Low levels of cardiorespiratory fitness (CRF) and physical inactivity have been associated with increased risk of several chronic diseases, including coronary artery disease, stroke, hypertension, diabetes, and some forms of cancer (4). Blair et al. (2) have suggested that measures of CRF may be more strongly associated with disease outcomes than physical inactivity. For instance, cardiovascular disease mortality rates are 8-fold higher in the lowest CRF quintile compared to the upper quintile, whereas rates are only 2-fold higher when the extremes of the physical activity distribution are contrasted (2,6,19). This disparity has been partially attributed to exposure misclassification of activity as a result of using self-reported physical activity instruments (2). Direct measurement of CRF, or its close approximation using standardized exercise testing protocols, requires considerable expense in terms of equipment, medical supervision, and exercise testing personnel (1,8). Therefore, the usefulness of CRF as an exposure variable in large scale epidemiologic studies is limited by the expense and impracticality of obtaining the measure.
Several recent reports have demonstrated that valid estimates of maximal oxygen consumption (O2max), a measure of CRF, can be obtained without exercise testing using personal data gathered from a questionnaire (9,10,14,26). The accuracy of questionnaire-based estimates of O2max have been evaluated in cross-validation analyses. These estimates have been reported to be less accurate than estimates of CRF using treadmill endurance performance, but of comparable accuracy to submaximal exercise tests suitable for field testing (8,14,18). Furthermore, these models appear to provide accurate estimates of CRF in all subject subpopulations except for young and highly fit groups (14).
Although these investigations have demonstrated the ability of questionnaire-based O2max prediction models to derive point estimates and group O2max means, they have not evaluated the prediction models for use in epidemiologic research. For these instruments to be useful in epidemiologic research they should be able to index the CRF of an individual so that accurate classifications of fitness within the cohort can be obtained. Estimates of disease risk could then be calculated by comparing rates of disease between fitness categories. Predicted fitness levels could be particularly useful in prospective investigations examining the relationship between fitness and relatively rare health outcomes, such as cancer, in which the size of the study population clearly precludes any form of exercise testing.
The present study sought to further our understanding of the classification accuracy of questionnaire-based CRF prediction models. Accordingly, the primary purpose of this investigation was to evaluate the ability of a nonexercise based CRF prediction instrument to classify an individual's fitness level. A secondary goal was to develop a generalized CRF prediction model from our data using height and body mass variables rather than percent body fat. This is advantageous since height and body mass are more readily obtained via questionnaire.
Subjects. The subjects for this investigation were 799 healthy adults (409 female and 390 male) between the ages of 19 and 79 yr. They were apparently healthy at the time of exercise testing, having no history of cardiovascular disease, orthopedic problems that would limit their exercise performance, and were not taking medication known to affect the heart rate or blood pressure response to exercise (1). Subjects were recruited from Amherst and Worcester, MA, for projects aimed at developing CRF tests using home exercise equipment. They were recruited through ads in local newspapers, health clubs, community centers, "Senior centers," retirement homes, and advertisements in university gymnasiums. Subjects who entered this sample were predominately Caucasian (>95%), well educated, and from middle to upper socioeconomic backgrounds. Before exercise testing, each subject signed an informed consent form which explained the risks and benefits of their participation in the project in accord with Human Subjects Committee protocol.
Physical activity and body composition assessment. Physical activity status was evaluated using an instrument developed by NASA's Johnson Space Center and used by Jackson et al. in prediction models of O2max(14,22). This instrument has subjects rate their last month of physical activity participation on a 0-7 scale. Responses of 0 and 1 represented no regular physical activity, whereas a response of 2 or 3 represented moderate intensity activities, and responses of 4 to 7 represented regular vigorous physical activity participation of increasing exercise time (see Appendix 1). Body density was estimated from skin-fold measures using the three site Jackson, Pollock, and Ward equations, and the Siri equation was used to convert estimated body density to percent fat (15,16,25). Body mass was measured using a standard physician's scale and height was measured in stocking feet. Body mass index (BMI) was calculated as kg·m−2.
