The Cross-Sectional Study:

(1)

The Cross-Sectional Study:

Investigating Prevalence and Association

Ronald A. Thisted

Departments of Health Studies and Statistics The University of Chicago

CRTP Track I Seminar, Autumn, 2006

(2)

Lecture Objectives

1. Understand the structure of the cross-sectional study design, 2. Understand the advantages and disadvantages of this design,

and

3. Understand the kinds of questions that can be addressed using cross-sectional data.

(3)

Outline

A Clinical Scenario: Are Kidney Stones and Hypertension Connected?

The Cross-Sectional Study The Logic

The Structure Pros and Cons

Conducting a Cross-Sectional Study

Steps in building a cross-sectional study Practical Issues

Discussion of clinical example The data set

The questions

Confounding and interaction Data analysis issues

Suggested Reading

(4)

Outline

The questions

Suggested Reading

(5)

Are Kidney Stones and Hypertension Connected?

Background

I Women age 34–59 with history of kidney stones more likely to also have prior diagnosis of hypertension. (Madore, 1998)

I Men age 40–75 with stone history: same direction of association, but weaker. (Madore, 1998b)

I Multiple studies report relationship between BP and stones;

wide variability in size of relationship.

The Gap: What is really going on?

(6)

Hypotheses:

I Some subgroups may be more susceptible to increased BP when stones are present than others.

I Heterogeneity of previous results may be due to differences in representation of high-risk subgroups.

I Association of BP and stones varies with (a) sex and (b) body size.

How to test these hypotheses?

I Need to be able to identify and study subgroups of interest.

I Need to avoid selection by stone status or blood pressure (so clinic-based samples are probably out)

I Would like to say something about subgroups as they exist in the adult population.

Investigated by Gillen, Coe, and Worcester using NHANES III: The Third National Health and Nutrition Examination Survey.

(7)

Hypotheses:

(8)

Hypotheses:

(9)

Outline

The questions

Suggested Reading

(10)

The Logic of Cross-Sectional Studies

I Looks at a “slice” of the population at a single point in time.

I If the selected sample is appropriately selected, composition of the sample reflects that in the population.

I Simple random sample

I Cluster sample

I Stratified random samples

I Multi-stage sample

I Perform pre-defined measurements and ascertainments.

Often include questionnaire/survey questions

I What can we do with this sample? We can estimate

I Prevalence. What fraction of the population has a particular characteristic?

[History of kidney stones? Diagnosed hypertension?

SBP > 140?]

I Association. What is the correlation between an “exposure”

and an “outcome?”

[Relationship of kidney stone history to HTN history]

(11)

The Structure of a Cross-Sectional Study

[People with disease/outcome]

[People without disease/outcome]

Begin Compare

Now

Risk Factor +

Risk Factor -

Risk Factor - Risk Factor + Study Population

(12)

The Structure of a Cross-Sectional Study, Continued

Begin Compare

[People with risk factor]

[People without risk factor]

Now

Outcome +

Outcome -

Outcome - Outcome + Study Population

(13)

Pros and Cons

I Advantages

I Cheaper/easier than longitudinal study: no follow-up required!

I Afford good control over the measurement/ascertainment process

I Can maximize completeness of key data (compared to retrospective study)

I Have greater control over precision of estimates in subgroups (stratified sampling)

I Often can be accomplished as secondary data analysis, that is, data collected by someone else (possibly for another purpose)

(14)

Pros and Cons, Continued

I Disadvantages

I In secondary data analysis, no control over purpose, choice, or method of data collection

I Cannot tell us about causal relationships (only correlation)

I Generalizability limited by sampled population, population definition

I Sample size requirements may be very large (especially when looking at rare outcomes or exposures)

I Potential for selection bias.

Example. “Length-biased sampling” results from the fact that individuals with long courses of a disease are more likely to be the ones identified as prevalent cases than people with courses of short duration.

(15)

Outline

The questions

Suggested Reading

(16)

How to conduct a cross-sectional study

I Identify (and define) population of interest [Adults aged 17–90 with knowledge of stone hx].

I Define outcomes

[Previous diagnosis of HTN; SBP].

I Define exposures (for correlational analysis) [Lifetime history of kidney stones]

I Create data collection “instruments”

[Surveys, interviews, physical measurement procedures: SBP]

Data collection forms are very useful for

I Standardization

I Protocol adherence

I Insure consistent ascertainment [Training of staff]

(17)

How to conduct a cross-sectional study, Continued

I Sample from population appropriately [Multistage sample of households].

I Obtain consent, then “measure”

I Analyze using appropriate statistical methods

I May require special techniques to account for sampling design.

I Statistical adjustment for confounders is usually necessary; we are (usually) after relationships that remain after adjusting for other factors

[Age, race, sex are differentially associated with both stone formation and blood pressure]

OR. . . Ask NORC.

