Tuesday, April 7, 2015

Computer Simulations

Approximations

Mathematical models, even very good ones, provide estimates that are approximately true.  The world isn’t perfectly round, but a mathematical model that assumes that it is will still tell you about how far you will need to travel to get to the other side.  A flat screen representation of a little part of the world isn’t correct either, but it can still get you around town.

Characterizing a risk is theoretically not difficult either.  Plugging an estimated dose into a simple dose-response model will yield an estimate of the magnitude of the undesirable response.  But, for most food safety issues that are usually concerned with effects that are too small to measure reliably there is a problem:  The simple dose-response model probably isn’t a very good approximation.  In fact, there usually are range of estimates that could be correct, and sometimes that range can be very wide.  Therefore, an uncertainty analysis may be necessary to show just how wide the estimates can be.  Public health risk models often must have another dimension to them as well.  Because of differences in food consumption patterns and individual differences in what a chemical does after it is ingested, a dietary risk model will usually incorporate statistical components that describe population variability.  As a result of all that added complexity, getting an approximately correct estimate for a problem that is worthy of quantitation may require a complex model or simulation.

Monte Carlo Simulation

When two or more mathematical probability “input” distributions are present in a model, the most common method for estimating the probability of an estimated value is called Monte Carlo simulation.  This “iterative” method uses a computer to repeatedly sample random draws from each input distribution to calculate the model output value; the resulting distribution is used to characterize the distribution of the final estimate. 

A one-dimensional Monte Carlo simulation can be used to calculate the uncertainty associated with a population estimate.  For example, the impact of the uncertainty of a per capita exposure assessment and the uncertainty associated with a dose-response estimate can be combined to estimate the uncertainty associated with a disease frequency.  Uncertainty simulations may include both statistical uncertainties (e.g. sampling error) as well as probability trees that represent theoretical uncertainty. 

A one-dimensional Monte Carlo simulation can also be used to characterize the uncertainty of the magnitude of an effect in an individual person.  In this case, statistical distributions that describe population characteristics must either be replaced with individual values (i.e. a distribution describing body weights can be replaced with the actual body weight for the individual) or be treated as statistical uncertainties when individual values are unknown.

But, to characterize the uncertainty associated with an effect of varying magnitude in a population, then one dimension is not enough.  Two dimensional public health computer simulations that employ one iterative loop embedded inside another one.  The inner loop typically randomly samples from distributions that describe the frequency of occurrence of a value in a population (e.g. a distribution of body weights), while the outer loop samples probability distributions and probability trees that are intended to represent uncertainty.  The end result is a two dimensional array of values that characterize the uncertainty associated with a frequency distribution.   In order to simplify presentation, the uncertainty dimension is often represented with a central value (i.e. the median or arithmetic average) and confidence intervals (e.g. the 5th and 95th percentiles) for each population estimate.

Other Simulation Methods

An alternative method for generating estimates from uncertain and variable input values is to use stratified sampling.  If the distributions aren’t comprised of discrete values already (e.g. an empirical distribution), this involves breaking each continuous distribution down into a set of fixed intervals with an average value for each, and then calculating all possible permutations.  This technique is not as easy to implement, but since it doesn’t use random numbers it gives a simulation result that is the same every time.

Latin Hypercube sampling is a synthesis of Monte Carlo and stratified sampling.  As in stratified sample each distribution in broken down into fixed intervals, but then each interval is sampled with an equal number of random draws.  This method has the advantage of sampling the entire distribution, but also produces more reproducible results with fewer iterations.

Probability Trees vs. Sensitivity Analysis

A probability tree is a tool for including the uncertainty arising from alternative theoretical assumptions in a quantitative uncertainty analysis.   Sensitivity analysis is a widely used alternative method for representing theoretical uncertainty that depends on providing entirely different estimates for each alternative assumption. 

What almost always happens with sensitivity analysis is that one assumption is used to for the primary estimate that is actually used to inform subsequent decision making, while the others are recognized as “also rans”.  As a result, for all practical purposes, this means that one theory is assigned a probability of one, while the others have a probability of zero.  As a form of political rhetoric, a sensitivity analysis is often used as a technique for acknowledging scientific criticism without really giving it any credence.  If the other theories really are crackpot ideas that have no scientific credibility whatsoever anyway, that is not a bad thing.  Im amy case, including the alternative estimates does give a reader the option of reversing the probability assignment of the author.  Furthermore, a sensitivity analysis may be used to show that an estimate is not heavily influenced by a particular assumption.  But, if the alternative assumptions really are both plausible and influential, then not including them in the uncertainty analysis will result in confidence intervals that are unjustifiably narrow. 

General References

Cullen AC and Frye HC (1999).  Probabilistic Techniques in Exposure Assessment: A Handbook for Dealing with Variability and Uncertainty in Models and Inputs.  Plenum, New York.

Morgan MG, Henrion M, and Small M (1992).  Uncertainty: A Guide to Dealing with Uncertainty in Quantitative Risk and Policy Analysis.  Cambridge University Press, Cambridge.

Official Post Soundtrack

Kraftwerk (1981).  Computer World.  In: Computer World, Track 1

Post Notes

Thesis #28.  Part of Risk Assessment Paradigm Thread

No comments:

Post a Comment