Exercise testing and oxygen consumption measurement. All subjects completed continuous treadmill testing following the American College of Sports Medicine safety guidelines (1). A modified Balke protocol was employed in 724 subjects, and the standard Bruce protocol was used for 75 subjects (1). Previous research has demonstrated that different testing protocols that use the same test end-point criteria obtain comparable measures of O2max(20,21). Individuals tested using the modified Balke protocol walked at a self-selected walking speed (4.8-7.3 km·h−1) starting at 0% grade. Stages 1 and 2 lasted 4 min, and the remaining stages lasted 2 min. Grade in stage 2 increased to 5% and in 2.5% increments thereafter until volitional exhaustion. The standard Bruce protocol began at 2.7 km·h−1, and 10% grade and increased both speed and grade every three min until volitional exhaustion.
Subjects over 40 yr of age (N = 561) were tested in the Human Performance Laboratory at the University of Massachusetts Medical Center in Worcester. They were examined by a physician before testing (e.g., physical exam, resting blood pressure, and electrocardiogram), and were monitored via 12-lead electrocardiography during the exercise test. Subjects less than 40 yr of age (N = 238) were tested in the exercise physiology laboratory at the Amherst campus. Before testing, this group of subjects completed the Physical Activity Readiness Questionnaire to screen out individuals with cardiovascular disease or orthopedic problems (1). Heart rate was continuously recorded during the Amherst testing using a heart rate telemetry system (Polar Vantage Heart Rate Monitor, Polar CIC, Inc., Port Washington, NY).
Oxygen consumption was measured using standard indirect calorimetry procedures. The laboratory in Worcester used a Medgraphics CPX (St. Paul, MN) metabolic cart to obtain oxygen consumption values. The subjects tested in Amherst had expired gases continuously sampled from a 3-L mixing chamber and analyzed for oxygen and carbon dioxide concentration using Ametek oxygen (model S-3AI, Cambridge, UK) and carbon dioxide (model CD-3A) analyzers. Inspired gas volumes were continuously measured using a Rayfield Equipment (Waitsfield, VT) dry gas meter, and gas concentrations and inspired volumes were converted from analog to digital signals through a personal computer interfaced with an A/D board. O2 PLUS software (Exeter Research, Brentwood, NH) was used to calculate the respiratory exchange parameters. A formal reliability analysis of the two metabolic systems was not completed; however, consistent system calibration protocols and O2max criteria were used in each laboratory.
Peak heart rate during exercise was determined as the highest recorded heart rate during the exercise test. Maximal oxygen consumption (O2max) was determined as the highest value of oxygen consumption observed in the last minute of the test. All subjects satisfied two of the following three O2max criteria: 1) leveling off of oxygen consumption despite an increase in power output, 2) peak heart rate within 15 beats of age predicted maximum (220 - age), and/or 3) a respiratory exchange ratio = 1. 1.
Model development and cross-validation procedures. Model development was completed using traditional model development methods (17), and a final model was cross-validated using the predicted residual sum of squares (PRESS) method (11). The PRESS cross-validation method differs from the traditional split-sample approach (13) in that it allows for use of virtually the complete study sample in both the model development and cross-validation processes. Holiday et al. (11) have demonstrated that the PRESS cross-validation method produces slightly less optimistic estimates of future model performance as compared to the split-sample method.
The foundation of the PRESS method is the calculation of predicted residual values for each subject while that subject is deleted from the model fitting process. Specifically, PRESS residuals (rP) are calculated for each individual i as, (Equation 1) where [yi = measured O2max, yi′ = predicted O2max, and hi = the leverage for subject i (i.e., the diagonal element of the hat matrix, or hi = x(X′X)−1 x′) (24)]. The PRESS statistic is calculated as the sum of squares of rPi for all subjects, or (Equation 2) This value, along with data from the ANOVA table produced in the model fitting process, is used to calculate a PRESS adjusted R2 (RP2) and standard error of the estimate (SEEP) as follows (11), (Equation 3) The RP2 and SEEP values were used in cross-validation analyses.
