For this study:
The extent to which the sample is unrepresentative of the population, by chance alone.
Selection bias would occur if the sampling method to at least some extent missed certain segments of the population: the homeless, for example.
Ink flow | Row | |||
rate setting | Pressure | plate | setting | Mean |
Low | Medium | High | ||
Low | 25.300 | 27.100 | 19.700 | |
21.600 | 24.200 | 21.900 | ||
Mean | 23.450 | 25.650 | 20.800 | 23.300 |
High | 31.100 | 25.600 | 26.600 | |
29.500 | 23.100 | 23.900 | ||
Mean | 30.300 | 24.350 | 25.250 | 26.633 |
Column Mean | 26.875 | 25.000 | 23.025 | 215.25 |
If, on the other hand, the experimenters used two instruments, one on the low ink flow and one on the high ink flow, and if one reads 5 (mm)2 high and the other is accurate, there would be a bias in comparing the effect of ink flow setting.
Ink flow rate setting | Pressure | plate | setting |
Low | Medium | High | |
Low | 3.7 | 2.9 | -2.2 |
High | 1.6 | 2.5 | 2.7 |
Blocking factor: nuisance factor used to reduce variance.
Randomization: chance assignment of treatments to experimental units. Used to eliminate bias due to unsuspected nuisance factors.
The experimenters made sure the process was stationary before taking data. Specifically, the responses were measured over time prior to running the experiment and no instability was observed. To further guarantee stationarity, each time the settings were changed the experimenters ran several sheets to assure that conditions stabilized before taking data.
We have described how the experiment could be run in blocks on two machines. It could have been blocked in other ways as well. For instance, if a day-to-day effect was suspected, one replicate could have been run on one day, and the other on a second day. Or perhaps to account for operator-operator variation, one set of replications could have been run with one operator and the second with another operator. Of course, if there was no reason to block, it could have been run as a CRD.
Order of run, and anything else not blocked should be randomized.
Two reps were done.
To verify the results, confirmatory experiments were run.
In a retrospective study, the end result (e.g. lung cancer or not) is observed, and differences in the hypothesized causes (e.g. smoking) are sought.
In addition to sampling error and selection bias, studies of human populations in which individuals are asked to respond to questions, verbally or in writing, are subject to other nonsampling errors, such as