Industry-funded research and bias in food science

Abstract

Is industry-funded scientific research likely to be biased towards finding positive results? Is industry more likely to work on topics with likely positive outcomes? Using publication-level data and focusing on food groups that are typically considered healthy, I evaluate each article’s abstract using crowdsourcing tools. I find little evidence to support selection on topics with positive outcomes, but industry is less likely to work on topics classified as unrelated to health. Conditional on a topic, I find that industry-funded research is 3.2% more positive compared to non-industry funded research with grains that receive heavier funding responsible for most of the effect. Industry-funded research is also more likely to receive a mention in certain industry newsletters. Coupled with firm incentives to use science to further their marketing efforts, such increased trade press coverage might play a role in shaping consumers’ opinions on what is healthy.

Appendices

Appendix : A: Case study: Oats

Of all oatmeal abstracts, 13% are industry-funded. Moreover, oatmeal is monopolized by Quaker Oats (Pepsi) not only in market share, but also in industry funding. Among all industry-funded abstracts, Quaker Oats contributes a disproportionate 32%, with the next firm (General Mills) contributing only 7%. This pattern suggests Quaker might have incentives to convince consumers its product is healthy.

To examine if industry-funded articles are more positive, I run the following descriptive regression for the dependent variable ratings:

$$ y_{aw}=\alpha+\beta Ind_{a}+\varepsilon_{aw} $$

(9)

where y_aw is worker w’s rating for abstract a in oatmeal numerically coded as 1 for positive ratings, 0 for neutral or unrelated ratings, and -1 for negative ratings. Ind_a = 1 if the funding source for abstract a contains an industry participant such as Quaker. The coefficient α is a constant that represents the non-industry average of the dependent variable and β is the effect attributable to industry funding. All standard errors are clustered at the abstract level.

Table 12 presents the results of this regression showing that industry-funded articles are more positive. Industry-funded articles are 0.088 (β from Table 12) more positive. Because the rating scale ranges from -1 to + 1, this implies a bias of 4.4% (0.088/2). However, this finding could also occur if industry-funded articles focus more on certain types of research that non-industry might choose not to focus on. In the empirical analysis section of the paper, where I use all whole grains, I therefore conduct analysis at the keyword level, comparing all abstracts associated with the same keyword along with other controls.

Table 12 Industry Funding and Abstract Ratings: Oats

In the above regression, “unrelated” ratings are assumed to be the same as “neutral” ratings (i.e., are assigned a value of 0). Because “unrelated” ratings can be fundamentally different, I allow for such differences in the main empirical analysis where I absorb such unrelated ratings into a separate coefficient.

Appendix : B: Logit model of abstract ratings

I estimate a logit model where ratings are converted to a binary outcome of positive versus not-positive. The logit model is specified in Eq. 10:

$$ p\left( y_{k,aw,gy}=1\right)=\frac{exp\left( \alpha+\beta Ind_{a}+\alpha_{g}+\alpha_{y}\right)}{1+exp\left( \alpha+\beta Ind_{a}+\alpha_{g}+\alpha_{y}\right)} $$

(10)

where $p\left (y_{k,aw,gy}=1\right )$ if abstract a is rated as positive and $p\left (y_{k,aw,gy}=0\right )$ if the abstract is rated as negative or neutral (i.e., not positive) by worker w. Ind_a = 1 if the funding source contains an industry participant. The coefficient α is the effect attributable to non-industry articles, and β is the effect attributable to industry funding. The additional fixed effects α_g and α_y correspond to the whole grain food group and year of publication respectively. Due to the large number of keywords, keyword fixed effects are not estimated in this logit specification. In this estimation, I exclude those abstract-ratings that are rated as unrelated to health outcomes.

Table 13 presents the marginal effect of industry-presence from the logit models with an increasing set of controls. Using the estimates from Table 13, column (3), the results indicate that industry-funded articles are 7.3% more likely to be positive (as opposed to negative or neutral).

Table 13 Rating of Industry-Funded Articles: Marginal Effects