In the model development process, univariate prediction of O2max from each continuous independent variable using linear regression in the full study sample (N = 799) was completed using PROC REG in the PC-based SAS system (version 6.11) (24). Regression diagnostics were evaluated to detect potential nonlinear relationships between dependent and independent variables. Independent variables that were related to O2max in a nonlinear fashion were transformed to meet the assumptions of the linear regression model (17). A generalized equation predicting O2max (mL·kg−1·min−1) was generated using the following independent variables: age (yr), age2, gender (male = 1, female = 0), physical activity status (0-7), height (m), body mass (kg), and BMI (kg·m−2). BMI and height and body mass were evaluated in separate models to avoid problems associated with collinearity. The group of predictor variables for the present analyses were developed from previously published reports (5,14), from analyses in a smaller subset of the present population (10), and because they could be obtained from self-reported questionnaire responses. The best fitting model was selected by evaluating the adjusted R2 and standard errors of the models (SEE). The SEE as a proportion of the mean O2max was also calculated (SEE% = (SEE/mean O2max) * 100).
The final prediction equation was cross-validated using predicted O2max values derived from the PRESS residuals (rPi) for each subject when they were deleted from the model fitting process. That is, the prediction model was developed on 798 subjects and this model was employed to predict a PRESS residual for the deleted subject (i.e., predicted O2max = measured O2max − rPi). The stability of the predicted O2max values were evaluated in the total sample (N = 799), and in subsamples of this group, by evaluating the mean and standard deviation (SD) of the measured and predicted O2max values, their mean difference, correlations between measured and predicted O2max and the SEEP in each group.
In addition, a sensitivity analysis of the final model was performed to examine the effect of reporting errors that could be derived from the use of self-reported predictor variables. Specifically, the individual effects on predicted O2max of reporting errors in body mass (−5 kg), height (+ 0.076 m), physical activity (+ 1 U), and the combined effects of these predictors were examined.
Classification of cardiorespiratory fitness. This portion of the analysis evaluated the ability of the prediction model developed in the present investigation to properly classify subjects into the appropriate CRF category. Measured CRF categories were employed as the criterion measure. By using age and gender specific O2max quintile cut-points derived from the study population, subjects were classified into categories of CRF based on their measured and predicted O2max values. The quintile cut-points employed are presented in Table 1. Classification into the correct measured quintile by the prediction equation was evaluated by cross-classification of the measured and predicted quintiles (Q1-5). Similar to the method used by Willett et al. (28), the percentage of correct classification overall and in the two extreme quintiles of the CRF distribution were evaluated. Specifically, the correct classification between measured and predicted Q1 and Q5, and the correct classification between measured Q1 and predicted Q1 or Q2 and between measured Q5 and predicted Q4 or Q5 were calculated. These analyses provide an indication of the ability of the prediction model to accurately classify the least (Q1) and most highly fit (Q5) subjects into the correct tail of the CRF distribution.
Descriptive characteristics are presented for men and women in Table 2. In terms of physical activity status, the individuals in this study population were, on average, regularly active in the month preceding their exercise test. Men and women under 60 yr of age on average reported regular participation in vigorous physical activity. Individuals over 60 yr of age on average reported regular participation in moderate intensity physical activity.
The best fitting CRF prediction equation contained the following independent variables; age, age2, gender, physical activity status, height, and body mass (R2 = 0.74, SEE = 5.64 mL·kg−1·min−1) (Table 3). The SEE% of the mean O2max was 15.2% (SEE% = 5.64/37.2 mL·kg−1 * 100). The model containing BMI had a similar R2 but a slightly higher SEE (R2 = 0.73, SEE = 5.76 mL·kg−1·min−1) compared with the model containing height and body mass. Thus, the height-body mass model was employed in the cross-validation analyses.
Adjustment of the model R2 and SEE using PRESS methods did not alter the adjusted R2 and produced only a slight increase in the SEE for this study population (RP2) = 0.74, SEEP = 5.66). The mean differences between measured and predicted O2max was −0.002 mL·kg−1·min−1 in the total sample and 2.1 mL·kg−1·min−1 or less in 12 of 13 subgroups (Table 4). In the most fit group (>44.5 mL·kg−1·min−1), the mean difference between predicted and measured O2max was −4.6 mL·kg−1·min−1. The correlations between measured and predicted O2max were generally high (r = 0.73 to 0.85), except across the O2max categories (r = 0.36 to 0.52). The lower correlations in these categories may be partly attributable to the restricted range of O2max values in each category. The SEEP value for the total sample was 5.7 mL·kg−1·min−1 and was larger in men than women, 6.3 vs 5.0 mL·kg−1·min−1, respectively. SEEP values ranged between 4.6 and 6.4 mL·kg−1·min−1 across each subgroup, except in the upper category of O2max in which SEEP was 7.0 mL·kg−1·min−1.
