Bad data

An additional advantage derived from plotting the residuals is that it can aid in detecting a bad data point. If one of the points noticeably deviates from the trend line, it is probably due to a mistake in sampling, analysis, or reporting. The best action would be to repeat the measurement. However, this is often impractical. The alternative is to reject the datum if its occurrence is so improbable that it would not reasonably be expected to occur in the given set of experiments. 

Richardson, T. H. Reproducible Bad Data for Instruction in Statistical Methods, /. Chem. Educ. 1991, 68, 310-311. 

To determine if a process unit is at steady state, a program monitors key plant measurements (e.g., compositions, product rates, feed rates, and so on) and determines if the plant is steady enough to start the sequence. Only when all of the key measurements are within the allowable tolerances is the plant considered steady and the optimization sequence started. Tolerances for each measurement can be tuned separately. Measured data are then collec ted by the optimization computer. The optimization system runs a program to screen the measurements for unreasonable data (gross error detection). This validity checkiug automatically modifies tne model updating calculation to reflec t any bad data or when equipment is taken out of service. Data vahdation and reconciliation (on-line or off-line) is an extremely critical part of any optimization system. 

For example, many include a low-level alert that automatically alerts the technician when acquired vibration levels are below a pre-selected limit. If these limits are properly set, the alert should be sufficient to detect this form of bad data. 

Unfortunately, not all distortions of acquired data result in a low-level alert. Damaged or defective cables or transducers can result in a high level of low-frequency vibration. As a result, the low-level alert will not detect this form of bad data. However, the vibration signature will clearly display the abnormal profile that is associated with these problems. 

I know of no experienced practitioner of chemometrics who would blindly use the full spectrum when applying PLS or PCR. In the book Chemometrics by Beebe, Pell and Seasholtz, the first step they suggest is to examine the data. Likewise, Kramer in his new book has two essential conditions The data must have information content and the information in the data must have some relationship with the property or properties which we are trying to predict. Likewise, in the course I teach at Union Carbide, I begin by saying that no modeling technique, no matter how complex, can produce good predictions from bad data. ... 

The benefits of producing good data are therefore broad and impinge on all of our daily lives, whether it is food, environment, health or trade. Laboratories that produce valid measurements have a higher status in the analytical world, since they produce data that are demonstrably traceable to a reference standard and reliable, with the cost of correcting bad data being lower. This means that such laboratories have a better chance of competing in the open market. 

Once the occurrence of bad data is detected (through the previous procedure), we may either eliminate the sensor or we may assume simply that it has suffered degradation in the form of a bias. In the latter case, estimates of the bias may allow continued use of the sensor. That is, once the existence of a systematic error in one of the sensors is ascertained, its effect is modeled functionally. 

Checking the data quality is strongly recommended inspection of the Wilson plot and data reduction statistics is very useful in judging the extent of the resolution to which the data can realistically be used. Pathologically bad data, for example those from a split crystal, twinned data, systematically incomplete data, low resolution overloads will always make model building and refinement hard if not impossible. 

It is important to understand that all of these steps equally impact the success of the projea. Not understanding the system can lead to a problem definition that when solved does not help improve the system. An unreasonable hypothesis is by definition defining the wrong problem. Poorly designed experiments and ulty experimental technique leads to bad data. And improper data analysis can transform even good data into useless results. Careful attention to all steps in this process is therefore required in order to achieve optimal results. 

One problem which may sometimes be most easily detected using plots of the data is that of detecting "outliers , or "bad" data points. These may have resulted from improper application of experimental techniques, incorrect measurements, or other factors not accounted for in the experimental design. Such data may be excluded from the regression analysis. However, care should be taken to not exclude legitimate data points arising from random variation in a functional property or from variation due to the consistent Influence of variable factors which should have been Included in the analysis (factors the Influence Of which could not have been excluded). 

Before using your calculator or computer to find the least-squares straight line, make a graph of your data. The graph gives you an opportunity to reject bad data or stimulus to repeat a measurement or to decide that a straight line is not an appropriate function. Examine your data for sensibility. 

