Mini Mental State Examination: Influence of sociodemographic, environmental and behavioral factors, and vascular risk factors

0895.4356/96/$15.00 SSDI 0895-4356(95)00541-B

] Clin Epidemiol Vol. 49, No. 1, pp. 73-78, 1996 Copyright 0 1996 Elsevier Science Inc.

ELSEVIER

Mini Mental State Examination: Influence of Sociodemographic, Environmental and Behavioral Factors, and Vascular Risk Factors W. Freidl, ’ R. Schmidt,2 W. J. Stronegger, I A. Irmler,2 B. Reinhart, ‘INSTITUTE

OF SOCIAL

MEDICINE

AND

‘DEPARTMENT

OF NEUROLOGY,

UNIVERSITY

OF

and M. Koch’

GRAZ,

GRAZ,

AUSTRIA

ABSTRACT.

Age and education have been found to affect the Mini Mental State Examination (MMSE) score of elderly normals, but there have been no studies assessing the influence of environmental and behavioral factors on this test. We therefore administered the MMSE to 1437 normal elderly subjects in the setting of a stroke prevention study and correlated their results to 16 sociodemographic, environmental, and behavioral factors, and vascular risk factors. Study statistics composed of a multiple logistic regression analysis and graphical models revealed the relations between variables in greater detail. Logistic regression yielded education level, occupational status, living as a single, general life stress, physical strain, and physical inactivity to be independent predictors of the MMSE score. Age was not included in this model. Graphical models demonstrated similar results, but did not include living as a single and physical inactivity. As shown in our independence graph, general life stress is the crucial predictor and links other environmental and sociodemographic variables with the test performance of elderly normals. J CLIN EPIDEMIOL 49; 1:73-78, 1996.

KEY WORDS.

MMSE, dementia, environmental

and behavioral factors, graphical models

INTRODUCTION The Mini Mental State Examination (MMSE) is among the most commonly used screening tests for evaluating the cognitive status in clinical practice and research. The clinical diagnosis of dementia usually includes this cognitive test as a component [l-5]. It is a brief cognitive test with 12 items assessingorientation, attencion, concentration, memory, language, and constructional abilities. The MMSE results have been consistently found to correlate with age and education. Lower scores were associated with increasing age, lower socioeconomic status, and lower educational level [6-141. Several authors have therefore recommended the use of age- and education-specific cut-off scores in order to avoid overdiagnosis of dementia in low-education individuals [5,7]. The associations between the MMSE score and education is a matter of debate. Some have argued that individuals with low levels of education might not be able to take tests well in spite of their cognitive capacity [2,12]. However, Bassett and Folscein [15] showed that loweducation individuals were also less able to perform tasks of everyday living, suggesting that these subjects indeed might be more cognitively impaired. Other authors [9,16] suggested that low-education individuals are s?lbject to a greater variability of pathological conditions. This interpretation is consistent with an increasing prevalence of many diseases among less educated individuals and those of lower socioeconomic class. They have generally been found to have higher blood pressure, to be overweight, co have higher serum cholesterol levels, and a higher rate of cardiac disease than high-education persons [ 17-191. A second aspect is that people with different levels of education must cope with different environmental hazards, experience different social stresses, and have different patterns of alcohol and cigarette consumption, eating, and physical activity. Moreover, social resources and support (Received in revised form 23 February 1995)

may vary between different educational strata. As has been emphasized in an article by Berkman [9], sociodemographic factors, and/or socioenvironmental exposures and behaviors correlated with them, might potentially be regarded as risk factors of cognitive impairment per se. Berkman states that they ought to be investigated rather than be obscured by essentially overadjusting. Therefore our study, for the first time, attempts to explore the implications and intercorrelations of a large set of sociodemographic, environmental, behavioral, and vascular risk factors on the total MMSE score in normal elderly communitydwelling individuals participating in a stroke prevention study. For data analysis we used logistic regression analysis and applied graphical models thereafter to better demonstrate the intervariable relationships. SUBJECTS

AND METHODS

Between September 1991 and March 1994, a sample of 7028 individuals aged 50 to 80 years was randomly selected from the official register of residents of Graz, Austria, using a two-step procedure stratified on age and sex. All individuals received a letter of invitation to participate in the Austrian Stroke Prevention Study, a study on the prevalence and effects of risk factors in our community. A total of 1573 subjects agreed to participate, with 1443 individuals fulfilling the inclusion criteria. Six subjects did not complete the test procedure, therefore a total of 1437 subjects has been included in the ultimate data analysis. A random sample of 200 nonresponders was interviewed by telephone and did not differ from responders for age, sex, and educational level. All study participants underwent a structured clinical interview, a physical and neurological examination, three blood pressure readings, electrocardiogram (ECG), and laboratory testing including blood cell count and a complete blood chemistry panel. Inclusion criteria were as follows: (1) no history of neuropsychiatric disease, (2) no complaints of forgetfulness, (3) no evidence of alcohol or drug dependence disorders, and (4) no clinically significant laboratory abnormalities.

