survival analysis and logistic regression

[MonDie]

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1380930/

 

They investigate the health associations of frequent church attendance by doing a multiple logistic regression and a survival analysis that adjusts for other factors (Cox proportional hazards). In layman's terms, how do survival analyses and logistic regressions proceed? My understanding is that they're related techniques.

[Acme]

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1380930/

 

They investigate the health associations of frequent church attendance by doing a multiple logistic regression and a survival analysis that adjusts for other factors (Cox proportional hazards). In layman's terms, how do survival analyses and logistic regressions proceed? My understanding is that they're related techniques.

IANE (I Am No Expert) but the way I read that page, only logistic regression is a technique. It sounds to me like survival analysis simply refers to the overall aim/goal of looking at who lived and who died. Notice there is no mention of survival analysis in their section titled Methods. (I don't see the term 'survival analysis' at the page at all; where did you find it?)

METHODS: The association between frequent attendance and mortality over 28 years for 5286 Alameda Country Study respondents was examined. Logistic regression models analyzed associations between attendance and subsequent improvements in health practices and social connections.

Edited August 11, 2014 by Acme

[MonDie]

Cox proportional hazards models with time-dependent covariates were used to analyze the relationship between attendance and mortality.[34][35] This method takes into account changes in attendance and adjustment variables reported at each new survey during follow-up.

 

[...]

 

Multiple logistic regression was used to assess associations between attendance and 1965-through- 1994 changes in health practices, body mass index, and social connections.

---

 

https://en.wikipedia.org/wiki/Proportional_hazards_models

 

Proportional hazards models are a class of survival models [i.e. survival analysis] in statistics. Survival models relate the time that passes before some event occurs to one or more covariates that may be associated with that quantity of time.

[Acme]

---

 

https://en.wikipedia.org/wiki/Proportional_hazards_models

So if you know that much, or at least know how to look it up, why are you asking here? Is that information what constitutes your 'My understanding is...'? Isn't it possible that such complex analyses have no effective layman's terms? Perhaps in layman's terms, they proceed cautiously.

[Bignose]

Survival analysis is an optimization technique wherein the many different inputs into the optimization are treated like 'genes'. The survival part comes from the fact that various solutions swap genes with one another to make children solutions. The least optimal solutions are then killed off (i.e. don't survive). And the surviving solutions are again bred -- swapping genes of the solution. See also evolutionary algorithm, or genetic algorithm.

 

A logistic regression is a technique where you discover the relative weights of the various inputs to create the most accurate prediction of some kind of sorting or categorical determination. A good example is handwriting analysis for optical character recognition. I.e. you sort a digit into the 1, 2, 3, 4,..., 9, or 0 bin. This is compared to a linear regression on which the weights of the inputs are adjusted to best match a given number.

 

Both techniques fall into the broader category of algorithms called machine learning.

[MonDie]

For my response, I was looking for one of the powerpoints I viewed previously, then I found this.

 

stat.ufl.edu/~winner/sta6934/surv.ppt

 

It looks a lot more informative.

 

edit2: darn those duckduckgo redirects.

 

---

 

Oh, and here's BigNose! Yay!

Edited August 11, 2014 by MonDie

[Acme]

For my response, I was looking for one of the powerpoints I viewed previously, then I found this.

stat.ufl.edu/~winner/sta6934/surv.ppt

It looks a lot more informative.

edit2: darn those duckduckgo redirects.

---

Oh, and here's BigNose! Yay!

Making my attempt at answering your question Łame. Ya win a few, ya lose a few. Interesting study anyway; I saved the PDF & will read.

[MonDie]

Survival analysis is an optimization technique wherein the many different inputs into the optimization are treated like 'genes'. The survival part comes from the fact that various solutions swap genes with one another to make children solutions. The least optimal solutions are then killed off (i.e. don't survive). And the surviving solutions are again bred -- swapping genes of the solution. See also evolutionary algorithm, or genetic algorithm.

 

A logistic regression is a technique where you discover the relative weights of the various inputs to create the most accurate prediction of some kind of sorting or categorical determination. A good example is handwriting analysis for optical character recognition. I.e. you sort a digit into the 1, 2, 3, 4,..., 9, or 0 bin. This is compared to a linear regression on which the weights of the inputs are adjusted to best match a given number.

 

Both techniques fall into the broader category of algorithms called machine learning.

 

Could you run through some examples?

 

---

 

I find their methods suspicious.

 

"Table 3 presents the results comparing frequent and infrequent attenders on improvements in health practices, body mass index, and social connections between 1965 and 1994. (Strawbridge et al)"

 

"Health practices and social connections could either confound the relationship between attendance and mortality (persons with good health practices and stronger social connections are frequent attenders of religious services) or act as intervening variables on a causal pathway between attendance and mortality. (Strawbridge et al)"

 

Correct me if I'm wrong, but shouldn't they have investigated the reverse relationships? If they wanted to show that A causes B, they should have shown not only that A predicted increases in B, but also that B did not predict increases in A.

They compared high-attenders to low-attenders on changes in smoking and drinking habits, but as far as I can tell, they did not compare high-smokers to low-smokers on changes in attendance. If not smoking predicted rises in attendance and high attendance predicted reductions in smoking (B and A predict eachother), that would suggest that either they share a common cause or they cause eachother, both of which would still explain the correlations at baseline.

 

Am I not understanding something about their methods?

Edited August 19, 2014 by MonDie

[Bignose]

Could you run through some examples?

Not really. Whole books are written on these topics. I tend to write long replies, but not that long...