Here's the text with grammatical and spelling corrections:
Hello, I have a dataset where each row represents a different group (number of groups=15). The diagnostic test is the same for all groups, but the number of negatives is higher than the number of positives. The counts of negatives (TN, FN) and positives were extracted from separate models—one model for positives and another for negatives. Consequently, when I calculate sensitivity, the value is low due to the high number of negatives. However, upon examining the data, it appears that the diagnostic test is better at identifying positives than TNs. My question is, how can I manage this data imbalance?
I think your question is inherent to diagnostic test accutracy (DTA) studies where the study design is generally cross-sectional and where prevalence of target condition is not a priori fixed.
I make the exercise (attached xls file) of calculating the Se(=TP/(TP+FN)) and Sp(=TN/(TN+FP) of your test in the 3 lines and you get a variation of the Se and Sp which per se is not suprising in diagnostic test accuracy study. The 3rd study seems a little different due to the higher Se despite a lower prevalence (a little counterintuitive but this may due to specific characteristics of the study). However, it is difficult to say more only with these 3 of 15 rows you have.
Heterogeneity of the results can be associated with various study characteristics in DTA studies (study design with one-gate vs 2 gates studies, prevalence and risk of bias associated with study design and characteristics, threshold effect....) or study covariates (ex: differences in case definition between studies, difference in patients chacracteristics....
There are many different options under frequentist and bayesian framework to account for the bivariate nature of diagnostic test accuracy studies and to determine pooled Se and Sp accounting for the variation of the values between and within studies.
The default of the meta-analytical models in diagnostic test accuracy is to use random effect model (so the weight attributed to the study is not proportional to the study size but also account for the specific study effect). This is a more fair assumption than fixed models in most of cases.
It is difficult to know if I answered correctly to your question due to the limited context you give about your study but in short you can not correct directly imbalance of numbers but you should dig in meta-analytical models of DTA studies and then determine if a particular study has some impact on the main model (sensitivity analysis) as well as subgroup analysis, heterogeneity assessment and publication bias assessment.
You could also dig in Cochrane DTA group documents to help you in your research.
I hope this helps.
- Badges
- Science method