I'm trying to meaningfully interpret the findings from this meta-analysis which states: "Additional evidence supported the association between fever during pregnancy and increased risk for NDD in offspring."

They base this statement on the following "Maternal exposure to fever during pregnancy increased the risk of [Neurodevelopmental Disorders] in offspring with an OR of 1.24 [95% CI: 1.12–1.38]. Secondary analysis revealed an increased risk for NDD when fever occurred during the first trimester of gestation [OR 1.13–95% CI: 1.02–1.26]."

I am more used to working with Cohen's d effect sizes. Based on this converter, these OR effects sizes of 1.24 and 1.13 appear to be trivial in size when converted to Cohen's d. If that converter is accurate, then these effect sizes become 0.12 and 0.07 respectively.

From a statistical perspective, would it fair to state that these Odds Ratios are trivial and should not worry to-be mothers who experience fever during pregnancy?

Edit / Additions: The context of my question relates to some pregnant women's fear of getting the Covid-19 vaccine due to fear of developing a fever after the injection which some studies are suggesting is associated with neurodevelopmental disorders.

I think as @Noah below mentions, there is arguably no universal rule for what constitutes meaningful or trivial. If memory serves me, Cohen himself even stated that his small/medium/large distinction was ambiguous and that each discipline needs to develop their own reference points of what is meaningful based on their body of literature.

I work in the area of mental health. If a therapy versus placebo generated a Cohen's d of 0.12 favouring therapy in reducing anxiety, then there is no way that therapy would be put forward as warranting implementation in a service as that is well below what is generally accepted in this field to be seen as meaningful.

Using @AdamO's reference to lead below to my query. This paper found that those with ADHD were more likely to have been exposed to lead during childhood than non-ADHD controls. Those with increased levels of lead in their blood had an increased risk of ADHD with an OR 6.0 [95% CI = 4.10–8.77]. Taking this as an example of a reference study, then it would appear that exposure to lead being linked to ADHD has an OR of 6.0 versus maternal exposure to fever being linked to NDD having an OR 1.24.

Said another way (and to quote from @AdamO).

  • Prevention of fever during pregnancy reduces the risk of NDD by a factor of 24%.
  • Prevention of lead exposure during childhood reduces the risk of ADHD by a factor of 500%.

Using the ADHD study as one example, it would appear that exposure to maternal fever has a much smaller association with NDD than lead exposure does with ADHD. I know that "much smaller" is still an arbitrary statement, but it terms of potentially allaying some fears of covid-19 vaccine, it somewhat puts those fears into context.


1 Answer 1


The methods section says quite plainly:

Odds ratio (OR) were pooled using random-effects meta-analysis. Heterogeneity in effect size across studies was studied using random-effects meta-regression analysis. (PROSPERO CRD42020182801)

and provides a variety of references. Most useful is probably the guidance from the Cochrane group, an authority on meta-analyses:

  1. Higgins JPT, Green S. Cochrane handbook for systematic reviews of interventions. New York: Wiley; 2008. [Google Scholar].

In broad terms, this works by allowing for a random intercept to account for differing levels of disease prevalence in the unexposed sampling frame, estimating a "common" odds ratio across all the studies. This is, of course, dependent on running several analyses to assess the "homogeneity" of the odds ratio. The method is intended to generalize the Mantel Haenszel Odds Ratio.

The normal approximation to the odds ratio is usually taken on the efficient (i.e. log odds) scale, where symmetric confidence intervals are calculated. One could calculate the corresponding approximate "Z" statistic for the odds ratio and compare to Cohen's D. But the practice is not standard for a few reasons. Mainly, the odds ratio doesn't often translate into useful scales for understanding the potential of public health interventions.

You may be appealing to the following chart which refers to odds ratios of 1.5 or smaller as producing "small" effects. enter image description here https://www.ncbi.nlm.nih.gov/labs/pmc/articles/PMC3444174/

This kind of table does quite a bit to undermine the public health advances of the last century. Indeed, the associations between common exposures and rare outcomes tend to have highly variable odds ratio. The seminal text from Breslow and Day 1987 published by the IARC contains an extensive rationale for basing public policy on odds ratios.