Freidl et al.

74 The MMSE was administered to all individuals by two trained physicians under constant laboratory conditions. The MMSE score falls in a range from lowest possible (0) to highest (30). In addition to the administration of the MMSE, three categories of possible predictors of the MMSE score were established. These were (1) sociodemographic variables that were considered as objective and stable traits to describe the socioeconomic status or social class position of an individual, (2) environmental and behavioral factors assessing participant healthrelated behaviors as well as demands and resources of the social microenvironment, and (3) well-documented vascular risk factors. The frequency distributions of the different variables within each category are listed in Table 1. Sociodemographic data included age, sex, educational level, occupational status, and living as a single. Education was categorized by years of schooling completed, and each participant was assigned to the category corresponding to the highest grade achieved. Occupation was categorized as blue-collar worker, white-collar worker, housewife/man, and retired. Environmental and behavioral factors included general life stress, physical strain, social contacts, physical inactivity, and cigarette and alcohol consumption. General life stress was assessedby a semistructured interview covering the spheres of family, occupation, and financial power. Each sphere was coded as problems present (1) or absent (0). We derived the total general life stress score by summing up the present problems (O-3). The two highest grades of general life stress were collapsed for further statistical analysis since they applied to fewer than 40 persons. Physical strain and social contacts were evaluated by a three-point selfdescription rating scale (1, low; 2, moderate; 3, high). Physical activity of subjects was assessedby an interview evaluating leisure time and occupational activity status. Subjects were defined as physically inactive if they reported that “they read, watch television, and do things that involve physical activity for less than 4 hours a week,” and if their work was done mainly while sitting, without much walking. Study participants were defined as smokers if they currently smoked more than 10 cigarettes a day. Alcohol consumption was coded as daily or not daily. Vascular risk factors included hypertension, diabetes mellitus, cardiac disease, cholesterol level, and obesity. Their diagnosis relied on subject history and appropriate laboratory findings. Arterial hypertension was coded as present, if a subject had a history of arterial hypertension with repeated blood pressure readings above 160/95 mmHg, or if the readings at examination exceeded this limit. Diabetes was considered evident, if a subject was treated for diabetes at the time of examination or if the fasting blood glucose level at examination exceeded 140 mg/dl. Cardiac heart disease was assumed to be present, if there was evidence of cardiac abnormalities known to be a source for cerebral embolism [20], or evidence of coronary heart disease as per the Rose questionnaire [21] or appropriate ECG findings (Minnesota code I [ 1 to 31, IV [ 1 to 31, or V [ 1 to 2]), or if an individual presented signs of left ventricular hypertrophy on ECG (Minnesota code III, 1 to 3). Hypercholesterolemia was thought to be present if a subject received drug treatment for hypercholesterolemia or if the total cholesterol level at examination exceeded 250 mg/dl. Values of 201 up to 250 mg/dl were categorized as borderline. “Overweight” was defined on the basis of the body mass index (kg/m’). Men were categorized as overweight if the body mass index was 227.8. The cut-off point for women was 227.3. The MMSE score was dichotomized (high and low), and the median was used as the cut-off point (score 27; range, 19-30). Missing values on the predictor variables were excluded pairwise from data analysis. Statistical analysis was performed first with the personal computer version of the Statistical Package for Social Sciences (SPSS) [22]. Group differences were calculated with a x2 test and the Mann-

