對于行為方法研究bias的修正方法

1、Common Method Biases in Behavioral Research:A Critical Review of theLiterature and Recommended RemediesPhilip M. Podsakoff, Scott B. MacKenzie, andJeong-Yeon LeeIndiana UniversityNathan P. PodsakoffUniversity of FloridaInterest in the problem of method biases has a long history in the behavioral sci

2、ences. Despite this, a comprehensive summary of the potential sources of method biases and how to control for them does not exist. Therefore, the purpose of this article is to examine the extent to which method biases influence behavioral research results, identify potential sources of method biases

3、, discuss the cognitive processes through which method biases influence responses to measures, evaluate the many different procedural and statistical techniques that can be used to control method biases, and provide recommendations for how to select appropriate procedural and statistical remedies fo

4、r different types of research settings.Most researchers agree that common method variance (i.e.,variance that is attributable to the measurement method rather than to the constructs the measures represent is a potential problem in behavioral research. In fact, discussions of the potential impact of

5、common method biases date back well over 40years (cf.Campbell &Fiske, 1959, and interest in this issue appears to have continued relatively unabated to the present day (cf.Bagozzi &Yi, 1990; Bagozzi, Yi, &Phillips, 1991; Campbell &OConnell,1982; Conway, 1998; Cote &Buckley, 1987,

6、 1988; Kline, Sulsky, &Rever-Moriyama, 2000; Lindell &Brandt, 2000; Lindell &Whit-ney, 2001; Millsap, 1990; Parker, 1999; Schmitt, Nason, Whitney, &Pulakos, 1995; Scullen, 1999; Williams &Anderson, 1994; Williams &Brown, 1994.Method biases are a problem because they are one o

7、f the main sources of measurement error. Measurement error threatens the validity of the conclusions about the relationships between mea-sures and is widely recognized to have both a random and a systematic component (cf.Bagozzi &Yi, 1991; Nunnally, 1978; Spector, 1987. Although both types of me

8、asurement error are problematic, systematic measurement error is a particularly serious problem because it provides an alternative explanation for the observed relationships between measures of different constructs that is independent of the one hypothesized. Bagozzi and Yi (1991noted that one of th

9、e main sources of systematic measurement error is method variance that may arise from a variety of sources:Method variance refers to variance that is attributable to the measure-ment method rather than to the construct of interest. The term method refers to the form of measurement at different level

10、s of abstraction,such as the content of specific items, scale type, response format, and the general context (Fiske,1982, pp. 8184.At a more abstract level, method effects might be interpreted in terms of response biases such as halo effects, social desirability, acquiescence, leniency effects, or y

11、ea-and nay-saying. (p.426However, regardless of its source, systematic error variance can have a serious confounding influence on empirical results, yielding potentially misleading conclusions (Campbell&Fiske, 1959. For example, letsassume that a researcher is interested in studying a hypothesiz

12、ed relationship between Constructs A and B. Based on theoretical considerations, one would expect that the measures of Construct A would be correlated with measures of Construct B. However, if the measures of Construct A and the measures of Construct B also share common methods, those methods may ex

13、ert a systematic effect on the observed correlation between the mea-sures. Thus, at least partially, common method biases pose a rival explanation for the correlation observed between the measures. Within the above context, the purpose of this research is to (aexamine the extent to which method bias

14、es influence behavioral research results, (bidentify potential sources of method biases, (cdiscuss the cognitive processes through which method biases in-fluence responses to measures, (devaluate the many different procedural and statistical techniques that can be used to control method biases, and

15、(eprovide recommendations for how to select appropriate procedural and statistical remedies for different types of research settings. This is important because, to our knowledge, there is no comprehensive discussion of all of these issues avail-able in the literature, and the evidence suggests that

16、many re-searchers are not effectively controlling for this source of bias.Extent of the Bias Caused by Common Method VarianceOver the past few decades, a considerable amount of evidence has accumulated regarding the extent to which method variance influences (ameasures used in the field and (brelati

17、onships between these measures. Much of the evidence of the extent to which method variance is present in measures used in behavioral research comes from meta-analyses of multitraitmultimethodPhilip M. Podsakoff and Jeong-Yeon Lee, Department of Management, Indiana University; Scott B. MacKenzie, De

18、partment of Marketing, Indi-ana University; Nathan P. Podsakoff, Department of Management, Uni-versity of Florida.Journal of Applied Psychology Copyright 2003by the American Psychological Association, Inc.879studies (cf.Bagozzi &Yi, 1990; Cote &Buckley, 1987, 1988; Williams, Cote, &Buckl

19、ey, 1989. Perhaps the most comprehen-sive evidence comes from Cote and Buckley (1987,who exam-ined the amount of common method variance present in measures across 70MTMM studies in the psychology sociology, marketing, business, and education literatures. They found that approximately one quarter (26

20、.3%of the variance in a typical research measure might be due to systematic sources of measurement error like common method biases. However, they also found that the amount of variance attributable to method biases varied considerably by discipline and by the type of construct being investigated. Fo

