Hello to all, I’m currently reading with great interest a lot of papers about meta-epidemiology.
I see that a frequent statistical approach is by computing within a meta-analytic framework the ratio of odds ratio between studies with different characteristics. I was wondering if this approach could be extended when considering meta-analysis of studies including diagnostic odds ratios (diagnostic accuracy meta-analysis)?
In or out of a diagnostic context its the same odds ratio and thus I would think that the same sort of meta-epidemiological study should be possible. If extracting only the DOR from each original study within a meta-analysis (for meta-regression) then all is well as the subsequent meta-analysis of RORs will be a standard rather than a diagnostic meta-analysis.
Although off-topic, the concern that makes a DOR different from an OR is only when it is generated through meta-analysis and not when it comes from original studies (as in meta-epidemiology). This is because, for conducting a diagnostic meta-analysis, the bivariate approach (combining Se and Sp into the DOR) is commonly used and this makes it very difficult to distinguish systematic error from truly varying thresholds across studies. However there is a method for diagnostic meta-analysis that starts off with the OR and then splits it into Se and Sp and this should obviate the former concern.
I found this paper which uses a somewhat different approach from what I was thinking to do
DOI:10.1001/jama.282.11.1061
In particular in the materials and methods:
In summary, the dependent variable of the model was the logarithm DOR. Explaining variables were 2 parameters for each meta-analysis (the common DOR and the threshold parameter) and 9 covariates to examine the effect of the different study characteristics, 1 for each feature. All study characteristics were evaluated simultaneously in a multivariate model.
A weighted linear regression analysis was used, with weights proportional to the reciprocal of the variance of the log DOR. This weighted linear regression assumes fixed effects. […] The model was fitted using maximum likelihood estimation
I was wondering if instead of using the MLE approach one could fit a linear weighted model and reporting the coefficients of the covariate included as a study characteristic?
example log(DOR) = threshold_param + summary_DOR + study_characteristic
You describe meta-regression which is just an extension of subgroup analyses allowing continuous and categorical variables. In essence its just simple regression with error weights. The only difference from other simple regression models is that the outcome is an effect size from a set of studies and the explanatory variables are all study characteristics.
The dilemma has been what weights to use and Cochrane suggests “random effects” weights when studies are heterogeneous. I have argued against this for a long time because the latter weights tend towards equality (natural weights) as heterogeneity increases and thus you end up with an unweighted regression. My suggestion is to always use inverse variance weights (regardless of heterogeneity) but also use robust error variances.