PHLT 8074 Week 5 Crop-Exposure Matrix Analysis Example

Testing the Assumptions Inside a Crop-Exposure Matrix: Homogeneity, Completeness, Stability and Correct Assignment in the Valley Study's Herbicide Matrix

Student Name

Walden University

PHLT 8074: Environmental and Occupational Epidemiology

Instructor Name

Month Day, Year

Testing the Assumptions Inside a Crop-Exposure Matrix: Homogeneity, Completeness, Stability and Correct Assignment in the Valley Study's Herbicide Matrix

What the Matrix Does

The composite valley study's crop-exposure matrix gives each pairing of farm task, crop and decade an expected number of weighted herbicide handling days per season. A participant who reports working as an orchard applicator in the 1980s receives the value in that cell for each season worked, regardless of what he or she remembers about specific days. The matrix is meant to provide an exposure estimate that does not depend on recall and therefore cannot be distorted by differences in how cases and controls remember. That strength rests on assumptions, and this analysis tests four of them.

A Portion of the Matrix

Table 1

Selected Cells of the Valley Crop-Exposure Matrix: Expected Weighted Handling Days per Season (Composite Values)

Note. Composite values for illustration. Values combine days of handling with the intensity weights defined in Week 3. Declines after 1995 reflect the spread of enclosed cabs and changes in labeling.

Assumption 1: Homogeneity Within Cells

The matrix assumes that everyone in a cell shares one exposure level, so that an orchard applicator in the 1980s handled about as much herbicide as every other orchard applicator of that decade. This is the assumption most likely to fail. Within one cell, exposure differs with equipment, farm size, crew role, weather and personal practice. Two orchard applicators in 1988 might differ several-fold if one used a backpack sprayer on a small family orchard and the other drove a tractor with an air-blast sprayer across a large operation. Intensity algorithms used in large agricultural cohorts adjust for application method, equipment repair and protective practices precisely because these factors change exposure substantially (Coble et al., 2011). The matrix ignores them.

Consequence

Assigning the cell mean to everyone in it produces Berkson error: each person's true exposure scatters around the assigned value. In a linear model, Berkson error of this kind does not bias the slope, but it makes estimates less precise (Armstrong, 1998). The valley study uses logistic regression, where Berkson error can introduce some bias when the scatter is large, so the larger the within-cell variation, the less the matrix's protection from bias can be taken for granted.

Assumption 2: Completeness

The matrix assumes its cells cover every task and crop through which participants handled the herbicide. Gaps would assign zero to real exposure. In the valley, two gaps are likely: vineyard work, which was minor in early decades but grew later, and right-of-way spraying along irrigation canals, done by county crews rather than farm workers. Participants in those jobs would be misclassified as unexposed. The remedy is to add cells for both, using county spray records, before the matrix is applied.

Assumption 3: Stability Within Decades

Each cell holds one value for ten years. Practices changed faster than that in some periods: enclosed cabs spread rapidly in the 1990s, and a labeling change restricted handheld application of the herbicide partway through one decade. A single decade value averages across these changes and misstates exposure for people who worked only in the early or late years. Splitting the most volatile decades into five-year periods would reduce this error at modest cost.

Assumption 4: Correct Assignment

The matrix is only as good as the job history it is applied to. Participants still report which tasks they did, on which crops, in which years. If cases recall farm jobs more completely than controls, the matrix inherits that difference through assignment, even though the values in its cells are free of recall. This weakness is smaller than in a full exposure interview, since remembering a job title is easier and less open to bias than remembering days of handling, but it is not zero (Teschke et al., 2002).

Is Berkson Error Really Harmless?

A common defense of matrices holds that Berkson error does no harm. That defense needs three qualifications. Berkson error is unbiased for the slope only under linear models and when the cell means are correct; in logistic models with large within-cell variation, some bias can appear. If the cell means themselves are wrong, for example because the matrix's sources overestimated early-decade use, the error becomes systematic and biases results. And Berkson error always reduces precision, which in a study of modest size may hide a real association. Spiegelman (2010) describes methods for correcting exposure uncertainty using validation data, which would allow the study to estimate how much the matrix's error affects its results.

How Much Within-Cell Variation Is Too Much?

The valley study can estimate within-cell variation using its own data. For participants in the validation substudy, record-based estimates of handling days can be compared with the matrix value for their cell. If record-based values cluster tightly around the cell mean, homogeneity holds reasonably well; if they spread widely, the matrix is assigning a common value to people whose exposures differ greatly. A useful summary is the ratio of within-cell to between-cell variance. Where within-cell variance is small relative to differences between cells, the matrix separates people usefully even if it misstates each individual; where it is large, the matrix's ranking of participants is weak. The analysis will report this ratio for the most heavily populated cells, so readers can judge how much weight the matrix-based results deserve.

What the Analysis Recommends

Add cells for vineyard work and canal spraying. Split the 1990s into two five-year periods. Use the matrix alongside the interview rather than in place of it, and compare the two estimates separately among cases and controls in the validation substudy. Report results using the matrix and using the interview side by side, so readers can see whether conclusions depend on the method. If the two methods agree in direction and roughly in size, readers can place more confidence in either; if they disagree, the study will say so plainly and discuss which method's assumptions are more likely to have failed in the valley's circumstances.

References

Armstrong, B. G. (1998). Effect of measurement error on epidemiological studies of environmental and occupational exposures. Occupational and Environmental Medicine, 55(10), 651-656. https://doi.org/10.1136/oem.55.10.651

Coble, J., Thomas, K. W., Hines, C. J., Hoppin, J. A., Dosemeci, M., Curwin, B., Lubin, J. H., Freeman, L. E. B., Blair, A., Sandler, D. P., & Alavanja, M. C. R. (2011). An updated algorithm for estimation of pesticide exposure intensity in the Agricultural Health Study. International Journal of Environmental Research and Public Health, 8(12), 4608-4622. https://doi.org/10.3390/ijerph8124608

Spiegelman, D. (2010). Approaches to uncertainty in exposure assessment in environmental epidemiology. Annual Review of Public Health, 31, 149-163. https://doi.org/10.1146/annurev.publhealth.012809.103720

Teschke, K., Olshan, A. F., Daniels, J. L., De Roos, A. J., Parks, C. G., Schulz, M., & Vaughan, T. L. (2002). Occupational exposure assessment in case-control studies: Opportunities for improvement. Occupational and Environmental Medicine, 59(9), 575-594. https://doi.org/10.1136/oem.59.9.575