Elsevier

Annals of Epidemiology

Volume 7, Issue 3, April 1997, Pages 188-193
Annals of Epidemiology

Original report
A nested approach to evaluating dose-response and trend

https://doi.org/10.1016/S1047-2797(96)00159-7Get rights and content

Abstract

PURPOSE: Conventional dose-response and trend analysis fits either a linear or categorical logistic model and tests the resulting coefficients. These analyses, however, are based on implausible assumptions.

METHODS: We present an alternative approach that uses likelihood ratio tests to compare nested regression models and determine when a model is rich enough to capture the data trends.

RESULTS: For illustration, we apply this approach to data on diet and colorectal polyps.

CONCLUSIONS: Comparison of linear and quadratic spline logistic models indicates that the conventional approach of using only a linear logistic model would not appropriately describe the association between intake of fruits and vegetables and colorectal polyps in our data. Graphical checking further supports this conclusion.

References (14)

  • M Maclure et al.

    Tests for trend and dose-response: misinterpretations and alternatives

    Am J Epidemiol

    (1992)
  • S Greenland

    Dose-response and trend analysis in epidemiology: Alternatives to categorical analysis

    Epidemiology

    (1995)
  • S Greenland

    Problems in the average-risk interpretation of categorical dose-response analysis

    Epidemiology

    (1995)
  • JS Witte et al.

    Relation of vegetable, fruit, and grain consumption to colorectal adenomatous polyps

    Am J Epidemiol

    (1996)
  • DW Hosmer et al.

    Applied Logistic Regression

    (1989)
  • D Clayton et al.

    Statistical Models in Epidemiology

    (1993)
  • S Greenland

    Power loss associated with automatic use of percentile categories and ordinal scores in dose-response and trend analysis of continuous exposures

    Epidemiology

    (1995)
There are more references available in the full text version of this article.

Cited by (40)

  • Longitudinal study of acculturation and BMI change among Asian American men

    2015, Preventive Medicine
    Citation Excerpt :

    In the level 1 models, we regressed BMI on linear, continuous age at BMI measurement. We first evaluated 9-knot quadratic spline regression models (Witte and Greenland, 1997). Testing of successively simplified models showed that linear models were sufficient to describe the association between age and BMI across all ages in our study sample.

  • Body mass index at age 25 and all-cause mortality in whites and African Americans: The atherosclerosis risk in communities study

    2012, Journal of Adolescent Health
    Citation Excerpt :

    For the continuous analysis, we selected the most appropriate forms for BMI and weight change by fitting models with quadratic splines. The initial models used quadratic spline regression analysis [15] with five knots to examine the shapes of the association of BMI at age 25 and weight change from age 25 to visit 1 with all-cause mortality. More simple models examined BMI and weight change in quadratic or linear forms.

  • Racial/ethnic differences in the monthly variation of preeclampsia incidence

    2007, American Journal of Obstetrics and Gynecology
    Citation Excerpt :

    Spline regression more closely approximates nonparametric regression,18,19 which makes no assumptions about the exposure-disease relation. We specified the date of delivery as the day of the year (a continuous variable [ie, day 1-365]) and then used methods proposed by Witte and Greenland20 to determine the most appropriate specification of the delivery day of the year in our model. A restricted quadratic spline with a knot at day 200 was found to have the best fit in the models for the total population (P = .01) and among white women (P < .05).

View all citing articles on Scopus

An initial version of this paper was presented at the 28th annual meeting of the Society for Epidemiologic Research, Snowbird, Utah, 1995.

1

John S. Witte was supported by FIRST Award CA-73270, as well as by grant CA-51923 from the National Cancer Institute.

View full text