Hi Anonymous,

We would normally only calculate length of stay (LOS) as a performance indicator for recovered cases. There are several factors that might affect the LOS in this case such as how quickly the case was detected and admitted for treatment, the compliance of the caretaker in adhering to protocols, absenteeism and the type discharge criteria being used for the programme. The LOS should be compared to expected averages for the protocols in use.

For negative outcomes, the LOS might give results that are difficult to interpret depending on what is included. Deaths frequently occur early in treatment if the child has complications or is very malnourished on enrolment in the programme. Non-cured (non-responders) will frequently be defined by a fixed LOS in the programme (e.g. 3 - 4 months).

For defaulters, the pattern of defaulting is contextual and may occur shortly after treatment begins (for example if access to the programme is difficult for whatever reason) or may occur later on when the child looks well but has not reached discharge criteria (as may occur when the further benefits of treatment outweighs the effort or cost of the carer and child attending treatment). Often there is a mix of these.

Whatever association you find it would be important to describe the underlying mechanisms.

I hope I understand your question ...

It will be a complicated (perhaps impossible) analysis because the "not-recovered" group will consist of many different outcomes (e.g. death, default / lost, transfer, and non-response). Some of these outcome classes are not "crisp". There may (e.g.) be wrongly classified deaths in the default / lost class which and also in the transfer class. Getting crisp classes requires active follow-up of cases in those classes. This level of follow-up is not common even in well-funded NGO programs.

If we can get crisp sets we still have problems. For example, data collected in SQUEAC coverage assessments has shown that defaulting is not uniform in time. We often see a peak in defaulting after the initial visit. This is a particular problem in programs in which the costs to beneficiary households are high. Another set of defaults is early default due to poor initial response which is often due to poor compliance by clinics or at home. Late defaults do occur and may be due to non-response due to poor compliance or illness. Defaults may also be due to RUTF stock-outs or migration which may occur at random. One set that I have seen in follow-up are "self-discharged as cured". These are cases who respond well to treatment and would have been discharged as cured but do not attend the proof of cure visit. These will have slightly shorter length of stay than cured cases particular if response is rapid and marked. There may be other reasons for default.

Another issue is that to the case defintion for non-response usually has a time limit of about twice the expected recovery period. This means that average lengths of stay for non-response will be longer than in cured cases. This may be a particular problem if monitoring of response to treatment is poor. This may not fix the problem. If monitoring response to treatment is good then we might see many transfers to stabilisation centres or hospital early in the treatment episode. These may go on to die and not recorded properly. They may or may not return to OTP. If they return to OTP as a continuation of the initial admission (and we can recognise this) then we might expect to record a longer length of stay compared to those case who responded well to the OTP treatment. If they are treated as a new admission on return then we might expect a shorter or similar length of stay to cured cases. If they don't return we will see a shorter length of stay.

I'm sure you'll be able to think of other complications.

All these complications make it difficult to know what any association we find actually means and whether the association will be stable across settings.

I assume that you are looking at an analysis that models:

I'd be concerned about the complications above. Careful description and analysis may work.

I think the best way to work around these complications is to work prospectively and standardise case-definitions &c. and to have active follow-up of transfers and defaults with a view of correct classification and in looking at the reasons for negative outcomes. This would be operational research. If you are going to do operational research it may be better to concentrate on identifying and addressing causes for non-response and other not-recovered outcomes.

I hope this is of some use.

Oh ... I see Paul got there first. We are in agreement. Paul's response is less rambling than mine.