PHLT 8500 Week 4 Coefficient Reading Example

What 0.017 Percentage Points Per Doubling Does and Does Not Say: Reading Each Adjusted Coefficient in the Well Arsenic and Hemoglobin A1c Model at Exactly Its Proper Width

Student Name

Walden University

PHLT 8500: Advanced Biostatistics

Instructor Name

Month Day, Year

What 0.017 Percentage Points Per Doubling Does and Does Not Say: Reading Each Adjusted Coefficient in the Well Arsenic and Hemoglobin A1c Model at Exactly Its Proper Width

Why Coefficients Are Misread

A coefficient from a multivariable model answers one narrow question, and most misreadings come from giving it a broader one. This paper takes the fully adjusted model from Weeks 2 and 3 and states, coefficient by coefficient, what each number licenses. The model regresses hemoglobin A1c on log base 2 of well arsenic, centered at 6 micrograms per liter, with the six Week 2 covariates, in 1,612 adults without diabetes from the composite private-well cohort.

The Exposure, Read in Full

The arsenic coefficient is 0.017 (95% confidence interval 0.004 to 0.030). A complete reading is: comparing two people in this cohort who have the same age, sex, body mass index, smoking status, education and rice intake, the one whose well water contains twice as much arsenic has, on average, an A1c 0.017 percentage points higher. Every part of that sentence does work. "In this cohort" limits the population. "Same age, sex" and the rest name what is held constant. "On average" says the coefficient describes groups, not each person. And "twice as much" fixes the unit.

The coefficient is not a statement that lowering any one person's arsenic would lower their A1c by that amount. That would require the causal assumptions encoded in the Week 3 diagram to hold (Greenland et al., 1999), including no unmeasured confounding, and it would apply to the average over people, not to each individual.

Per Doubling, Not Per Unit

Because arsenic is on a log scale, the coefficient does not describe the effect of one more microgram per liter. A rise from 3 to 6 and a rise from 50 to 100 micrograms per liter are both one doubling and carry the same predicted difference. Over a larger contrast, the differences add: going from 3 to 24 micrograms per liter is three doublings, for a predicted A1c difference of about 0.05 points. Reporting results for a contrast readers care about, such as the 25th to the 75th percentile of wells, 2 to 14 micrograms per liter, makes the number concrete. That contrast is a little under three doublings and corresponds to about 0.048 points.

Table 1

Predicted A1c Differences for Selected Arsenic Contrasts (Composite Figures)

Note. Composite figures; intervals scale with the number of doublings.

Is the Difference Large?

An A1c difference of 0.05 points is small for an individual; clinicians rarely act on changes under 0.5. For a population, though, a small shift in the average moves many people across diagnostic thresholds. If the association is causal, a shift of 0.05 across the whole distribution would push some people near the 5.7% prediabetes cutoff over it. The coefficient alone cannot tell us how many, which is why Week 5 models prediabetes and diabetes directly. A rough sense of scale helps, though. About a fifth of the cohort has an A1c between 5.5% and 5.7%, just under the prediabetes line, and those are the people for whom a shift of a few hundredths of a point could change a clinical label. Whether that change would matter for their health is a separate question the label cannot answer.

The Covariate Coefficients

The model also reports coefficients for each covariate. Body mass index has 0.021 per unit, age 0.008 per year and current smoking 0.11 compared with never smoking. It is tempting to read these as effects of weight, age and smoking. Westreich and Greenland (2013) call this temptation the Table 2 fallacy. The model was built to estimate the arsenic association, and the covariates were chosen to close backdoor paths for arsenic, not for themselves. The age coefficient, for instance, is adjusted for body mass index and smoking, which may lie on paths from age to A1c, so it estimates something closer to a direct effect of age than its total effect. The smoking coefficient is adjusted for education, appropriate if education confounds smoking, but not for other confounders of smoking that the model never needed. None of these coefficients answers a well-defined causal question.

Fixed in the Model, Not in Life

"Holding age constant" is a statement about the comparison the model makes, not about the world. In the data, people with high-arsenic wells are older on average, and the model compares them with equally old people with low-arsenic wells. Such comparisons rely on there being people of every age at every arsenic level. Where that overlap is thin, for example very young adults with very high arsenic wells, the model fills the gap by assuming the linear relationship continues. The coefficient is therefore most trustworthy within the ranges where data are dense, roughly ages 40 to 70 and arsenic from 1 to 40 micrograms per liter.

The Intercept

The intercept of 5.58% is the predicted A1c for a 55-year-old woman who has never smoked, holds a college degree, has a body mass index of 28, eats no rice and drinks from a well at 6 micrograms per liter. It is a reference point, not a finding. Because the predictors are centered, the intercept describes a realistic person, which makes the model easier to communicate.

What a Reader Should Take Away

Read at its proper width, the exposure coefficient says: among adults in this composite cohort, after accounting for the measured confounders, higher well arsenic goes with slightly higher A1c, about 0.05 points across the middle half of wells. It is consistent with prospective evidence linking arsenic to incident diabetes (Kuo et al., 2015) but is cross-sectional and cannot by itself establish that arsenic raises A1c. The covariate coefficients are adjustment tools and should not be interpreted as effects.

References

Greenland, S., Pearl, J., & Robins, J. M. (1999). Causal diagrams for epidemiologic research. Epidemiology, 10(1), 37-48. https://doi.org/10.1097/00001648-199901000-00008

Kuo, C.-C., Howard, B. V., Umans, J. G., Gribble, M. O., Best, L. G., Francesconi, K. A., Goessler, W., Lee, E., Guallar, E., & Navas-Acien, A. (2015). Arsenic exposure, arsenic metabolism, and incident diabetes in the Strong Heart Study. Diabetes Care, 38(4), 620-627. https://doi.org/10.2337/dc14-1641

Westreich, D., & Greenland, S. (2013). The Table 2 fallacy: Presenting and interpreting confounder and modifier coefficients. American Journal of Epidemiology, 177(4), 292-298. https://doi.org/10.1093/aje/kws412