0277-953619053.00 + 0.00 Copyright c 1990 Pergamon Press plc
Sot. Sri. Med. Vol. 30. No. 7. pp. 81 l-818, 1990 Printed in Great Bntain. All rights reserved
PATIENT SOCIODEMOGRAPHIC CHARACTERISTICS AS PREDICTORS OF SATISFACTION WITH MEDICAL CARE: A META-ANALYSIS JUDITH A. HALL*
and MICHAEL C. DORNAN
Department of Psychology, Northeastern University, Boston, MA 02115 and Department of Social Medicine and Health Policy, Harvard Medical School, Boston, MA 02115, U.S.A. Abstract-A meta-analysis was performed to examine the relation of patients’ sociodemographic characteristics to their satisfaction with medical care. The sociodemographic characteristics were age, ethnicity, sex, socioeconomic status (three indices), marital status, and family size. Greater satisfaction was significantly associated with greater age and less education, and marginally significantly associated with being married and having higher social status (scored as a composite variable emphasizing occupational status). The average magnitudes of relations were very small, with age being the strongest correlate of satisfaction (mean r = 0.13). No overall relationship was found for ethnicity, sex, income, or family size. For all sociodemographic variables, the distribution of correlations was significantly heterogeneous, and statistical contrasts revealed the operation of several moderating variables. The meaning of the overall results and their relation to earlier reviews is discussed. Key words-consumer
satisfaction, sociodemographic
INTRODUCTION
Patients’ satisfaction with their medical care has been examined in relation to many antecedents and conse-
quences. Some relations are well established, according to reviews of the literature. Satisfaction is related to the affective quality of the provider’s manner, the amount of information conveyed by the provider, the provider’s technical and interpersonal skill, and the length of the medical visit [l]. Younger physicians (e.g. residents) receive higher satisfaction ratings than more experienced physicians [2], and patients with better health [3,4] and more continuous health care [3, 5,6] are relatively high in satisfaction. Patients higher in satisfaction are more adherent to medical recommendations [3, 7,8]. Methodological variables are also related to satisfaction; notably, scores are higher when patients evaluate their own care as opposed to medical care or physicians in general [2,31. Patients’ sociodemographic characteristics are the variables most often studied in relation to satisfaction but they are, paradoxically, the least well understood. Age, sex, and the like are easily collected and this no doubt explains their frequent appearance in studies of patient satisfaction. Their frequent appearance is definitely not explained by the dramatic or consistent nature of their relations with satisfaction, for authors often comment that patient background variables show relations that are weak, inconsistent, or nonexistent. Patients’ age is seen as a more consistent correlate than other sociodemographic variables. Most reviewers agree that older patients are more satisfied [3,6,9], but one recent reviewer appears to be *Address correspondence to: Judith A. Hall, Department of Psychology, Northeastern University, 360 Huntington Ave., Boston, MA 02115, U.S.A.
characteristics, meta-analysis, review
skeptical about this [S]. Female patients have been seen as more satisfied [3, 5,6]. An early review found higher education to be associated with greater satisfaction [6]. but a more recent one concluded that relations were completely inconsistent [S]. Social class, income, marital status, and race have also been seen as inconsistent [5,6,9], but larger family size has been associated with less satisfaction [6]. By and large, reviewers have failed to reach confident conclusions. At the extreme, Fox and Storms [IO] reached no conclusions at all and summarized the situation as follows: The literature on satisfaction with health care presents contradictory findings about sociodemographic variables.. The situation has grown so chaotic that some writers dismiss [sociodemographic] variables as reliable predictors of satisfaction [p. 5571. As a possible resolution of the chaos, Fox and Storms suggested that moderating variables may be operating that obscure an overall understanding. A
moderating variable is one that influences the nature of the relation between two other variables. The goal in this regard would be to uncover sample characteristics or methodological variables that can predict the nature of satisfaction-sociodemographic correlations. Fox and Storms argued that the focus should “shift from obtaining stability of results to understanding the conditions under which discrepant findings can be predicted” [p. 557. They also pointed to the policy implications of learning which settings and practice arrangements best satisfy various sociodemographic groups. The present article presents a meta-analysis of correlations between patients’ satisfaction with medical care and selected patient sociodemographic characteristics. Meta-analysis, or quantitative review of research findings [1 l-141, is well suited for summarizing this literature for several reasons. 811
JUDITH A. HALL and MICHAEL C.
