TL;DR
In this foundational 1946 paper, Frederick Mosteller argues that when data is cheap but analysis is expensive, "inefficient" statistics—based on simple sorting and counting rather than complex sums of squares—are often the superior choice. He provides a rigorous mathematical framework for using order statistics (like the median or range) to estimate means, standard deviations, and correlations with surprisingly high efficiency.
What Problem Does This Paper Address?
In the mid-1940s, mathematical statistics was obsessed with "efficiency"—finding estimators that extract the absolute maximum information from a sample (like the sample mean or standard deviation). However, Mosteller noticed a practical disconnect: in fields like industrial quality control or large-scale educational testing, researchers often had "mountainous" piles of data but limited time and budget for calculation.
At the time, calculating a standard deviation required squaring every number in a dataset—a laborious task before the age of digital computers. Mosteller sought to bridge this gap by proving that "crude" methods, such as using only a few specific points from a sorted list (order statistics), could provide results nearly as good as the "perfect" formulas, but at a fraction of the computational cost.
Methods
Mosteller focuses on systematic statistics, which are functions of order statistics (data points ranked from smallest to largest). The core of his approach involves:
- Asymptotic Normality: He proves that for large samples, the joint distribution of several order statistics tends toward a multivariate normal distribution. This allows him to treat these simple "counting" methods with the same mathematical rigor as traditional estimators.
- Mean Estimation: Instead of averaging all $n$ points, he proposes averaging $k$ specific observations (e.g., the 10th, 30th, 50th, 70th, and 90th percentiles).
- Standard Deviation Estimation: He utilizes "quasi-ranges"—the difference between two symmetric order statistics (like the 93rd percentile minus the 7th percentile)—to estimate $\sigma$ without ever squaring a number.
- Correlation Estimation: He introduces a "four-corner" counting method for bivariate data. By simply counting how many observations fall into specific outer regions of a scatter plot, one can estimate the correlation coefficient $\rho$ using a lookup table or chart.
Datasets
This is primarily a theoretical and methodological paper. It does not use modern training datasets. Instead, Mosteller:
- Derives results based on the Normal (Gaussian) Distribution.
- Uses a small illustrative sample of 25 observations to demonstrate his graphical correlation estimation method.
- References Fisher and Yates’ tables for expected values of order statistics.
Key Insights and Ideas
- The Economy of Information: Mosteller reframes "efficiency" not just as a mathematical property, but as a financial one. If it costs $100 to get 100% efficiency but only $1 to get 80% efficiency by using a larger, cheaper sample, the "inefficient" statistic is actually the more economical choice.
- The Power of Eight: One of the most striking findings is that using just eight suitably chosen order statistics to estimate the standard deviation yields an efficiency of 90%. You can ignore the vast majority of your data and still be remarkably accurate.
- The 27% Rule: For estimating correlation in large samples where means and variances are unknown, Mosteller finds that the optimal "cut" for his counting method involves looking at the top and bottom 27% of the observations—a heuristic that still appears in various psychometric and statistical contexts today.
- Punched-Card Optimization: The methods were specifically designed to be performed by a "counting sorter," the high-speed data processing hardware of the era, bypassing the need for slow tabulating machines that performed addition or multiplication.
Key Results
- Mean Estimation: Using just 10 order statistics achieves 97.3% efficiency compared to the sample mean. Even using just two (the 27th and 73rd percentiles) gives 81% efficiency.
- Standard Deviation: A simple quasi-range (using 2 points) provides 65% efficiency. Stepping up to 8 points hits 90%.
- Correlation: The "four-corner" counting method achieves roughly 52% efficiency compared to the Pearson correlation coefficient ($r$). While lower than the other estimators, the massive reduction in labor (no sums of products) made it highly attractive for "mountainous" data.
Limitations and Future Work
- Assumption of Normality: The efficiencies calculated are specific to data drawn from a normal distribution. The paper notes that if the underlying distribution is not normal, these specific percentiles might not be optimal.
- Computational Bottlenecks: Mosteller admits that finding the absolute optimum spacing for more than three order statistics ($k > 3$) was computationally difficult at the time, leading him to suggest "good enough" arbitrary spacings.
- Small Samples: While the theory is "asymptotic" (aimed at large samples), Mosteller notes that the exact behavior for very small samples ($n < 10$) requires further tabulated data, which he mentions was in preparation at the time of writing.
- Tests of Significance: The paper focuses on estimation; Mosteller flags "inefficient" hypothesis testing as a subject for a future paper.
Summary updated