No Calculation When Observation Can Be Made

Abstract

For a long time, when data were needed by a nation, a census was often carried out in an attempt to make measurements on every household or every person in that nation. The move from conducting a census to conducting a sample survey where only a subset of households or persons was measured was met with opposition. Sampling requires that weights be calculated and applied to the observations in the sample, while observations of all units in a census do not require such calculation. Wanting to preserve censuses, the opposition to sampling was best expressed in the translated firm phrase, “No calculation when observation can be made”. With advances in probability sampling theory and applications, this resistance movement eventually failed. Societies have become more complex and people more mobile; data collection costs have risen; and the public’s response rates to government sample surveys continue to fall, though not as sharply as with non-government sample surveys. Other sources of data (e.g., administrative records, big data, commercial data) to measure human behavior and condition are being investigated for increased use in production of official statistics. Though there are technical concerns (e.g., representativeness, data quality, privacy), other features of big data (e.g., relatively inexpensive; largely digital measurement; minimize respondent burden; lot of data; rich, complex, diverse, flexible) and the growing demand for data gathering and managing tools suggest a very useful source of data. However, with big data and the implicit promise of observation of everything, there is a risk that users may move to question the need for sampling or statistical calculation. Indeed, the new phrase might be “No statistical calculation when big data observation can be made”, a phrase lacking scientific merit. In this paper, we take a brief look at the move from censuses to samples. Specifically, we (i) highlight Kiaer (1895, 1897) nonrandom representative method which laid seeds for the current use of survey sampling methodology to make measurements for official statistics; (ii) highlight Neyman’s 1934 contribution with stratified random sampling and optimal sample allocation to achieve representativeness, (iii) note some recent developments (Wright 2012, 2014, 2016, 2017) to improve Neyman’s optimal allocation, and (iv) offer some personal observations looking forward.

Tommy Wright is Chief of the Center for Statistical Research and Methodology, U. S. Census Bureau, Washington, D.C. 20233 (E-mail: tommy.wright@census.gov) and adjunct faculty in statistics at Georgetown University. The views expressed are the author’s and not necessarily those of the U. S. Census Bureau.