/(1 + A) = B$.
Proof. $S/(1 + A)$ is chi-squared with $N - 1$ degrees of freedom, by the theorem on normal samples of The Chi-Squared and t Distributions. By Lemma 1 with $\nu = N - 1 > 2$, $E[(1 + A)/S] = 1/(N - 3)$; multiply by $(N - 3)/(1 + A)$. $\square$
Plugging this into Theorem 1 gives
$$\hat\theta_i^{\mathrm{JS}} = \bar y + \Big(1 - \frac{N - 3}{S}\Big)(y_i - \bar y),$$
which is the James-Stein estimator toward the mean. In a series of papers in the early 1970s Efron and Morris emphasized this empirical Bayes motivation of the rule (Efron and Morris, 1973; Efron and Hastie, 2016, §7.5). Its section below proves that it beats $y_i$ without any assumption about where the $\theta_i$ came from. Efron and Hastie (2016, eq. 7.14) give its Bayes risk under the hierarchical model as $NB' + 3(1 - B')$ with their $B' = A/(A + 1)$, against $NB'$ for the Bayes rule with known hyperparameters and $N$ for $y$. For $N = 20$ and $A = 1$ that is 11.5, 10 and 20. The script simulated it: 30 seeds (1,001 to 1,030), each 2,000 draws of 20 means and their estimates, gave a median total squared error of 11.50 (interquartile range 11.43 to 11.56).
Tweedie's formula gives the same rule. Selective Inference proves Tweedie's formula: if $\theta$ has any prior and $y \mid \theta$ is normal with mean $\theta$ and variance $\sigma^2$, with marginal density $f$, then $E[\theta \mid y] = y + \sigma^2\frac{d}{dy}\ln f(y)$. When the marginal is normal with mean $\mu$ and variance $\sigma^2 + \tau^2$, as in Theorem 1(2), $\frac{d}{dy}\ln f(y) = -(y - \mu)/(\sigma^2 + \tau^2)$, and the formula gives $y - B(y - \mu) = (1 - B)y + B\mu$: Theorem 1(1). Tweedie's formula is the general empirical Bayes rule, of which linear shrinkage is the normal case. When the marginal is estimated without assuming it is normal, the correction can be nonlinear. In Efron and Hastie's prostate example (2016, §20.3) it shrinks the values near 0 almost to 0, but the most extreme, $z = 5.29$, only to 4.03; the straight-line rule of the normal prior, with slope 0.22, would have taken it to about 1.2.
The winner's curse. Theorem 1 of Selective Inference proves that the largest of several unbiased estimates overstates its target on average. Shrinkage is a correction: the group with the largest $y_i$ is usually one whose error is large and positive, and the posterior mean pulls it back. Efron and Hastie (2016, §20.3) note that Bayes estimators are immune to selection bias, which offers some hope that empirical Bayes estimates might cure the winner's curse. The state rates below measure how much of the curse an estimated prior removes.
Death rates by state: the gamma-Poisson model
Deaths are counts, and a count model avoids the normal approximation. Let $d_i$ be the deaths in area $i$ during the window and $t_i$ its person-years. The two levels are:
- $d_i \mid \lambda_i$ is Poisson with mean $t_i\lambda_i$, independently, where $\lambda_i$ is the area's death rate;
- $\lambda_1, \ldots, \lambda_N$ are independent draws from Gamma$(a, b)$, with shape $a$, rate $b$, mean $a/b$ and variance $a/b^2$.
Proposition 3 (gamma-Poisson shrinkage). The posterior of $\lambda_i$ is Gamma$(a + d_i, b + t_i)$, and its mean is
$$E[\lambda_i \mid \mathbf{d}] = \frac{a + d_i}{b + t_i} = (1 - B_i)\frac{d_i}{t_i} + B_i\frac{a}{b}, \qquad B_i = \frac{b}{b + t_i}.$$
Marginally, $d_i$ is negative binomial,
$$P(d_i = d) = \frac{\Gamma(a + d)}{d!\,\Gamma(a)}\Big(\frac{b}{b + t_i}\Big)^{a}\Big(\frac{t_i}{b + t_i}\Big)^{d}, \qquad d = 0, 1, \ldots$$
Proof. The likelihood of $\lambda_i$ is proportional to $\lambda_i^{d_i}e^{-t_i\lambda_i}$, and the prior density to $\lambda_i^{a - 1}e^{-b\lambda_i}$, so the posterior is proportional to $\lambda_i^{a + d_i - 1}e^{-(b + t_i)\lambda_i}$: Gamma$(a + d_i, b + t_i)$, the gamma-Poisson update of Priors, Posteriors and Conjugate Families with exposure $t_i$. Its mean is $(a + d_i)/(b + t_i)$, and $$\frac{a + d_i}{b + t_i} = \frac{t_i}{b + t_i}\cdot\frac{d_i}{t_i} + \frac{b}{b + t_i}\cdot\frac{a}{b}.$$ For the marginal, apply Theorem 2 of Bayes Factors and Posterior Model Probabilities: $P(d_i = d) = f(d \mid \lambda)\pi(\lambda)/\pi(\lambda \mid d)$ at any $\lambda > 0$, with $f(d \mid \lambda) = (t_i\lambda)^de^{-t_i\lambda}/d!$, $\pi(\lambda) = b^a\lambda^{a - 1}e^{-b\lambda}/\Gamma(a)$ and $\pi(\lambda \mid d) = (b + t_i)^{a + d}\lambda^{a + d - 1}e^{-(b + t_i)\lambda}/\Gamma(a + d)$. The powers of $\lambda$ and the exponentials cancel, leaving $t_i^db^a\Gamma(a + d)/(d!\,\Gamma(a)(b + t_i)^{a + d})$, which is the formula. $\square$
The hyperparameter $b$ has a direct reading: it is the number of person-years of data that the prior is worth. An area with $t_i = b$ person-years in the window is pulled halfway to the prior mean. Overdispersion in Count Data derives the same negative binomial as a Poisson count whose rate varies by a gamma distribution, which is exactly the population level here.
