Statistics | Monte-Carlo for distributions
Statistics | Monte-Carlo for integrated weight-average s values
This function is for studying the statistical confidence limits in distribution analyses. The idea of the Monte-Carlo analysis is the generation of a large number (e.g. 100 - 1000) synthetic data sets that are similar to the experimental data set, but each with different random normally distributed noise. Each of these new data sets is analyzed, and the distributions are stored. The resulting set of distributions can then be studied, point by point, and the mean and probability contours can be calculated.
In essence, the Monte-Carlo statistics allows to build up a probability distribution function, and allows to study how much the noise of the data is translated into the distribution. As a consequence, the Monte-Carlo procedure can be useful for investigating what features of the distribution are a result of noise in the data, and what features are essential for describing the data. Two main features are highlighted in the implementation of Sedfit: The confidence contours of the distribution (i.e., it's shape), and the integral (weight-average s-values and loading concentrations across a peak or in a region).
The underlying assumption made in the current implementation is that the best-fit distribution represents the 'true' data well enough, so that the effects of noise in the simulated data and in the real data are the same. A second assumption made is that the error distribution of each data point is Gaussian.
The procedure is implemented here as follows:
1) with a distribution model, the best fitted values for all data points are stored. Important: The best-fit distribution has to be found before executing the Monte-Carlo analysis, and the residuals should be random and show no systematics (this may be assessed, for example, by the runs test, or by looking at the residuals bitmap). The best-fit found (automatically without regularization) will be taken as reference for generating new distributions.
2) the number of iterations N must be specified. It is suggested to perform a ‘test’ Monte-Carlo with a small (e.g. 10-100) number of iterations, first, to check the performance of the procedure, and then make a large statistics (100-1000) overnight.
3) The confidence level: After calculating the N new distributions based on the N generated data sets, the distribution plot will show 3 curves, the mean curve in black, and a blue and a red curve, which enclose a specified percentile of the data. For example, if the confidence level is 0.7, then the middle 70% is enclosed by the 0.15 (in blue) and the 0.85 (in red) quantile. Setting this confidence level changes these percentiles of the display and storage.
4) A directory for the storage of the calculated distributions has to be specified. This will contain N distributions, named 1.dat,2.dat,...,N.dat, as well as a distribution called "mean.dat", and different percentiles called "lim***.dat", e.g. lim160.dat for the 16% percentile.
5) The calculation of the Monte-Carlo iterations follows. Synthetic data sets are generated based on the previously found best-fit data (see point 1 above), with normally distributed noise added in the same magnitude as was obtained in the previous best-fit. Each of these new data sets is analyzed in the same way as the original distribution analysis. Each new calculated distribution is stored in memory, and in a file (see point 3 above).
6) The statistics of the curves is analyzed: As described in point 3, the mean and specified quantiles of the distributions are shown. For example, if the confidence level was 0.68, the curves shown in the display are the mean and 1 standard deviation contour for each point in the distribution. A continuous loop allows to probe new confidence levels, each of which will subsequently be displayed and stored in the specified directory. This way, a complete statistical analysis can be performed, without any need to re-calculate the Monte-Carlo iterations. This loop is stopped by pressing the ‘cancel’ button.
New in version 8.7: Monte-Carlo for integrated weight-average s values
This is a better statistics to use for determining the precision of the weight-average s-values across a peak or in a region. For example, when analyzing low concentrations of aggregate contaminations, the concentration may be too low to precisely locate the peak, yet the presence of material in a given s-range may be highly significant. In such a case, the confidence contours for the distribution may reach zero everywhere in the s-interval, but the integral across the region may be precisely determined and non-zero.
In this variation, after specifying the confidence level, the lower limit for the integration range and then the upper limit for the integration range needs to be specified. The output message box the will report the average sw value, it's 68% confidence interval, the average deviation, and the standard deviation of the samples. It will also report the average concentration in the integrated interval and its confidence interval. After closing the results message box, different integration limits can be chosen (no need to recalculate the Monte-Carlo simulations, just different statistical analysis of the result).