Logistic regression for prediction of outcome?

#1
I am designing a project and am inexperienced with the statistical analysis to say the least. I am trying to determine the predictive factors associated with the risk of developing an outcome. Its a complex project as their are multiple variables with the potential to impact the incidence of the outcome occurring.
Example outcome: pleural effusion
Variables: Fluid administration, surgery type, heart failure...the list goes on. I am assuming a logistic regression is the best way to tackle this but would appreciate some input
 

ondansetron

TS Contributor
#2
What question are you trying to answer more specifically? Just if a patient will have any pleural effusion as found (by what method)? Xray? Ultrasound? Only the effusions that get tapped and drained? Are these effusions measured for volume estimates or literally just we identify it or not on (insert method here)?
 
#4
What question are you trying to answer more specifically? Just if a patient will have any pleural effusion as found (by what method)? Xray? Ultrasound? Only the effusions that get tapped and drained? Are these effusions measured for volume estimates or literally just we identify it or not on (insert method here)?
specifically trying to determine correlational factors that might lead to the outcome of pleural effusion requiring drainage after surgery. So for example, do those with surgery times over 100mins have more likelihood, or do those that have a fluid balance over a certain value in the ICU have more pleural effusions prior to discharge
 
#5
I would guess they have a pool of patients with binary defined plural effusion and want model predictors. Per @ondansetron how is outcome defined and formatted.
yes essentially. pleural effusion requiring drainage after surgery. The argument currently is that drainage tubes are removed too early which leads to effusion formation, solution: leave tubes in longer. my hypothesis is, that is nonsense, earlier drain removal makes no difference. All these other factors effect effusion formation/development.
 

hlsmith

Not a robit
#6
Well yes this sounds like multiple logistic regression. Let us know if you have particular questions.

What is your sample size and what are the proportion with drain needed yes/no?

Also, turning time into a binary variable can mean results are sample specific. another approach is to leave them in the model and report them as say a standard deviation increase or decrease in time.
 

ondansetron

TS Contributor
#7
yes essentially. pleural effusion requiring drainage after surgery. The argument currently is that drainage tubes are removed too early which leads to effusion formation, solution: leave tubes in longer. my hypothesis is, that is nonsense, earlier drain removal makes no difference. All these other factors effect effusion formation/development.
From a standard logistic regression you won’t be able to conclude that “earlier removal makes no difference”.
 
#8
I am trying to determine the predictive factors associated with the risk of developing an outcome.
Example outcome: pleural effusion
I don't know what "pleural effusion" is. But it might be a continuous varaiable, so that it is on a ratio scale. If you recode that into high and low of 1 and 0, that is to dichotomize, you will loose information. (Just like you will loose information if you recode a persons length in centimeter into tall and short). It mighth be as to throw away 1/3 of your sample.

(To dichotomize an explanatory variable as time, will also loose infomation.)

It also makes the statistical analysis more difficult.

Why throw away information and make things more difficult?