Health Insurance and Mammography: Would a Medicare Buy-In Take Us to Universal Screening?

Abstract

Objective

To determine whether health insurance expansions via a Medicare buy-in might plausibly increase mammography screening rates among women aged 50–64.

Data Sources

Two waves of the Health and Retirement Study (HRS) (1994, 1996).

Study Design

A longitudinal study with most explanatory variables measured at the second wave of HRS (1994); receipt of mammography, number of physician visits, and breast self exam (BSE) were measured at the third wave (1996).

Data Extraction

Our sample included women aged 50–62 in 1994 who answered the second and third HRS interview (n = 4,583).

Principal Findings

From 1994 to 1996, 72.7 percent of women received a mammogram. Being insured increased mammography in both unadjusted and adjusted analyses. A simulation of universal insurance coverage in this age group increased mammography rates only to 75–79 percent from the observed 72.7 percent. When we accounted for potential endogeneity of physician visits and BSE to mammography, physician visits remained a strong predictor of mammography but BSE did not.

Conclusion

Even in the presence of universal coverage and very optimistic scenarios regarding the effect of insurance on mammography for newly insured women, mammography rates would only increase a small amount and gaps in screening would remain. Thus, a Medicare buy-in could be expected to have a small impact on mammography screening rates.

Mammography is an important part of prevention efforts and has been shown to reduce mortality among women older than 50 (Shapiro et al. 1982; Roberts et al. 1990; Tabar et al. 1992; Nystrom et al. 1993; Kerlikowske et al. 1995; Salzman, Kerlikowske, and Phillips 1997). Universal mammography screening in the 50–64 age group would save nearly two hundred lives per 100,000 women, or three to four times more lives than universal cervical or colorectal cancer screening (Law, Morrisand Wald 1999). Current mammography guidelines of the National Cancer Institute (NCI), United States Preventive Services Task Force (USPSTF), and the American Cancer Society (ACS) advocate mammography for all women older than age 50 every one to two years (National Institutes of Health 1997; Leitch et al. 1997; United States Preventive Services Task Force 1996).

Mammography screening rates lag behind national recommendations in spite of recent increases (Cook et al. 1998). Health insurance has consistently been shown to increase mammography (Ayanian et al. 2000; O'Malley et al. 2001; Hsia et al. 2000; Blustein 1995). In 1998, Medicare expanded its coverage to include annual mammograms for all Medicare eligible women. Thus, at age 65, virtually all women have insurance coverage for yearly mammography, though 13 percent of women between ages 50 and 64 were uninsured in 1995 (Health Care Financing Administration 2000).

Expanding health insurance coverage is one policy for increasing mammography screening in the United States. However, other factors are important. Minorities, the poor, older women, and women who do not visit a physician regularly are less likely to receive mammography (O'Malley et al. 2001;Hawley et al. 2000; Makuc, Breen, and Fried 1999;Phillips et al. 1998; Burns et al. 1996; Marwill, Freund, and Barry 1996; Schoen, Marcus, and Braham 1994). Women who perceive themselves to be at risk of breast cancer are more likely to receive mammography screening (Aiken et al. 1994; Lauver et al. 1999; Pearlman et al. 1997). Other important predictors of women's mammography use is recommendation of a mammogram by a physician, patient/physician interaction, and trust of the physician (O'Malley et al. 2001; Hynes et al. 1998; Simon et al. 1998; Stoner et al. 1998; Andersse and Urban 1997; Cole et al. 1997; Rimer 1997; O'Malley, Earp, and Harris 1997; Taplin et al. 1997; Beaulieu et al. 1996; Hawley et al. 2000).

The objective of this study was to determine whether a Medicare buy-in would be likely to increase mammography screening among women aged 50–64. Our work differs from that of others primarily because we accounted statistically for potential endogeneity between mammography, breast self-exam (BSE), and physician visits. Endogeneity may exist because data on physician visits and BSE were collected at the same time, and for the same period, as mammography. As highlighted in a recent issue of Health Services Research(2000:5, part 2), potential endogeneity is often neglected in health services research and attention to potential endogeneity is often needed to better understand true associations. However, the choice of what variables to treat as endogenous is somewhat arbitrary; therefore, we have chosen to treat the two variables we thought most likely to be endogenous to mammography because of the time period over which both were collected.

Study Hypotheses

Our primary study hypotheses were that higher self-reported use of mammography would be associated with: (1) having health insurance; (2) having more physician visits; and (3) regular conduct of BSE. Our goal was to determine the impact of expanded health insurance coverage on mammography rates for women aged 50–64.