The sensitivity analysis examining the effects of reporting errors on predicted O2max values demonstrated that relatively large reporting errors did not inflate predicted O2max dramatically. Underreported body mass (−5 kg), overreported height (+ 0.076 m), and overreported physical activity (+ 1 U) produced only small increases in predicted O2max of 1.45, 0.70, and 1.48 mL·kg−1·min−1, respectively. The combined effects of underreported body mass and over reported physical activity would increase predicted O2max values by 2.93 mL·kg−1·min−1. Finally, predicted O2max would only be increased by 3.63 mL·kg−1·min−1 if all three predictors were reported with this level of error.
Classification into the correct CRF category was examined by cross-classification of measured and predicted O2max quintiles (Table 5). Correct classification occurs in the main diagonal of this table. In the total sample, 290 of 799 subjects (36%) were correctly classified into the appropriate fitness quintile. Among the 509 subjects who were not correctly classified, 369 of them (73%) were classified into an adjacent quintile. Thus, 659 of 799 subjects (83%) were classified correctly or within one fitness quintile. In gender specific analyses, men and women were classified correctly 38 and 35% of the time, or within one fitness quintile 83 and 82% of the time, respectively (data not shown).
Correct classification of the least fit individuals was evaluated by calculating the percentage of subjects in measured Q1, who were predicted to be in Q1 or Q2. Thus, 103 of 151 (68%) of the least fit subjects were correctly classified into the lower end of the CRF distribution. Similarly, in the upper end of the distribution, 134 of 157 (85%) of the most fit subjects were classified into Q4 or Q5 by the prediction equation (Table 5). In the tails of the distribution (Q1 and Q5), classification errors more than 2 quintiles from the correct measured CRF quintile occurred 8 and 2% of the time, respectively. Extreme misclassification into the lowest and highest quintile occurred only rarely [1 of 799 (0.13%)].
Correct classification into the upper and lower ends of the CRF distribution across subgroups of gender, age, physical activity status, and BMI are presented in Table 6. Slightly greater classification accuracy was observed in the upper as compared with the lower end of the CRF distribution. In addition, reduced classification accuracy was observed in the low end of the CRF distribution among younger, highly active subjects, and those with a BMI < 25 kg·m−2. In the upper end of the distribution, a slightly lower classification accuracy was observed among the least active (PAS = 0-1) and oldest subject (60-79 yr) groups. Relatively high classification accuracy (within one quintile) was observed among the least active and least fit (98%), and among the most active (PAS ≥ 5) and most fit individuals (91%).
The primary finding of this investigation was that a questionnaire-based prediction model can reasonably categorize CRF within a population. Although the overall correct classification across the quintiles of CRF was modest (36%), misclassified individuals were predicted to be in the quintile adjacent to the measured quintile 75% of the time. Therefore, 83% of all subjects were classified either correctly or within one quintile of measured CRF. Importantly, misclassification errors more than 2 quintiles from the truth were rare and prediction errors derived from the use of self-reported data were small. Therefore, these findings support the concept that CRF can be reasonably characterized in a population without exercise testing by using data obtained from a questionnaire.
Estimates of CRF may be obtained by combining standard subject information (age, gender, height, and body mass) with a brief assessment of an individual's physical activity status. The activity questionnaire can be completed quickly and may easily be integrated into a mailed questionnaire or an interview process (see Appendix 1). Investigators examining the effects of CRF on health outcomes should select the "measure" of CRF that most appropriately balances their requirements for CRF measurement precision with the feasibility of obtaining the measure. The most accurate but least feasible means of estimating CRF is by maximal treadmill testing [SEE = 3-4 mL·kg−1·min−1(8)]. Submaximal exercise testing provides slightly less precise estimates of fitness as compared to maximal treadmill testing [SEE = 4.4 to 6.2 mL·kg−1·min−1(14,18)]. Nonexercise estimates of CRF using data that may be obtained by questionnaire are similar in precision to submaximal exercise prediction models [SEE = 3.6-5.3 mL·kg−1·min−1(9,14)]. Interestingly, recent reports have demonstrated that the precision of CRF estimates may be enhanced when perceived functional ability and more refined physical activity assessments are utilized as predictors (9). In addition, Blair et al. (3) have demonstrated that sedentary behaviors add importantly in predicting fitness levels. Future research should focus on improving the precision of nonexercise prediction models since an improved precision of fitness prediction would be expected to produce better classification accuracy.
