The effect of behavioural changes over time on Cox proportional hazards estimates
(2018) STAN40 20172Department of Statistics
- Abstract
- This thesis considers bias when studying behaviours in a Cox proportional hazards model. In Cox proportional hazards regressions and cohort studies in general, measurements are often made during a limited period of time. Behaviours may, however, change rather dramatically over time, and if these changes are unknown, they will distort the results in the regression models. We study this problem in the context of the effects of smoking and physical activity on cardiovascular disease by simulating Cox proportional hazards models. Changes in behaviour are simulated with Markov chains in four scenarios. In each scenario we perform ten sets of simulations where each set has a different transition probability.
The first scenario considers a... (More) - This thesis considers bias when studying behaviours in a Cox proportional hazards model. In Cox proportional hazards regressions and cohort studies in general, measurements are often made during a limited period of time. Behaviours may, however, change rather dramatically over time, and if these changes are unknown, they will distort the results in the regression models. We study this problem in the context of the effects of smoking and physical activity on cardiovascular disease by simulating Cox proportional hazards models. Changes in behaviour are simulated with Markov chains in four scenarios. In each scenario we perform ten sets of simulations where each set has a different transition probability.
The first scenario considers a dichotomous variable indicating physical inactivity. We find that an increasing probability of changing behaviour will eventually completely dilute the baseline estimates. In the second, third, and fourth scenario we instead look at a smoking status variable containing the categories smoker, ex-smoker, and non-smoker. The three-category variable was in the regressions decomposed into the two dichotomous variables Smoker and Ex-smoker. In the second scenario we only allow transitions from smoker to ex-smoker. That leads to the hazard ratio estimates of Smoker going towards the hazard ratio of Ex-smoker. In the third scenario transitions are also allowed from ex-smoker to smoker. This results in the hazard ratios of the two variables moving towards each other, as the transition probabilities become larger. Lastly, the fourth scenario have Markov chains where non-smokers are additionally allowed to transition to smokers. There we find that the hazard ratios of Smoker and Ex-smoker go towards 1.0 when the transition probability of going from non-smoker to smoker is large. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/8938892
- author
- Hedberg, Jonatan LU
- supervisor
- organization
- course
- STAN40 20172
- year
- 2018
- type
- H1 - Master's Degree (One Year)
- subject
- language
- English
- id
- 8938892
- date added to LUP
- 2018-06-15 14:03:49
- date last changed
- 2018-06-15 14:03:49
@misc{8938892,
abstract = {{This thesis considers bias when studying behaviours in a Cox proportional hazards model. In Cox proportional hazards regressions and cohort studies in general, measurements are often made during a limited period of time. Behaviours may, however, change rather dramatically over time, and if these changes are unknown, they will distort the results in the regression models. We study this problem in the context of the effects of smoking and physical activity on cardiovascular disease by simulating Cox proportional hazards models. Changes in behaviour are simulated with Markov chains in four scenarios. In each scenario we perform ten sets of simulations where each set has a different transition probability.
The first scenario considers a dichotomous variable indicating physical inactivity. We find that an increasing probability of changing behaviour will eventually completely dilute the baseline estimates. In the second, third, and fourth scenario we instead look at a smoking status variable containing the categories smoker, ex-smoker, and non-smoker. The three-category variable was in the regressions decomposed into the two dichotomous variables Smoker and Ex-smoker. In the second scenario we only allow transitions from smoker to ex-smoker. That leads to the hazard ratio estimates of Smoker going towards the hazard ratio of Ex-smoker. In the third scenario transitions are also allowed from ex-smoker to smoker. This results in the hazard ratios of the two variables moving towards each other, as the transition probabilities become larger. Lastly, the fourth scenario have Markov chains where non-smokers are additionally allowed to transition to smokers. There we find that the hazard ratios of Smoker and Ex-smoker go towards 1.0 when the transition probability of going from non-smoker to smoker is large.}},
author = {{Hedberg, Jonatan}},
language = {{eng}},
note = {{Student Paper}},
title = {{The effect of behavioural changes over time on Cox proportional hazards estimates}},
year = {{2018}},
}