Whitney U-test. A multiple logistic regression analysis was used to assessthe relative contribution of sociodemographic, environmental, and behavioral factors, and vascular risk factors on the MMSE score. Backward selection stepwise regression was applied to create a model of predictors of test performance. Because results of statistical procedures can vary considerably depending on model building we decided to perform a second analysis utilizing the so-called graphical models, which allow the critical assessment of the results of the first analysis and reveal the relations between variables more comprehensively. The statistic program DIGRAM [23] was used, which is designed for discrete graphical models. Graphical models are a multivariate statistical method particularly useful for explaining and describing direct and indirect interrelationships between several variables. Causal pathways and intervening variables can be identified. Conditioning on subsets of variables is the key theoretical concept underpinning graphical models. Graphical models were introduced by Darroch et al. [24] and Lauritzen [25], who showed that the family of models for discrete variables defined in terms of conditional independence between pairs of variables is a subset of the broader class of log linear models and that the class of decomposable models is a subclass of the graphical models. Whittaker [26] published a detailed and readable review. Graphical models are characterized by an independence graph, where vertices represent variables and edges or arrows symbolize conditional relationships. There will be no edge or arrow between two variables if they are conditionally independent given the remaining variables of the model. The independence graph is a visualization of the statistical model in the same way that path diagrams are used to illustrate causal models. An independence graph is an object of graph theoretical methods whereas a graphical model is a statistical procedure. The most characteristic feature of the graph, the global Markov property, entails an explicit set of rules for interpreting the independence graph. This property means that any two subsets of variables separated by a third subset are conditionally independent only on variables in the third subset [26]. To illustrate these concepts by a simple example, take for instance the question that two variables A and B may be associated only through a confounder variable X. Figure 1 (case i) depicts the case of no association between A and B after controlling for X, which means A and B are conditionally independent given X, whereas Fig. 1 (case ii) illustrates the case of an association remaining after controlling for X. Depending on the context, in case i X also could be an intervening variable in the causal pathway from A to B, indicating no direct effect of A on B. For recursive graphical models the variables are grouped into a linearly ordered set of recursive levels building up block recursive independence graphs. Within the same level connections between vari-

O)*\,/”

wi‘\x/’

FIGURE 1. Independence graphs. Case i, A and Bare conditionally independent given X; case ii, A and B remain associated after controlling for X.

MMSE: Environmental TABLE

75

and Behavioral Factors

1. Sample characteristics

for MMSE predictors MMSE score High (%)

Low (%)

Predictors

n

Group differences/ p values

1. Sociodemographic variables Age (years) SO-55 56-60 61-65 66-70 71-80 Sex Female Male Education (years) 58 9-10 11-13 14-18 Occupational status Blue-collar worker, 1 White-collar worker, 2 Housewife/man, 3 Retired, 4 Living as a single

39 50

307 306 382 338 104

p < 0.01

2; 60

61 50 47 43 40

54 46

46 54

830 607 i

NS

40

p < 0.01

40 25

2 75

443 581 315 98

52 34 50 57 59

48 66 50 43 41

62 308 234 825 1 381

2. Environmental General life stress Low 0 1 High 2 Physical strain Low 1 2 High 3 Social contacts Low 1 2 High 3 Physical inactivity Cigarette consumption Alcohol consumption

p < 0.01 p < 0.01

and behavioral factors

39 45 57

61 55 43

242 374 816

p < 0.01

46 41 58

54 59 42

108 543 781 I

p < 0.01

57 50 46 45 52 42

43 50 54 55 48 58

453 531 453 388 194 129

p < 0.01 p < 0.01 NS NS

3. Vascular risk factors Hypertension Diabetes mellitus Cardiac disease Cholesterol level 1200 201-250 >250 Obesity Abbreviation:

52 50 53

48 50 47

573 118 508

48 51

52 49 48 45

338 694 403 157

::

NS NS NS NS NS

NS, Not significant; n, Number of subjects.

ables are undirected-symbolized by edges-and between levels they are directed from the lower to higher level-symbolized by arrows. Assumptions on recursive structure must be formulated before the statistical analysis. This partitioning of all variables into levels must be done on the basis of theoretical foundations reflecting usually causal ordering. Using subject matter knowledge one can set up this ordering, but the procedure is not designed to infer a causal ordering from the data.

The basic tests of conditional independence include test statistics occurring naturally in the context of log linear models, but considering in addition alternative test statistics that in some cases have considerably greater power and greater credibility. Two basic tests of conditional independence are available in DIGRAM: 1. Pearson’s x2 statistic 2. The partial y coefficient suggested by Davis [27]

76

Freidl et al.