Results from our cross-validation analyses suggest that estimates of CRF are valid and that misclassification tended to be randomly distributed across most cross-validation subgroups, although slightly higher classification accuracy in the upper end of the CRF distribution was noted in this population. Although a high degree of classification accuracy would be desirable, a degree of exposure misclassification is often tolerated due to the difficulty of measuring many exposures in epidemiologic research (23). By comparison, the classification accuracy of the present CRF prediction model is similar to that of instruments which have proven to be useful in examining the effects of diet on several health outcomes (27).
A recent report suggesting that the effects of exposure misclassification secondary to measurement error can be complex resulting in risk estimates that can be biased toward or away from the null underscores the importance of detailed examination of the potential for misclassification error within a given study (7). Our finding of a slightly greater classification accuracy in our more fit subjects as compared with our less fit subjects is somewhat problematic. In this population and with this prediction equation, classification errors occurred more frequently in the least fit subjects. It is unclear whether this pattern of misclassification would be observed in other populations and using other CRF prediction equations. Since low-fitness is clearly associated with mortality and cardiovascular disease end-points (2), such classification errors could result in differential misclassification of exposure. Differential misclassification is of concern because it has the potential to bias estimates of disease risk either toward or away from the null value (23). That is, spurious associations with an outcome may be created through measurement error in the CRF exposure. Clearly, if CRF prediction instruments were to be employed in an epidemiologic investigation, further model development and/or cross-validation within the specific subject population being investigated could provide estimates of misclassification patterns in the target population. This information may then be employed to make inferences regarding the influence of misclassification on the risk estimates that were calculated.
We know of only a single investigation that has evaluated the ability of nonexercise-based CRF prediction models to classify fitness, and the investigators did so indirectly. Whaley et al. (26) developed a nonexercised based prediction model in a population of 702 men and 473 women. For a middle-aged man and woman in their study population, the investigators calculated a 95% confidence interval for their predicted CRF values. The investigators then compared the width of the 95% confidence interval (∼10.6 mL·kg−1·min−1) with the treadmill time predicted CRF quintile cut-points used by Blair et al. (2) to examine mortality in the Aerobics Center Longitudinal Study. Because the range of the 95% confidence interval was included within the upper and lower quintile of CRF in the Aerobics Center data, Whaley et al. concluded that nonexercise-based CRF prediction models were not sufficiently accurate to discriminate between individuals in the low and high CRF categories. Jackson et al. (12) have noted that the distribution of the residual error values (predicted - observed) have a bell-shaped distribution, and that the likelihood of inaccuracy is reduced at the tails of the residual distribution. That is, the true classification errors are likely to be less than those reported by Whaley et al. Our data support this interpretation and show that CRF prediction models are more accurate than the Whaley et al. results would suggest. Our data clearly show that predicted CRF values are only rarely misclassified in the extreme (0.13%, e.g., beyond three quintiles from measured values). Thus, the CRF prediction model appears to discriminate between individuals of high and low fitness reasonably well.
The accuracy of predicted CRF values from the height-body mass model presented here are comparable with previously published investigations that employed BMI in their prediction equation. The SEE values in the model of Jackson et al. was 5.3 mL·kg−1·min−1, compared with the present SEEP value of 5.7 mL·kg−1·min−1 in cross-validation analyses. Additionally, in subgroup analyses the range of SEE values reported by Jackson et al. (14) was 4.5-6.9 mL·kg−1·min−1, except among the most aerobically fit (SEE = 12.1 mL·kg−1·min−1). The range of SEEP values across subgroups in the present study were quite similar (Table 4).
Our finding of a significant underestimation of CRF among individuals in the upper end of the CRF distribution has been consistently observed in several of the published investigations that cross-validated their prediction models (10,14,26). Jackson et al. (14) point out that this fact may not limit the utility the prediction models because the prediction accuracy begins to diminish substantially only in the upper tail of the distribution (>55 mL·kg−1·min−1). In their data, this represented only 4% of their population, and in our data set only about 6% of the population would be affected. More importantly, the CRF prediction models have consistently underestimated, rather than overestimated, CRF values in the upper end of the distribution. Therefore, the models still classified highly fit subjects appropriately into the upper CRF categories.