Empirical Bayes fit. The script maximizes the product of the 51 negative binomial probabilities over $(a, b)$. For the four weeks from week 50 of 2016, $\hat a = 42.7$ and $\hat b = 21{,}887$ person-years: the fitted distribution of state rates has prior mean 194.9 per 100,000 and standard deviation 29.8. The prior mean is the center of the distribution of the 51 areas' rates, close to the unweighted average of their raw rates (195.0 in this window), and not the population-weighted pooled rate of the window, 187.9. The five areas with the least person-time and the two with the most:
| Area | Deaths | Raw rate | $\hat B_i$ | Empirical Bayes | Six-year rate |
|---|---|---|---|---|---|
| Wyoming | 62 | 138.8 | 0.329 | 157.2 | 152.9 |
| Vermont | 116 | 242.6 | 0.314 | 227.6 | 216.8 |
| District of Columbia | 114 | 216.1 | 0.293 | 209.9 | 187.4 |
| Alaska | 83 | 146.1 | 0.278 | 159.7 | 123.3 |
| North Dakota | 105 | 181.5 | 0.275 | 185.2 | 184.4 |
| Texas | 3,134 | 146.0 | 0.010 | 146.5 | 145.1 |
| California | 4,703 | 156.4 | 0.007 | 156.7 | 152.1 |
Rates are per 100,000 person-years; the six-year rate uses all the weeks of 2014 to 2019 outside the window. In 2019 Wyoming had 578,759 residents and California 39,512,223.
When does shrinkage help? One window can mislead, so the script drew 30 windows for each length, the start chosen at random, with replacement, among the windows with no suppressed week (seeds 1 to 30; 29, 29, 28 and 25 distinct windows for the four lengths), and computed the total squared error of the raw and empirical Bayes rates against the six-year rates outside each window. Ratios below 1 favor shrinkage. "Smallest ten" restricts the sum to the ten areas with the least person-time:
| Window | Windows available | Error ratio, all 51 | Error ratio, smallest ten | $\hat B$ Wyoming | $\hat B$ California |
|---|---|---|---|---|---|
| 1 week | 303 | 0.649 (0.577 to 0.779) | 0.588 (0.463 to 0.714) | 0.713 | 0.0354 |
| 4 weeks | 272 | 0.864 (0.794 to 1.032) | 0.845 (0.711 to 1.137) | 0.346 | 0.0077 |
| 13 weeks | 204 | 0.978 (0.960 to 1.018) | 0.969 (0.910 to 1.069) | 0.136 | 0.0023 |
| 52 weeks | 62 | 0.945 (0.922 to 1.014) | 0.870 (0.846 to 1.100) | 0.036 | 0.0006 |
Each cell is the median over the 30 windows with the interquartile range; the shrinkage factors are medians. In this experiment shrinkage reduced the total error in 29 of the 30 one-week windows and in 21 of the 30 four-week windows. The gain fell as the window grew and the shrinkage factors fell with it: with a year of data even Wyoming's estimate moved by less than 4% of its distance to the center. Complete pooling did worse than the raw rates in all 120 windows drawn (111 distinct); its median error ratio ranged from 2.06 for one week to 74.8 for a year, and was 6.8 for four weeks. Two cautions apply to the comparison. The six-year rate is itself an estimate, with its own small sampling error. It also differs from a window's true rate by real changes over time, which no estimate from the window can know, and which is one reason the 13-week and 52-week rows do not improve steadily. And requiring a window with no suppressed week means that Alaska's and Wyoming's weekly counts in every window are at least 10 by selection; the effect is small (11 suppressed area-weeks in all) and weakens as weeks are summed, but it slightly favors the raw rates of those two areas, most in the one-week windows.