Methods

Data

Our data came from the Health and Retirement Study (HRS), a nationally representative sample of persons who were 51–61 years of age in 1992 and their spouses, who were given an identical interview regardless of their age. We used waves two (1994) and three (1996) to test our models because questions on mammography and BSE were first included in the survey at wave three.

Study Subjects

The analysis sample included women who: (1) were 50–62 years of age at the second HRS interview (1994); (2) survived to the third HRS interview (1996); and (3) had useable information on mammography use and physician visits between the second (1994) and third (1996) HRS interviews. Of the 4,583 observations that met these criteria, only 4,075 were available for multivariate models because of item nonresponse.

Dependent Variable

The primary dependent variable was the self-reported receipt of mammography that occurred during a two-year period between waves two and three of HRS.

Explanatory Variables

Explanatory variables included health insurance, individual perceptions, health behaviors, and demographics.

Health Insurance. We used three binary variables to represent health insurance (Medicaid, private health insurance, and CHAMPUS/other), with uninsured the omitted group.

Individual Perceptions. A woman's perceived likelihood of living to 75 or older was measured using a scale from 0 (absolutely no chance) to 10 (absolutely certain) with the following question: “What do you think the chances are that you will live to 75 or more?” We rescaled the variable to range from 0–1.0. Self-reported health status was measured with two binary variables: excellent/very good and fair/poor, with good the omitted reference group. Risk tolerance was measured by a single binary variable representing whether a woman was very risk tolerant, measured using a series of hypothetical job change questions (Barsky et al. 1997). Being risk tolerant meant the respondent was willing to accept the highest salary variance offered them.

Health Behaviors. The number of physician visits between waves two and three was considered endogenous to mammography, as was a woman's report of conducting BSE monthly (yes/no) between waves two and three. Current smoking, and vigorous exercise at least three times per week were both included as binary (yes/no) variables.

Demographic Variables. Education was represented as two binary variables (less than high school, and college or more). Race was measured as white versus other. We divided age into three groups: 50–54, 55–59, and 60–62 at the wave two HRS interview. Income was divided into quartiles, with the first including women with family incomes up to $12,621; the second with an upper limit of $30,588; the third $56,815; and the top quartile consisting of persons with income above this level.

Statistical Methods. A chi-square test was used to test the association between mammography use and BSE and explanatory variables. For continuous measures, a t-test for the difference in means between those receiving a mammogram or conducting BSE and those who did not was used. Initial multivariate analyses based on equation 1 were conducted to inform the final specification used to test our study hypotheses (not shown).

(1)

Because physician visits and BSE were collected at the same time (wave three) as self-reported mammography use, and over the same period of time (1994–1996), these variables are plausibly endogenous to mammography. Endogeneity could occur if information received (e.g., bad news) from a mammogram relatively early in the period 1994–1996 caused a woman to increase physician utilization and/or to increase her performance of BSE. Concluding that BSE or physician visits increased the likelihood of mammography would be a faulty attribution in such a case. Likewise, results from BSE could influence physician visits and/or mammography. For this reason, we estimated a final model (equations 2a, 2b and 3) below in which number of physician visits and BSE were considered endogenous explanatory variables, using two-stage least squares instrumental variable regression. In this approach, variables that are correlated with the use of physician services and BSE but not with mammography are called instrumental variables and are used only in a first-stage regression. A predicted value of the endogenous explanatory variable is then used as an explanatory variable for mammography in the second-stage equation, providing unbiased estimates of the effect of physician visits and BSE on mammography use. We used the IVREG (Instrumental Variable Regression) procedure in STATA to estimate this system of equations (Statacorp 2001).

(2a)

(2b)

(3)