The CRF prediction model developed in this investigation may have a limited generalizability because it was developed in a population that was healthy, highly fit, highly active, and very lean. However, the comparability of the prediction error (SEEP) obtained in the present analyses with other published nonexercise prediction models (14,26) suggests that the classification accuracy reported in the present analyses would be similar using other published equations. It is generally understood that prediction models are generalizable to populations that are similar to those from which they were developed (17). The stability of predicted CRF values using the present model is unknown in groups of individuals whose characteristics vary substantially from the range of characteristics in our cross-validation sample [e.g., children and adolescents and the obese (BMI > 30 kg·m−2)]. Furthermore, the present model was developed in a healthy and active population of primarily Caucasians, and physical activity status was evaluated only for leisure-time physical activity. It is unknown how stable CRF estimates would be in ethnically diverse populations and populations who have different physical activity behaviors or CRF levels. Whaley et al. (26) provided evidence that CRF can be accurately estimated in less active, fatter, and less fit subjects; Jackson et al. (14) provided evidence that CRF may be accurately predicted in subjects with cardiovascular disease; and George et al. (9) demonstrated that CRF can be predicted accurately in college students. Investigators planning to use nonexercise CRF prediction models should select a model that was developed and cross-validated on a population that most closely resembles their own study population.
In conclusion, the results of this investigation support the concept that CRF can be reasonably characterized in a population without exercise testing using data obtained from a questionnaire. In this investigation, a large proportion (83%) of all subjects were classified correctly or within one quintile of the true CRF category by the prediction model. Moreover, a large proportion (68 and 85%) of the least and most fit individuals were predicted to be in the correct tail of the CRF distribution, and extreme misclassification errors were rare. Therefore, predicted CRF values may be useful in etiologic research examining the effect of fitness level on health outcomes that require such large sample sizes that exercise testing is not feasible. Further research should focus on improving the precision of the predicted CRF values obtained from the nonexercise prediction models, thereby improving their classification accuracy. In addition, future research should examine the ability of predicted CRF values to provide additional insight into disease risk above and beyond that provided by the individual components of the prediction models (e.g., age, gender, activity status, height, and weight).
APPENDIX 1. Code for determining physical activity status.
Use the appropriate number (0 to 7) which best describes your general ACTIVITY LEVEL in the PREVIOUS MONTH.
Did not participate regularly in programmed recreation sport or heavy physical activity.
0 Avoid walking or exertion, e.g. always use elevator, drive whenever possible instead of walking.
1 Walk for pleasure, routinely use stairs, occasionally exercised sufficiently to cause heavy breathing or perspiration.
Participated regularly in recreation or work requiring modest physical activity, such as golf, horseback riding, calisthenics, gymnastics, table tennis, bowling, weight lifting, yard work.
2 10 to 60 minutes per week.
3 Over 1 hour per week.
Participate regularly in heavy physical exercise such as, running or jogging, swimming, cycling, rowing, skipping rope, running in place or engaging in vigorous aerobic activity type of exercise such as tennis, basketball, or handball.
4 Run less than 1 mile per week or less than 30 minutes per week in comparable physical activity.
5 Run 1 to 5 miles per week or spend 30 to 60 minutes per week in comparable physical activity.
6 Run 5 to 10 miles per week or spend 1 to 3 hours per week in comparable physical activity.
7 Run more than 10 miles per week or spend over 3 hours per week in comparable physical activity.
1. American College of Sports Medicine. Guidelines for Exercise Testing and Prescription.
Philadelphia: Lee & Febiger, 1991.
2. Blair, S. N., H. W. Kohl, R. S. Paffenbarger, D. G. Clark, K. H. Cooper, and L. W. Gibbons. Physical fitness and all-cause mortality: a prospective study of healthy men and women. J.A.M.A.
3. Blair, S. N., W. B. Kannel, H. W. Kohl, N. Goodyear, and P. F. Wilson. Surrogate measure of physical activity and physical fitness. Am. J. Epidemiol.
4. Bouchard, C., R. J. Shepard, and T. Stephens. The Consensus Statement: Physical Activity, Fitness, and Health.
Champaign, IL: Human Kinetics Publishers, 1994, pp. 9-76.
