American Journal of Epidemiology Vol. 152, No. 3 : 289
Copyright © 2000 by The Johns Hopkins University School of Hygiene and Public Health
LETTERS TO THE EDITOR |
RE: "USE OF TWO-SEGMENTED LOGISTIC REGRESSION TO ESTIMATE CHANGE-POINTS IN EPIDEMIOLOGIC STUDIES"
Institut für Medizinische Statistik und Epidemiologie Technische Universität München München, Germany
Statistisches Beratungslabor Universität München München, Germany
 |
INTRODUCTION |
|---|
 TOP INTRODUCTION REFERENCES |
|---|
Pastor and Guallar (1
) have described an interesting problem in epidemiology, the use of two-segmented logistic regression. In the introduction to their paper, they stated that none of the usual methods provides inference procedures for estimating the location of the change-point. However, a method for estimating change-points was described several years ago (2). The situation considered there described a threshold with no effect below a certain level.
There has been much discussion about appropriate test statistics (3–6). More recently, an exact algorithm for estimating breakpoints in segmented generalized linear models was described (7). Finally, the results were compared by using different statistical models (8) that showed how the estimation can depend on the model used for the analysis. Careful modeling and interpretation seems to be very important.
When consequences are important, such as in the assessment of threshold values in occupational medicine, it is obvious that only one value is required. The work of Pastor and Guallar (1) is a first step, but a lot more must be done to enable this method to be used in practice.
Regarding the example used in the paper (1), several questions remain unanswered. How can a threshold be established? Is a formal test available? No value of any likelihood function (with and without a threshold) was given. What is the interpretation of a threshold or change-point if the corresponding parameter (ß2) is nonsignificant? In the example considered, it is unclear whether the parameters ß1 and ß2 are indeed significantly different from zero (table 2 (1)). In the paper, the authors compared five models that led to different estimates of the change-point. How can we discriminate between these models?
|
NOTES |
|---|
Editor's note: In accordance with Journal policy, Drs. Pastor and Guallar were asked if they wished to respond to this letter but chose not to do so.
|
REFERENCES |
|---|
|
TOPINTRODUCTIONREFERENCES |
|---|
-
Pastor R, Guallar E. Use of two-segmented logistic regression to estimate change-points in epidemiologic studies. Am J Epidemiol 1998;148:631–42.
[Abstract/Free Full Text] - Ulm K. A statistical method for assessing a threshold in epidemiological studies. Stat Med 1991;10:341–9.[Web of Science][Medline]
- Cox C. Threshold dose-response models in toxicology. Biometrics 1987;43:511–23.[Web of Science][Medline]
- Ulm K. On the estimation of threshold values. (Letter). Biometrics 1989;45:1324–6.
- Cox C. On the estimation of threshold values. (Letter). Biometrics 1989;45:1327–8.
- Silvapulle JM. On testing for threshold values. (Letter). Biometrics 1991;47:1629.
- Küchenhoff H. An exact algorithm for estimating breakpoints in segmented generalized linear models. Comput Stat 1997;12:235–47.
- Küchenhoff H, Ulm K. Comparison of statistical methods for assessing threshold limiting values in occupational epidemiology. Comput Stat 1997;12:249–64.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  R. Pastor and E. Guallar RE: "USE OF TWO-SEGMENTED LOGISTIC REGRESSION TO ESTIMATE CHANGE-POINTS IN EPIDEMIOLOGIC STUDIES" Am. J. Epidemiol., March 15, 2001; 153(6): 615 - 615. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||