3 THE SCIENTIFIC REALITY
More significantly, the method of post-stratification does not really fit in the design-
based framework. Recall that the design-based conception of sample representation
relies exclusively on the power of the selection process to justify sample-to-population
inference. The idea is supposed to be that, as long as researchers adopt adequate
sampling procedures, they should not feel the need to also analyze sample composition.
Besides, the design-based framework does not provide guidance for how sample
composition should be analyzed. To see this, we can compare the NCS with similar
sampling efforts from other countries. The NCS of America post-stratified against sex,
age, marital status, race, education, region, and urbanicity (Mickelson et al., 1997,
p.1095); the German National Health (GNH) survey post-stratified against sex, age
(with a different range), marital status (in finer categories), and employment status
(Jacobi et al., 2002); the Australia National Mental Health Survey (ANMHS), however,
decided to not post-stratify at all (Henderson et al., 2000).
In addition to the inconsistencies across similar survey efforts, those that do post-
stratify provide very little reasoning as to why they decide on the characteristics that
they do. Post-stratification as a method depends on the existence of full enumeration
demographics data like the NHIS, which is often called “auxiliary data” or “organic
data” in this context. The existence of such data limits whether and how a sample
survey can afford to post-stratify. That said, post-stratification also reflects conscious
choices on the part of the research team. For example, the NCS chose to post-stratify
against the NHIS rather than the US Census because “[the NHIS] includes a much wider
array of sociodemographic variables for the purposes of poststratification” (Mickelson
et al., 1997, p.1095). This is certainly not because the US Census did not gather a lot
of data. In 1989, the Census Bureau gathered information as diverse as age differences
between bride and groom, prevalence of AIDS, immigrative status, and average weekly
expenditure (US Census Bureau, 1989). Instead, the US Census gathered the wrong
sorts of data, at least from the perspective of the NCS.
It is clear that researchers make judgments about which characteristic imbalance
is worth correcting in a randomly selected sample, and yet these judgments are rarely
explicitly stated or argued for. Indeed, there is no theoretical space within the design-
based framework for such corrections, so it only makes sense that corrections like these,
when they do occur, are guided more by intuition than by arguments.
The discussion concerning post-stratification has highlighted two important ob-
servations. First, even the best random selection efforts result in sample imbalances
deemed worthy of correction by researchers. The elimination of “subjective selec-
tion bias” guaranteed by random selection is clearly insufficient. Second, while post-
stratification is frequently used to correct for chance bias, the practice is not principled.
This is because post-sampling corrections of this form do not fit into the design-based
understanding of how sampling is supposed to work.
Worse still, large-scale survey efforst like the NCS are relatively uncommon; most
research teams within the social sciences do not have nearly as much resources to em-
ploy anything like an area probability sample over a nation. This is compounded by
the fact that many research projects within psychology, anthropology, and economics
8