The partial y is a rank correlation based on the Goodman and Kruskal y, a modification of the Kendall T for use in contingency tables. These and other statistics are discussed in Agresti [28,29]. As an alternative to the asymptotic p values, DIGRAM supports Monte Carlo approximations of exact p values from the conditional distribution of test statistics given sufficient marginals under the hypothesis [30]. For this procedure a 1% significance level was assumed and we again used the backward elimination method.

with moderate physical strain, as intervening variable, leading to a high MMSE score. High educational level, occupational status “blue-collar or whitecollar worker,” high physical strain, and low social contacts are related with high general life stress. Moreover, low social contacts are associated with low educational level, living as a single, and high age. These associations clearly demonstrate that general life stress keeps a central position as intervening variable being inversely correlated with the MMSE score.

RESULTS

DISCUSSION

Table 1 displays the distribution of predictor variables according to the dichotomized MMSE score. Table 2 gives the results of logistic regression analysis predicting the likelihood of scoring high or low in the MMSE. The final model included educational level, occupational status, living as a single, general life stress, physical strain, and physical inactivity as independent predictors of the MMSE score. Individuals who were highly educated, white-collar workers, were not living as a single, and sustained low general life stress and moderate physical strain while being highly physically inactive were more likely to obtain high MMSE scores. In a second step the data were analyzed using the above-described method of graphical models. A set of recursive levels was defined. These were

The major goal of our study was to identify factors that may explain the variance of MMSE results in a normal population. This warrants a study among mentally healthy subjects, as the inclusion of patients suffering from dementia or memory impairment would bias our analysis. To exclude the possibility of any interrelations between sociodemographic variables, environmental, behavioral, and vascular risk factors, and the MMSE score being attributable to the prevalence of dementia in our sample, such cases were removed from data analysis. Most previous investigations included sociodemographic variables such as sex, age, educational level, and occupational status to predict the MMSE score. Environmental and behavioral factors were discussed as other possible influencing variables in some studies but were never integrated into the empirical models [9,16]. To overcome this weakness we here added these hypothetically relevant determinants. To guarantee the soundness of the empirical model we have chosen two different statistical procedures for data analysis. This study indicates that educational level, occupational status, physical strain, and general life stress are statistically relevant predictors for the dementia test score by both methods used. Some of our findings seem to be in the focus of interest. As has already been pointed out, it was possible to confirm educational level as a determinant in the vast majority of studies. It seems remarkable, however, that in our study the MMSE score was found to be independent of age. The reason for this independence might stem from the association of occupational status “retired” and high age, meaning that the information of high age was included in the predictor variable “occupational status.” The logistic regression procedure obscures this

1. 2. 3. 4.

Socio-demographic characteristics Environmental and behavioral factors Vascular risk factors MMSE score

The results are shown in Fig. 2, where the arrows indicate directed associations between different levels and the lines show relations of variables within the same level (all at p < 0.01). The bivariate associations within and between levels are specified in the key to Fig. 2. The main results of the independence graph are seen in Fig. 2. High educational level, occupational status “white-collar worker,” low general life stress, and moderate physical strain are directly associated with a high MMSE score. In addition, occupational status “white-collar worker” is interrelated

TABLE

2. Final model of logistic regression analysis backward Coefficient

Variable Education 9-10 11-13 14-18 Occupation

4 Single (1) Stress 1

2 Physical 5 Physical

SE

elimination

procedure

df

P

3

0.0000 0.0490 0.0000 0.0000 0.0000 0.2090 0.7275 0.0676 0.0051

(1)

95%CI

1.3179 2.3770 4.4283

1.00-1.74 1.73-3.28 2.61-7.51

1.4636 0.9011 0.5999 0.6952

0.81-2.65 0.50-1.62 0.35-1.04 0.54-0.90

0.2760 0.8659 1.4880

0.1402 0.1642 0.2694

0.3809 -0.1041 -0.5109 -0.3635

0.3032 0.2988 0.2796 0.1299

-0.4577 -0.7950

0.1804 0.1841

0.0001 0.0112 0.0000 0.0130

0.6328 0.4516

0.44-0.90 0.32-0.65

0.2301 0.2384 0.1306

0.0248 0.5851 0.0452

1.6762 1.1390 1.2989

1.07-2.63 0.71-1.82 1.00-1.68

1

strain

inactivity

Odds ratio

0.5165 0.1302 0.2615

“The lowest grade of each variable was used as reference category for computation of the odds ratios. Abbreviations: SE, standard error; df, degrees of freedom; Cl, confidence interval.

