SEE_ALSO
NAME
errorpars  Edit parameters related to data and coefficient errors.
USAGE
errorpars
DESCRIPTION
The chisquare computation used in all tasks in the fitting package depends on the availability of error estimates for each input datum. These can be defined using a variety of sources. Also, when using the amoeba fitting method, some knowledge of data errors is needed for properly computing estimates for the nonlinear fit coefficient errors.
Individual error bars for each datum can be input from tables or lists. If the input data come from a list, then the third column specifies the error bars' amplitude. If the input data are taken from a table, then the errcolumn parameter specifies the error source. If errcolumn contains a single word, it is taken as the name of the column that contains amplitudes of symmetrical errors in the input table. If the X and Y data come from different tables, then the table containing Y data is also assumed to contain error data. If errcolumn is left as an empty string (""), it is assumed that there is no input error column. In this case, a default error bar of length 1.0 is applied to all data points. This makes the chisquare minimization to be equivalent to the unweighted rootmeansquare (rms) residual minimization. The chisquare value output by the fitting tasks in the case of constant 1.0 error bars will be related to rms by:
d  1 chisq =  rms**2 d  cwhere d is the number of data points and c the number of fitted coefficients.
If an error bar in the error column is set to INDEF, the data point will be discarded from the fit. No individual error bars can be input from image data.
If no error column is available, or when inputing image data, an estimate for the data error can nonetheless be input to the tasks trough parameter sigma. This error will be applied equally to all data points, thus having the same effect on the minimization routine as the one described above for a "no error column" input (that is, unweighted fit), but making the chisquare values output by the tasks look more realistic. If sigma is set to anything different from INDEF and there is an input error column, the value of sigma will supersede the column input.
The errtype parameter has dual meaning. It is used for both building input data error bars, as well as to control nonlinear coefficient error computation. In its first role, if set to either "uniform" or "bars", input data error bars will be constructed as described above. If however errtype is set to "ccd", each individual datum error bar will be defined as a combination of Poisson and readout errors. This will supersede any eventual input error bar, either coming from a tabel column or from parameter sigma. This feature might be useful e.g. when inputing image data that otherwise can't be properly weighted.
The second use of the errtype parameter is related to the internal algorithms used by the nonlinear fitting tasks to estimate the coefficient errors. The LevenbergMarquardt method computes the coefficient errors directly from the covariance matrix and these errors are output by the tasks when the resample parameter is set to "no". But the downhill simplex method (amoeba) is unable, by itself, to calculate errors for the fitted function coefficients. The tasks thus resort to an independent Monte Carlo resampling technique to allow estimation of those errors. The resampling technique can be used also with the LevenbergMarquardt method. In that case the coefficient errors thus computed supersede the internal errors obtained from the covariance matrix.
This resampling technique works as follows: after fitting the function, the independent variable vector is replicated N times, each data point being replaced by the fitted function value plus noise with dispersion given by one of three criteria: (i) directly the dispersion of the original data around the fit; (ii) error bar of each point, or (iii) ccdtype error. For each replication, the function is refitted, and the error estimate for each coefficient is the dispersion of the N obtained values. Parameter errtype tells the tasks which one of the above ways must be used to generate the replicas's scatter. The "uniform" option uses the rms of the fit. The "bars" option uses the Y error bar of each datum, taken either from column specified by the errcolumn parameter, or from the sigma parameter. The "ccd" option assumes that the independent variable values are subject to counting statistics and may have an added component of readout noise.
Both "bars" and "ccd" modes use independent error information, so the coefficient errors thus computed should be representative of the "true" errors. The "uniform" mode must be used as a last resort, since it in some way "reuses" the information already present in the data (and model) to generate the replicas. Thus, coefficient errors computed with "uniform" error type should be taken with a little more caution.
It must be emphasized that coefficient errors obtained in that way are estimates which have a different nature from errors obtained from a covariance matrix. In particular, the error values will show a dependency on N, being more close to the "true" value the larger the N. Usually the errors estimated by resampling are larger than errors derived from the covariance matrix.
Since the process of refitting the function several times can be time consuming, two more parameters are provided to control the error computation. Parameter resample allows enabling/disabling of the tasks' resampling error feature. Parameter replicas specifies the number of replicas to use in resampling. It is good practice to disable error computation until an acceptable solution is found, and then refit with error computation enabled. Extensive testing with artificial data showed that the fitted coefficient's statistical distribution is reasonably well sampled when replicas is set to ~1520.
PARAMETERS
 (errcolumn = "" ) [string]
 Column name for the error column, if fitting from a table.
 (errtype = uniform) [string, allowed values: uniform  bars  ccd]

Error type. Used to both generate replicas in the resampling for coefficient error estimation (nonlinear fit only), and to interpret input error bars.
 (resample = no) [boolean]
 Estimate coefficient errors on each nonlinear fit ?
 (sigma = INDEF) [real, min=1.E10]
 Error bar to be applied to all data points. If set to anything different from INDEF, it supersedes eventual error bars defined by errcolumn parameter.
 (epadu = 1.) [real, min=0.]
 Gain, in electrons per ADU, for ccd noise model.
 (readnoise = 0.) [real, min=0.]
 Readout noise in electrons, for ccd noise model.
 (replicas = 15) [integer, min=2]
 Number of replications in nonlinear fit resampling.
BUGS
A detailed error analysis using artificial data sets showed that the coefficient error estimates generated by the resampling technique may under or overestimate the "true" errors measured in the data set ensembles by factors of up to 20 percent, in particular when Poisson noise is involved. On the other hand, the error estimates generated by the linear fitting task gfit1d, for powerseries polynomials and when using either ccdtype data or inputing data error bars to properly compute chisquare, underestimate the true errors by large factors (25). When error information is not input to gfit1d, however, the powerseries coefficient errors show the same behavior as in the nonlinear case.
Some bias, in the range of 0.52 percent, may also affect the coefficients. This seems to be highly dependent on functional form and signaltonoise ratio.
REFERENCES
"Error and bias in the STSDAS fitting package", I. Busko, ADASS VI, PASP Conf. Series 1997.
SEE ALSO
nfit1d, ngaussfit, n2gaussfit, gfit1d