The winner's curse in the windows. In each four-week window, take the area with the highest raw rate. Its raw rate exceeded its six-year rate by a median of 12.4 per 100,000 (interquartile range −2.8 to 19.5) over the 30 windows; its empirical Bayes rate exceeded it by 0.8 (−11.8 to 12.5). In the one-week windows the highest raw rate exceeded the six-year rate by 38.8 and its empirical Bayes rate fell short by 3.7. The selection that Theorem 1 of Selective Inference describes is largely undone by shrinkage here.
What the model leaves out. The rates are crude, and part of the spread between areas is age: Utah's young population and West Virginia's old one explain much of the difference between 106.0 and 254.6. A better population level gives each area its own center from known characteristics, such as the share of residents aged 65 and over, by making $\ln(a/b)$ a linear function of them. That is the hierarchical version of Poisson regression, and the next section shows on NHANES how much a better center can matter.
Hypertension in NHANES cells
The 5,182 NHANES adults fall into 36 cells by age group (18 to 39, 40 to 59, 60 and over), sex and race and Hispanic origin (the six NHANES categories: Mexican American, other Hispanic, non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, and other race including multiracial). The cells hold 31 to 421 adults, and the share with hypertension in a cell runs from 0.094 to 0.915. The script drew samples of 400 and of 1,000 adults (seeds 1 to 30) and estimated each cell's share from the sample, comparing four estimates with the cell's share among all 5,182:
- the raw share in the sample;
- a beta-binomial model (the binomial analogue of Proposition 3: cell shares drawn from one Beta$(a, c)$ distribution, estimated by maximizing the marginal beta-binomial likelihood, with posterior mean $(a + x_i)/(a + c + m_i)$ for $x_i$ cases among $m_i$ sampled adults), which shrinks toward one common share;
- the normal hierarchical model of Theorem 1 with one common mean, $\theta_i$ normal with mean $\mu$ and variance $\tau^2$, using the same variances $\sigma_i^2$ as the next model, so that the two normal models differ only in their center;
- the same normal model with a separate mean for each age group, $\theta_i$ normal with mean $\mu_{g(i)}$ and variance $\tau^2$, where $g(i)$ is the age group of cell $i$ and $\sigma_i^2 = \bar p_g(1 - \bar p_g)/m_i$ uses the sample share of the age group; $\mu_1, \mu_2, \mu_3$ and $\tau^2$ are estimated by maximum likelihood as in Proposition 1, with each $\hat\mu_g$ the precision-weighted mean of its own cells.
Ratios of total squared error to that of the raw shares, median (interquartile range) over 30 samples, over the cells with at least one sampled adult:
| Sample | Beta-binomial, one share | Normal, one mean | Normal, age-group means |
|---|---|---|---|
| 400 adults | 0.608 (0.506 to 0.733) | 0.592 (0.495 to 0.677) | 0.304 (0.222 to 0.354) |
| 1,000 adults | 0.897 (0.787 to 0.944) | 0.864 (0.766 to 0.897) | 0.425 (0.377 to 0.580) |
The two normal models use the same likelihood and the same variances and differ only in their center, so the comparison between their columns isolates the center. Hypertension depends strongly on age, so a single center is a poor guess for most cells: its estimated spread is large (in every sample, a larger $\hat\tau^2$ than the age-group model's in the normal model, and a smaller $a + c$ than a beta-binomial fitted within any one age group), and the shrinkage it permits is small. Centers by age group leave a much smaller spread within groups, so small cells can be shrunk hard. In the median sample the age-group model cut the total error to less than half of the raw error at both sample sizes; it did so in 26 of the 30 samples of 400 and 19 of the 30 samples of 1,000. In these samples, a center that carries what is known about the groups (here the age group) mattered more than the form of the model: the beta-binomial and the normal model with one center did about equally well.
The James-Stein theorem
Everything so far assumed a model for how the true values were drawn. Stein (1956) proved that the separate estimates of several normal means are inadmissible (some other estimator does at least as well in total squared error for every set of true means and better for some), and James and Stein (1961) gave an explicit estimator that does strictly better for every set (Efron and Hastie, 2016, §7.5). Theorem 2 below shows that three means suffice when shrinking toward a fixed point. Efron and Hastie (2016, §7.1) call the theorem "a rude shock to the statistical world of 1961", because the defeat came in the simplest setting of all.
The setting: $\mathbf{X} = (X_1, \ldots, X_N)$ has independent normal coordinates with means $\boldsymbol\theta = (\theta_1, \ldots, \theta_N)$ and variance 1 (any known common variance can be scaled to 1). The loss of an estimate $\boldsymbol\delta$ is $\lVert\boldsymbol\delta - \boldsymbol\theta\rVert^2$, and its risk is the expected loss for fixed $\boldsymbol\theta$, $R(\boldsymbol\theta) = E_{\boldsymbol\theta}\lVert\boldsymbol\delta(\mathbf{X}) - \boldsymbol\theta\rVert^2$. The separate estimate $\boldsymbol\delta = \mathbf{X}$ has risk $N$ for every $\boldsymbol\theta$.
