Maximum Likelihood Estimation (MLE)


MLE in Practice

Analytic MLE

Sometimes we can write a simple equation that describes the likelihood surface (e.g. the line we plotted in the coin tossing example) that can be differentiated. In this case, we can find the maximum of this curve by setting the first derivative to zero. That is, this represents the peak of a curve, where the gradient of the curve turns from being positive to negative (going left to right). In theory, this will represent the maximum likelihood estimate of the parameter.

Numerical MLE

But often we cannot, or choose not, to write an equation that can be differentiated to find the MLE parameter estimates. This is especially likely if the model is complex and involves many parameters and/or complex probability functions (e.g. the normal probability distribution).

In this scenario, it is also typically not feasible to evaluate the likelihood at all points, or even a reasonable number of points, in the parameter space of the problem as we did in the coin toss example. In that example, the parameter space was only one-dimensional (i.e. only one parameter) and ranged between 0 and 1. Nonetheless, because p can theoretically take any value between 0 and 1, the MLE will always be an approximation (albeit an incredibly accurate one) if we just evaluate the likelihood for a finite number of parameter values. For example, we chose to evaluate the likelihood at steps of 0.02. But we could have chosen steps of 0.01, of 0.001, of 0.000000001, etc. In theory and practice, one has to set a minimum tolerance by which you are happy for your estimates to be out. This is why computers are essential for these types of problems: they can tabulate lots and lots of values very quickly and therefore achieve a much finer resolution.

If the model has more than one parameter, the parameter space will grow very quickly indeed. Evaluating the likelihood exhaustively becomes virtually impossible - even for computers. This is why so-called optimisation (or minimisation) algorithms have become indispensable to statisticians and quantitative scientists in the last couple of decades. Simply put, the job of an optimisation algorithm is to quickly find the set of parameter values that make the observed data most likely. They can be thought of as intelligently playing some kind of hotter-colder game, looking for a hidden object, rather than just starting at one corner and exhaustively searching the room. The 'hotter-colder' information these algorithms utilise essentially comes from the way in which the likelihood changes as the they move across the parameter space. Note that it is precisely this type of 'rate of change' information that the analytic MLE methods use - differentiation is concerned with the rate of change of a quantity (i.e. the likelihood) with respect to some other factors (i.e. the parameters).

Other Practical Considerations

Briefly, we shall look at a couple of shortcuts and a couple of problems that crop up in maximum likelihood estimation using numerical methods:

Removing the constant

Recall the likelihood function for the binomial distribution:

In the context of MLE, we noted that the values representing the data will be fixed: these are n and h. In this case, the binomial 'co-efficient' depends only upon these constants. Because it does not depend on the value of the parameter p we can essentially ignore this first term. This is because any value for p which maximises the above quantity will also maximise

This means that the likelihood will have no meaningful scale in and of itself. This is not usually important, however, for as we shall see, we are generally interested not in the absolute value of the likelihood but rather in the ratio between two likelihoods - in the context of a likelihood ratio test.

We may often want to ignore the parts of the likelihood that do not depend upon the parameters in order to reduce the computational intensity of some problems. Even in the simple case of a binomial distribution, if the number of trials becomes very large, the calculation of the factorials can become infeasible (most pocket calculators can not represent numbers larger than about 60!). (Note: in reality, we would quite probably use an approximation of the binomial distribution, using the normal distribution that does not involve the calculation of factorials).

Log-likelihood

Another technique to make life a little easier is to work with the natural log of likelihoods rather than the likelihoods themselves. The main reason for this is, again, computational rather than theoretical. If you multiply lots of very small numbers together (say all less than 0.0001) then you will very quickly end up with a number that is too small to be represented by any calculator or computer as different from zero. This situation will often occur in calculating likelihoods, when we are often multiplying the probabilities of lots of rare but independent events together to calculate the joint probability.

With log-likelihoods, we simply add them together rather than multiply them (log-likelihoods will always be negative, and will just get larger (more negative) rather than approaching 0). Note that if

a = bc

then
log(a) = log(b) + log(c)

So, log-likelihoods are conceptually no different to normal likelihoods. When we optimise the log-likelihood (note: technically, we will be minimising the negative log-likelihood) with respect to the model parameters, we also optimise the likelihood with respect to the same parameters, for there is a one-to-one (monotonic) relationship between numbers and their logs.

For the coin toss example above, we can also plot the log-likelihood. We can see that it gives a similar MLE for p (note: here we plot the negative of the log-likelihood, merely because most optimisation procedures tend to be formulated in terms of minimisation rather than maximisation).

Model identification

It is worth noting that it is not always possible to find one set of parameter values that uniquely optimises the log-likelihood. This may occur if there are too many parameters being estimated for the type of data that has been collected. Such a model is said to be 'under-identified'.

A model that attempted to estimate additive genetic variation, dominance genetic variation and the shared environmental component of variance from just MZ and DZ twin data would be under-identified.

Local Minima

Another common practical problem when implementing model-fitting procedures is that of local minima. Take the following graph, which represents the negative log-likelihood plotted by a parameter value, x.
Model fitting is an iterative procedure: the user has to specify a set of starting values for the parameters (essentially an initial 'first guess') which the optimisation algorithm will take and try to improve on.

It is possible for the 'likelihood surface' to be any complex function of a parameter value, depending on the type of model and the data. In the case below, if the starting value for parameter x was at point A then optimisation might find the true, global minimum. However, if the starting value was at point B then it might not find instead only a local minimum. One can think of the algorithm crawling down the slope from B and thinking it has reached the lowest point when it starts to rise again. The implication of this would be that the optimisation algorithm would stop too early and return a sub-optimal estimate of the parameter x. Avoiding this kind of problem often involves specifying models well, choosing appropriate optimisation algorithms, choosing sensible starting values and more than a modicum of patience.



Return to front page
Site created by S.Purcell, last updated 20.05.2007