812
First, when prevailing relations are small or inconsistent in direction, the narrative (traditional) reviewer finds it difficult to perceive an overall trend, whereas the quantitative reviewer has more powerful tools for detecting statistically significant patterns [ 151. Second, when many studies are available, the narrative reviewer becomes overwhelmed by the sheer quantity of data. In contrast, the methods of the quantitative review are easily applied to many results. Because some previous reviews of the satisfaction literature have been highly selective, there is a great need to examine as large a proportion as possible of the available studies. can assess statistically meta-analysis Third, whether results from different studies are more variable (heterogeneous) than one would expect from sampling variation alone. The presence of significant heterogeneity suggests, in turn, that moderating variables are at work. Meta-analysis provides statistical tools for detecting the effects of moderating variables. Our meta-analysis examined the overall direction, strength, and statistical significance of correlations between patient satisfaction and eight commonly measured sociodemographic variables. Heterogeneity of results for each of these variables was assessed, and statistical contrasts were performed to test for the operation of several moderating variables. Finally, Fox and Storms noted that poor reliability could account for small correlations in the literature, and speculated that a clearer picture would emerge from looking only at studies based on psychometrically sound instruments [lo]. We approached this issue by relating the estimated reliability of satisfaction instruments to the size of the correlations obtained with them. Actual reliability coefficients were rarely available in the literature. However, owing to the positive relation between internal consistency reliability and scale length [13], we used the number of satisfaction items on a scale as an estimate of its reliability. METHOD Literature search and coding
Previous reports [2, 16) describe the compilation of the meta-analysis bibliography and the details of coding. Briefly, studies were included if: (a) they were stated by their authors to measure patients’ satisfaction with medical care or used measures that were indistinguishable from satisfaction measures used in other studies; (b) satisfaction was measured quantitatively; (c) at least one potential correlate of satisfaction was analyzed; (d) sample size exceeded IO (most were far in excess of this); and (e) they appeared in an English-language book or journal. ‘Medical care’ was defined broadly to include inpatient and ambulatory health care, excluding psychiatric and dental visits [2]. Online retrieval and other methods, followed by screening using the criteria mentioned above, produced 221 independent studies reported in 230 books or articles.* *The complete bibliography, as well as the bibliography for the present analyses, can be obtained from the first author.
DORNAN
All sociodemographic variables analyzed in relation to satisfaction with medical care were extracted from this database. However, some sociodemographic variables were measured very infrequently (e.g. community type, religion, and employment status). In the present analyses we included only those sociodemographic variables that were measured in 10 or more studies. Altogether, 110 studies were included. The sociodemographic variables included were patient’s age (or parents’ age in the case of pediatric samples), ethnicity (white vs black or Hispanic). sex, social status, income, education, marital status (married vs not), and family size. Social status refers to all definitions of social class that were not simply education or income; these included white collar vs not, indices based on occupational prestige along with other indicators, occupational prestige by itself, and ‘sociodemographic status’ left undefined. In this cluster of definitions called social status, occupational prestige was therefore a common theme. For each correlation extracted from a study, a code indicated which of 11 aspects of medical care it pertained to. If a correlation was instead based on global satisfaction ratings or on a composite made by combining separate ratings of different aspects, this too was noted. The 11 aspects were access, cost, overall quality of care, humaneness of providers, competence of providers, information given by providers, bureaucracy, physical facilities, providers’ attention to psychosocial problems, continuity of care, and outcome of care. To describe study attributes, a coding form was developed. The following items from it were tested as possible moderating variables: age of sample (pediatric vs adult), ethnicity of sample (white vs black or Hispanic), and percentage of the sample that was male. Extraction of results from the literature
For each sociodemographic variable in a study, the following information was extracted wherever possible: (1) direction of the relation of the sociodemographic variable to satisfaction (because authors varied in their coding of sociodemographic variables and satisfaction, it was sometimes necessary to change the sign of the reported relation for consistency of interpretation), (2) magnitude of the relation, in terms of the product-moment correlation (r), a useful index of effect size [13] (for marital status, ethnicity, and sex, these were point biserial correlations). and (3) standard normal deviate (2) based on this correlation and its associated sample size (these Zs are typically converted to P-values using standard tables). Sometimes in a study, several correlations were reported for the same sociodemographic variable and the same aspect of satisfaction; for illustration, several satisfaction items in a study might concern the humaneness of the provider and each might be correlated with the patient’s sex. We considered each such result to be an estimate of the true relation for that sociodemographic variable and that aspect of satisfaction, and therefore we randomly selected one of the results for entry in the analysis.