We prove a general form that shrinks toward any linear subspace: toward 0, toward the common mean, or toward a regression line. Let $P$ be an $N \times N$ projection matrix (symmetric, $P^2 = P$) of rank $r$. The vectors with $P\mathbf{v} = \mathbf{0}$ form the subspace shrunk toward, and $P\mathbf{X}$ is the part of $\mathbf{X}$ off it. Define
$$\boldsymbol\delta^{\mathrm{JS}}(\mathbf{X}) = \mathbf{X} - \frac{r - 2}{\lVert P\mathbf{X}\rVert^2}\,P\mathbf{X}.$$
With $P = I$ ($r = N$) this is $(1 - (N - 2)/\lVert\mathbf{X}\rVert^2)\mathbf{X}$, the original estimator, shrinking toward 0. With $P = I - \frac1N\mathbf{1}\mathbf{1}^{\mathsf T}$ ($r = N - 1$), $P\mathbf{X} = \mathbf{X} - \bar X\mathbf{1}$ and the estimator is the empirical Bayes rule of the last section, shrinking toward $\bar X$. With $P = I - H$ for the hat matrix $H$ of a regression on $p$ columns ($r = N - p$), it shrinks toward the fitted regression values; the age-group model above shrinks toward group means in the same spirit, with unequal variances and an estimated $\tau^2$, so it is not exactly this estimator.
The idea of the proof. The risk is $N$ minus a cross term plus a squared term. The cross term contains $E[(X_i - \theta_i)g(\mathbf{X})]$, the covariance of the noise with the correction, which looks impossible to compute without knowing $\boldsymbol\theta$. Stein's lemma turns it into the expectation of a derivative, which does not involve $\boldsymbol\theta$, and the two terms then combine into a negative quantity.
Lemma 2 (Stein's lemma in one dimension). Let $Z$ be normal with mean $\eta$ and variance 1, and let $g$ be continuously differentiable with $g$ and $g'$ bounded. Then $E[(Z - \eta)g(Z)] = E[g'(Z)]$.
Proof. Let $\phi$ be the standard normal density, so $Z$ has density $\phi(z - \eta)$ and $\frac{d}{dz}\phi(z - \eta) = -(z - \eta)\phi(z - \eta)$. Integrating by parts, $$\int_{-\infty}^{\infty} g(z)(z - \eta)\phi(z - \eta)\,dz = \Big[-g(z)\phi(z - \eta)\Big]_{-\infty}^{\infty} + \int_{-\infty}^{\infty} g'(z)\phi(z - \eta)\,dz.$$ The boundary term is 0 because $g$ is bounded and $\phi(z - \eta) \to 0$ as $z \to \pm\infty$; both integrals converge because $g$ and $g'$ are bounded and $|z - \eta|\phi(z - \eta)$ is integrable. $\square$
Efron and Hastie (2016, §7.5) report that Stein (1981) derived a simpler proof of the James-Stein theorem than the original. The proof below rests on Lemma 2. Optimism and Effective Degrees of Freedom proves the lemma for unbounded functions with finitely many corners (its Lemma 1) and derives from it Stein's unbiased risk estimate for a general fitting rule (its Theorem 4).
Theorem 2 (James-Stein). If $r \ge 3$, then for every $\boldsymbol\theta \in \mathbb{R}^N$,
$$E_{\boldsymbol\theta}\lVert\boldsymbol\delta^{\mathrm{JS}}(\mathbf{X}) - \boldsymbol\theta\rVert^2 = N - (r - 2)^2\,E_{\boldsymbol\theta}\Big[\frac{1}{\lVert P\mathbf{X}\rVert^2}\Big] < N.$$
Proof. Step 1: rotate. Choose an orthogonal matrix $Q$ whose first $r$ rows are an orthonormal basis of the column space of $P$, so that $QPQ^{\mathsf T} = D$, the diagonal matrix with $r$ ones followed by $N - r$ zeros. Let $\mathbf{Y} = Q\mathbf{X}$ and $\boldsymbol\eta = Q\boldsymbol\theta$. By Theorem 1 of Sampling Distributions from the Normal, the coordinates of $\mathbf{Y}$ are independent normal with means $\boldsymbol\eta$ and variance 1. Rotation preserves lengths, $\lVert P\mathbf{X}\rVert = \lVert QP\mathbf{X}\rVert = \lVert D\mathbf{Y}\rVert$, and $Q\boldsymbol\delta^{\mathrm{JS}} - \boldsymbol\eta = \mathbf{Y} - \boldsymbol\eta - (r - 2)D\mathbf{Y}/\lVert D\mathbf{Y}\rVert^2$. So it suffices to treat $P = D$: with $S = Y_1^2 + \cdots + Y_r^2$, the estimator replaces $Y_i$ by $Y_i - (r - 2)Y_i/S$ for $i \le r$ and leaves the other coordinates alone. ($S > 0$ with probability 1.)
