Abstract
When the observed values fluctuated (we say that they have ‘Response Errors’) with considerably large variance, so that we can not ignore the variance of errors in observation in comparison with the variance of observation within groups, we face the possibility of committing errors in the analysis of data.
To deal with the distortion in the analysis of data with response errors, a model for correcting response errors of subjects was presented. In blief our model is a probability model of response errors, considering that subjects who reacted + in a given test may react-in the re-test, when the reliability of the test is low.
The Bayesian probability was employed to estimate the true distribution of each probability of response from the results of test and re-test.
When the response is+, ±, or-, the model for correcting the error is as follows:
We estimate the probability of p, q, r, s, t, u, in Table 4 from the results of test and re-test. Table 2 shows the true values, whereas Table 3, observed results. Unknown quantities are p, r, t, n(+), n(±), n (-).
Then, we compute the Bayesian probability of these unknown quantities, and construct Table 5, in which the occurrence of response is considered as probabilistic, not as the {1, 0} pattern.
This model was applied to an experiment on the impression of a photograph of a young woman to illustrate the distortion in the pattern analysis of data.
Further, the distortion in the coventional scale analysis and its reliability were discussed.