21、r example, Cote and Buckley (1987found that, on average, method variance was lowest in the field of marketing (15.8%and highest in the field of education (30.5%.They also found that typical job performance measures contained an average of 22.5%method variance, whereas attitude measures contain an av

22、erage of 40.7%.A similar pattern of findings emerges from Williams et al. s (1989study of just the applied psychology literature.In addition to these estimates of the extent to which method variance is present in typical measures, there is also a growing body of research examining the extent to whic

23、h method variance influences relationships between measures (cf.Fuller, Patterson, Hester, &Stringer, 1996; Gerstner &Day, 1997; Lowe, Kroeck, &Sivasubramaniam, 1996; Podsakoff, MacKenzie, Paine, &Bach-rach, 2000; Wagner &Gooding, 1987. These studies contrasted the strength of th

24、e relationship between two variables when com-mon method variance was controlled versus when it was not. They found that, on average, the amount of variance accounted for when common method variance was present was approximately 35%versus approximately 11%when it was not present. Thus, there is a co

25、nsiderable amount of evidence that common method variance can have a substantial effect on observed relationships between measures of different constructs. However, it is important to rec-ognize that the findings suggest that the magnitude of the biasproduced by these method factors varies across re

26、search contexts (cf.Cote &Buckley, 1987; Crampton &Wagner, 1994.Not only can the strength of the bias vary but so can the direction of its effect. Method variance can either inflate or deflate observed relationships between constructs, thus leading to both Type I and Type II errors. This poi

27、nt is illustrated in Table 1, which uses Cote and Buckley s (1987estimates of the average amount of trait variance, the average amount of method variance, and the average method intercorrelations and inserts them into the equation below to calculate the impact of common method vari-ance on the obser

28、ved correlation between measures of different types of constructs (e.g.,attitude, personality, aptitude:R x , y (trueR ti , tj x y (trueR mk , ml x y ,(1where true R ti, tj is the average correlation between trait i and trait j , t x is the percent of trait variance in measure x , t y is the percent

29、 of trait variance in measure y , true R mk, ml is the average correlation between method k and method l , m x is the percent of method variance in measure x , and m y is the percent of method variance in measure y .For example, the correlation .52in the second row of the first column of Table 1was

30、calculated by multiplying the true corre-lation (1.00times the square root of Cote and Buckley s (1987estimate of the percent of trait variance typically found in attitude measures (times the square root of their estimate of the percent of trait variance typically found in personality measures (.391

31、 plus the average of their estimates of the typical correlation between methods for attitude (.556and personality (.546con-structs multiplied by the square root of their estimate of the percent of method variance typically found in attitude measures (times the square root of their estimate of the pe

32、rcent of method variance typically found in personality measures (.247.Table 1Relationship Between True and Observed Correlation for Average Measures by Type of ConstructTrue R ti, tj correlation (R ti, tj 2Type of Constructs.52(.27.35(.12.28(.08.21(.04.18(.03Attitude job performance and satisfactio

33、n .51(.26.32(.10.25(.06.17(.03.13(.02Personality personality .53(.28.33(.11.25(.06.17(.03.13(.02Personality aptitude.53(.28.34(.12.26(.07.18(.03.14(.02Personality job performance and satisfaction .53(.28.32(.10.23(.05.15(.02.10(.01Aptitude aptitude.54(.29.34(.12.26(.07.18(.03.14(.02Aptitude job perf

34、ormance and satisfaction.54(.29.32(.10.24(.06.15(.02.11(.01Job performance and satisfaction job performance and satisfaction.54(.29.31(.09.21(.04.12(.01.07(.00Note. Values within the table are the observed correlations R x, y (andsquared correlations R x, y 2 calculated using Cote and Buckley s (198

35、8formula shown in Equation 1of the text. For the calculations it is assumed that (athe trait variance is the same as that reported by Cote and Buckley (1987for each type of construct (e.g.,attitude measures .298, personality measures .391, aptitude measures .395, and job performance and satisfaction

36、 measures .465, (bthe method variance is the same as that reported by Cote and Buckley (1987for each type of construct (e.g.,attitude measures .407, personality measures .247, aptitude measures .251, and job performance and satisfaction measures .225, and (cthe correlation between the methods is the

37、 average of the method correlations reported by Cote and Buckley (1987for each of the constructs (e.g.,method correlations between attitude attitude constructs .556, personality attitude constructs .551, personality personality constructs .546, aptitude attitude constructs .564, aptitude personality

38、 constructs .559, aptitude aptitude constructs .572, job performance and satisfaction attitude constructs .442, job performance and satisfaction personality constructs .437, job performance and satisfaction aptitude constructs .450, and job performance and satisfaction job performance and satisfacti

39、on constructs .328. These calculations ignore potential Trait Method interactions.880PODSAKOFF, M AC KENZIE, LEE, AND PODSAKOFFThere are several important conclusions that can be drawn from Table 1. For example, the entry in the first column of the first row indicates that even though two attitude c