Step 2: $E[1/S]$ is finite. The joint density of $(Y_1, \ldots, Y_r)$ is at most $(2\pi)^{-r/2}$. So $E[1/S] \le 1 + E[\mathbf{1}\{S < 1\}/S] \le 1 + (2\pi)^{-r/2}\int_{\lVert\mathbf{u}\rVert < 1}\lVert\mathbf{u}\rVert^{-2}\,d\mathbf{u}$. In polar coordinates the last integral is $A_r\int_0^1\rho^{-2}\rho^{r - 1}\,d\rho = A_r/(r - 2)$, with $A_r$ the surface area of the unit sphere in $\mathbb{R}^r$; it is finite because $r \ge 3$.
Step 3: Stein's identity for the correction. Fix $i \le r$ and condition on the other coordinates. Let $s = \sum_{j \le r, j \ne i}Y_j^2$, which is positive with probability 1 because $r \ge 2$. As a function of $y = Y_i$, $g(y) = y/(y^2 + s)$ is continuously differentiable, $|g(y)| \le 1/(2\sqrt{s})$, and $g'(y) = (s - y^2)/(y^2 + s)^2$, so $|g'(y)| \le 1/s$. Lemma 2, applied conditionally, gives $E[(Y_i - \eta_i)Y_i/S \mid \text{others}] = E[(S - 2Y_i^2)/S^2 \mid \text{others}]$. Both sides are integrable: $|(Y_i - \eta_i)Y_i/S| \le |Y_i - \eta_i|/\sqrt{S}$, whose mean is at most $\sqrt{E[(Y_i - \eta_i)^2]\,E[1/S]}$ by the Cauchy-Schwarz inequality (Probability Inequalities), finite by Step 2; and $|S - 2Y_i^2|/S^2 \le 1/S$. So the unconditional expectations are equal (by Fubini's theorem, cited as in Random Variables and Distribution Functions). Summing over $i \le r$, $$E\Big[\sum_{i \le r}(Y_i - \eta_i)\frac{Y_i}{S}\Big] = E\Big[\frac{rS - 2S}{S^2}\Big] = (r - 2)\,E[1/S].$$
Step 4: expand the loss. $$\lVert Q\boldsymbol\delta^{\mathrm{JS}} - \boldsymbol\eta\rVert^2 = \lVert\mathbf{Y} - \boldsymbol\eta\rVert^2 - 2(r - 2)\sum_{i \le r}(Y_i - \eta_i)\frac{Y_i}{S} + (r - 2)^2\frac{S}{S^2}.$$ Taking expectations with Step 3, the risk is $N - 2(r - 2)^2E[1/S] + (r - 2)^2E[1/S] = N - (r - 2)^2E[1/S]$, and $E[1/S] > 0$. $\square$
No prior appears in the theorem: the $\theta_i$ are fixed, unrelated numbers. Yet each estimate uses all the other coordinates through $\lVert P\mathbf{X}\rVert^2$. In the language of decision theory, Theorem 2 with $P = I$ makes $\mathbf{X}$ inadmissible for $N \ge 3$: another estimate has risk at most as large everywhere and smaller somewhere. (Efron and Hastie, 2016, §7.1, state the theorem for the version that shrinks toward the mean, which needs $N \ge 4$.)
Example (an unrelated mean helps). The theorem does not say that each coordinate improves. It says that the sum of squared errors improves on average. Combining the rate of a state with, say, the price of tea and a batting average still lowers the total risk, but the gain shrinks as $\lVert P\boldsymbol\theta\rVert$ grows, and the gain on unrelated quantities is the gain of a total that nobody needs. Efron and Hastie (2016, §7.5) describe the discomfort the theorem caused (why should one player's estimate depend on the other players' performances?) and frame it as a question of direct against indirect evidence. In practice, shrinkage is worth having when the groups are alike enough that $\lVert P\boldsymbol\theta\rVert^2$ is small compared with $N$, which is what a hierarchical model assumes and what its estimated $\tau^2$ measures.
The exact risk. By Theorem 2 of Sampling Distributions from the Normal, $\lVert P\mathbf{X}\rVert^2$ has the noncentral chi-squared distribution with $r$ degrees of freedom and noncentrality $\lambda = \lVert P\boldsymbol\theta\rVert^2$. By its Theorem 5 that distribution is a Poisson mixture of central chi-squared distributions with $r + 2J$ degrees of freedom, $J$ Poisson with mean $\lambda/2$. Taking expectations term by term (all terms are nonnegative) and applying Lemma 1,
$$E\Big[\frac{1}{\lVert P\mathbf{X}\rVert^2}\Big] = \sum_{j = 0}^\infty e^{-\lambda/2}\frac{(\lambda/2)^j}{j!}\cdot\frac{1}{r + 2j - 2}, \qquad R(\boldsymbol\theta) = N - (r - 2)^2\sum_{j = 0}^\infty e^{-\lambda/2}\frac{(\lambda/2)^j}{j!\,(r + 2j - 2)}.$$
The risk depends on $\boldsymbol\theta$ only through $\lambda$. At $\lambda = 0$ (all means on the target subspace) it is $N - (r - 2) = N - r + 2$: 2 when shrinking toward 0, 3 when shrinking toward the mean, whatever $N$ is. As $\lambda \to \infty$ the sum tends to 0, because the Poisson probability of $J < k$ tends to 0 for each $k$ and the terms with $j \ge k$ are at most $1/(r + 2k - 2)$, so the risk tends to $N$ without reaching it. It also increases with $\lambda$: for $\lambda_1 < \lambda_2$, a Poisson variable with mean $\lambda_2/2$ has the distribution of $J_1 + K$, with $J_1$ Poisson with mean $\lambda_1/2$ and $K$ an independent Poisson variable with mean $(\lambda_2 - \lambda_1)/2$ (independent Poisson variables add, The Poisson Distribution). Since $1/(r + 2j - 2)$ decreases in $j$, $1/(r + 2J_1 + 2K - 2) \le 1/(r + 2J_1 - 2)$ for every outcome, so $E[1/\lVert P\mathbf{X}\rVert^2]$ is smaller at $\lambda_2$, and the risk larger.
