Epidemiology and Biostatistics

Abstract

Epidemiology and biostatistics are the foundations of public health science and practice. Together, these fields focus on planning and executing research and evaluation studies and using information about risk factors and cause to shape programs and policies. Disability should be defined as a state that is largely independent of health and health status. As described in the World Health Organization model of disability (the International Classification of Functioning, Disability and Health), disability is a state that is inexorably connected to the environment in which people live. That is, with a properly functioning social, built, and policy environment, a person with disability would also function well and would enjoy improved health outcomes. Therefore, disability epidemiology provides a rich opportunity to apply the concepts and methods of epidemiology and biostatistics since it requires attention to pragmatic problems in field research, precise definitions of disability and health, and utilizing a theoretical model that considers risk factors across multiple levels of personal and environmental influences. Several federal data sources in the United States are available with which to conduct descriptive and analytic studies of disability. Ideally, disability epidemiology will occur within a participatory action framework that incorporates the preferences and perspectives of people living with disability from study design through dissemination. Researchers also should consider the communication or cognitive impairments that study participants may experience when designing the study and consent process to assure both broad inclusion and the ethical conduct of studies within disability epidemiology.

2.9. Appendix: Incorporating Disability Examples into Epidemiology Coursework: Questions and Examples

• 1. Prevalence, Odds Ratios, Stratified Analysis (Confounding or Effect Modification), Study Design, and Causal Inference

  As a project for a disability epidemiology course, M.P.H. students Tori Vahle and Mary Gould analyzed data from special surveys in Missouri adapted from the Behavioral Risk Factor Surveillance System (BRFSS). These were random-digit-dialed telephone surveys conducted in six Missouri counties between 1995 and 1997. The sample consisted of a total of 3343 adults: 1380 from rural and 1963 from nonrural areas. Disability was defined as “limited in any way in any activities because of any impairment or health problem.”

  • (a) The Table 2.5 shows the crude results of the study. Is there an increased prevalence of disability in rural areas compared to urban areas?

Table 2.5 Risk of disability by urban/rural residence

Answer: The odds ratio for this table is 1.14 (95% confidence interval [CI] = 0.96, 1.34).

• (b) The data were also stratified by age. Table 2.6 shows these data (note: age is missing for 31 subjects). Accounting for age, how does the odds ratio change compared to your answer based on Table 2.5 and answer a. above? Why?

Table 2.6 Risk of disability by residence and age

Answer: The rural population is older than the urban population; we can see that 20% of the urban group is over age 65 and 29% of the rural population is over age 65. Since disability increases with age (due to increasing chronic conditions), the crude odds ratio is biased high just because of the older age of rural adults. The age-adjusted OR is 0.98 (95% CI = 0.83, 1.17). However, if one examines the stratum-specific results, it appears that the results are reversed for the two age groups. For younger adults, the effect of living in a rural area is protective (OR = 0.62, 95% CI 0.50, 0.78), but for older adults, it is associated with an increased risk (OR = 2.19, 95% CI 1.63, 2.96). This kind of effect modification finding (age modifies the effect of rural residence) would need to be checked for statistical significance (we can tell it would be, since the 95% CI of the two stratum-specific estimates do not overlap) and also to be sure the difference makes sense (preferably, it would have been hypothesized in advance). If we had not hypothesized this finding in advance, we’d be cautious about interpreting it as effect modification. Perhaps we’d recommend this for further follow-up in future studies. See Tables 2.7 and 2.8 for the stratum-specific results.

Table 2.7 Younger adults (aged 18–64): risk of disability by urban/rural residence

Table 2.8 Older adults (aged 65+): risk of disability by urban/rural residence

• (c) The results of the cross-sectional study above seem to suggest that rural residence is associated with an increased risk of disability in older adults and protective in younger adults. What kind of study would make the causal inference stronger for this hypothesis? Since we cannot assign people randomly to their geographic residence, this will have to be an observational design.

  Answer: These cross-sectional data raise concerns about causal inference because we don’t know if residence preceded disability. There may be reasons that that people may have moved to a different kind of setting. It is plausible that a person might move to a city for its healthcare, social, or transportation services if they had a disability. If so, the prevalence odds ratio we calculated above underestimates the association of rural living and disability. Or, perhaps, people with disability might stay in the rural area and be less likely to move because they had strong social support in the rural area. In this case, the lower prevalence of disability in urban areas might be falsely high and the rural excess inflated because of the exodus of people without disability. Whether or not either is true, they both pose potential problems to thinking that rural residence causes disability in older adults because of inconsistent temporality of the exposure (residence) and outcome. A stronger study design would be to examine cohorts of people, initially free of disability, to see which group was more likely to incur a disability. This is likely to be a question also that could benefit from a less heterogeneous outcome: for example, maybe sensory impairments are more likely to occur in one setting or another, but mobility impairments may be similar. Cohorts are, unfortunately, very expensive study designs; a very specific hypothesis about residence (the components of rural living that increase the risk of disability, e.g.) would probably take place over a long time period and require a large sample size and many different types of rural and urban settings.

