There are several research works published with methodological rigors because of the relevance of the research question in medicine. Is it necessary to follow the rules of statistics?
Managing/ Chief Editor Nepal Journal of Epidemiology.
http://www.nepjol.info/index.php/NJE/article/view/6151
No evidence is not ideal, but spurious evidence is potentially dangerous. Multivariate regression analysis is essential in non-randomised studies to adjust for inevitable confounding. Without this vital adjustment results can be very misleading. In this sense statistics is essential. However when deciding on inclusion of variables in the multivariate model do not rely on statistical significance of variables on univariate testing - use variables that have a good biological basis for inclusion.
The main reason is that because hypothesis testing is designed for comparisons between interventions, it is sensibly geared to avoiding type I errors and being more permissive to type II errors, as the former might lead to a new drug replacing an effective drug even if the new drug does not really work on a population level, whilst the latter would merely lead to the relatively safe continuation of the status quo. However when we are trying to see if confounders are present we do NOT want to be overly cautious about type I errors or permissive about type II errors. Picking up a confounder as having an effect on the outcome when it may, in a population sense, not have a real effect (a type I error) is not a major problem in this particular context as it will not affect the multivariate model unduly. On the other hand, missing a potential confounder (type II) could be very serious as this may lead to an invalid model, that could lead to spurious conclusions. This is why p
It is easy to see why the Managing Editor of a scientific journal is concerned mostly with “the quality of a published article". For a researcher ( in particular, for a researcher in epidemiology) the question should be re-phrased as follows:
“Whether Univariate or Multivariate regression analysis is important for establishing associations between a disease (or a condition, or even a clinically important symptom – such as cough or wheezing, or some lab test) and risk factor(s)?”
We live in the multi-factorial World and development of our diseases, even when induced by a specific agent, is promoted or attenuated by a galore of environmental, socio-economic, racial and individual (genetic included) factors. So the answer seems to be unqualified: the multivariate regression analysis (MVRA) is certainly more important!
Nevertheless, if a study neglects the univariate regression analysis (UVRA) it can be, and rather often is weak in two respects.
First, even if you are interested but in a particular risk factor (which position is, I am sure, erroneous but still typical) your MVRA-based model should be adjusted for a number of confounders. It is virtually impossible to include into such model all potential risk factors. So an epidemiologist chooses a short list of so-called “a priory confounders” and everybody is happy although, I am afraid, nobody should be as in each particular situation ranking of risk factors can be different and even some unexpected factors characteristic just of the situation under study can be strongly influencing. Besides, there are well known criteria of choosing a variable as a possible confounder, and one and the same factor can meet these criteria in one situation but not meet them in another.
So the UVRA (with adjustment for age and for gender) should be performed before the MVRA. Afterwards one may choose this or that algorithm of selecting explanatory variables for the MVRA.
Second. A final MVRA-based model usually approximates real impacts-outcome association based on the principle “necessary and sufficient”. For instance, if two risk factors strongly correlate. one of them can be not included into the final model as redundant. It may seem absolutely correct to your biostatistician but not for you if you are a conscientious epidemiologist looking both for understanding of all the whys and because of-s and for substantiating effective scenarios of risk management. Mind you that it may be that, out of those two variables, just that excluded because of a somewhat lower odds ratio or a lesser p-value or based on some other formal criterion corresponds to the risk factor which is manageable in practice (or is really important
For instance, in a small town with unusually high cancer prevalence the predominant dwelling without modern conveniences correlated with using coal ranges in kitchens. In the UVRA of lung cancer risk both factors were found to have the adjusted odds ratio significantly higher than 1.0, but only the said type of dwelling was included into the MVRA-based model. If I ignored the UVRA results I would neither understand why the absence of central water supply and modern sanitation enhances lung cancer risk nor know that to install electric or even well-vented gas ovens would made the trick.
At last, there are problems outside the analytical epidemiology that can be satisfactorily solved with the UVRA – first of all, in experimental studies design of which usually excludes or equalizes all concurrent influences. For instance, if I am dealing with rats of equal gender, age and initial body mass, recruited from one and the same animal clinic, fed the same fodder and so on, and so on and they were repeatedly injected with the same doses of some lead salt, and if I am interested in association between blood lead level and urinary excretion of delta-aminolevulinic acid (or some other toxic effect) the UVRA would be my choice while there would be no field at all for the MVRA.
Thus my answer is: both are important.
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India