Patients’ characteristics and satisfaction
813
Table I. Adeouacv of data reoortina for eiabt sociodemoaraohic variable-s Sociodemographic variable Number (%) of studies for which
Age
Ethnicity
Sex
Magnitude and direction of correlation available Direction only of correlation available
(258%) ,3:;,
Neither magnitude nor direction of correlation available Total N of studies
Social status
(20&
(4&,
(&
(2& 75
40
Because sociodemographic variables could be correlated with up to 11 different aspects of satisfaction, as well as global or total satisfaction scores, a given study could have a number of different correlations assigned to it, even after the random selection process just described. For some purposes it was desirable to reduce these further to a single aggregate correlation per study to maintain statistical independence of sampling units in subsequent analysis and also to increase statistical power. This was done by averaging all the correlations already entered (one per aspect, total, or composite; see above) for a given study for a given sociodemographic variable. These correlations could number as many as 12, that is, 11 possible aspects plus a total or composite. However, in most studies only a few aspects were measured (the average number was less than 4). In many studies, the authors did not report the results in exactly the form needed for analysis. When necessary we made calculations ourselves using standard formulas [ 131; examples include calculating the standard normal deviate when it was not given, and calculating the correlation coefficient from proportions or when only means and standard deviations were reported. Data analysis
For description of overall trends we employed frequencies and mean correlations (unweighted and weighted by sample size). For testing the heterogeneity of correlations, the chi-square statistic of Rosenthal and Rubin was used [17]. Calculations of combined probabilities for a group of studies testing the same relationship was done following the Stouffer method [18]. To calculate the number of additional null results required to bring a significant combined P into nonsignificance, the ‘file drawer’ formula of Rosenthal [ 131was applied. ‘File drawer’ refers to the imaginary storage place of these hypothetical null results not known to the reviewer. To assess effects of
59
Marital status
Family size
(5&
7 (54%)
Income
Education
13 (54%) s (33%) 3 (12’;)
26 (45%) II (194/o) 21 (36%)
cst, 4 (33%)
(u2%,
24
58
I2
13
32
moderating variables, and Rubin [17] were weights are applied for the heterogeneity file drawer test, and Appendix.
contrasts derived by Rosenthal used in which a priori contrast to the correlations. Formulas test, combined probabilities, contrasts are supplied in the
RESULTS
Adequacy of reporting results
Before describing the nature of relations between sociodemographic variables and satisfaction, we present figures on the completeness with which results were reported in the studies reviewed. Table 1 shows the availability of correlational results for the eight sociodemographic variables (a result was considered available if it was reported directly or could be calculated as described above). The bottom row shows the total number of studies that related each sociodemographic variable to satisfaction, and the rows above break up these totals according to whether the magnitude and direction of the association were available (at least once in a given study), whether the direction but not the magnitude was available (at least once in a given study), and whether neither was ever available in a given study. Table 1 reveals plainly that correlation coefficients were available less than half the time, with a low of 25% for social status and a high of 58% for ethnicity and marital status. Another quarter, approximately, of the studies made the direction of the relation available but not the magnitude, and roughly a third provided neither. Overall trend9
Tables 2 and 3 address the overall nature of the relations between the eight sociodemographic variables and satisfaction. Table 2 is based on correlations as the unit of analysis; thus, correlations for several different aspects of satisfaction could come
Table 2. Descriptive statistics: frequency analysis Sociodemonraohic variable Social status
Income
Education
Marital status
Family size
(27%)
Higher 61% (22136)
Higher 57% (20 35)
Higher* 52% (32162)
Married 86% (12114)
t 50% (6112)
61% (lWl5)
75% W/16)
78%
37% (10/27)
100% (l/l)
60%‘1 (3/S)
Age
Ethnicity
Sex
Modal direction % of Correlations in modal direction:
Older 88% (7l/81)
White 73% (38152)
Male
% of Significant correlations in modal direction5
95% (39/4l)
61% (lWl5)
Index of outcome
*The modal direction is inconsistent with other indicators (see Table 3). tNo modal direction. $Excludes relations of unknown direction. #Excludes relations that are not significant. ~[Sec Table 3 for direction of trend.
(7 9)
814
JUDITH A. HALL and MICHAEL
C.