MMSE: Environmental

MMSE

77

and Behavioral Factors

3 VASCULAR RISK FACTORS

2 ENVIRONMENTAL AND BEHAVIORAL FACTORS

1 SOCIO-DEMOGRAPHIC FACTORS

FIGURE 2. The independence graph. Abbreviations: MMSE, Mini Mental State Examination score; cardiac, cardiac disease; diabe, diabetes; ht, hypertension; cholest, cholesterol; obesity, obesity; stress, life stress; sot cant, social contacts; phys str, physical strain; phys ina, physical inactivity; smoking, cigarette smoking; alcohol, alcohol consumption; age, age; sex, sex; educatio, educational level: occupati, occupational status; single, living as a single. Key to dependencies: Within levels 1. Sociodemographic factors Sex female-educational level low Occupational status housewife Living as a single present Educational level low-occupational status blue collar Age high-occupational status retired Living as a single present 2. Environmental and behavioral factors Physical strain high-life stress high Physical inactivity low Social contacts low-life stress high Life stress high-physical inactivity low Cigarette smoking high-alcohol consumption high 3. Vascular risk factors Diabetes present-hypertension present Cardiac disease present Obesity present Hypertension present-cardiac disease present Obesity present Between levels 1 and 2 Educational level high-social contacts high Life stress high

pathway, while the relationship between age and occupational status was demonstrated by the graphical model. The ability to assessintervariable relationships is a clear advantage of graphical over regression models. In the graphical modeling procedure numerous plausible and welldocumented relationships appeared within and between the levels, as was the case in many other social-scientific studies and in studies on medical risk factors. A closer look at the interdependencies between the different levels reveals that sociodemographic variables such as

Occupational status blue/white collar-life stress high Occupational status white collar-physical strain moderate Living as a single present-social contacts low Sex male-alcohol consumption high Cigarette smoking high Age high-cigarette smoking low Social contacts low Between levels 1 and 3 Sex female-cholesterol high Age high-hypertension high Diabetes present Cardiac disease present Educational level low-obesity present Between levels 2 and 3 Physical inactivity present-obesity present Between all levels and the MMSE high Educational level high Occupational status white collar Life stress low Physical strain moderate

educational level, occupational status, sex, and age are associated with environmental, behavioral, and vascular risk factors. As an example, a high educational level itself is associated with a high MMSE score, but a high educational level is also related to high general life stress of an individual, which in turn has a negative effect on the MMSE results. However, in our study high education is a directly protective factor, but considering the moderator general life stress it has a negative effect on cognitive performance. The interrelations of the educational level with environmental fac-

78 tors suggest that not only the educational level itself must be considered a factor influencing the cognitive performance of the elderly, but also concurring environmental demands and resources. In our study the variable of general life stress is of pivotal importance. This stress variable links the variables of quality of social relationships, physical strain, occupational status, and educational level with the level of cognitive performance. Therefore stress must be considered as an intervening variable. .This demonstrates the impact of environmental factors in the life style of an individual on the judgment of cognitive performance. As an overall conclusion it can be stated that the evaluation through recursive graphical models allows a much more refined constitution, analysis, and interpretation of the model than the logistic regression backward elimination procedure. Although the variable of general life stress turned out to be a significant predictor in both methods, the first method allowed a much more elaborate interpretation of this factor. In comparison to the logistic regression analysis the recursive graphical models method offers the decisive advantage of pointing out both the associations of predictors on the same hierarchical level as well as the intervening function of variables between different levels. References 1 Folstein MF, Folstein SE, McHugh PR. Mini-Mental State: A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975; 12: 189-198. 2 O’Connor DW, Pollitt PA, Treasure FP, Brook CPB, Reiss BB. The influence of education, social class and sex on Mini-Mental State scores. Psychol Med 1989; 19: 771-776. 3 Brayne C, Galloway P. The association of education and socioeconomic status with the Mini-Mental State Examination and the clinical diagnosis of dementia in elderly people. Age Ageing 1990; 19: 91-96. 4 Gagnon M, Letenneur L, Dartigues J-F, Commenges D, Orgogozo J-M, Barberger-Gateau P, Alp&witch A, D&s A, Saiamon R. Validity of the Mini-Mental State Examination as a screening instrument for cognitive impairment and dementia in French elderly community residents. Neuroepidemiology 1990; 9: 143-150. 5. Bleecker ML, Molla-Wilson K, Kawas C, Agnew I. Age-specific norms for the MMSE. Neurology 1988; 38: 1565-1568. 6. Fillenbaum 0, Heyman A, Williams K, Prosnitz B, Burchett B. Sensitivity and specificity of standardized screens of cognitive impairment and dementia among elderly black and white community residents. J Clin Epidemiol 1990; 43: 651-660. 7. Crum RM, Ar.thony JC, Bassett SS, Folstein MF. Population-based norms for the Mini-Mental State Examination by age and educational level. JAMA 1993; 269: 2386-2391. 8. Magaziner J, Bassett SS, Hebel RJ. Predicting performance on the MiniMental State Examination. Use of age and education-specific equations. J Am Geriatt Sot 1987; 35: 996-1000. 9. Berkman LF. The association between educational attainment and mental status examination: Of etiologic significance for senile dementias or not? J Chron Dis 1986; 39: 171-174.

