Are U-shaped relationships between risk factors and outcomes artifactual?

Abstract

Background: The objective of this study was to evaluate whether the observed nadir in a U- or J-shaped relationship between a particular risk factor and a future health outcome is a function of the distribution of the risk factor in the sample being analyzed.

Methods: Data from the ORIGIN trial were used to assess the relationship between three risk factors (weight, systolic blood pressure, and serum insulin) and the hazard of a major cardiovascular event comprising a nonfatal myocardial infarction, nonfatal stroke, or cardiovascular death. Three spline curves were generated for each risk factor. The first was based on all available data, the second for a subgroup with a higher mean risk factor level, and the third for a subgroup with a lower mean risk factor level. Nadir levels of the risk factor (i.e., risk factor levels predicting the lowest hazard) were then identified for each spline curve.

Results: When compared to the nadir values based on all available data, nadir values for all three risk factors were higher for the subgroups with higher mean levels and lower for those with lower mean levels.

Conclusions: The distribution of a risk factor in the population is an important determinant of its nadir value. Populations with high or low values may have high and low nadirs, respectively. Identification of a nadir for a modifiable risk factor from epidemiologic relationships may therefore arise from this distribution bias and is therefore unrelated to therapeutic targets.

Keywords: J shape; J型; U-shape; U型; cardiovascular outcome; insulin; risk factor; systolic blood pressure; weight; 体重; 心血管结局; 收缩压; 胰岛素; 风险因素.

FIGURE 1

Spline curve showing the relationship between 3 risk factors and the hazard ratio for composite cardiovascular outcome (nonfatal MI, stroke, and cardiovascular death). Panel 1 shows the relationship for weight, Panel 2 shows the relationship for systolic blood pressure, and Panel 3 shows the relationship for winsorized insulin. Each Panel is divided in 3 showing the relationships (A) in the entire study sample, (B) in the high subgroup, and (C) in the low subgroup.