DORSAN
Table 3. Descriptive statistics: effect sizes and significance levels So&demographic Index of outcome Direction of trend Mean correlation* Mean correlation. weighted* .S of studies for means Combined 2 Combined P (2 tail)
Aee
Ethnicitv
Sex
Social status
Older 0.13 0.13 27 X.68 lo-‘
White -0.02 -0.02 23 1.54 0. I2
Male -0.04 -0.01 19 I.25 0.22
Higher 0.04 0.1 I 8 I .94 0.06
variable
Income
Education
Marital status
Family size
Higher 0.05
Lower -0.04 -0.03 26 2.41 0.02
Married 0.03 0.03 I 1.78 0.08
Smaller -0.02 0.00 7 0.66 0.52
0.01 13 I .23 0.22
*Based on aggregated correlations per study (see text for explanation). Excludes relations of unknown magnitude
from a given study. Table 3, on the other hand, is based on studies as the unit of analysis. using the aggregation method described earlier (i.e. where one average correlation is derived for each study for a given sociodemographic variable). Table 2 shows the modal-that is, most frequently occurring-direction of the correlations for each sociodemographic variable, as well as the percentages of all correlations occurring in the modal direction and the percentages of significant correlations that were in the modal direction. In terms of modal direction, the more satisfied patients tended to be older, white, male, higher in social class (according to social status, income, and education), and married. It is apparent, however. that these trends are not equally strong. For age, ethnicity, and marital status, the percentages of correlations in the modal direction are at least 76%, but for the remaining variables the percentages are much closer to 50% (that is, the level to be expected if there were no overall trend in terms of the signs of the correlations). Indeed, for education only 52% of the correlations go in the modal direction. and as the next row of Table 2 shows, most of the significant correlations for education go in the opposite direction. Table 3 confirms that the net relation of satisfaction to education is negative. Table 3 presents average correlations (effect sizes) both unweighted and weighted by sample size. The weighted correlation is probably more accurate than the unweighted one, since correlations from larger studies have less sampling error and can justifiably be awarded more weight. The weighted and unweighted average correlation between age and satisfaction is 0.13 and is the largest effect in the table. The next largest is for social status, with an average weighted correlation of 0. I I. The remaining correlations are very small in magnitude. The last two rows of Table 3 present combined 2s and their associated P-values, that is, the probability associated with the entire set of effects for a given sociodemographic variable. These combined Pvalues were achieved by calculating the standard normal deviate (2) for each of the correlations that was available after aggregation (one per study for a given sociodemographic variable, as described earlier) and applying the formula provided in the Appendix. N was the total number of studies known to have examined the relation of satisfaction to a given sociodemographic variable, regardless of whether any quantitative information on the relation was provided. It was conservatively decided to estimate the standard normal deviate to be zero whenever we did not have any more exact information.
The combined P-value is significant, in the modal direction, for age; it is significant in the direction opposite to the modal direction for education. The combined P-value also closely approaches significance, in the modal direction, for marital status and social status. Satisfaction is thus significantly (or nearly significantly) associated with being older, having higher social status, being married, and having less education. As described earlier, the ‘file drawer’ formula was applied to ascertain the number of additional null results required to bring these significant combined Ps into nonsignificance. This calculation is valuable for it provides the reviewer with a sense of how easily overturned the combined P is by the addition of previously unknown studies. For age, the number of additional null results required is over 2000; we would conclude that it is extremely unlikely that this many unknown studies with null results exist. For education. marital status, and social status, however, these figures are much smaller; only 67, 2, and 12 additional null results, respectively, would be required to bring the combined P to exactly P = 0.05, one-tail, meaning that just one more null result would push the combined P above this value. It is possible that this many unknown (unretrieved or unpublished) null results do exist, and therefore our confidence in the robustness of our combined Ps for these predictors is reduced proportionately. After completion of the meta-analysis, several more results for sex of patient appeared (these references are available from the authors). All three of these correlations suggested greater satisfaction among women and therefore are not consistent with the overall result shown in Tables 2 and 3. However, when these new results are added to the other 19 studies shown in Table 3, the overall result is essentially unchanged. The weighted mean correlation becomes 0.00 and the unweighted mean correlation becomes -0.02, with a new combined Z of -0.42 (i.e. farther from statistical significance than it was without the new results). In addition, a fourth new study reported that patient sex was ‘not related’ to satisfaction. On balance, therefore, it is still the case that there is no average difference in the satisfaction of men and women. Variability in effect sizes
Another set of tests addressed the heterogeneity of the relations between satisfaction and each sociodemographic variable. The P-values associated with these tests were all highly significant (0.01 or smaller), indicating that the correlations were more variable than one would expect from normal sampling
Patients’ characteristics and satisfaction Table
4. Results
of planned
contrasts
relating
direction
and magnitude
815
of correlations
Sociodemographic
to study
characteristics
variable
Social
Marital
Family
status
Income
Education
status
size
-
NS
0.01
0.001
-
-
(12)
(25) 0.001
-
-
(17) NS
-
0.00
Potential moderating
variable
Age (pediatriunonpediatric) Ethnicity (white/black ?‘a of Male
0.05
(23) NS
(18) -
0.05
(8) 0.0 (6) NS
(‘;6”
NS
(8) NS (IO)
(6)
(8)
(12) NS
patients
trend)
Entries
are P-values
not significant.