Identification of the model was obtained by excluding certain variables from the second-stage equation (equation 3). We determined the reasonableness of these exclusions by testing whether these instrumental variables were jointly significant predictors of the first-stage equations (e.g., Were the sum of the instrumental variable coefficients statistically different from 0?). If so, then they are considered valid instruments. We chose two instrumental variables for the sake of parsimony: (1) lagged physician visits in the two years prior to the baseline interview, and (2) lagged physician visits between waves one and two. We chose two forms of lagged physician visits as instrumental variables because literature suggests they may be predictors of the endogenous variables in our structural model, and they were not significant predictors of mammography use in our data. Physician visits are a major source of information regarding the uptake of BSE, and studies have found that having a regular physician and being taught BSE in a physician's office are both significantly associated with the performance of BSE (Champion and Menon 1997; Danigelis et al. 1995). While we have no information on the content of physician visits, or about whether BSE was recommended or demonstrated, more physician visits would plausibly increase the opportunity for women to learn about BSE; more visits should translate into more BSE, all else equal. In addition, first-stage regressions show that lagged physician visits do increase the probability of BSE. Lagged physician visits are an obvious choice as a predictor of subsequent visits, as past health behavior is a good predictor of future health behavior (Bastani, Maxwell and Bradford 1996). Further, in our model, these variables were not significantly associated with mammography use between waves two and three, making them plausible instruments.

We tested the validity of the instrumental variables in the following manner. First, we estimated a reduced-form regression with BSE and physician visits as the dependent variable with all nonendogenous variables and instrumental variables included as explanatory variables. Next, we conducted an F-test to determine whether the instrumental variables were jointly significant or not. If significant, this suggests that the instrumental variables as a whole have substantial predictive power in the first-stage regression, and are therefore likely to have an effect on the final results.

Results

Descriptive Results

Of females aged 50–64, 72.5 percent received a mammogram during the two-year period, 1994–1996; 62.8 percent reported conducting BSE. More than 90 percent of women reported at least one physician visit between waves two and three, with a mean number of visits of 7.9 (SD=8.9). Of the sample, 78.2 percent had private health insurance, while only 13.3 percent of women were uninsured. Respectively, 3.9 percent and 4.6 percent of women were covered by Medicaid and CHAMPUS or other governmental insurance. Roughly three-fourths of our sample was white and married, and 15.3 percent had a college degree or more, while 25.4 percent did not complete high school. The mean age at wave two was 56.1.

Mammography and BSE were related to women's health insurance, health behaviors, individual perceptions, and demographic factors (Table 1). Women who received mammography had more physician visits between waves 2 and 3 (8.6 versus 6.1, p <0.001), than those who did not. Those who reported BSE also had more physician visits than women who did not (8.1 versus 7.5, p = 0.02). Women who received mammography were less likely to be uninsured compared to women who did not (7.9 percent versus 25.4 percent, p <0.001). A smaller difference in rates of uninsurance was seen for women who conducted BSE compared to those who did not (11.8 percent versus 14.3 percent, p = 0.01). Women who received mammography and who conducted BSE were more optimistic regarding their perceived chance of living to age 75, and differed from those who did not on most health behavior and demographic characteristics, showing them to have a better overall risk factor profile and higher SES.

Table 1.

Characteristics of the Sample

	Received a Mammogram, 1994–96			Conducted BSE, 1994–96
	Yes	No	p	Yes	No	p
Health Insurance
Uninsured	7.9	25.4	<0.001	11.8	14.3	0.01
Insured
Medicaid	3.5	5.0	0.017	4.1	3.4	0.26
Private	83.3	64.9	<0.001	78.7	77.4	0.28
Champus/other	5.2	3.0	0.001	5.0	3.9	0.10
Perception Variables
Perceived probability of living to 75
Actual value, 1994*	0.67	0.62	<0.001	0.66	0.64	0.003
Risk tolerant	11.2	11.1	0.96	11.3	11.1	0.77
Self reported health, excellent/very good	52.8	44.7	<0.001	50.2	51.2	0.51
Self reported health, fair/poor	16.1	23.8	<0.001	18.2	18.3	0.93
Health Behaviors
Number of physician visits, 1994–96*	8.6	6.1	<0.001	8.1	7.5	0.02
Breast self examination at least once per month, 1994–96	66.8	52.5	<0.001	NA	NA	NA
Current smoker	19.1	32.8	<0.001	23.4	21.9	0.23
Former smoker	32.4	25.5	<0.001	30.8	29.9	0.54
Heavy drinker	1.4	1.8	0.32	1.4	1.8	0.30
Body mass index*	27.1	27.5	0.03	27.3	26.9	0.02
Vigorous exercise	21.8	19.0	0.04	22.9	17.9	<0.001
Demographics
Age
50–54	18.0	15.1	0.02	17.4	17.0	0.74
55–59	43.2	45.3	0.20	44.2	43.2	0.52
60–64	38.7	39.6	0.60	38.4	39.8	0.36
Family income
1st quartile	18.3	32.1	<0.001	21.6	22.9	0.30
2nd quartile	22.6	28.5	<0.001	24.0	24.5	0.34
3rd quartile	25.6	21.4	<0.001	25.3	23.0	0.34
4th quartile	33.6	18.0	<0.001	29.6	29.1	0.34
Mean	52,650	30,094	<0.001	44,451	49,775	0.29
Married	77.1	66.4	<0.001	75.6	71.7	0.003
Education
College or more	17.6	9.0	<0.001	14.2	17.1	0.01
Less than high school	21.6	35.7	<0.001	25.1	26.1	0.43
White	73.9	71.3	0.07	72.3	74.5	0.11
N	4,583			4,583