Effect sizes are not a universal lens from which one can view all effects from clinical and observational studies. Odds ratios are problematic because they only approximate the risk ratio when the incidence is rare. Assuming NDD is rare and well characterized, we have that prevention of fever during pregnancy reduces the risk of NDD by a factor of either 24% or 13% or in the vicinity, depending which analysis is least biased by observational design. You may not be aware but fever during pregnancy is not benign, a mother's immune system is suppressed to prevent a host-vs-graft type reaction, and common infections can have serious consequences, especially listeria. This is women who go through pregnancy are advised against cruises, fountain drinks, processed lunch meats, and sushi; all (mostly) fine for an adult with a normal immune function.

Some argue a risk difference or number-needed-to-treat is a more important scale for understanding potential impact of a preventative measure. The problem is that for very rare events the risk difference is necessarily very small, but if they are events of grave consequences, a binary (yes/no) comparison may not suffice to present the value of a policy or treatment.

In public health cost-benefit analyses, adverse effects are often translated into QALYs - quality adjusted life years. QALYs are a unit that allow one to quantify and compare the loss of dying 10 years early, versus dying 5 years early but spending the last 20 years of your life under terrible living conditions. Neurodevelopmental delay reduces your quality of life forever, and most likely reduces life span. An analysis of lead exposure on NDD using QALYS deemed household lead exposure such an adverse health outcome that public health experts suggested it was cost effective to perform lead abatement on all American homes due to the risk of lead ingestion and inhalation.

As a last point, one must consider the cost of implementing an intervention; both at the outset and to sustain it according to the initial vision. This is an endless point of debate in economics - since government spending is kind of a mystifying subject. Nevertheless, you can ask questions like "I have $500 to spend on something, what's the most impact?" Lead abatement is obviously a highly cost intensive process in terms of hours and materials; demolition, HEPA negative air pressure filters, containment, etc. On the other hand, training protocols for safety in immunosuppressed individuals is kind of standard and inexpensive, not perfectly effective, but individuals are usually motivated to follow the code. Prior to the COVID19 pandemic, for anyone who cared for a loved one going on a chemo that cause neutropenia, you might full well remember mask wearing, hand washing, and social distancing. These are relatively inexpensive things to do that produce a big effect for cancer patients, and as it seems may benefit pregnant mothers.

  • 1
    $\begingroup$ I'm not sure this answers the question of whether odds ratios of 1.24 and 1.13 should be considered "meaningful" or "trivial", which seems to be the core of OP's post. I read the question to mean "I understand how Cohen's D values can be interpreted as meaningful vs. trivial; can odds ratios be interpreted in the same way using this converter?" Nothing to do with the methodology of pooling effect sizes across studies. $\endgroup$
    – Noah
    Commented Jan 11, 2022 at 21:20
  • $\begingroup$ @Noah good point. It's a densely layered question to the point it's hard to pinpoint exactly where the line of reasoning fails. Consider the edit, any other input needed by your (or OP's) reckoning? $\endgroup$
    – AdamO
    Commented Jan 11, 2022 at 22:49
  • 3
    $\begingroup$ (+1) looks good to me. I think the main issue is that the meaningfulness of odds ratios is context-dependent; there is no universal rule for what constitutes as meaningful or trivial (this is true of other effect sizes, too). I think your edit explains that quite vividly. $\endgroup$
    – Noah
    Commented Jan 11, 2022 at 23:54
  • $\begingroup$ Thanks for both of your comments. @Noah, you are correct that it is the meaningfulness of the ORs that I am trying to understand. I have added more to the original question above to potentially add any further clarity to what I am attempting to understand. $\endgroup$
    – Strooby
    Commented Jan 12, 2022 at 16:16
  • 1
    $\begingroup$ Strooby: suppose that in one group of 1000 there are ten deaths and in another of 1000 there are 15 deaths. The odds ratio is 1.33 which is close to the ones which you are bothered about. Now suppose these are two different treatments. Would you recommend to your patient that they take the treatment from the second group or the first? $\endgroup$
    – mdewey
    Commented Jan 12, 2022 at 17:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.