(15) for the contrasts. IV of studies
If a contrast
entered
Sex
0.0001
or Hispanic)
(linear
Ethnicity
Age
into
(9) was not calculated
each contrast
appears
Table
NS
due to too limited
data. -
(17) appears
I
(4)
in the table. The
term NS
means
in parentheses.
variation. When this occurs, it is appropriate to ask whether the correlations come from different distributions that can be identified on the basis of known study characteristics. Any such study characteristic would thus be operating as a moderating variable. Accordingly, we selected three sample characteristics to be tested as possible moderating variables and subjected them to the meta-analytic contrast procedure described earlier. These contrasts can be thought of in the same way as contrasts in the analysis of variance; that is, they consist of apriori weights which reflect a prediction about the relation of a grouping variable to the actual correlations obtained. Metaanalytic contrasts are very powerful for they draw on the statistical power of the original studies and not just on the N of available studies [13]. As explained under the Method Section, the variables tested as grouping or moderating variables were age, ethnicity, and percentage of male patients in the sample. Table 4 shows the P-values for each of these contrasts. As the table reveals, about half of the contrasts were significant. Table 5 offers a verbal description of each of these significant effects. It is interesting to note that, due to the high statistical power of these contrasts, the actual differences in the sizes of correlations obtained for a given contrast were often quite small. As an example, the contrast in the size of the satisfaction-ethnicity correlation between adult and pediatric samples, which was significant at the P = 0.05 level, reveals an average correlation of -0.04 for adult samples and an average correlation of -0.01 for pediatric samples. Meta-analytic contrasts are therefore sensitive to small differences between subsets of studies.
Sociodemographic
I
5. Description
A final analysis addressed the association between satisfaction scale reliability and the correlations obtained with different satisfaction instruments. As mentioned earlier, Fox and Storms [lo] hypothesized that ‘noise’ or inconsistency in the correlations could arise from studies that used psychometrically inadequate instruments. We could rarely ascertain the reliability of a satisfaction instrument, but we could almost always discover the number of items it included. Because internal consistency reliability is a joint function of the number of items and the average intercorrelation among them [13], we decided to treat the number of items as an approximation of reliability. For each sociodemographic variable we correlated the number of items on which each correlation was based with the absolute value of that correlation (absolute values were used because greater reliability should affect the magnitude of a correlation but not its sign). These new correlations ranged from -0.27 for education (N = 42) to 0.79 for marital status (N = 7). The mean and median of the correlations obtained for the eight sociodemographic variables were both 0.18. Thus, there is a weak positive relation, on average, between estimated reliability and the size of the correlation coefficients obtained. Unfortunately, such a positive relation does not explain why some correlations are strongly positive and some strongly negative for the same sociodemographic variable; it merely says that the strongest correlations, of either sign, tended to have the larger number of items and presumably the greater reliability. of sianiticant
variable
contrasts
Description The
Age
tendency
for older patients
to be more satisfied
is stronger
for adult
(r = 0.14)
than pediatric
(I = 0.05)
samples Ethnicity
The
tendency
for
whites
to be more
satisfied
is stronger
for adult
( I = -0.04)
than
pediatric
(I = -0.01)
samples Sex
The
tendency
for men to be more satisfied
is stronger
for black/Hispanic
(r = -0.12)
than white
(I = 0.01)
samples Social
status
The
tendency
black/Hispanic The
Income
tendency
(I = 0.00) The
Education
Family
The
size
Weighted
mean correlations
(r)
for higher
class
patients
to
be more
satisfied
is
stronger
for
white
(I = 0.14)
than
samples income
patients
for less educated patients
samples;
tendency
percentage
higher
(I = 0.08)
to be more satisfied
is stronger
for pediatric
(I = 0.05)
than adult
samples
tendency
(I = 0.06)
for
stronger
for patients
of females
are provided
for
for
with
white smaller
to be more satisfied (I = -0.03) family
in it.
all two-group
contrasts.