(18)

How to conduct a cross-sectional study, Continued

I Sample from population appropriately [Multistage sample of households].

I Obtain consent, then “measure”

I Analyze using appropriate statistical methods

I May require special techniques to account for sampling design.

I Statistical adjustment for confounders is usually necessary; we are (usually) after relationships that remain after adjusting for other factors

[Age, race, sex are differentially associated with both stone formation and blood pressure]

OR. . . Ask NORC.

(19)

Practical Issues for the Primary Cross-Sectionalist

I Non response

I Representativeness

I Logistics issues: cluster sampling, household contact

I Defining eligibility (target population)

I Defining measures in advance; respondent burden

I Data collection forms

(20)

Outline

The questions

Suggested Reading

(21)

Kidney Stones and BP

Gillen, et al, used NHANES III data (publicly available) to address their research questions.

I Conducted 1988–1994 by NCHS

I National population-based sample

I Noninsitutionalized persons aged > 2 months

I n = 33 994

I Stone history available on n = 20 029.

(22)

Answering key questions

Have you ever had a kidney stone?

919 answer Yes (4.6%)

This is a simple prevalence calculation BP outcomes:

1. Have you ever been told you have high blood pressure?

2. SBP, DBP, Pulse pressure

Relationship of sex to stone formation:

Kidney stone Hx −: 54% women

Kidney stone Hx +: 40% women p < 0.001

Odds of stone history 1.73 higher in men than women Relation of stone formation to hypertension:

“SFs were more likely to report a previous diagnosis of hypertension compared with non-SFs (32.7% vs 24.6%;

P = 0.001).”

[Univariate relationship]

(23)

Answering key questions

P = 0.001).”

(24)

Answering key questions

P = 0.001).”

(25)

Answering key questions

P = 0.001).”

(26)

Answering key questions

P = 0.001).”

(27)

Confounding: An Issue

I African-Americans less likely to form stones

I African-Americans more likely to have hypertension

I Want to “hold constant” AA effects when exploring stone-HTN effects

I Similarly for other confounders

Regression is one approach to confounders

(28)

Regression analysis

therapy, and 19% were estimated to have under- gone surgery to remove a stone.

Prior Diagnosis of Hypertension

Model-based estimates of the association between history of nephrolithiasis and self-reported diagnosis of hypertension are listed in Table 2. After adjustment for age, race, smoking status, history of CVD, and diabetes, a margin- ally significant interaction between stone history and sex was found (P ! 0.056). In women, it was estimated that SFs experienced a 69% increase in odds of reporting a prior diagnosis of hyperten- sion (95% CI, 1.33 to 2.17; P " 0.001). In men, odds of reporting a previous hypertension diagnosis were 20% greater for SFs compared with non-SFs; however, this association was not statis- tically significant (95% CI, 0.88 to 1.64; P ! 0.237). No additional interactions between stone history and age, race, or diabetes were observed.

In addition, subanalyses of SFs did not show significant associations between stone treatment and self-reported hypertension.

SBP, DBP, and Pulse Pressure

Figures 1and2show estimates of mean differences in SBP and DBP comparing SFs with non-SFs. Given potential differences in the asso-

Table 2. Logistic Regression Results Modeling Self-Reported Diagnosis of Hypertension in the

NHANES III Sample Covariate

Adjusted Odds

Ratio* (95% CI) P

History of renal stones (yes v no)†

Women 1.69 (1.33-2.17) "0.001

Men 1.20 (0.88-1.64) 0.237

Age (/5 y) 1.22 (1.20-1.24) "0.001 African-American race 1.48 (1.29-1.69) "0.001 BMI (/kg/m²) 1.10 (1.09-1.12) "0.001 Ever smoker (yes v no) 1.04 (0.89-1.22) 0.600 History of CVD (yes v no) 2.00 (1.66-2.41) "0.001 Diabetes (yes v no) 1.51 (1.25-1.82) "0.001

*Adjusted for all covariates listed.

†P for interaction between history of renal stones and sex ! 0.056.

Table 1. Population Characteristics by History of Stone Disease From NHANES III Characteristic

History of Renal Stones (n ! 919)

No History of Renal

Stones (n ! 19,110) P*

Mean age at time of interview (y) 54.1 # 0.7 43.1 # 0.4 "0.001

Women (%) 40.4 # 2.4 54.0 # 0.5 "0.001

African American (%) 3.8 # 0.5 10.7 # 0.6 "0.001

Smoking (ever v never) 60.3 # 2.3 53.8 # 0.8 0.007

CVD† (%) 6.2 # 0.8 4.0 # 0.3 0.001

Diabetes (%) 6.4 # 1.0 5.0 # 0.2 0.112

Mean BMI (kg/m²) 27.7 # 0.3 26.6 # 0.1 0.001

Dietary intake (mg/d)