Freidl et al. 10. lorm AF, Scott R, Henderson AS, Kay DWK. Educational level differences bn the Mini-Mental State: The role of the test bias. Psycho1 Med 1988; 18: 727-731. 11 Zhang M, Katzman R, Salmon D, Jin H, Cai G, Wang Z, Qu G, Grant 1, Yu E, Levy P, Klauber MR, Liu WT. The prevalence of dementia and Alzheimer’s disease in Shanghai, China: Impact of age, gender, and education. Ann Neural 1990; 30: 428-437. 12. O’Connor DW, Pollitt PA, Treasure FP. The influence of education and social class on the diagnosis of dementia in a community population. Psychol Med 1991; 21: 219-224. 13. Fratiglioni L, Grut M, Forsell Y, Grafstr(im M, Homen K, Eriksson K, Viitanen M, Bsckman L, Ahlbom A, Winblad B. Prevalence of Alzheimer’s disease and ocher dementias in an elderly urban population: Relationship with age, sex and education. Neurology 1991; 41: 1886-1892. 14. Uhlmann RF, Larson EB. Effect of education on the Mini-Mental State Examination as a screening test for dementia. J Am Geriatr Sot 1991; 39: 876-878. 15. Bassett SS, Folstein MF. Cognitive impairment and functional disability in the absence of psychiatric diagnosis. Psycho1 Med 1991; 21: 77-84. 16. Snowdon DA, Ostwald SK, Kane RL, Keenan LN. Years of life with good and poor mental and physical function in the elderly. J Clin Epidemiol 1989; 42: 1055-1066. 17. Pincus T, Callahan LF, Burkhauser RV. Most chronic diseases are reported more frequently by individuals with fewer than 12 years of formal education m the age 18-64 United States population. J Chron Dis 1987; 40: 865-874. 18. Pincus T, Callahan LF. Associations of low formal education level and poor health status: Behavioral, in addition to demographic and medical, explanations! J Clin Epidemiol 1994; 4: 355-361. 19. Townsend P, Davidson N. Inequalities in Health: The Black Report. Penguin Books, London, 1988. 20. Kittner SJ, Sharkness CM, Price TR. Infarcts with a cardiac source of embolism in the NINCDS Stroke Data Bank: Historical features. Neurology 1990; 40: 281-284. 21. Rose GA. The diagnosis of ischemic heart pain and intermittent claudicanon in field surveys. Bull WHO 1962; 27: 645-658. 22. Norusis MJ (for SPSS, Inc.). SPSS Ad vanced Statistics User’s Guide. SPSS, Inc., Chicago, 1990. Program for Discrete Graphical 23. Kreiner S. User Guide to DIGRAM-a Modeling. Research Report 89-10. Statistical Research Unit, University of Copenhagen, Copenhagen, Denmark, 1989. 24. Darroch JN, Lauritzen SL, Speed TP. Markov fields and log linear interactions model for contingency tables. Ann Stat 1980; 8: 522-539. 25. Lauritzen SL. Lectures on Contingency Tables, 3rd Ed. University of Aalborg Press, Aalborg, Denmark, 1989. 26. Whittaker J. Graphical Models in Applied Multivariate Analysis. John Wiley & Sons, New York, 1990. 27. Davis ]A. A partial coefficient for Goodman and Kruskal’s gamma. J Am Stat Assoc 1967; 69: 174-180. 28. Agresti A. Analysis of Ordinal Categorical Data. John Wiley & Sons, New York, 1984. 29. Agresti A. Categorical Data Analysis. John Wiley & Sons, New York, 1990. 30. Kreiner S. Analysis of multidimensional contingency tables by exact conditional test: Techniques and strategies. Stand J Stat 1987; 14: 97-112.