• 2. Relative Risk, Classification, and Confounding.

  Andresen, Fouts, Romeis, and Brownson (1999) analyzed the risk of disability among different ethnic groups of women in the United States. The data were based on a national stratified random-digit-dialed sample of women aged 40 and older; most questions were derived from the modules of the Behavioral Risk Factor Surveillance System, including disability definitions. One measure of disability was defined from a woman’s report that she was “limited” and also that she needed personal care assistance with activities such as eating, bathing, dressing, or getting around the house (classified as having activities of daily living dependence, or ADL). In another analysis, women were asked to describe their overall health status as excellent, very good, good, fair, or poor. Tables 2.9 and 2.10 show the responses of women who were Native American or Alaskan natives compared to white women. Participants included 774 white and 739 Native American women; because of some missing responses, not all analyses included all women who were interviewed. Are Native American women more likely to be disabled according to ADLs? Are they more likely to be in fair/poor health (for this problem, answers are grouped as fair/poor versus good-excellent)? Can you think of reasons that these results might differ? Consider that the average age of white women was 57.3 and that of Native American women was 54.4 (t-test for difference p < 0.01). What additional analysis would you recommend to make sure these results were not biased by age? Why?

Table 2.9 Risk of ADL dependency for ethnic minority women

Table 2.10 Risk of lower health status for ethnic minority women

Answers: In the first table (Table 2.9), the relative risk (RR) for ADL dependency for Native American women is 3.26 (95% confidence intervals [CI] of 1.69, 6.32; statistical significance,${\chi}_{\mathrm{MH}}^2<0.01$) compared to white women. In the second table (Table 2.10), Native American women are more likely to be in fair/poor health, but the estimate is not as large (RR = 1.70 and 95%CI 1.41, 2.06; statistical significance, ${\chi}_{\mathrm{MH}}^2<0.01$). Since these are prevalence data, some might argue for analysis using the odds ratio (OR); however, race/ethnic group is clearly temporally prior to the health status/disability determination, so we are less concerned about the problem of cross-sectional data here.

It is hard to compare directly the two RR estimates; while ADL dependency is one appropriate classification for disability, it is not synonymous with health status. Therefore we might expect that disability and health status have a relationship, but that they would not be overlap** definitions. A woman might require ADL assistance for mobility impairment, but consider her health status to be excellent or very good.

The problem of confounding by age is very likely here. Since various chronic conditions and mobility impairments increase with age (a prime example is arthritis), and Native American women are, on average, younger than white women in this sample, our calculated estimates of the risk of poor outcomes may be confounded by age. If we conducted a stratified analysis with age, we would expect that the RRs we calculated may be biased low and that the age-adjusted RRs should be somewhat larger.

• 3. Classification and Proxy Response.

  In a reliability study of adults with disability and proxy respondents, we found that people with disability (PWD) and family members—who also answered for them—disagreed about the level of functional impairment of the PWD themselves (Andresen et al., 2001). The tables below show how their answers compare for one measure of dependence (needing help bathing) and the report of pain as a secondary condition. Calculate the percent agreement, κ, and difference in the response (proportion) of the proxy from the person with disability (Table 2.11).

Table 2.11 Agreement of people and their family member proxies on dependence in bathing (needs any help)

Answers: The summary answer is listed below (and calculation of κ also described). Proxies do agree with PWD about needing assistance with bathing, although the κ is not in the “excellent” (above 0.75). Overall, their responses over-estimated the need for assistance compared to the person with a disability. These results are common to tests of proxy response: the reliability is better for objective compared to subjective variables. The direction of differences also is common: proxies consider the PWD to be more “disabled” than they are.

Variable	% Agreement	κ	% yes responses		Proxy difference
			PWD	Proxy
Need assistance bathing?	84.3	0.66	33.8	36.4	+ 2.6%

Calculating formulas for percent agreement and κ (for a completed treatment of these methods, see Fleiss, Levin, and Paik, 2004):

• (a) Calculating overall percent agreement.

PWD response	Proxy response		Total
	Yes	No
Yes	P11	P12	P1.
No	P21	P22	P2.
Total	P.1	P.2	Total = 1.0

1. Percent agreement, or percent observed, is P₀ = P₁₁ + P₂₂

But this is misleading because some agreement is due to chance alone. We therefore calculate the expected agreement (by multiplying the column and row totals for each cell, as in calculating chi-square statistics) and then summarize the agreement that is beyond chance, as a proportion of all that is possible. That defines the κ statistic (below). The examples are calculated below (Table 2.12).

Table 2.12 Agreement of people and their family member proxies on dependence in bathing (needs any help)

Observed agreement is P₀ = 0.272 + 0.571 = 0.843

Or about 84% of the proxy-PWD sets agree on whether or not the PWD needs assistance in bathing.

• (b) Calculating expected agreement (Table 2.13).