Calcium 816.1 # 23.7 817.6 # 9.9 0.951

Sodium 3,504.9 # 92.7 3,414.3 # 35.1 0.372

Potassium 2,950.7 # 68.6 2,852.3 # 22.5 0.184

Magnesium 307.3 # 7.1 299.5 # 2.5 0.322

Alcohol intake (g/d) 7.7 # 1.3 9.5 # 0.5 0.169

Mean SBP (mm Hg) 128.8 # 0.8 127.2 # 0.3 0.049

Mean DBP (mm Hg) 77.4 # 0.6 75.7 # 0.3 0.023

Pulse pressure (mm Hg) 51.4 # 0.7 51.4 # 0.3 0.977

Self-report of hypertension (%) 32.7 # 2.3 24.6 # 0.7 0.001

Treatment for stones (%)

Medication 39.4 # 2.7 — —

Extracorporeal shock-wave lithotripsy 7.7 # 1.5 — —

Surgery to remove stones 18.9 # 2.1 — —

NOTE. Statistics expressed as mean or percentage # SE and weighted to account for the stratified multistage sampling design of NHANES III. With the exception of age, all estimates are age adjusted and represent the estimated mean for a 50-year-old participant.

*P are age adjusted and account for the stratified sampling design of NHANES III.

†CVD is defined as any history of myocardial infarction, stroke, or congestive heart failure.

KIDNEY STONES, HYPERTENSION, AND OBESITY 265

What if degree of association depends on values of another variable?

(29)

Interaction effects

I SFs have higher BP than non-SFs

I This is stronger for heavier people

I This is more true for women than for men

ciation between stone history and BP, presented results are stratified by sex (P ! 0.002 and 0.005 for SBP and DBP outcomes for the interaction between SF and sex, respectively). The association between BP (both SBP and DBP) and SFs also varied with BMI (P ! 0.002 for SBP; P ! 0.065 for DBP). Point estimates are presented for each BMI quintile. All estimates were adjusted for age, race, BMI, smoking status, history of CVD, and diabetes.

No differences in SBP or DBP were apparent comparing SFs with non-SFs in nonoverweight individuals. However,Figs 1and2suggest an increasing trend in BP differences between SFs and non-SFs with respect to BMI. As shown in Fig 1, it was estimated that for women in BMI quintiles 4 and 5, history of nephrolithiasis was associated with 7.62–mm Hg (95% CI, 1.04 to 14.2; P ! 0.024) and 4.36–mm Hg (95% CI, 0.30 to 8.42; P ! 0.036) increases in SBP,

respectively. As shown inFig 2, it was estimated that women SFs in BMI quintiles 4 and 5 had DBPs 4.67 mm Hg (95% CI, 0.41 to 8.93; P ! 0.032) and 2.92 mm Hg (95% CI, –0.60 to 6.43;

P ! 0.102) higher than those of similar women non-SFs. Although similar trends in BP differences between SFs and non-SFs were observed in men, the magnitude of these differences was smaller and not statistically significant. No significant difference in pulse pressure was observed, regardless of sex or BMI level (data not shown).

DISCUSSION We compared BP in persons with a history of nephrolithiasis with those without this history and found that the relationship varied with respect to sex and body size. After adjustment for potential confounders, we observed that a history of stone disease was associated with significantly increased odds of self-reported hypertension in Fig 1. Multiple linear regression estimates (and corresponding 95% CIs) of the average difference in SBP comparing SFs with non-SFs by sex and BMI quintile. All estimates are adjusted for age, race, BMI, smoking status, history of CVD, and diabetes.

GILLEN, COE, AND WORCESTER 266

(30)

Outline

The questions

Suggested Reading

(31)

Data Analysis/Statistics/Limitations

I Causality cannot be directly assessed

I Do stones cause high BP, or does high BP predispose to stones?

I Can view as “hypothesis generating” studies

I Causality implies time course assessment

I No experimental intervention

I Generalizability always an issue

I Less so for population-based studies

I Sampling method is key to generalizability

I Relevance of defined population

[NHANES: noninstitutionalized population]

I Particularly important for smaller-scale cross sections: “Cross section of what?”

I Convenience samples are especially problem-prone since it is hard to know how people “select in” to the population: chart review, clinic waiting rooms, callers to a help line.

I Confounders

(32)

Data Analysis/Statistics/Limitations

I Confounders

(33)

Data Analysis/Statistics/Limitations

I Confounders

(34)

Data Analysis/Statistics/Limitations, Continued

I Confounders

I Statistical adjustment possible, if confounders are measured!

I Regression methods are most common, although stratification can also be used

I Unmeasured confounders

I A major worry, but need to identify

I Ingenuity may lead to suitable proxies

I Unrecognized confounders = land mine

(35)

Outline

The questions

Suggested Reading

(36)