You have only 20 out of 100 dead and predicted 30 and that's it? or you have more data?
Hello guys:
I have two sets of data, actual outcome (live vs. dead), which is categorical, and a risk adjustment predicted outcome, which is represented in continuous variables, i.e. percentages.
Now the actual mortality is 20% in a population of 100 subjects and the predicted mortality is 30% for the same population.
What statistical test can I use to compare the actual mortality with the predicted one to see if there is a significant difference between the actual and predicted values.
Thanks in advance guys.
You have only 20 out of 100 dead and predicted 30 and that's it? or you have more data?
Please forgive me if my last post was confusing, English is a second language to me.
Allow me to elaborate; I have a predicted APACHE IV mortality score for every patient in my cohort. So I have 100 predicted mortality rates. Also, the actual outcome after patient follow-up was documented. So in the duration of follow-up I have the actual outcome of all 100 enrolled patients (live vs. dead).
The 30% represents the median predicted mortality score, since the predicted mortality was not normally distributed, I calculated the median and IQR.
I want to see if there is a statically significant difference between these two variables.
Since one is categorical and the other is continuous, I don't know which test to use.
Could any please tell me what test is appropriate here?
Thanks in advance and sorry for the confusion.
Ok, let me get this straight, you are not Apache, but you calculated something with Apache. Ok.
You have 100 patients, some of them died and some of them are still alive. But you also made predictions about them dying or living. An you also have 100 APACHE IV mortality rates. And you want to see whether the predicted apache scores are close to the observed apache scores. Correct?
Then I suggest you do G=Sum{[(exp-obs)^2]/exp } and compare it with chi-square at 99 degrees of freedom at 95%.
Let me give you an example of my data maybe that will elaborate:
Patient out come --- APACHE predicted mortality rate
live --- 13.4
deceased --- 15.2
live --- 9.7
live --- 10.7
live --- 9.7
deceased --- 18.7
The overall mortality rate in my cohort is 20%. The median APACHE predicted mortality rate is 30%.
What test should I use to assess the variance between these two variables?
Again I am very sorry for the confusion and thank you very much for the prompt replies.
A ok, I see now. Ok, this is logistic regression. the patient outcome is the response variable, the dependent and the apache the independent, right?
do you know SPSS or R?
Yes I have SPSS V-16.
I did the logistic regression both using SPSS and Excel (using XLSTAT).
In SPSS, the dependant variable was patient outcome; the covariates were the predicted scores.
The SPSS gave me a lot of tables where can I find the significance (P-value)?
On XLSTAT, in the ''Evaluating the goodness of fit of the model'' table; R²=0.065, and Pr. > L.R. Chi-square=0.007.
In the ''Estimates of the parameters of the model (maximum likelihood)'' Pr. > Chi-square< 0.0001
How can I interrupt these results?
in spss the significance is in the significance column of the table with estimates of the parameters. XLSTAT no idea. But R^2=0.065 is very bad.
scerab (08-09-2011)
So according to this table, from SPSS, the difference between the actual and predicted rates are insignificant, right?
Omnibus Tests of Model Coefficients:
---Chi-square ---df --- Sig.
Step -124.810---112---.192
Block -124.810---112---.192
Model -124.810---112---.192
? can you tell me how you performed the logistic regression? the steps i mean. But i think yes, insignificant.
1-analysis
2-regression
3-Binary logistic
4-I selected hospital outcome (live vs. dead) as the ''dependant'' variable.
5-APACHE IV predicted mortality rate added to the covariates.
6-options ----> (selected) classification plot.
7-finally, click ''ok''.
In the output sheet I got a number of tables:
* Block 0: Beginning block
a- Variables in the equation
b- Variables not in the equation
* Block 1: Method = Enter:
a- Omnibus Tests of Model Coefficients
b- Model summery
c- Classification table
d- Variables in the equation
Was my protocol correct?
Thank you very much for you assistance.
Now i do not have spss in front of me, but i think try another thing also. I remeber there was something different with this approach.
Analyze? generalised linear models? generalized linear models and then select binomial distribution, response (dependent) variable is binary (0-1) and link function logit. Go to model and choose independent in the covariate(s) white box and then then model transfer the independent variable on the right hand side. Then press ok and see what you get.
Thank you very much for the directions.
I tried them but, did not get any results at all.
Were the steps I followed on SPSS not correct?
''Evaluating the goodness of fit of the model'' table; R²=0.065, and Pr. > L.R. Chi-square=0.007.
In the ''Estimates of the parameters of the model (maximum likelihood)'' Pr. > Chi-square< 0.0001
what is R^2 for logistoc regression?
I think you did something wrong then. As i said i do not have spss now. I remeber there was something with the approach you followed, but I do not remember what. Probably you did something wrong. The binary variable is scale like the apache right?
Tweet |