To the Editor—Li et al^{Reference Li, Kang-Birken, Matthews, Kenner and Fitzgibbons1} recently published a notable article evaluating the frequency of patients who had a viral respiratory infection (VRI) confirmed via rapid polymerase chain reaction (PCR) and were prescribed an antibiotic prior to discharge at 3 distinct emergency departments (EDs) within a single healthcare network. Antibiotic use is associated with adverse events and leads to antibiotic resistance.^{Reference Fleming-Dutra, Hersh and Shapiro2,3} In the United States, >2.8 million antibiotic-resistance infections occur annually, which highlights the importance of research in appropriate antibiotic prescribing.³ In their study, Li et al^{Reference Li, Kang-Birken, Matthews, Kenner and Fitzgibbons1} conducted an exploratory analysis using multivariable logistic regression to evaluate the predictors influencing antibiotic prescriptions in patients with PCR-confirmed VRI prior to discharge from the emergency department. Their study reveals 3 notable predictors that we believe are likely to bear out as meaningful predictors. However, the study is also an excellent example of an underappreciated design challenge. We describe the challenge and provide some guidance to overcome it.

These authors selected covariates that had univariate associations with P values <.25 and they then used a stepwise backward elimination to determine the final multivariable model. The initial covariates selected were age; immunocompromised status (ie, chronic kidney disease, diabetes mellitus, human immunodeficiency virus, or an active malignancy); receipt of antibiotics within 7 days prior to their ED visit; duration of symptoms; hemoglobin levels; abnormal chest x-ray results; and discharge diagnoses of influenza, upper respiratory tract infection, pneumonia, and otitis media.^{Reference Li, Kang-Birken, Matthews, Kenner and Fitzgibbons1} In table 3 of this article, the final model reported effect estimates for the following covariates: receipt of antibiotics within 7 days prior to the ED visit, immunocompromised state, abnormal chest x-ray results, and discharge diagnosis of pneumonia.^{Reference Li, Kang-Birken, Matthews, Kenner and Fitzgibbons1} Readers should use caution when interpreting results like these because they may be biased due to what is known as the “Table 2 Fallacy.”^{Reference Bandoli, Palmsten, Chambers, Jelliffe-Pawlowski, Baer and Thomspon5} The “Table 2 Fallacy” is when effect estimates for multiple exposures and their confounders are estimated from the same statistical model, results that are often presented in an article’s “Table 2.”^{Reference Bandoli, Palmsten, Chambers, Jelliffe-Pawlowski, Baer and Thomspon5,6} Specifically, the covariates included in the final model of this study are secondary exposures of interest and may be serving as confounders to VRIs and/or antibiotic prescriptions. Therefore, the interpretation of the effect estimates across the modeled covariates may change if any of those covariates themselves are confounded by other covariates that are not included in the model. This can happen even when the primary study exposure is not confounded.

To illustrate, we have calculated crude odds ratios for the covariates included in the final model. Based on table 3 of Li et al,^{Reference Li, Kang-Birken, Matthews, Kenner and Fitzgibbons1} VRI serves as the primary exposure and the prescription of antibiotics is the outcome. In this model, all of the covariates are adjusted for the exposure of viral respiratory infection. However, when creating the final model through the backward stepwise method, the potential confounder(s) influencing the other covariates in the final model may have been dropped. For instance, the variable immunocompromised status is likely influenced by the variable age. Because the final model does not adjust for the confounder age, the results reported do not provide an unbiased effect estimate for immunocompromised status. Additionally, the other covariates reported in table 3 may have confounders that were not conditioned on in the final model, thus misrepresenting the true, unbiased effect estimate. With respect to the criteria the authors used to classify immunocompromised status (ie, chronic kidney disease, diabetes mellitus, human immunodeficiency virus, and an active malignancy), with the exception of HIV, these conditions are directly affected by age. The odds ratio (OR) reported for immunocompromised status was 3.51 (95% confidence interval [CI], 1.44–8.81). However, when we calculated the crude odds ratio of immunocompromised status, the ratio decreased to 1.72 (95% CI, 0.91–3.16) (Table 1). If age were adjusted for, we could potentially see a different effect estimate altogether, even if age was not directly associated with the outcome.

Table 1.

Adjusted Versus Unadjusted Effect Estimates, Highlighting the “Table 2 Fallacy”

Note. CI, confidence interval.

^a Crude unadjusted odds ratios were calculated in R studio software with the EpiTools package (R Foundation for Statistical Computing, Vienna, Austria).

With the increase of antibiotic resistance, research evaluating components that contribute to antibiotic prescribing practices is vital to assist in formulating interventions for limiting inappropriate antibiotic use. Overall, we believe Li et al provide important contributions toward understanding antibiotic prescribing practices in the ED. To circumvent the misinterpretation of multiple effect estimates in future research, we suggest using multiple models that are tailored to generate unbiased effect estimates for the exposure(s) of interest.^{Reference Westreich and Greenland4,Reference Bandoli, Palmsten, Chambers, Jelliffe-Pawlowski, Baer and Thomspon5} In addition, we suggest explicitly specifying between the unadjusted and adjusted effect measures from a single logistic regression model in the footnotes of the table.^{Reference Westreich and Greenland4,Reference Bandoli, Palmsten, Chambers, Jelliffe-Pawlowski, Baer and Thomspon5} Conditioning on covariates should be firstly based on a priori knowledge of how the variables influence the primary exposure and/or outcome; and secondly on the observed effect sizes. A statistically nonsignificant covariate may still meaningfully reduce bias, and a statistically significant one with a small effect size may have trivial impact on bias.