The Problem of Bias

Differences between the comparison and treatment groups can result In a biased estimate of program impacts in two ways  

The only way to eliminate both sources of bias is to randomly assign individuals or households that volunteer to participate in the program into treatment and control groups. This experimental design ensures that, with a large enough sample, the two groups are statistically similar in terms of observable and unobservable characteristics. 

Note that this method only uses observations made at one point in time and therefore assumes that the outcomes of the treated and the counterfactual populations evolve in a similar way over time. 

Double diffetetice or di ference-in-dififereiices method is another estimation method that can be used with experimental, quasi-experimental, and nonexperimental designs. Impact is estimated by comparing the outcomes for the treatment and the comparison groups (first difference) before and after the intervention (second difference). This method requires baseline and follow-up data from the same treatment and control groups, ideally as panel data. If the samples for the follow-up survey differ from the baseline survey, they should be from the same geographic clusters or strata in terms of some other variable. 

Instrumental variables are used with nonexperimental design to control for selection bias. These variables determine program participation, but do not affect outcomes. Evaluators can often use geographic variations in program availability and program characteristics as instruments, especially when endogenous program placement seems to be a source of bias. 

---

However, it is not always so easy to get two medications, therapies, or procedures to be indistinguishable to patients, much less to all the nurses and doctors involved. For instance, one drug may require intravenous administration, while the other is an oral medication. The two drugs may have very different durations of action or very different side effects, so that both the patients and all the clinicians know which patients are on which medication. The action of the active agent may be quite obvious, and the effect of the placebo (nothing) is nothing, so that even if patients do not deduce what arm of the study they are in, the clinicians will probably catch on, and the problem of bias in the observations surfaces. 

We will begin by taking a look at the detailed aspects of a basic problem that confronts most analytical laboratories. This is the problem of comparing two quantitative methods performed by different operators or at different locations. This is an area that is not restricted to spectroscopic analysis many of the concepts we describe here can be applied to evaluating the results from any form of chemical analysis. In our case we will examine a comparison of two standard methods to determine precision, accuracy, and systematic errors (bias) for each of the methods and laboratories involved in an analytical test. As it happens, in the case we use for our example, one of the analytical methods is spectroscopic and the other is an HPLC method. 

In this chapter, the data reconciliation problem for dynamic/quasi-steady-state evolving processes is considered. The problem of measurement bias is extended to consider dynamic situations. Finally in this chapter, an alternative approach for nonlinear dynamic data reconciliation using nonlinear programming techniques will be discussed. 

Bias estimation in variables that are known a priori to be biased can be incorporated into the previous problem by incorporating the bias as a parameter (McBrayer and Edgar, 1995). The new Objective Function in Eq. (8.55) in the presence of bias becomes... 

We have seen that Lagrangian PDF methods allow us to express our closures in terms of SDEs for notional particles. Nevertheless, as discussed in detail in Chapter 7, these SDEs must be simulated numerically and are non-linear and coupled to the mean fields through the model coefficients. The numerical methods used to simulate the SDEs are statistical in nature (i.e., Monte-Carlo simulations). The results will thus be subject to statistical error, the magnitude of which depends on the sample size, and deterministic error or bias (Xu and Pope 1999). The purpose of this section is to present a brief introduction to the problem of particle-field estimation. A more detailed description of the statistical error and bias associated with particular simulation codes is presented in Chapter 7. 

As to 07, the value of 0.67 for NO2 seems to be of wide application (Section III.C), but a slightly lower value, 0.65, may also be considered (Section V.A). Here also the possibility of solvent effects and the dangers of bias in correlations must be borne in mind. With or or 0 the range of values obtained by different methods presents a problem. A generally useful value for both these parameters appears to be 0.10, but it must be borne in mind that there is a school of thought which favours an appreciably higher value, at least 0.15 (Section V.A). 

However, when identical sets of compounds are tested against both (or all) targets of interest, this (while still not entirely addressing the problem of undersampling the huge chemical space) partly removes the target-specific bias of each individual activity set related to a particular receptor, as shovm in Figure 13.3. 

While only -10% of microarray datasets address the problem of batch effects (48), the degree of error contributed by batch effects may be significant. Batch effects may include experimental variations introduced due to multiple types of technical bias (e.g., time, laboratory, reagents, handling). Analysis of multiple methods to address batch effects has been addressed for precision, accuracy, and overall performance (48). Once probe set raw intensities have been processed via normalization and possible additional corrective measures, values can be used for downstream analyses in identifying differentially expressed genes and corresponding functional associations. 

Every one of us working in the held has some bias or other. One of mine concerns the question of macromolecular sequences. The bias is, that the bottom-up approach to the origin of life will never be close to a solution - both conceptually and experimentally - unhl the problem of the onset of macromolecular sequences is clarihed. Obviously the origin of the specihc macromolecular sequences (as opposed to simple polymerizahon) is not an easy queshon to answer, as it is linked to the general problem of structure regulahon. 

However, rather than attempt to remove all bias, the aim is to reduce the bias to acceptable levels that do not, in each case, exceed a designated magnitude. Then the test for bias can be designed to confirm the presence of bias when the probability of a bias of that magnitude exists. Indeed, the nature of the problem is such that the absence of bias cannot be proven. 

There are a number of technical solutions to the problems of positivity bias. One path for enhancing satisfaction studies lies in using expanded scales, such as 0 to 10 rating scales rather than 1 to 5 scales. While the expanded scales are not normally used in attitude assessment the careful labelling of each scale point can offer a route for enhancing the discriminatory power of a single satisfaction item. Typical scale point labels for a 0 to 10 scale would read ... 

The problem of publication bias can have a profound impact on the messages conveyed by systematic reviews and meta-analyses. When preparing to write these types of papers, authors typically conduct a computer search for articles that meet certain criteria. Examples of these criteria might be ... 

---