1
$\begingroup$

I have dichotomous data where some of my independent variables are categorical, some are continuous and some are binary (0/1).

My dependent is a binary response (Fail/NoFail, 0/1).

The data is some readings collected every day over a period of time.

The goal is to use this data and see if we can figure out cause of failure, the end response.

Example data format

Date, Type, Mileage, S1, S2, S3, ... , Response

03/02/2013,A,32000,1,0,1,..., 1

03/03/2013,B,32400,0,0,0,...,0

03/04/2013,C,45000,0,1,1,...,1

Can we do time series modeling? Any other modeling techniques that can be used?Any suggestions on what type of other exploratory analysis can be used to figure out patterns in data?

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ Dichotomous is just another word for binary. So, you have a binary response. The question still seems very broad to me. Yes, given a date variable time series might make sense, but an aim of establishing cause implies emphasis on variables related to cause. $\endgroup$
    – Nick Cox
    Commented Feb 27, 2014 at 14:42
  • $\begingroup$ Any Packages in R that can be used to model the data using time series, since the input variables are a mixture of continuous,Binary and categorical. $\endgroup$
    – shi
    Commented Feb 27, 2014 at 14:48
  • $\begingroup$ This sounds like some sort of survival analysis to me. $\endgroup$
    – RioRaider
    Commented Feb 27, 2014 at 15:09
  • $\begingroup$ @RioRaider's guess looks good to me. If you are following individuals to failure, keeping track of those individuals is essential. $\endgroup$
    – Nick Cox
    Commented Feb 27, 2014 at 15:57

1 Answer 1

Reset to default
2
$\begingroup$

Logit model would work here. Look up "logit" command in your statistical package. Your dependent variable would be converted to odds ratio. Independent continuous variables are obvious. Categorical variables will be converted into dummy/indicator variables. If you have a variable $x_i$ with M categories (unordered), you'll end up with M-1 dummies. You can find this all in the help for your command, such as the one in Stata.

Yes, this can be a time series model too, if your independent variables are time dependent. You have to be careful with the underlying assumption of independence of the error terms, of course. The failure reason in this period may be correlated to the failure reason in the next period, if this dependence is not fully captured by your independent variables, then it'll pop up in the error terms, which will be an issue.

Improve this answer
$\endgroup$
2
  • $\begingroup$ we used logistic regression, but since my output response has more non fails than fails. About 97000 non fails and 8000 fails. over a period of 30 days.The model we fit does not explain any variability in the model. $\endgroup$
    – shi
    Commented Feb 27, 2014 at 14:56
  • 1
    $\begingroup$ i used logit regression to model default probabilities, which could be in the order of 0.01% per time period. the fact that the fail/success rates are 10% should not be a problem per se. yoor issue could be with a functional form of or the selection of independent variables. did you look at Cox hazard models? $\endgroup$
    – Aksakal
    Commented Feb 27, 2014 at 15:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.