Language Biases in Pediatric Emergency Department

Summary

Our objective was to determine if a family’s preferred language of Spanish versus English was associated with differences in management of bronchiolitis in the PED.

Rates of PED testing, interventions, and disposition among children whose families’ preferred language was Spanish were compared to children whose families’ preferred language was English. Primary outcomes were frequencies of chest x-ray and bronchodilator orders.

Logistic regression was used to calculate adjusted odds ratios after controlling for age, emergency severity index, prior visit, and nesting within attending physicians.

Among children diagnosed with bronchiolitis, Spanish-speaking families were more likely to have chest x-rays, complete blood counts, and blood cultures ordered compared to English-speaking families.

patients from families with a preferred language of Spanish were more likely to receive diagnostic testing that did not align with the American Academy of Pediatrics bronchiolitis guidelines.

Initial thoughts:

• how does one differentiate the difference in treatment that spanish-speaking families should have? I assume that they picked something whereby there is no prescribed difference in treatment.11 Though paper and practice obviously differ. But then it’s a question of whether or not these variations across language groups is a bias, or leads to a good outcome.
• another key confounder is going to be spanish-speaking families having different symptoms coming in, but I assume that the logistic regression tries to capture that, by controlling for the other variables

Body

It seems like they’re trying to argue specifically for the language spoken, as opposed to the ethnicity. This makes me feel like they’re slicing up the cake a little too finely. Basically, they want to claim something about the language barrier actually hampering communication and leading to potential miscommunication and therefore mis-diagnoses?

However, it is unclear if these differences were due to language barriers or other racial/ethnic factors.

Therefore, the objective of this study was to determine if a family’s preferred language of Spanish versus English was associated with differences in diagnostic testing and management of bronchiolitis in the PED. We hypothesized that there would be a higher frequency of diagnostic tests, medications, treatment orders, and admission rates for Spanish-speaking families compared to English-speaking families who presented to the PED with a child with bronchiolitis.

Data:

had a preferred language of English, or if they had a preferred language of Spanish and requested an interpreter.

We used generalized linear mixed models for each outcome and nested for attending physician correlation to account for similar practice patterns among the same provider.

So, for a given outcome (say whether or not a chest x-ray was ordered in the PED, which I assume is binary), they then look at the two language variables as the covariates (and I assume, they also include other covariates to make sure they don’t explain the differences). The mixed model is for the nesting, so as to account for provider variation.22 Does that mean at the actual attending physician level?? That seems a little too fine-grained.

At the end, they have 13000 English and 440 Spanish. Very unbalanced!

Figure 1: flow diagram

Okay, here comes the sensitivity analysis.

To account for the possibility that the presence of pneumonia may contribute to the increased odds of ordering chest x-rays, CBCs, and blood cultures, we conducted 2 sensitivity analyses to account for the diagnosis of pneumonia.

Looks like they basically include a covariate of the presence of pneumonia.

Odds Ratio

• great explanation of odds vs probability on CV.

If you run a logistic regression with \(Y\) as the response, \(X\) as your group, and \(Z\) as additional covariates, then the coefficient in front of \(X\) corresponds to the conditional odds ratio, given \(Z\). The math is pretty straightforward: the form is \(\log p/(1-p) = \beta_x X + Z \beta_z\), giving

\[ \begin{align*} \frac{\P{Y=1 \given X=1}}{\P{Y=0 \given X=1}} &= \exp\left\{ \beta_x X \right\} \exp\left\{ Z \beta_z \right\}, \\ \frac{\P{Y=1 \given X=0}}{\P{Y=0 \given X=0}} &= \exp\left\{ Z \beta_z \right\} \end{align*} \]

and solving for these two equations gives the result. Thus, the exponential of the coefficient of \(X\) gives the conditional odds-ratio between \(X\) and \(Y\). Since we’re dealing with conditional statements, it’s important to pick what is \(X\) and what is \(Y\). In this particular case, I’m pretty sure that \(X\) is the english/spanish speaking variable, while \(Y\) are the various outcomes.

Multiple Testing

What’s confusing here is that usually when I think of multiple testing/comparison I’m thinking about having many covariates (\(X\)’s), and searching for a relationship there with a single outcome \(Y\). But here it’s reversed: we have multiple outcomes and really only a single covariate (plus the rest of the covariates are also fixed). But I think the problems with multiple testing continue to hold. The general point is that if you run a bunch of hypothesis tests that are based on some null distribution, then regardless of the configuration of the variables, so long as you’re doing multiple tests, you’re going to have to account for this.

My general recommendation for these kinds of things is, first and foremost, you have to disclose how many tests you’ve run (i.e. did you pre-register your primary and secondary outcomes). Then, if you have a few outcomes, it’s probably best to just adjust for multiple testing. There are papers out there that claim that, for expository research, it’s not as big of a deal, and perhaps that’s the way it is in Medicine, but my predisposition is always to err on the side of caution.