The script checked the formula by simulation for $N = 10$ with $\boldsymbol\theta$ a multiple of a fixed vector with mean 0 (so that $\lVert\boldsymbol\theta\rVert^2$ is also $\lVert P\boldsymbol\theta\rVert^2$ for the shrinkage toward the mean): 30 seeds, each 2,000 draws of $\mathbf{X}$. The table gives the exact risks and the simulated medians; "+" is the positive-part rule, which replaces a negative shrinkage coefficient $1 - (r - 2)/\lVert P\mathbf{X}\rVert^2$ by 0.
| $\lVert\boldsymbol\theta\rVert^2$ | Toward 0, exact | simulated | Toward mean, exact | simulated | Toward mean, + | Separate |
|---|---|---|---|---|---|---|
| 0 | 2.00 | 2.01 | 3.00 | 3.01 | 2.28 | 9.97 |
| 2.5 | 3.65 | 3.67 | 4.58 | 4.59 | 4.07 | 9.97 |
| 10 | 6.21 | 6.21 | 6.90 | 6.90 | 6.82 | 9.97 |
| 40 | 8.62 | 8.62 | 8.92 | 8.92 | 8.92 | 9.97 |
The separate estimates' simulated risk is 9.97 in every row because the same noise was used in each. The positive-part rule was never worse than the plain rule in this simulation and much better near the target. Efron and Hastie (2016, §7.5) note that the James-Stein estimator is itself inadmissible, improved everywhere by the positive-part change, and that further improvements are minuscule; that result is cited, not proved, here.
The baseball players. Efron and Hastie (2016, §7.2, Table 7.1) apply the rule to 18 Major League players' batting averages over their first 90 times at bat in the 1970 season, predicting their averages over the rest of the season. The sum of squared prediction errors was 0.0425 for the players' own averages and 0.0218 for the James-Stein estimates, about half. Their notes explain that the data are partly artificial: the original 45 at bats recorded for each player were augmented by 45 simulated binomial draws. The state death rates above are the same design on public data: an early window to estimate, the rest of the record to check.
Where else shrinkage appears
- Ridge regression is the posterior mean under a normal prior centered at 0 (Ridge and Lasso Regression, Proposition 3), and Efron and Hastie (2016, ch. 7) treat it in the same chapter as the James-Stein estimator, as shrinkage applied to regression coefficients.
- Smoothing of word probabilities by a Dirichlet prior shrinks each document's estimates toward the collection's (Choosing a Prior).
- Local false discovery rates apply the empirical Bayes idea to many tests at once (q-Values and the Share of True Nulls).
- Model comparison: the hierarchical model is the continuous family that Gelman et al. (2013, §7.4) prefer to a Bayes factor between "all groups equal" and "all groups unrelated" (Bayes Factors and Posterior Model Probabilities).
What is proved here, what is cited, and what is computed
Proved on this page: the posterior and marginal distributions of the normal hierarchical model and its Bayes risk (Theorem 1); the weighted-mean estimate of $\mu$ (Proposition 1); the mean of a reciprocal chi-squared variable (Lemma 1); the unbiasedness of $(N - 3)/S$ (Proposition 2); the gamma-Poisson posterior, its shrinkage form and its negative binomial marginal (Proposition 3); Stein's lemma in one dimension (Lemma 2); the James-Stein theorem for shrinkage toward any subspace of codimension $r \ge 3$ (Theorem 2); its exact risk as a Poisson mixture, its values at the target and its limit; the reduction of Tweedie's formula to linear shrinkage for a normal marginal.
Taken from other pages: the conjugate updates (Priors, Posteriors and Conjugate Families); the marginal likelihood identity (Bayes Factors and Posterior Model Probabilities); rotation, projections and the noncentral chi-squared mixture (Sampling Distributions from the Normal); the chi-squared density and the distribution of a normal sample's sum of squares (The Chi-Squared and t Distributions); Tweedie's formula and the winner's curse (Selective Inference); the Cauchy-Schwarz inequality (Probability Inequalities). Fubini's theorem and termwise integration of nonnegative series are cited, as in the Probability Theory unit.
