100 likes | 262 Views
A novel approach to quantify random error explicitly in epidemiological studies. Imre Janszky • Johan Hakon Bjørngaard • Pal Romundstad • Lars Vatten Eur J Epidemiol (2011) 26:899–902. Background to the paper.
E N D
A novel approach to quantify random error explicitlyin epidemiological studies Imre Janszky • Johan HakonBjørngaard • Pal Romundstad • Lars Vatten Eur J Epidemiol (2011) 26:899–902
Background to the paper Random errors – errors in measurement that lead to measurable values being inconsistent when repeated measures of a constant attribute are taken. Caused by unpredictable fluctuations in the readings of a measurement apparatus or interpretation of instrumental readings. Most frequently used methods for handling random error are largely misunderstood or misused by researchers. I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902
Statistical significance • Myth is that the P value is a direct measure of random error or statistical variability. • CIs are also poorly understood and frequently misused • Theoretically confidence intervals from 0 to 1 has exactly the same imprecision as a study with a confidence interval from 1 to infinity. • The P value is probably the most misunderstood statistical concept. • Often dichotomise results on whether P values exceeds 0.05 • Nearly all articles in medical related journals report statistical significance I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902
Aims of the paper They propose a simple approach to quantify the amount of random error which does not require a solid background in statistics for its proper interpretation. Hope the method may help researchers refrain from over simplistic interpretations relying on statistical significance (p value). I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902
Alternative solutions to handle random error Bayesian methodology Likelihood Intervals presenting the likelihood function I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902
Proposal: 1) to present the random error in units analogous to the ‘‘meter”. 2) to use the amount of the random error present in a hypothetical study as the unit of random error. 3) Create a hypothetical study free from systematic error, including 1 million people, OR 1 for association for a dichotomous exposure. 4) Half population exposed, and half would have outcome of interest.
Creating the gold standard unit of random error Amount of random error in hypothetical study = Number of random error units =(SE/0.004)2 SE – standard error of log OR in actual study we want to assess precision 0.004 – standard error for the log OR in the hypothetical gold standard study 0.004 calculate from: SE = (1/a + 1/b + 1/c +1/d) a – both outcome and exposure (250,000 participants) b – without outcome, but exposed (250,000 participants) c – with outcome, not exposed (250,000 participants) d – without outcome, not exposed (250,000 participants)
I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902
What the paper shows • Presenting the number of random error units provides • direct and comparable information on the amount of • random error in each study. • Precision is greater the lower the number of random error units. • The random error can be low even if study lacks statistical significance. • Principles could be extended to other measures e.g. hazard ratios. I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902
Limitations of the paper Number of random error units is correct only if the underlying statistical model is correct. Provides no information on systematic errors like biases or confounding factors—that may often be more important to consider than the random error. Is it really any better than the other methods we already have? I Janszky, J Bjørngaard, P Romundstad, L Vatten. A novel approach to quantify random error explicitly in epidemiological studies. Eur J Epidemiol (2011) 26:899–902