THE AUTHORS REPLY

We read with interest the letter from Dr. Karp (1) on our paper on descriptive epidemiology (2). Dr. Karp argues that “unlike in inquiries in epidemiologic practice, there is no place- and time-specific target population in either causal or descriptive epidemiologic research, where the domain of inference is a theoretical/abstract (super-)population, infinite in size” (1, p. 682). Statistical inference for descriptive and causal questions often relies on the idea of the theoretical infinite superpopulation referenced by Karp. However, the domain of inference is not abstract or theoretical. Unless otherwise stated, the target population is assumed to be the superpopulation that would have given rise to the study sample using simple random sampling (3). Therefore, the target population has a well-defined distribution of patient covariates and a clear context, all of which determine the value of the final estimate (even in the absence of bias).

Dr. Karp’s view seems to be echoing the viewpoint in the debates about the value of “representativeness” (4) that “biological relationships are generalizable” (5, p. 1014) and therefore a study sample should be representative of a meaningful target population for descriptive epidemiologic inquiries but that representativeness of a meaningful target population is unnecessary or undesirable for causal epidemiologic inquiries. We disagree.

Questions in epidemiology, and their answers, are meaningless without context in time and space (6). Epidemiology informs clinical and public health decision-making, and these decisions apply to specific populations. Even in the absence of bias, estimates of descriptive measures will vary between populations if any predictor of the outcome being described differs across those populations. For causal measures, estimates will vary if the distribution of any effect modifier varies between populations (7, 8). Therefore, if a study does not specify a target population, its results are uninterpretable for any purpose unless one assumes that there are no differences in the prevalence of these predictors/modifiers across populations and, accordingly, that results apply universally to (or at the very least, are averaged over) all humans (past, present, and future). If the measure of interest varies across populations, failing to report the target population for an epidemiologic study is at best inadequate and at worst misleading.

Imagine a study describing a high incidence rate of new human immunodeficiency virus (HIV) infections. Interpretation of the results would be very different if the target population for the study was the people of South Africa (where high HIV incidence is a known public health problem), residents of a small town in North Carolina (where high HIV incidence would represent a previously unknown public health problem), or persons referred for HIV testing at a local clinic (where high HIV incidence might indicate that appropriate groups are being screened for HIV). Furthermore, imagine that investigators in a subsequent study report a large rate ratio for the effect of an intervention designed to reduce the rate of new HIV infections. If the study was conducted in South Africa, its results are unlikely to be generalizable to a sexually transmitted disease clinic in San Francisco, California, because of the different baseline rates of HIV infection and because of the very different contexts (in South Africa, adolescent women and young girls represent the majority of new HIV cases, whereas the majority of new HIV cases in San Francisco are among men who have sex with men).

Having a clear case definition (minimizing measurement error), a clear “domain of inference” (defining the target population), a clear understanding of sampling/diagnostic biases (minimizing selection bias), and a clear analytical plan for describing the occurrence of cases over time (avoiding overadjustment) are all crucial for understanding the evolution of health outcomes in a population and for understanding the effect of exposures or interventions on the evolution of that outcome in a population.

ACKNOWLEDGMENTS

This work was supported by grants K01 AA028193, K01 AI125087, and R01 AI157758 from the National Institutes of Health.

Conflict of interest: none declared.

REFERENCES