than
is stronger
black/Hispanic
size to be more satisfied
for adult
(r = -0.04)
(I = 0.04) is stronger
than pediatric
samples
as the sample has a higher
816
JUDITH
A. HALL and MICHAELC. DORNAN
DlSCtiSSlON
The present review helps clarify relations between satisfaction with medical care and the patient’s sociodemographic characteristics. In general, relations were extremely small even when statistically significant. Sociodemographic characteristics are a minor predictor of satisfaction, at best. It is important, nevertheless, to place these small correlations in proper perspective by noting that established correlates of satisfaction such as the patient’s health status [3,4], the physician’s communication behaviors [I], and the physician’s technical competence [I] achieve average magnitudes of quite modest size. Indeed, small effect sizes are the rule not the exception in much health sciences research; for example, well established risk factors for cardiovascular disease account for only about 2% of variation in occurrence of the disease [19]. Finding weak relations between satisfaction and background variables lends support to the validity of the satisfaction instruments, since they were designed to measure events (or perceptions of events) in the process of medical care and not simply be a reflection of response biases in the patient. In spite of the application of meta-analytic methods, many questions remain, among them the nature of the many associations that were examined by primary investigators but never made available in their published reports, as revealed in Table 1. It is likely that these omissions were made because the correlations were very small and did not reach statistical significance. However, it will not be possible to
gain a complete understanding of satisfaction, nor any other important issue in medical care, unless investigators are more thorough in their description of results. In terms of overall trends, it appears that patients’ age and education bear a significant relation to *One could argue this more forcibly if the education and social status correlations were uniformly based on the same samples of subjects. Unfortunately this was not often the case. tFrom reading the three reviews in question [3,5,6], one would conclude that the tally is six showing more satisfaction in women, nine showing ‘no difference’, and none showing more satisfaction in men. However, we chose to include two studies acknowledged by Pascoe [3] to show more male satisfaction but which he dismissed as atypical. In his words, such “discrepant results tend to be reported in studies investigating either a unique type of organizational setting or a narrow range of diagnosed health problems” [p. 1971. Because Pascoe offered no evidence to justify discounting these two ‘discrepant’ results, we chose to include them in the tally. Weiss [5], for his part, made three errors in reporting the studies he cited. One study he cited as showing ‘no difference’ actually showed significantly greater male satisfaction, and two of the studies he cited as showing greater female satisfaction actually showed the opposite (both significantly). One of the latter errors [2l] is understandable. The direction of scaling of the satisfaction instrument in that study was unclear, but was the opposite of what Weiss inferred (Thomas R. Zastowny, personal communication, 2 Dec., 1986). After including Pascoe’s two rejected studies and correcting Weiss’s errors, we arrived at the tally reported in the text.
satisfaction, with patients’ social status and marital status having nearly significant relations. The trends indicate that greater satisfaction is associated with being older, having less education, having higher social status (mainly an index of occupational status), and being married. It is perplexing, to say the least, that the net results for social status and education go in opposite directions; one can see how previous reviewers might have reached the conclusion that the literature was completely inconsistent. At the present time we can offer no confident explanation for these disparate trends. One possibility, of course, is that the apparent differences are not real and would disappear if more studies were located or conducted. The ‘file drawer’ analysis suggested, indeed, that the nearly-significant overall result for social status would be overturned by the addition of only 12 null results. But we cannot rule out the possibility that the variation is meaningful. A point to note is that different indices of sociodemographic status are not perfectly correlated with each other (201. There is therefore no srafisricul necessity for correlations between different sociodemographic indices and satisfaction to have congruent signs. correlations between indices of Imperfect social class mean that groups can be identified who experience status inconsistency-eg. being high in education but low in occupational prestige. Our finding of opposite-direction correlations for education and social status with satisfaction suggests that the lowest level of satisfaction occurs in precisely this group.* One can imagine that these individuals feel the most deprived and resentful toward others, perhaps especially toward privileged groups such as physicians. Such attitudes could translate into dissatisfaction with medical care as well as with other facets of life. Further research could address this question. Our review did not sustain earlier conclusions, based on more limited reviews, that women [3, 5,6] and individuals with smaller families [6] and more education [6] are on average more satisfied. The absence of a male-female difference in satisfaction is especially interesting since the three previous reviewers reached a different conclusion [3, 5,6]. Surprisingly, when we examined the studies these reviewers cited, we found four showing women to be more satisfied, eight showing ‘no difference’, and five showing men to be more satisfied. Thus, it seems that our meta-analysis reached the appropriate conclusion regarding patient sex.7 An additional important finding of the metaanalysis is the demonstration of significant variability in the relations of all eight sociodemographic variables to satisfaction. Although contrasts revealed several moderating factors to be significant, this analysis actually raises more questions than it answers. First, although the correlations were heterogeneous for all sociodemographic variables, we have not accounted well for this variation. Second, it is far from clear how the moderating relations should be interpreted. Why, for example, are men more satisfied than women among black or Hispanic samples but not among white samples? And why is the relation of ethnicity to satisfaction stronger for adult than