40、onstructs are perfectly correlated, the observed correlation between their measures is only .52because of measurement error. Similarly, the entry in the last column of the first row indicates that even though two attitude constructs are completely uncorrelated, the observed correlation between their

41、 measures is .23because of random and systematic measurement error. Both of these numbers are troubling but for different reasons. The entries in the entire first column are trou-bling because they show that even though two traits are perfectly correlated, typical levels of measurement error cut the

42、 observed correlation between their measures in half and the variance ex-plained by 70%.The last column of entries is troubling because it shows that even when two constructs are completely uncorrelated, measurement error causes the observed correlation between their measures to be greater than zero

43、. Indeed, some of these numbers are not very different from the effect sizes reported in the behav-ioral literature. In view of this, it is disturbing that most studies ignore measurement error entirely and that even many of the ones that do try to take random measurement error into account ignore s

44、ystematic measurement error. Thus, measurement error can in-flate or deflate the observed correlation between the measures, depending on the correlation between the methods. Indeed, as noted by Cote and Buckley (1988,method effects inflate the observed relationship when the correlation between the m

45、ethods is higher than the observed correlation between the measures with method effects removed and deflate the relationship when the correlation between the methods is lower than the observed cor-relation between the measures with method effects removed.Potential Sources of Common Method BiasesBeca

46、use common method biases can have potentially serious effects on research findings, it is important to understand their sources and when they are especially likely to be a problem. Therefore, in the next sections of the article, we identify several of the most likely causes of method bias and the re

47、search settings in which they are likely to pose particular problems. As shown in Table 2, some sources of common method biases result from the fact that the predictor and criterion variables are obtained from the same source or rater, whereas others are produced by the measure-ment items themselves

48、, the context of the items within the mea-surement instrument, and/orthe context in which the measures are obtained.Method Effects Produced by a Common Source or RaterSome methods effects result from the fact that the respondent providing the measure of the predictor and criterion variable is the sa

49、me person. This type of self-report bias may be said to result from any artifactual covariance between the predictor and criterion variable produced by the fact that the respondent providing the measure of these variables is the same.Consistency motif. There is a substantial amount of theory (cf.Hei

50、der, 1958; Osgood &Tannenbaum, 1955 and research (cf.McGuire, 1966 suggesting that people try to maintain consistency between their cognitions and attitudes. Thus, it should not be surprising that people responding to questions posed by research-ers would have a desire to appear consistent and r

51、ational in theirresponses and might search for similarities in the questions asked of them thereby producing relationships that would not other-wise exist at the same level in real-life settings. This tendency of respondents to try to maintain consistency in their responses to similar questions or t

52、o organize information in consistent ways is called the consistency motif (Johns,1994; Podsakoff &Organ, 1986; Schmitt, 1994 or the consistency effect (Salancik&Pfeffer, 1977 and is likely to be particularly problematic in those situa-tions in which respondents are asked to provide retrospec

53、tive accounts of their attitudes, perceptions, and/orbehaviors.Implicit theories and illusory correlations. Related to the no-tion of the consistency motif as a potential source of common method variance are illusory correlations (cf.Berman &Kenny, 1976; Chapman &Chapman, 1967, 1969; Smither

54、, Collins, &Buda, 1989, and implicit theories (cf.Lord, Binning, Rush, &Thomas, 1978; Phillips &Lord, 1986; Staw, 1975. Berman and Kenny (1976have indicated that illusory correlations result from the fact that “raters often appear to possess assumptions concern-ing the co-occurrence of r

55、ated items, and these assumptions may introduce systematic distortions when correlations are derived from the ratings ”(p.264; Smither et al. (1989have noted that these “illusory correlations may serve as the basis of job schema or implicit theories held by raters and thereby affect attention to and

56、 encoding of ratee behaviors as well as later recall ”(p.599. This suggests that correlations derived from ratees responses are com-posed of not only true relationships but also artifactual covariation based on ratees implicit theories.Indeed, there is a substantial amount of evidence that implicit

57、theories do have an effect on respondents ratings in a variety of different domains, including ratings of leader behavior (e.g.,Eden &Leviatin, 1975; Lord et al., 1978; Phillips &Lord, 1986, attributions of the causes of group performance (cf.Guzzo, Wag-ner, Maguire, Herr, &Hawley, 1986;

58、 Staw, 1975, and perceptions about the relationship between employee satisfaction and perfor-mance (Smitheret al., 1989. Taken together, these findings indi-cate that the relationships researchers observe between predictor and criterion variables on a questionnaire may not only reflect the actual co

59、variation that exists between these events but may also be the result of the implicit theories that respondents have regarding the relationship between these events.Social desirability. According to Crowne and Marlowe (1964,social desirability “refers to the need for social approval and acceptance and the belief that it can be attained by means of culturally acceptable and appropriate behaviors ”(p.109. It is generally viewed as the tendency on the part of individuals to present themselves in a favorable light, regardless of their true feelings about an issue or to