topic Re: Procedure for predicting a dichotomous outcome and output predicted values in Statistical Procedures

Procedure for predicting a dichotomous outcome and output predicted values

seisel — Wed, 23 Jan 2019 21:52:31 GMT

I would like to get a model to predict the dichotomous yes/no outcome (yes sleep disturbance/ no sleep disturbance) and then output that to a new dataset for each individual at each time point (is this possible?). There are many individuals that do not have all 4 time points and we would like to be able to include everyone. So far I have only been able to get the model to output means/probabilities and they are the same for each participant and each time point (or just for individuals with data at that time point). Any suggestions on what model/code I should be using? For example, if my data and model look like this:

data sleep;

input ID Months SleepDist;

cards;

1 0 1

1 10 1

1 25 0

1 39 1

2 0 1

2 10 .

2 25 1

2 39 0

3 0 1

3 10 .

3 25 0

3 39 .

4 0 1

4 10 .

4 25 0

4 39 0

5 0 1

5 10 1

5 25 1

5 39 0

6 0 0

6 10 .

6 25 0

6 39 1

;

proc genmod data= sleepdata;

class ID /missing;

model SleepDist = Months / dist=bin expected;

repeated subject=ID / type=un covb corrw;

output out = try

pred= pred;

run;

What I get:

ID	Months	SleepDist	Predicted Value
1	0	1	0.164
1	10	1	0.269
1	25	0	0.485
1	39	1	0.694
2	0	1	0.164
2	10	.	0.269
2	25	1	0.485
2	39	0	0.694
3	0	1	0.164
3	10	.	0.269
3	25	0	0.485
3	39	.	0.694

...

I would like:

ID	Months	SleepDist	Predicted Value
1	0	1	1
1	10	1	1
1	25	0	0
1	39	1	1
2	0	1	1
2	10	.	1
2	25	1	1
2	39	0	0
3	0	1	1
3	10	.	0
3	25	0	0
3	39	.	1

…

Thanks!

Re: Procedure for predicting a dichotomous outcome and output predicted values

Reeza — Wed, 23 Jan 2019 22:14:39 GMT

You need to decide what your cut off is. Most classification algorithms give you a probability and then you use a cutoff to convert that to a 0/1. Ie if predicted value > 0.7 then 1, else 0. Or 0.5, or 0.3. You usually pick the cutoff by looking at the classification tables or the ROC graphs.

https://www.researchgate.net/post/How_do_I_calculate_the_best_cutoff_for_ROC_curves

Re: Procedure for predicting a dichotomous outcome and output predicted values

PGStats — Wed, 23 Jan 2019 22:48:10 GMT

Also, be careful what probability you are modelling. For a neutral cutoff (0.5) you would be doing:

proc genmod data=sleep;
   class ID;
   model SleepDist(event='1') = Months / dist=bin expected;
   repeated subject=ID / type=un covb corrw;
   output out = sleepPred pred=pred;
run;

data sleepProb;
set sleepPred;
SleepDistPred = pred > 0.5;
run;

(event='1') tells the procedure which probability you are modelling.