Patients’ characteristics and satisfaction pediatric samples? An additional complexity in interpreting these results is that the moderating variable itself may not be causally responsible for its apparent moderating effect, but may be related to another variable which is causally responsible. Thus, issues of substantive theory and causation remain very unclear even after gaining a degree of empirical insight that far exceeds our earlier understanding. For individual moderating effects, it is sometimes possible to come up with post hoc interpretations. For illustration, it may easily be the case that the age-satisfaction relation is stronger for adult than pediatric studies (Table 5) because there is not as much variation in the ages of parents who bring their children for medical care as there is in a general adult sample. At this stage, however, we believe caution should be exercised in interpreting these effects without a more comprehensive theoretical framework. Moreover, the outstanding theoretical question in our view is not why certain moderating effects exist, but why relations exist at all between sociodemographic variables and satisfaction. Two distinct possibilities can be suggested. First, the relations may arise out of response tendencies on the part of patients in different groups [IO]. This pattern of causation implies that satisfaction is independent of the actual care received by patients. For example, older patients may be more satisfied because they become generally mellow and accepting, or because they feel more reluctant than younger patients to pass negative judgment on their care. In a related vein, one can speculate that more educated patients are less satisfied because they have heightened expectations or apply stiffer standards in their evaluations of care (regardless of the nature of that care), and are consequently disappointed compared to less educated patients. A second theoretical possibility is that the relation of satisfaction to a sociodemographic variable is mediated by events that occur in the process of medical care. This could happen if, for example, older patients happened to be treated in a more thorough or responsive manner than younger patients. ‘Ageist’ stereotypes to the contrary, at least some research indicates that physicians have more negative attitudes toward younger patients [22]. And a recent videotape study found physicians to be “less communicatively dominant, more nonverbally responsive as listeners, and more egalitarian in their interactions with middle-aged and older patients relative to their encounters with younger clients” [23, p. 811. Thus, older patients’ greater satisfaction could stem from more positive treatment received, and not from their response tendencies. Similarly, one can suggest that patients higher in social status are more satisfied due to better treatment by physicians. Indeed, it has been repeatedly shown that higher class patients receive better care along a number of dimensions, even within the same health care sites, than lower class patients [24-261. Such speculation makes the negative relation for education especially perplexing. Perhaps more educated patients apply such stiff standards in evaluating their care that even the objectively better care they receive does not seem as subjectively satisfying as the lower quality care rendered to less educated patients.
817
We must not be content with intriguing theorizing; we must confirm the net directions of sociodemographic variables to satisfaction, and we must test the different causal hypotheses. Clearly, one way to do this is to measure directly the nature of the medical care given to patients in different groups, along with their satisfaction, and see if variation in medical care can account for the correlations between sociodemographic variables and satisfaction. Doing this will enable investigators to clarify the role of nonmedical factors in patient satisfaction and develop a better understanding of the concept of satisfaction and the factors that influence it. research was supported by core funds from the Institute for Health Research, a joint program of the Harvard Community Health Plan and the Harvard School of Public Health. The authors would like to thank Howard S. Frazier for his encouragement on this project and Arnold M. Epstein for his helpful comments on the manuscript. Acknowledgements-This
REFERENCES I. Hall J. A., Roter D. L. and Katz N. R. Meta-analysis of correlates of provider behavior in medical encounters. Med. Care 26, 657, 1988. 2. Hall J. A. and Dornan M. C. Meta-analysis of satisfaction with medical care: description of research domain and analysis of overall satisfaction levels. Sot. Sci. Med. 27, 637, 1988. 3. Pascoe G. C. Patient satisfaction in primary health care: a literature review and analysis. Euol. Proa. Plan. 6. 185. 1983. 4. Hall J. A., Feldstein M., Fretwell M. er al. Health status and satisfaction with medical care in an older HMO population. Med. Cure. In press. 5. Weiss G. L. Patient satisfaction with primary medical care: evaluation of sociodemographic and predispositional factors. Med. Care 26, 383, 1988. 6. Ware J. E. Jr, Davies-Avery A. and Stewart A. L. The measurement and meaning of patient satisfaction. Hlth Med. Care Sem. Ra-. 1, 2, 1978.
7. DiMatteo M. R. and DiNicola D. D. Achieving Purienr Compliance: The Psychology of the Medical rioner’s Role. Pergamon, New York, 1982.
8. Ley P. Satisfaction, Br. J. clin. Psychol.
Pracri-
compliance and communication. 21, 241, 1982.