Numbers in cells are proportions, and p values are from chi square test. Numbers noted with ^* are mean values, and the p value in those cases are for a t-test of difference in means. All variables are from HRS wave 2 (1994) unless otherwise noted.

Interaction between Health Insurance and Physician Visits. Women with health insurance were generally more likely to receive mammography, except among the relatively small group who did not have any physician visits during the period (Table 2). Overall, 76.3 percent of the women who had health insurance received a mammogram, while only 45.4 percent of those who were uninsured did so (p <0.001). However, among women with no physician visits, there was no health insurance effect, with only 19.9 percent of those who were insured receiving a mammogram compared to 16.7 percent of uninsured women (p = 0.50). For women who had at least one physician visit there was a strong insurance effect, with health insurance increasing the likelihood of having a mammogram (79.9 percent versus 51.0 percent, p <.001).

Table 2.

Influence of Physician Visits and Being Uninsured on Mammography, 1994–96

	Percent Receiving Mammography
Total Sample,n = 4,583
	Yes
Insured	76.3
Uninsured	45.4	p <0.001
No Physician Visits,n = 312
	Yes
Insured	19.9
Uninsured	16.7	p = 0.50
At Least 1 Physician Visit,n = 4,271
	Yes
Insured	79.9
Uninsured	51.0	p <0.001

Numbers are row percentages. P value is for the chi square test.

Adjusted Mammography Screening Rates without Accounting for Endogeneity. Health insurance, physician visits, and BSE were all important predictors of mammography (Table 3). Women who had private health insurance coverage were more than two times as likely to receive a mammogram compared to uninsured women (OR 2.54 95% CI 2.06–3.14, p <0.001); those covered by CHAMPUS or other types of coverage were also more likely to have gotten a mammogram. However, women who were covered by Medicaid did not differ from uninsured women in terms of their likelihood of receiving a mammogram. Physician visits had a positive impact on the likelihood of receiving a mammogram, with each visit increasing the probability by around 4 percent (OR 1.04 95% CI 1.03–1.06, p <0.001). In an alternative specification not shown in Table 3, women who had at least one physician visit were about four times more likely to receive mammography compared to those who did not have a physician visit. Women who reported completing BSE were nearly twice as likely to receive a mammogram compared to women who did not conduct BSE (OR 1.86 95% CI 1.60–2.16, p <0.001). Other statistically significant findings were generally consistent with the literature and our preliminary descriptive analyses.

Table 3.

Determinants of Mammography Receipt, 1994–96

	Logistic Regression
	OR	95% CI	p
Health Insurance
Insured
Medicaid	1.48	0.97–2.26	0.07
Private	2.54	2.06–3.14	<0.001
CHAMPUS/other	2.16	1.42–3.29	<0.001
Perception Variables
Perceived probability of living to 75	1.05	0.80–1.39	0.71
Risk tolerant	1.17	0.92–1.49	0.20
Self reported health, excellent/very good	1.12	0.93–1.34	0.23
Self reported health, fair/poor	0.72	0.56–0.93	0.01
Health Behaviors
Number of physician visits 1994–96*	1.04	1.03–1.06	<0.001
Breast self examination at least once per month, 1994–96	1.86	1.60–2.16	<0.001
Current smoker	0.53	0.45–0.63	<0.001
Body mass index	0.98	0.97–0.99	0.004
Demographics
Age
50–54	0.96	0.76–1.20	0.71
55–59	0.90	0.77–1.07	0.23
Family income
1st quartile	0.97	0.78–1.20	0.77
3rd quartile	1.24	1.00–1.53	0.05
4th quartile	1.92	1.52–2.42	<0.001
Married	1.30	1.08–1.57	0.006
Education
College or more	1.34	1.05–1.70	0.02
Less than high school	0.75	0.61–0.90	0.003
White	0.73	0.60–0.88	0.001
N	4,075
Wald Chi2	443.76
Pseudo R2	0.12
Log Likelihood	–2098.2

Table contains adjusted odds ratios and 95% confidence intervals and p values obtained from multivariate logistic regression.