The BOD value would be more accurate and reliable because it is based on the valid data points only, while the outlining bad data points are discarded. 

The role of validation is to write and execute protocols to collect data and to verify that the process is repeatable and reproducible. It is not its responsibility to make the process work. The axiom of bad data in, bad data out is very pertinent to validation. Also bear in mind that just because the process in the batch record has been validated does not mean that a failure or deviation cannot occur during the execution of the batch record. It is not the FDA s belief that validation will prevent failures It is its belief that validation will show that a successful process can be repeated when key steps (as should be listed in the batch record) are repeated from batch to batch within a specified variation. Some key responsibilities for the validation group to consider... 

Of course, some bad data are bad data because a measurement or experiment is demonstrably wrong and clearly not representative of the claimed invention, in which case the patent applicants need to use their sound judgment. The applicant doesn t want to drown the examiner in meaningless data points if it obfuscates what could or should be a clear scientific or legal conclusion. 

Statistics must not be viewed as a method of making sense out of bad data, as the results of any statistical test are only as good as the data to which they are applied. If the data are poor, then any statistical conclusion that can be made will also be poor. 

The development of a QSAR model with high quality or predictivity depends on many factors. One important point is the quality of the dataset, bad data points will... 

Differentiation is, at best, less accmate than integration. This method also clearly indicates bad data and allows for compensation of such data. Differentiation is only valid, however, when the data are presumed to differentiate smoothly, as in rate-data analysis and the interpretation of transient diffusion data. 

You may have seen a monitor that has a bad data cable. This is the monitor nobody wants everything has a blue (or red, or green) tint to it, and it gives everyone a headache. This monitor could also have a bad gun, but more often than not, the problem goes away if you wiggle the cable (indicating a bad cable). 

The key to a successful indexing is not a complete absence of impurity peaks (a few may be present) but it is the accuracy with which peak positions have been determined and the absence of significant systematic errors. Yet another important piece of advice, given in the manual, should always be followed do not waste computer time on bad data. Since the cost of computer time continuously lowers, but the cost of labor continuously rises, this statement could be rephrased do not waste your time on bad data. The latter is indeed applicable to any type of data analysis. 

This procedure has been found to successfully remove bad data without altering the valid data points. It is important to note that the filtering process does not smooth or average out the time series it only looks for and corrects spurious data points. An example of the procedure is shown in Figures 3.12-3.15 taken from Gill et al. (2002). 

Finally, let s conclude this chapter with an extremely important discussion of statistics and the value of data replication and confirmation. We all know that all statistics have an associated probability that goes along with them. Without going into a long discussion on the subject, what this means to us is that there is always a chance that a wrong conclusion may be drawn from a given data set. There is always a chance that we may obtain some bad data or even a statistical outlier in our final response data. This is especially true for small sets of data. 

The goodness of fit criteria that are used in CURFIT are as follows (1) If the data are not described by selected equation, CURFIT returns the conclusion that the data is described by two or more equations (of the selected form) with overlapping domains. In this case the domains, parameters, and the associated maximum errors for each equation are given. (2) If the data are described by the selected equation, CURFIT computes its parameters and their associated maximum errors. Bad data points are automatically rejected. Thus the number of equations returned by CURFIT determines whether or not the data is described by the proposed reaction model. In those cases where the model is not described by all the data, the information returned by CURFIT can be used to specify what subset (s) of the data fits the reaction model. 

The ability of diode array spectrometers to acquire data rapidly also allows the use of measurement statistics to improve the quantitative data. For example, 10 measurements can be made at each point in one second, from which the standard deviation of each point is obtained. The instrument s computer then weights the data points in a least-squares fit, based on their precisions. This maximum-likelihood method minimizes the effect of bad data points on the quantitative calculations. 

The following is a set of data for the vapor pressure of ethanol. Plot these points by hand on graph paper, with the temperature on the horizontal axis (the abscissa) and the vapor pressure on the vertical axis (the ordinate). Decide if there are any bad data points. Draw a smooth curve nearly through the points, disregarding any bad points. Use Excel to construct another graph and notice how much work the spreadsheet saves you. 

---