Limits...
A Simulation Study of Categorizing Continuous Exposure Variables Measured with Error in Autism Research: Small Changes with Large Effects.

Heavner K, Burstyn I - Int J Environ Res Public Health (2015)

Bottom Line: Variation in the odds ratio (OR) resulting from selection of cutoffs for categorizing continuous variables is rarely discussed.Cutoffs chosen for categorizing continuous variables can have profound effects on study results.When measurement error is not too great, the shape of the OR curve may provide insight into the true shape of the exposure-disease relationship.

View Article: PubMed Central - PubMed

Affiliation: Department of Environmental and Occupational Health, School of Public Health, Drexel University, Philadelphia, PA 19104, USA. karynkh@aol.com.

ABSTRACT
Variation in the odds ratio (OR) resulting from selection of cutoffs for categorizing continuous variables is rarely discussed. We present results for the effect of varying cutoffs used to categorize a mismeasured exposure in a simulated population in the context of autism spectrum disorders research. Simulated cohorts were created with three distinct exposure-outcome curves and three measurement error variances for the exposure. ORs were calculated using logistic regression for 61 cutoffs (mean ± 3 standard deviations) used to dichotomize the observed exposure. ORs were calculated for five categories with a wide range for the cutoffs. For each scenario and cutoff, the OR, sensitivity, and specificity were calculated. The three exposure-outcome relationships had distinctly shaped OR (versus cutoff) curves, but increasing measurement error obscured the shape. At extreme cutoffs, there was non-monotonic oscillation in the ORs that cannot be attributed to "small numbers." Exposure misclassification following categorization of the mismeasured exposure was differential, as predicted by theory. Sensitivity was higher among cases and specificity among controls. Cutoffs chosen for categorizing continuous variables can have profound effects on study results. When measurement error is not too great, the shape of the OR curve may provide insight into the true shape of the exposure-disease relationship.

No MeSH data available.


Related in: MedlinePlus

Odds ratios (ORs) for different cutoffs (between mean ± 3 standard deviations) used to dichotomize a mismeasured exposure (W1).
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4555337&req=5

ijerph-12-10198-f002: Odds ratios (ORs) for different cutoffs (between mean ± 3 standard deviations) used to dichotomize a mismeasured exposure (W1).

Mentions: The characteristics of the population are similar to what was expected based on the simulation parameters (data not shown). The effect of changing the cutoff for dichotomizing W1 is shown in Figure 2 (and Figure A1). Each of the three exposure-outcome relationships had a distinctly shaped OR curve. Even with such a large population (n = 1,000,000), at the extreme cutoffs, there was non-monotonic oscillation in the effect estimates that is open to various interpretations. (In all scenarios, the greatest cutoff for the mismeasured exposure yielded more than 750 exposed cases, which is rarely attained even in ‘large’ epidemiology studies. For example, in the linear model with a weak true association and small measurement error, there were 846 cases with a mismeasured exposure at least as great as the highest cutoff.) The linear models had U-shaped graphs with slightly lower ORs for very large cutoffs compared to very small ones. For all strengths of the true association, the OR was increasingly underestimated (compared to the scenario with no measurement error) and the OR curve became flatter as measurement error increased. The highest estimates of OR tend to occur when cutoffs are made at the extremes of the exposure distribution, with changes from an easily dismissed OR of <2 tending to occur with the cutoff in the center of the distribution and a clear “signal” of OR >2 apparent with categorizations at the extremes when there was a moderate underlying association.


A Simulation Study of Categorizing Continuous Exposure Variables Measured with Error in Autism Research: Small Changes with Large Effects.

Heavner K, Burstyn I - Int J Environ Res Public Health (2015)

Odds ratios (ORs) for different cutoffs (between mean ± 3 standard deviations) used to dichotomize a mismeasured exposure (W1).
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4555337&req=5

ijerph-12-10198-f002: Odds ratios (ORs) for different cutoffs (between mean ± 3 standard deviations) used to dichotomize a mismeasured exposure (W1).
Mentions: The characteristics of the population are similar to what was expected based on the simulation parameters (data not shown). The effect of changing the cutoff for dichotomizing W1 is shown in Figure 2 (and Figure A1). Each of the three exposure-outcome relationships had a distinctly shaped OR curve. Even with such a large population (n = 1,000,000), at the extreme cutoffs, there was non-monotonic oscillation in the effect estimates that is open to various interpretations. (In all scenarios, the greatest cutoff for the mismeasured exposure yielded more than 750 exposed cases, which is rarely attained even in ‘large’ epidemiology studies. For example, in the linear model with a weak true association and small measurement error, there were 846 cases with a mismeasured exposure at least as great as the highest cutoff.) The linear models had U-shaped graphs with slightly lower ORs for very large cutoffs compared to very small ones. For all strengths of the true association, the OR was increasingly underestimated (compared to the scenario with no measurement error) and the OR curve became flatter as measurement error increased. The highest estimates of OR tend to occur when cutoffs are made at the extremes of the exposure distribution, with changes from an easily dismissed OR of <2 tending to occur with the cutoff in the center of the distribution and a clear “signal” of OR >2 apparent with categorizations at the extremes when there was a moderate underlying association.

Bottom Line: Variation in the odds ratio (OR) resulting from selection of cutoffs for categorizing continuous variables is rarely discussed.Cutoffs chosen for categorizing continuous variables can have profound effects on study results.When measurement error is not too great, the shape of the OR curve may provide insight into the true shape of the exposure-disease relationship.

View Article: PubMed Central - PubMed

Affiliation: Department of Environmental and Occupational Health, School of Public Health, Drexel University, Philadelphia, PA 19104, USA. karynkh@aol.com.

ABSTRACT
Variation in the odds ratio (OR) resulting from selection of cutoffs for categorizing continuous variables is rarely discussed. We present results for the effect of varying cutoffs used to categorize a mismeasured exposure in a simulated population in the context of autism spectrum disorders research. Simulated cohorts were created with three distinct exposure-outcome curves and three measurement error variances for the exposure. ORs were calculated using logistic regression for 61 cutoffs (mean ± 3 standard deviations) used to dichotomize the observed exposure. ORs were calculated for five categories with a wide range for the cutoffs. For each scenario and cutoff, the OR, sensitivity, and specificity were calculated. The three exposure-outcome relationships had distinctly shaped OR (versus cutoff) curves, but increasing measurement error obscured the shape. At extreme cutoffs, there was non-monotonic oscillation in the ORs that cannot be attributed to "small numbers." Exposure misclassification following categorization of the mismeasured exposure was differential, as predicted by theory. Sensitivity was higher among cases and specificity among controls. Cutoffs chosen for categorizing continuous variables can have profound effects on study results. When measurement error is not too great, the shape of the OR curve may provide insight into the true shape of the exposure-disease relationship.

No MeSH data available.


Related in: MedlinePlus