Cited: the three pooling choices and their priors, and the view of empirical Bayes as an approximation (Gelman et al., 2013, §§5.1, 5.4); the Bayes risk $NB' + 3(1 - B')$, the positive-part improvement, the history (Stein, 1956; James and Stein, 1961; Robbins, 1956; Stein, 1981; Efron and Morris, 1973) and the baseball example (Efron and Hastie, 2016, ch. 6, 7 and 20).
Computed: every rate, factor, ratio and table on the page, by tools/experiments/statistics/hierarchical-models-and-shrinkage.py, from the NCHS weekly counts, the Census Vintage 2019 populations and NHANES 2017–2018. The window experiments use seeds 1 to 30 for each length, the NHANES samples seeds 1 to 30, the James-Stein table seeds 1 to 30 and the Bayes risk seeds 1,001 to 1,030.
Worked problems
These problems are posed for this article, on the state data and on small exact cases. Every number is recomputed by the article's script.
Problem 1 (Wyoming by hand). In the four-week window from week 50 of 2016, Wyoming had 62 cancer deaths in 44,685 person-years. With $\hat a = 42.66$ and $\hat b = 21{,}887$ person-years, find the raw rate, the shrinkage factor, the empirical Bayes rate and its posterior standard deviation.
Solution. The raw rate is $62/44{,}685 = 138.8$ per 100,000 person-years. The shrinkage factor is $\hat B = 21{,}887/(21{,}887 + 44{,}685) = 0.3288$. The prior mean is $42.66/21{,}887 = 194.9$ per 100,000. By Proposition 3 the estimate is $(1 - 0.3288)(138.8) + 0.3288(194.9) = 157.2$, which is also $(42.66 + 62)/(21{,}887 + 44{,}685)$ per person-year. The posterior is Gamma$(104.66, 66{,}572)$ (shape and rate, rate in person-years), with standard deviation $\sqrt{104.66}/66{,}572$ per person-year, or 15.4 per 100,000. Wyoming's six-year rate, 152.9, is within one posterior standard deviation of the estimate; the raw rate was 14.1 below it.
Problem 2 (five means by hand). Five independent observations, each normal with variance 1, are 2.1, −0.3, 0.9, −1.4 and 0.7. Compute the James-Stein estimates toward 0 and toward the mean, and their risks if all five true means are equal.
Solution. Toward 0, $r = 5$ and $S = 2.1^2 + 0.3^2 + 0.9^2 + 1.4^2 + 0.7^2 = 7.76$, so the factor is $1 - 3/7.76 = 0.6134$ and the estimates are 1.288, −0.184, 0.552, −0.859 and 0.429. Toward the mean, $\bar X = 0.4$, $r = 4$ and $S = \sum(X_i - 0.4)^2 = 6.960$, so the factor is $1 - 2/6.960 = 0.7126$, and the estimates $0.4 + 0.7126(X_i - 0.4)$ are 1.611, −0.099, 0.756, −0.883 and 0.614. If all five means are equal, $P\boldsymbol\theta = \mathbf{0}$ for the shrinkage toward the mean, and by the exact risk formula its risk is $5 - 4 + 2 = 3$, against 5 for the separate estimates. Shrinkage toward 0 has risk 2 only if the common mean is 0; otherwise $\lambda = 5\theta^2$, and its risk lies between 2 and 5.
Problem 3 (James-Stein as empirical Bayes). In Theorem 1 with all $\sigma_i^2 = \sigma^2$ known, show that plugging in $\hat\mu = \bar y$ and the unbiased estimate of $1/(\sigma^2 + \tau^2)$ gives the James-Stein estimator toward the mean, scaled to variance $\sigma^2$.
Solution. Marginally the $y_i$ are independent normal with mean $\mu$ and variance $\sigma^2 + \tau^2$ (Theorem 1(2)). Then $S/(\sigma^2 + \tau^2)$ is chi-squared with $N - 1$ degrees of freedom, with $S = \sum(y_i - \bar y)^2$, and by Lemma 1, $E[(N - 3)/S] = 1/(\sigma^2 + \tau^2)$ when $N \ge 4$. The shrinkage factor is $B = \sigma^2 \cdot 1/(\sigma^2 + \tau^2)$, estimated without bias by $\hat B = (N - 3)\sigma^2/S$. Theorem 1(1) with these estimates gives $\bar y + (1 - (N - 3)\sigma^2/S)(y_i - \bar y)$. Dividing every $y_i$ by $\sigma$ turns this into the estimator of Theorem 2 with $P = I - \frac1N\mathbf{1}\mathbf{1}^{\mathsf T}$ and $r = N - 1$, applied to variance-1 data, since $(N - 3)\sigma^2/S = (r - 2)/\lVert P\mathbf{y}/\sigma\rVert^2$. So the same estimator is the plug-in Bayes rule under the hierarchical model and, by Theorem 2, better than the separate estimates for every set of true means.