PWD response	Proxy response		Total
	Yes	No
Yes	P1. P.1	P1. P.2	P1.
No	P2. P.1	P2. P.2	P2.
Total	P.1	P.2	1.00

1. Percent expected is P_e = P_1. P_.1 + P_2. P_.2

Table 2.13 Agreement of people and their family member proxies on dependence in bathing (needs any help)

Percent expected agreement is 0.123 + 0.421 = 0.544 = P_e

Or over 50% of the agreement is expected by chance alone!

• (c) Calculating κ. κ is the measure of “beyond chance” agreement. That is, accounting for chance, how much more agreement do we observe (as a proportion of 100 better than chance)?

$$ \kappa =\left({P}_o\hbox{--} {P}_e\right)/\left(1.0-{P}_e\right) $$

Agreement about dependence in bathing, this would be:

(0.843–0.544) / (1.0–0.544) = 0.66. The κ is 0.66, or agreement of PWD and their proxies is 65% better than chance. This would be considered “good” agreement.

• 4. Study Designs

For each of the following research questions, consider the issues of (1) frequency of the outcomes; (2) feasibility and practicality, in measuring the outcome and/or exposures; (3) the issues of resulting causal inference; (4) the stage at which the question is directed (descriptive, hypothesis generating, hypothesis testing); and (5) the potential for sources of existing data. Suggest the best study design and explain the reasons for your choice.

• (a) Cleft lip (with or without cleft palate) occurs in about 1 to 2 births/1000 in Northern European countries and in people of these backgrounds in the United States. How would you investigate the hypothesis that a woman’s exposure to certain medications during early pregnancy may increase the risk of these birth outcomes?

• (b) Among adults with high-level spinal cord injury (SCI—affecting motor control of upper limbs), there is a large amount of variability in the incidence of upper respiratory infections (URI). While there is evidence that URI is increased compared to the general population, it is not clear if certain kinds of URIs are more common or if URI is just more serious when it does occur (bringing it to medical attention). Because you are part of a large, pre-paid healthcare plan, you have access to a large clinical group of people with the appropriate SCI classification and (1) can assume each person will attend a general medical clinic at least once a year and (2) their other clinical, emergency, and hospital visits are available in the same medical care system; what kind of study would you perform to determine more exactly the nature and risk factors for URI in people with SCI compared to the other plan enrollees?

Answers

• (a) This fairly uncommon outcome may best be studied early on by a case-control study. Medications taken during pregnancy may be recalled with some accuracy by mothers in this design. Considerable care would need to be taken to assure that case mothers were not better at recalling their histories (or telesco** in exposures at other times); a validity study, perhaps using medical records and prescription records, would assist in finding out the overall accuracy of reports and if it were “differential” by case status. Because pregnancy is of short duration, it also may be possible to do a cohort study; this would be especially true for exposures of interest that are uncommon (e.g., a specific medication used to treat infections). Any of these designs would be somewhat difficult if asking about over-the-counter medications; however, diligent work on obtaining exposure information and/or getting women to record such information (specifically based on interviews at their first prenatal visit, e.g.) might overcome these difficulties.

• (b) This might be a good opportunity to use the records and billing data of healthcare plan in a cohort study; potentially much of this work could be accomplished by database analysis, with no further data collection from individuals. If substantial data cleaning or database construction is required (e.g., combining pharmacy records, office visit records, hospital billing, etc.), you might want to use data on all enrollees with SCI and a sample of others. An alternate or auxiliary effort could be a survey of enrollees classified by exposure, as (1) all enrollees with SCI, compared with (2) a sample of other enrollees, matched possibly for age and gender. They would be surveyed, perhaps on several occasions, about their incidence of URI, symptoms severity, and office and/or hospital visits. In either design, the potential for differential misclassification and identification of URIs exists. One would want to ascertain data accuracy (e.g., respiratory infections noted and coded accurately and the same for SCI and others?), and a validity sub-study may be needed. The direction probably is in the direction of better ascertainment for enrollees with SCI, but you would want to confirm this.

• 5. Utilizing Publicly Available Data Sources.

  Most of the surveillance data sources that include disability mentioned in this chapter are publicly available. The BRFSS disability data are easily accessible through the Disability and Health Data System (https://www.cdc.gov/ncbddd/disabilityandhealth/dhds/index.html). The data portal provides tools for users to run basic analyses without downloading the data itself, and results can be output as figures or tables. Use the Disability and Health Data System to answer the questions below. Include table(s) or figure(s) to summarize your findings.

  • (a) Identify the overall prevalence of disability and disability by type (i.e., vision, hearing, cognitive, mobility, self-care, or independent living disability) in your state.

  • (b) Explore demographic subgroups within the population to assess whether there are differences in disability prevalence. For example, you might look at specific age groups, people with different race or ethnicity classification, or people with different levels of socioeconomic status.

  • (c) Examine at least one health outcome and evaluate whether there are differences by disability status or type in the state.

  Answers

  Answers will, of course, depend on the state in which students live, the year(s) of data available through the Disability and Health Data System at the time the assignment is completed, and the demographic characteristics the student chooses. As the instructor, you may choose to limit these parameters to assure you can easily confirm their work.