desperate for help
Message posted by care (via 134.50.237.225) on August 26, 2001 at 11:34 AM (ET)

Can anyone tell me what PROPORTIONAL HAZARDS regression means?! I atleast have a clue of what regression entails.

thanks in advance

READERS RESPOND:
(In chronological order. Most recent at the bottom.)

Re: desperate for help
Message posted by Josie LaPlante (via 207.5.236.168) on August 26, 2001 at 4:58 PM (ET)

Although linear regression can be used with a dichotomous dependent variable (or a polychotomous ordinal), when you are interested in assessing risk it isn't the best approach because there is nothing to constrain the risk to >=0 and <1, the range of probabilities. So other methods are used.

Conceptually, proportional hazards analysis is more similar to logistic regression (logit) than to linear regression.

Logistic regression, a variant used in the health sciences and some other fields, models the logarithm of the odds of an outcome, which is called the logit. Once you derive the logit you can convert it to an odds ratio. The odds ratio tells you the increase in risk that something will happen, given a particular independent variable. If the odds ratio is one, the odds or chance is equivalent. So for example, if the odds of getting lung cancer is 1.8 among smokers (I made up this number) it means the risk is 80% higher than for a non-smoker. A critical issue with odds ratios is who the comparison group is, because the risk is translated as the increase over that group.

Proportional hazards analysis is used in epidemiology primarily. It is based on the natural logarithm of the relative hazard. The relative hazard is the ratio of time to outcome when a particular risk factor (e.g., smoking)is present to time to outcome when the risk factor is not present. So it is used to model the change in the length of time to cancer if you smoke compared with someone who does not, rather than just the increase in risk of getting cancer if you smoke(which the logit would give you.) Proportional hazards analysis is a survival analysis method.

Another regression-based technique that you may hear about is discriminant [function] analysis. This technique is used to sort people into groups based on a set of characteristics, e.g., who is likely to drop out of high school and who is not. It is a classification technique.

I have probably over-simplified but I hope this gives you a sense of why alternative methods are used beyond linear regression.
If you want to pick up a good book that helps you decide between multivariable statistical approaches, I would recommend Michael H. Katz, Multivariable Analysis: A Practical Guide for Clinicians, Cambridge Univ Press, 1999. It's a paperback & not too expensive.