Summary
| Quantity | Formula |
|---|---|
| Normal hierarchical model | $y_i \mid \theta_i$ normal$(\theta_i, \sigma_i^2)$; $\theta_i$ normal$(\mu, \tau^2)$ |
| Shrinkage | $E[\theta_i \mid \mathbf{y}] = (1 - B_i)y_i + B_i\mu$, $B_i = \sigma_i^2/(\sigma_i^2 + \tau^2)$; Bayes risk $(1 - B_i)\sigma_i^2$ |
| Empirical Bayes | $y_i$ normal$(\mu, \sigma_i^2 + \tau^2)$; $\hat\mu = \sum w_iy_i/\sum w_i$, $\hat\tau^2$ by marginal likelihood |
| Gamma-Poisson | $E[\lambda_i \mid \mathbf{d}] = (a + d_i)/(b + t_i)$, $B_i = b/(b + t_i)$; $d_i$ negative binomial |
| Unbiased factor | $E[(N - 3)/S] = 1/(1 + A)$; $E[1/\chi^2_\nu] = 1/(\nu - 2)$ |
| Stein's lemma | $E[(Z - \eta)g(Z)] = E[g'(Z)]$ |
| James-Stein | $\mathbf{X} - (r - 2)P\mathbf{X}/\lVert P\mathbf{X}\rVert^2$; risk $N - (r - 2)^2E[1/\lVert P\mathbf{X}\rVert^2] < N$ for $r \ge 3$ |
| Exact risk | $E[1/\chi^2_r(\lambda)] = \sum_j \text{Poisson}(j; \lambda/2)/(r + 2j - 2)$; $N - r + 2$ at $\lambda = 0$ |
What we have learned
When many related quantities are estimated at once, the data on all of them say how far apart they are, and that spread decides how far each estimate should move toward the others. The normal hierarchical model makes the rule exact: each estimate is pulled toward the center by the share of its variance that is noise. Estimating the center and the spread from the data gives empirical Bayes estimates, and in the equal-variance case the James-Stein estimator, which beats the separate estimates for every set of true means once there are three or more. On state cancer death rates the pull is about a third for the smallest states from four weeks of data and almost nothing for the largest, and in the experiments it reduced the error most when the data were thinnest. On NHANES cells, with the model held fixed, shrinking toward age-group means instead of one mean roughly halved the remaining error: the choice of center mattered as much as the shrinkage.
Empirical Bayes plugs in point estimates of the hyperparameters and ignores their uncertainty. A full Bayesian analysis puts a prior on them and averages over their posterior, which seldom has a closed form. Drawing from such posteriors is the subject of Markov Chain Monte Carlo.
References
- Efron, B., and Hastie, T. (2016). Computer Age Statistical Inference: Algorithms, Evidence, and Data Science. Cambridge University Press. Chapter 6 (§6.5), Chapter 7 (§§7.1–7.3, 7.5, Table 7.1, equations 7.12–7.14) and §20.3. record · 9 pages
- Efron, B., and Morris, C. (1973). Stein's estimation rule and its competitors: an empirical Bayes approach. Journal of the American Statistical Association, 68(341), 117–130. https://doi.org/10.1080/01621459.1973.10481350 Cited as Efron and Hastie (2016) describe the series of papers it belongs to. record
- Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. (2013). Bayesian Data Analysis (3rd ed.). CRC Press. §§5.1, 5.4 and 7.4. record · 7 pages
- James, W., and Stein, C. (1961). Estimation with quadratic loss. In Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, vol. 1, 361–379. University of California Press. Cited as Efron and Hastie (2016) cite it. record
- National Center for Health Statistics (2020). National Health and Nutrition Examination Survey 2017–2018 Data Documentation, Codebook, and Frequencies: Demographic Variables and Sample Weights (DEMO_J) and Blood Pressure (BPX_J). Centers for Disease Control and Prevention. https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/DEMO_J.htm record · 10 pages
- National Center for Health Statistics (2021). Weekly Counts of Deaths by State and Select Causes, 2014–2019. Centers for Disease Control and Prevention, data.cdc.gov dataset 3yf8-kanr. https://data.cdc.gov/d/3yf8-kanr record · 13 pages
- Robbins, H. (1956). An empirical Bayes approach to statistics. In Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, vol. 1, 157–163. University of California Press. Cited as Efron and Hastie (2016) cite it. record
- Stein, C. (1956). Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. In Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, vol. 1, 197–206. University of California Press. Cited as Efron and Hastie (2016) cite it. record
- Stein, C. (1981). Estimation of the mean of a multivariate normal distribution. The Annals of Statistics, 9(6), 1135–1151. https://doi.org/10.1214/aos/1176345632 Cited as Efron and Hastie (2016) cite it. record · 2 pages
- U.S. Census Bureau, Population Division (2019). Annual Estimates of the Resident Population for the United States, Regions, States, and Puerto Rico: April 1, 2010 to July 1, 2019 (Vintage 2019), file nst-est2019-alldata.csv. https://www2.census.gov/programs-surveys/popest/datasets/2010-2019/national/totals/ record · 2 pages