9. Locker D. and Dunt D. Theoretical and methodological issues in sociological studies of consumer satisfaction with medical care. Sot. Sci. Med. 12, 283, 1978. 10. Fox J. G. and Storms D. M. A different approach to sociodemographic predictors of satisfaction with health care. Sot. Sci. Med. lSA, 557, 1981. II. Green B. F. and Hall J. A. Quantitative methods for literature reviews. A. Rev. Psychol. 35, 37, 1984. 12. Louis T. A., Fineberg H. V. and Mosteller F. Findings for public health from meta-analyses. A. Ra. publ. Hlth 6, 1, 1985. Procedures /or Social 13. Rosenthal R. Mera-analyric Research. Sane. Beverly Hills, Calif., 1984. 14. Glass G. V., McGaw B: and Smith M. L. Meta-anulysis in Social Research. Sage, Beverly Hills, Calif., 1981. 15. Cooper H. M. and Rosenthal R. Statistical versus traditional procedures for summarizing research findings. Psychol. Bull. 87, 442, 1980. 16. Hall J. A. and Doman M. C. What patients like about their medical care and how often they are asked: a meta-analysis of the satisfaction literature. Sot. Sci. Med. 27, 935, 1988.
818
JUDITH
A. HALLand MICHAELC.
17. Rosenthal R. and Rubin D. B. Comparing effect sizes of independent studies. Psychol. Bull. 92, 500, 1982. 18. Rosenthal R. Combining results of independent studies. Psychol. Bull. 85, 185, 1978. 19. Friedman H. S. and Booth-Kewley
S. The ‘diseaseprone personality’; a meta-analytic- view of the construct. Am. Psycho/. 42, 539, 1987. 20. Jencks C. Ine&ali~y: i Re&sessment of rhe Effect of Family and Schooling in America. Basic Books, New York, 1972. 21. Zastowny T. R., Roghmann K. J. and Hengst A. Satisfaction with medical care: replications and theoretic reevaluation. Med. Care 21, 294, 1983. 22. Harris I. B., Rich E. C. and Crowson T. W. Attitudes of internal medicine residents and staff physicians toward various patient characteristics. J. med. Educ. 60, 192, 1985. 23. Street R. L. and Buller D. B. Patients’ characteristics
affecting physician-patient nonverbal communication. Hum. ~okkm. Res. 15, 60, 1988. 24. Waitzkin H. Information aivine in medical care. J. Hlfh sot. Behau. 26, 81, 1985. 25. Ross C. E., Mirowsky J. and DutT R. S. Physician status
characteristics and client satisfaction in two types of medical practice. J. Hlrh SOC.Behao. 23, 317, 1982. 26. Wasserman R. C., Inui T. S., Barriatua R. D. et al. Pediatric clinicians’ support for patients makes a difference: an outcome-based analysis of clinician-parent interaction. Pediatrics 74, 1047, 1984.
APPENDIX I. Hetereogeneiry
Test
The ingredients are (I) a set of independent
correlation coefficients (r), each expressed as its associated Fisher z (here called r;), (2) N - 3, where N is the number of sampling units on which each r is based, and (3) K, or the number of correlations in the set. The quantity c (N, - 3)(r, - i:)’
is distributed as x2 with K - I dJ A nonsignificant P-value for this test indicates that the correlations are no more
DORNAN
variable than one would expect given sampling from a common population. II. Combined Probabilities The ingredients are (I) Z for each independent result to be combined. or the standard normal deviate associated with each P-value, with signs attached to reflect the direction of the relationship (e.g. two results of P = 0.05, twotail. that go in opposite directions would both be assigned Zs of I .96 but one would have a minus sign), and (2) K, or the number of P-values to be combined. The quantity
is itself a Z whose associated P-value can be looked up in standard tables. A nonsignificant combined P indicates that the null hypothesis of no overall relation cannot be rejected. III. ‘File Drawer’
Formula
The ingredients are the same as for the combined probabilities. The formula
(CZY
K
--
2.706 yields X, or the number of unretrieved, truly null results required to bring any significant combined P to exactly P = 0.05, one-tail. IV. Conrrasrs
The ingredients are the same as for the heterogeneity test, plus a set of contrast weights (j.,) that are determined from theory (chosen such that their sum will be zero). Contrast weights may compare two groups, test a linear trend, etc. The quantity
-_ Z i.,r,,
Jc “:
N-3 is distributed as a Z whose assoiiated P-value can be looked up in standard tables. A nonsignificant contrast P indicates that the null hypothesis of no relation between the contrast weights and the observed correlations cannot be rejected. Source: Rosenthal, 1984 [ 131.