Confidence Levels Used in Simulation

[MarkMackenzie]

When using a simulation model that involves reliability which should be used, the point estimate for the MTBF or the LCL on the MTBF?
If a point estimate is used then small and large samples with the same point estimate would be treated as equal. This seems inappropriate as there would be more confidence in parameter estimate using the data from a large sample than a small sample.
If a LCL is used then the distribution would be centred around the LCL not the MTBF.

[tarik]

I don't understand what are you trying to simulate and for what purpose? What are the variables and what is the ouput?

[MarkMackenzie]

I can use the simulation model for a number of different purposes.
One example is the use of blocksim to determine the reliability/availability of a system.
If you have the raw failure data (and it is exponential) do you use a point estimate for the MTBF or the LCL on the MTBF?
I can see arguments for both.

[tarik]

My recommendation is to use the estimated parameters for your blocks, especially if you have many of them in your RBD model. When you obtain the final results, you can then see how close you are to your requirement and observe how making changes in the parameters impact the results.

[MarkMackenzie]

Thanks tarik your answer is practical but I was hoping for a theoretically correct answer.
When using a simulation model that generates random numbers to represent a failure distribution of the population the estimated parameters are based of failure data collected from on a sample, generally tested in some way that is indicative of its actual use.
From the data, estimates of the distribution and its parameters can be made.
These parameter estimates are then used in a simulation model - such as Blocksim.
In most cases when inputting the parameters into a model only the mean is required for an exponential distribution. If you have a degree of confidence in the data (eg LCL on the mean) how can this be included in the model?
Hence, my question do you use the mean or the LCL.
In many cases there is not a given requirement to meet. The task is to quantify the reliability of a system. Using the correct parameter in a model is important. Where there is a large failure data set the LCL and the Mean are relatively close. Where the failure data is limited there is a significant difference between the LCL and the Mean.

[tarik]

I agree with the issues you brought up. The appropriate thing to do would be to use the Fisher Matrix information of each model. We are working on improvements in BlockSim to accommodate for uncertainties in the block's model parameters. Until the improvement is implemented, I can only recommend the 'practical' answer I gave you earlier.