Two-Stage Least Squares Results Accounting for Endogeneity. Findings on the linkage between health insurance and mammography had similar sign and significance when we accounted for the potential endogeneity of physician visits and BSE to mammography (Table 4). After accounting for endogeneity, persons with private health insurance had their probability of having a mammogram increased by 0.18 compared to women who were uninsured. Those with CHAMPUS or other insurance were 0.11 more likely compared to the uninsured. Each physician visit increased the probability of having a mammogram by 0.018, an effect that was about half as large as that found when not accounting for endogeneity. However, performance of BSE was not a significant predictor of mammography after accounting for potential endogeneity, in contrast to the findings when these two variables were treated as exogenous explanatory variables (Table 3). The relationship of the other explanatory variables to receipt of mammography were generally consistent with the results in Table 3.

Table 4.

Determinants of Mammography with Endogenous Explanatory Variables

	Two-Stage Least Squares
	Structural Model			First Stage			First Stage
	Dependent=Mammography			Dependent=Physician Visits			Dependent=BSE
	Coeff.	s.e.	p	Coeff.	s.e.	p	Coeff.	s.e.	p
Endogenous Explanatory Variables
Number of physician visits, 1994–96	0.019	0.0057	0.001
Conducts monthly self breast exam	−0.057	0.77	0.94
Exogenous Explanatory Variables
Health Insurance
Private health insurance	0.18	0.027	<0.001	1.52	0.38	<0.001	0.026	0.024	0.27
Medicaid	0.085	0.073	0.25	2.03	0.74	0.006	0.083	0.046	0.072
Other governmental	0.11	0.058	0.05	1.42	0.58	0.01	0.073	0.036	0.04
Perception Variables
Perceived probability of living to 75, 1994	0.031	0.084	0.71	−0.31	0.46	0.50	0.10	0.029	<0.001
Risk tolerant	0.027	0.022	0.25	0.033	0.39	0.93	0.0088	0.024	0.71
Self reported health,
excellent/very good	0.040	0.019	0.03	−1.62	0.30	<0.001	−0.00097	0.018	0.96
fair/poor	−0.11	0.041	0.01	3.91	0.41	<0.001	−0.010	0.025	0.68
Health Behaviors
Current smoker	−0.11	0.032	<0.001	0.045	0.30	0.88	0.035	0.019	0.06
Body mass index, 1994	−0.0041	0.0022	0.06	0.10	0.023	<0.001	0.0029	0.0014	0.04
Demographics
Age
50–54	−0.0096	0.020	0.63	0.092	0.37	0.80	−0.0025	0.023	0.91
55–59	−0.018	0.015	0.23	0.21	0.27	0.44	0.0037	0.017	0.83
Family income, 1994
1st quartile	−0.0085	0.023	0.72	−0.19	0.38	0.62	−0.0080	0.024	0.74
3rd quartile	0.044	0.024	0.06	−0.22	0.35	0.53	0.012	0.022	0.57
4th quartile	0.10	0.021	<0.001	−0.30	0.36	0.41	−0.0092	0.022	0.68
White	−0.070	0.045	0.12	0.44	0.31	0.15	−0.051	0.019	0.007
Married	0.057	0.040	0.19	0.072	0.31	0.82	0.051	0.019	0.009
Education
Less than high school	−0.056	0.023	0.01	−0.46	0.33	0.16	−0.017	0.020	0.40
College or more	0.023	0.059	0.69	0.56	0.36	0.12	−0.069	0.022	0.002
Instrumental Variables
Lagged physician visits, prior to wave 1	Excluded			0.19	0.017	<0.001	0.0025	0.0011	0.02
Lagged physician visits, prior to wave 2	Excluded			0.16	0.011	<0.001	0.00042	0.00070	0.55
N	4,075			4,075			4,075
F (21, 4053)	21.80			49.6			2.91
Prob>F	<0.001			<0.001			<0.001
Adj. R squared	0.11			0.20			0.01

Table shows two-stage least squares results from instrumental variable regression. The first three columns show coefficients, standard errors, and p values of the structural or second stage regression with dependent variable mammography between 1994 and 1996. Variables noted as excluded from the structural model were used only as instrumental variables in the first stage regressions. Columns 4–6 are for the first stage regression with dependent variable physician visits; columns 7–9 are for the first stage regression with dependent variable BSE.