5. Bruce, R. A., F. Kusumi, and D. Hosmer. Maximal oxygen intake and nomographic assessment of functional aerobic impairment in cardiovascular disease. Am. Heart J.
6. Ekelund, L. G., W. L. Haskell, J. L. Johnson, F. S. Whaley, M. H. Criqui, and D. S. Sheps. Physical fitness as a predictor of cardiovascular disease mortality in asymptomatic North American men: the Lipid Research Clinics Mortality Follow-up Study. N. Engl. J. Med.
7. Flegal, K. M., P. M. Keyl, and F. J. Nieto. Differential misclassification arising from nondifferential errors in exposure measurement. Am. J. Epidemiol.
8. Foster, C., A. S. Jackson, M. L. Pollock, et al. Generalized equations for predicting functional capacity from treadmill performance. Am. Heart J.
9. George, J., W. Stone, and L. Burkett. Non-exercise O2max
estimation for physically active college students. Med. Sci. Sports Exerc.
10. Heil, D. P., P. S. Freedson, L. E. Ahlquist, J. Price, and J. M. Rippe. Non-exercise regression models to estimate peak oxygen consumption. Med. Sci. Sports Exerc.
11. Holiday, D., J. Ballard, and B. McKeown. PRESS-related statistics: regression tools for cross-validation
and case diagnostics. Med. Sci. Sports Exerc.
12. Jackson, A., L. Weir, and R. Ross. The validity of non-exercise cardiorespiratory fitness
models (letter). Med. Sci. Sports Exerc.
13. Jackson, A. S. Research design and analysis of data procedures for predicting body density. Med. Sci. Sports Exerc.
14. Jackson, A. S., S. N. Blair, M. T. Mahar, L. T. Weir, R. M. Ross, and J. E. Stuteville. Prediction of functional aerobic capacity without exercise testing. Med. Sci. Sports Exerc.
15. Jackson, A. S., and M. L. Pollock. Generalized equations for predicting body density for men. Br. J. Nutr.
16. Jackson, A. S., M. L. Pollock, and A. Ward. Generalized equations for predicting body density for women. Med. Sci. Sports Exerc.
17. Kleinbaum, D. G., L. L. Kupper, and K. E. Muller. Applied Regression Analysis and Other Multivariable Methods.
Belmont, CA: Duxbury Press, 1988.
18. Kline, G. M., J. P. Pocari, R. Hintermeister, et al. Estimation of O2max
from a one mile track walk, gender, and body weight. Med. Sci. Sports Exerc.
19. Paffenbarger, R. S., R. T. Hyde, A. L. Wing, and C. C. Hsieh. Physical activity, all-cause mortality, and longevity of college alumni. N. Engl. J. Med.
20. Pollock, M., R. Bohannon, K. Cooper, et al. A comparative analysis of four protocols for maximal treadmill stress testing. Am. Heart J.
21. Pollock, M., C. Foster, D. Schmidt, C. Hellman, A. Linnerud, and A. Ward. Comparative analysis of physiologic responses to three different maximal graded exercise test protocols in healthy women. Am. Heart J.
22. Ross, R., and A. Jackson. Exercise Concepts, Calculations, and Computer Applications.
Carmel, IN: Benchmark Press, 1990.
23. Rothman, K. Modern Epidemiology.
Boston, MA:Little Brown and Company, 1986.
24. SAS Institute. SAS/STAT User's Guide,
Version 6, Vol. 2, 4th Ed. Cary, NC: SAS Institute, Inc., 1989.
25. Siri, W. E. Body composition from fluid spaces and density. In: Techniques for Measuring Body Composition,
J. Brozek and A. Henschel (Eds.). Washington, DC: National Academy of Science, 1961, pp. 223-244.
26. Whaley, M. H., L. A. Kaminsky, G. B. Dwyer, and L. H. Getchell. Failure of predicted O2
peak to discriminate physical fitness in epidemiological studies. Med. Sci. Sports Exerc.
27. Willett, W. Reproducibility and validity of food-frequency questionnaires. In: Nutritional Epidemiology,
W. Willett (Ed.). New York: Oxford University Press 1990, pp. 92-126.
28. Willett, W. C., L. Sampson, M. J. Stampfer, et al. Reproducibility and validity of a semi-quantitative food frequency questionnaire. Am. J. Epidemiol.