|
Size: 881
Comment:
|
← Revision 20 as of 2015-02-02 14:34:05 ⇥
Size: 2972
Comment:
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 1: | Line 1: |
| Describe FAQ/criteria here. For 2 x 2 tables there are four terms used to summarise the classification table of observed and predicted group membership outputted by discriminant procedures such as binary logistic regression. |
== Two group discriminant diagnostics == |
| Line 6: | Line 3: |
| Let’s call the two groups positive (+) and negative (-) then | For 2 x 2 tables there are four terms used to summarise the classification table of observed and predicted group membership outputted by discriminant procedures such as binary logistic regression. Let’s call the two groups positive (+) and negative (-) with classification table given below. ||||||||<style="TEXT-ALIGN: center"> || ||||<style="TEXT-ALIGN: center">True || ||||||||<25% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> || || + || - || ||||||||<25% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> Pred || + || a || b || ||||||||<25% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> || - || c || d || |
| Line 9: | Line 14: |
| * Sensitivity = a/(a+c)= proportion of those who are really + who are predicted to be + * Specificity = d/(b+d) = proportion of those who are really - who are predicted to be - |
|
| Line 10: | Line 19: |
| ||||||||<style="TEXT-ALIGN: center"> ||||<style="TEXT-ALIGN: center">True || |||||||| || + || - || Sensitivity = a/(a+c)= proportion of those who are really + who test + Specificity = d/(b+d) = proportion of those who are really - who test - |
* Positive Predictive Value (PPV) = a/(a+b) = proportion of those predicted as + who really are + |
| Line 19: | Line 21: |
| PPV = a/(a+b) = proportion of those who tested + who really are + | * Negative Predictive Value (NPV) = d/(c+d) = proportion of those predicted as - who really are - |
| Line 21: | Line 23: |
| NPV = d/(c+d) = proportion of those who tested - who really are - | * [[FAQ/wavelength_eg| An example]] is available using wavelength. * [[FAQ/gallowayeg| An example of high sensitivity and low positive predictive value]] * [[FAQ/likrat | Positive and Negative Likelihood ratios for diagnostic effectiveness]] __Producing a classification table in a Cross-validation sample__ Output from a discriminant procedure such as logistic regression from one data set (design set) can easily be tested on another one (test set) when there are two groups. Simply compute the sum of the product of the regression coefficients from the __design set__ multiplied by the predictor variables in the __test set__ to produce a score for each subject. If the score is positive allocate to one group (usually the lower coded group) and if this score is negative allocate to the other group (usually the higher coded group but check on these codings). The classification table comparing predicted groups to true group membership can then be produced using, for example, the CROSSTABS procedure in SPSS. In general for k groups (k-1) multiple discriminant scores are produced with one group having a zero score. The predicted group for each individual corresponds to the group with the highest discriminant score. __References__ Fleiss, JL (2003) Statistical Methods for Rates and Proportions, 3rd Edition. Wiley:New York. (in South Wing 5 of University library). Warrens MJ (2011) Chance-corrected measures for 2 x 2 tables that coincide with weighted kappa. ''British Journal of Mathematical and Statistical Psychology'' '''64''' 355-365. This quotes the above indices and others showing they are special cases of the weighted Cohen's kappa measure of agreement. |
Two group discriminant diagnostics
For 2 x 2 tables there are four terms used to summarise the classification table of observed and predicted group membership outputted by discriminant procedures such as binary logistic regression.
Let’s call the two groups positive (+) and negative (-) with classification table given below.
|
|
True |
||||
|
|
+ |
- |
|||
Pred |
+ |
a |
b |
|||
|
- |
c |
d |
|||
The following four quantities are often quoted and asked for by journals as a means of evaluating the sharpness of the model fit in the two group case.
- Sensitivity = a/(a+c)= proportion of those who are really + who are predicted to be +
- Specificity = d/(b+d) = proportion of those who are really - who are predicted to be -
- Positive Predictive Value (PPV) = a/(a+b) = proportion of those predicted as + who really are +
- Negative Predictive Value (NPV) = d/(c+d) = proportion of those predicted as - who really are -
An example is available using wavelength.
An example of high sensitivity and low positive predictive value
Positive and Negative Likelihood ratios for diagnostic effectiveness
Producing a classification table in a Cross-validation sample
Output from a discriminant procedure such as logistic regression from one data set (design set) can easily be tested on another one (test set) when there are two groups. Simply compute the sum of the product of the regression coefficients from the design set multiplied by the predictor variables in the test set to produce a score for each subject. If the score is positive allocate to one group (usually the lower coded group) and if this score is negative allocate to the other group (usually the higher coded group but check on these codings). The classification table comparing predicted groups to true group membership can then be produced using, for example, the CROSSTABS procedure in SPSS. In general for k groups (k-1) multiple discriminant scores are produced with one group having a zero score. The predicted group for each individual corresponds to the group with the highest discriminant score.
References
Fleiss, JL (2003) Statistical Methods for Rates and Proportions, 3rd Edition. Wiley:New York. (in South Wing 5 of University library).
Warrens MJ (2011) Chance-corrected measures for 2 x 2 tables that coincide with weighted kappa. British Journal of Mathematical and Statistical Psychology 64 355-365. This quotes the above indices and others showing they are special cases of the weighted Cohen's kappa measure of agreement.