We conducted a test to determine the power of the instrumental variables used in the first stage equations of our three-equation system to determine whether the exclusionary restrictions necessary for identification that we made in Table 4 were reasonable (Staiger and Stock 1997). We estimated a model of physician visits regressed on all of the exogenous and instrumental variables, and then conducted an F-test of whether the two instrumental variables (physician visits during the two years prior to wave one, and number of visits between waves one and two) were jointly significant, which they were: F (1, 4,063) 82.3, p <0.001. A similar test was run after estimating a model with conduct of BSE as the 0/1 dependent variable which also showed the instruments to be strong: chi-square (1) 5.24, p = 0.02. We ran conceptually the same test but estimated the 0/1 BSE dependent variable using OLS (ordinary least squares regressions) and also found the instruments used to be strong F (1, 4,055), p = 0.01.

Ability of Insurance Expansion to Increase Mammography. Of the 1,253 women who did not receive a mammogram, only 318 (25.4 percent) were uninsured. If all 318 women who were uninsured and who did not report a mammogram had been insured and gotten a mammogram, the proportion complying with current mammography recommendations between 1994–1996 would have risen from 72.5 percent to 79.0 percent. This represents an absolute upper bound effect of insurance expansions on mammography that could not be expected in practice. Using the coefficients shown in Table 4, we simulated the effect of providing health insurance coverage to all women; doing so would increase mammography rates from 72.5 percent to 75 percent.

Discussion

Health insurance coverage has generally been found to be a key determinant of utilization of mammography and other health care services. However, we found that the ability of health insurance expansions to substantially increase mammography screening rates above those now observed is limited even under very optimistic scenarios regarding the effect of insurance on newly insured women in the age range 50–64. Therefore, a Medicare buy-in is no panacea for increasing mammography screening among such women.

The major benefit of our study was the estimation of a two-stage least squares, or instrumental variable approach to assess the relationships between mammography, physician visits, and BSE while treating physician visits and BSE as endogenous explanatory variables. Hawley et al. (2000) estimated a two-stage least squares model of mammography use, treating physician recommendation for mammography as an endogenous variable and found that health insurance was not a strong predictor of mammography. While we did find health insurance to be a significant predictor of mammography, we also found that insurance expansions would not have a big impact on mammography screening rates.

Mammography and BSE have generally been studied separately, sometimes in relation to other health behaviors (Pearlman et al. 1996; Champion and Menon 1997; Cummings et al. 2000). Lauver et al. (1999) found variables associated with mammography differed from those associated with BSE, but did not consider the effect of one on the other. Our results were consistent with findings that women who conducted BSE were more likely to receive mammography (Rakowski et al. 1995; Pearlman, Rakowski, Ehrich 1996), though this effect went away when we treated BSE as an endogenous explanatory variable of mammography. Our results also line up with findings that show that mammography is increased among women with visits to a regular physician (O'Malley et al. 2001). While we had no measure of content of visit, our work is suggestive of a pattern whereby the content of the office exam and rapport between patient and physician is an important determinant of uptake of mammography (Hynes et al. 1998; Simon et al. 1998; Rimer 1997; O'Malley, Earp, and Harris 1997; Taplin et al. 1997; Hawley et al. 2000; Flocke,Stange and Zyzanski 1998;Fox,Siu and Stein 1994).

Limitations include the self-reported nature of mammography, BSE, and physician visits, although women's self-report of mammography use has been found to be accurate (Zapka et al. 1996; Degnan et al. 1992). We had no information on the type of physician visited (specialty, race, gender), the setting (private practice, clinic, emergency room, usual source of care), or the content of the visit (was mammography discussed, encouraged?). Nor did we have information concerning the extent of coverage for mammography in women's health insurance plans that has been shown to be important (Blustein 1995). We lacked other important factors including language barriers, lack of transportation, and discomfort of the procedure (Woloshin et al. 1997; Crane et al. 1996; Bobo et al. 1999).

Future work should consider the plausible endogeneity between mammography, BSE, and physician visits. Longitudinal studies with more precise timing of these events could help disentangle the causal relationships that are key to understanding the uptake of mammography, and perhaps improve policy designed to increase screening (Bastani, Maxwell, and Bradford 1996). However, no strategy as straightforward as expanding health insurance coverage emerges from our results that would increase mammography rates closer to universal screening.