I am not sure what the problem is, but survival analysis should give you results. You will need to specify the independent variables eg different methods of contraception or age or parity. The dependent variable is, I think any contraceptive use at the end of follow up. The time metric will be start of contraception to dropout or sterilization or end of follow up. kaplan-Meier graphs will give you a visual representation and Cox's regression will give you Hazard Ratios.

Have you done a Kaplan-Meier Curve? Each participant will have a time variable called 'time-to-event' over a 24 month period, which indicates their time contribution in the study, and a variable called censor, which indicates whether they stop using contraceptives, loss to follow-up, or nothing happen.

Dear Adnan,

if You investigate different reasons of contraceptive discontinuation You may try competing risk analysis. It explores different outcomes (not one like survival).

The analog of log-rank test is Gray's test. The analog of Cox's regression is Fine & Gray regression model.

Also the problem with data may be short follow-up.

Good luck!

Adnan, I find your question very vague. I don't know what it means that you get no results at all. Also I apologize that I am not looking at the DHS web site. From what I get from your question, you do not have a straightforward survival analysis. You have at least two kinds of status changes initiation and stoppage. I don't know if your subjects enter in the two different statuses as well. At the very least this is a multispell design. This type of analysis allows for changing from one state to the other and back again, for example, as one might see with smoking cessation (although in that case everyone usually enters as a smoker, but may quit, relapse, quit again, etc.). Again without knowing the DHS method, I don't know if you also have to consider contraceptive type or reasons for initiating or stopping, in which case instead of a normal logit at each time point you would need a multinomial logit. You didn't mention what your frequency of measurement is either, except it is implicit you have multiple time points, but if for example, you only measured annually, you would have just three time points, and thus very little data for the complexities of the multispell design. Assuming more than one status change is possible you actually need the multispell design because otherwise you will be censoring subjects before they are through and thus causing all the problems survival analysis was meant to solve. Bob

Thank you everyone. You are right, as I read it, my question was unclear.

The problem is the following. The database that we are working with is not arrayed so that we can simply plug in the query in SPSS. The way it is arrayed is that while each respondent is a line (row), each month over a 24 month period is a variable. For each variable, the choices can be using no method, pregnant/ immediate post partum, or using a specific method (actually named).

SPSS gives the option of specifying time variable but wont accept our variables.

Perhaps if anyone can suggest where we can look at a data set (even with a few values) so that we can re-code our data to fit them to SPSS.

thanks

Based on the information provided in the 24 variables indicating contraceptive use status at each of the observation month, you will need to create the following two variables: 1) time, 2) failure

Time is the duration of contraceptive use (analysis time or time at risk). For example, someone started using a contraceptive at month 5 continued using it until month 10 (i.e., month 11 did not use the contraceptive) then value for your time variable is 6.

Failure should be coded ‘1’ if you know the actual duration of contraceptive use and 0 otherwise. For the above example the value for failure should be ‘1’. For segments of contraceptive use you are not sure when it started (left censoring) or when was it discontinued (right censoring) the value for failure should be ‘0’. For example if a person was using contraceptive during the first moth of observation then it is left censoring. Similarly if the person was using contraceptive during the last month of follow-up then it is right censoring. Note that a person can have both right and left censoring. Censored observations are included in the ’at risk’ pool but their outcome is not counted.

For method specific contraceptive discontinuation rate if the person switches between method it can be treated as right censoring or failure, which would depend on your study objective.

For individuals who have more than one segment (duration) of contraceptive use then there more complex syntax (for instance in Stata); however you can have several rows of data for a single individual, each row indication one segment of contraceptive use with its time and failure values.

I hope this helps!

From what I'm reading it sounds like you have a discrete time survival analysis problem, which really resolves to a logistic regression model, with month dummies that can have odds associated with them. I'm not sure SPSS supports this approach, although Stata and R do. This will also solve the problem of analyzing stopping and starting. In the month records you can also include time-dependent covariates --e.g. marital status change during one of the months?

Dear Ali and William. Thanks, both of your answers sound good. Will try them out. Our stats person knows stata, so this should be doable. Much appreciate it.

I had one additional thought. When doing repeated events analyses, it's probably safest to stratify the sample into "repeaters" and"non-repeaters" and do smoething like a Cox regression model for each and then comare the effects of the variables. The reason I say this is that repeaters and non-repeaters may be different, and variables may have different effects in these tow groups. Also, it's important to look at the numbers of repeaters. It's sometimes hard to do separate analyses if repeaters are few in number. Anyway, good luck. Keep us posted.

Thanks Marianne