Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
Lifting the Curtain Backstage Cognition, Frontstage
Behavior, and the Interpersonal Transmission of Culture
Amir Goldberg Stanford Graduate School of Business
Behavioral Insights from Text Conference 2018
Computa(onal Culture Lab
www.comp-culture.org
Jennifer Chatman
Sameer Srivastava
Richard Lu
Culture Ma1ers in Organiza(ons
What makes people fit in culturally?
Why Is Cultural Fit Consequen(al?
• It sustains social order: – It facilitates tacit and complex coordina(on
Weber & Camerer 2003, Crémer Garicano & Prat 2007, Kreps 1990
– It is a complement to formal control Kunda 1986, Van den Steen 2010
• It facilitates membership: – It serves as a signal about iden(ty and commitment
Barth 1969, Eliasoph & Lichterman 2003, Harrison & Carroll 2006
– It affects self percep(on, a1achment and mo(va(on Akerlof & Kranton 2005, Kreps & Baron 2014
Formal organiza(ons are not unique: the pressures of cultural fit pervade all walks of social life.
What is Cultural Fit?
• Cogni&ve cultural fit: shared beliefs, values and schemas
• Behavioral cultural fit: norma(ve compliance
Cogni7on
Behavior
The Front and the Backstage
Following Goffman, we think of behavior and cogni(on as the front and backstage of cultural fit: • People can only observe others’ behaviors, and rely on these
observa(ons to draw inferences about others’ underlying cogni(on
• Most studies implicitly assume that the front and backstage are aligned
Liaing the Curtain We dis(nguish between two dimensions of the backstage: • Perceptual Accuracy:
Correspondence between and individual’s percep(on of the prevalent norms, and peers’ readings of the norma(ve environment
• Value Congruence: Correspondence between an individual’s preferred values, and those prevalent among peers
The Frontstage
Who achieves behavioral cultural fit? • Previous work has almost exclusively focused on value congruence.
• But having certain preferences does not mean having the ability to signal them.
• Perceptual accuracy is essen(al for behavioral alignment.
• People can only learn from others’ behaviors.
Hypotheses
We make two complementary arguments: • Hypothesis 1:
Perceptual accuracy (but not value congruence) is predic(ve of behavioral fit
• Hypothesis 2: Culture is learned from peers: perceptual accuracy and behavioral fit are therefore both suscep(ble to behavioral peer influence
Argument Summary Lifting the Curtain
Figures
Backstage Frontstage
a b c d e a b c d e a b c d ea b c d e a b c d e a b c d ea b c d e a b c d e a b c d e
V P B A
B
C
DValues Perceptions Behaviors
PublicPrivate
a b c d e a b c d e a b c d e
B
a b c d e a b c d e a b c d e
B
B
a b c d e a b c d e a b c d e
a b c d e a b c d e a b c d ea b c d e a b c d e a b c d e
a b c d e a b c d e a b c d ea b c d e a b c d e a b c d e
V P
V P
V P a b c d e a b c d e a b c d e
a b c d e a b c d e a b c d e
a b c d e a b c d e a b c d e
a b c d e a b c d e a b c d e
a b c d e a b c d e a b c d e
a b c d e a b c d e a b c d e
FIGURE 1: A schematic illustration of our theory. Four individuals (A-D) are each characterized by theirvalues (V), perceptions (P) and behavioral probabilities (B). Arrows correspond to causal relationships.
FIGURE 2: Conceptual Overview of the Machine Learning Process
38
H1: within person
H2: interpersonal transmission
Empirical Seeng Midsized U.S. technology firm: • 7 years of email data: metadata and content • One (me administra(on of the Organiza(on Culture Profile
(OCP) O’Reilly, Chatman and Cadwell 1991
• Personnel data • Reorganiza(on producing quasi-‐exogenous shias in the
structure of interac(on
Measurement Strategy We measure three different constructs: On the frontstage: • Behavioral fit:
Email: Interac(onal Language-‐Use Model Goldberg et al. 2016, Srivastava et al. 2017
On the backstage: • Perceptual Accuracy:
OCP: Match between own and peers’ view of prevailing values • Value Congruence:
OCP: Match between own preferences and peers’ view of prevailing values
We use a machine-‐learning random forest model to transform cross-‐sec(onal survey responses into longitudinal data
Interac(onal Language-‐Use Model of Behavioral Fit
We define person i’s behavioral fit at (me T as the similarity between:
We define person i’s behavioral fit at (me T as the similarity between: • Distribu(on of linguis(c categories in i’s outgoing messages
l1 l2 l3 … lk
Outgoing, Oi,T
Interac(onal Language-‐Use Model of Behavioral Fit
We define person i’s behavioral fit at (me T as the similarity between: • Distribu(on of linguis(c categories in i’s outgoing messages • Distribu(on of linguis(c categories in i’s incoming messages
l1 l2 l3 … lk
Outgoing, Oi,T
l1 l2 l3 … lk
Incoming, Ii,T
Interac(onal Language-‐Use Model of Behavioral Fit
We define person i’s behavioral fit at (me T as the similarity between: • Distribu(on of linguis(c categories in i’s outgoing messages • Distribu(on of linguis(c categories in i’s incoming messages
l1 l2 l3 … lk
Outgoing, Oi,T
l1 l2 l3 … lk
Incoming, Ii,T
Jensen-Shannon Divergence
We use the LIWC lexicon as linguis(c units
Interac(onal Language-‐Use Model of Behavioral Fit
Backstage: OCP
Amplified Asking We use a random forest machine learning model applied to email communica(on to transform one-‐shot OCP results into longitudinal variables: • Imputed value congruence • Imputed perceptual accuracy
Assessing Imputed Measures Imputed values exhibit different trajectories: • Value congruence is mostly stable • Perceptual accuracy increases over (me
Preliminary Analyses—Evaluating the Variables of Interest
Before turning to our main results, we summarize three preliminary analyses that soughtto evaluate the validity of the backstage and frontstage cultural fit measures, particularlythe backstage measures that were imputed using the procedure described in Appendix A.First, given that we theorized that value congruence is relatively stable over time, whereasperceptual accuracy is more susceptible to change, we traced the two imputed measures overa person’s tenure in the organization. We restricted this analysis to the first 36 months ofemployment given that only about 10% of employees had tenure exceeding 36 months duringour observation period. We separately estimated OLS and fixed effect regressions of the twobackstage variables using indicators for each month (up to month 36 of employment). Theseresults are depicted in Figure 3. According to both models, when employees first enter theorganization, they have relatively high value congruence and relatively low perceptual ac-curacy. Through approximately the first year of employment, however, perceptual accuracyincreases sharply and continues a more gradual ascent thereafter. In contrast, value con-gruence increases—albeit not as steeply—in the first four months of employment and thenremains mostly stable over the remaining months. These results support our contentionthat value congruence is relatively stable, while perceptual accuracy is more malleable.
3 6 9 12 15 18 21 24 27 30 33 36Tenure
-0.6
-0.4
-0.2
0
0.2
0.4
Within-person
Perceptual AccuracyValue Congruence
3 6 9 12 15 18 21 24 27 30 33 36Tenure
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
OLS
Perceptual AccuracyValue Congruence
Figure 3: OLS and fixed effect regressions of perceptual accuracy and value congruence,with indicators for each tenure month up to 36 months in the company.
Second, in Table ?? we report the results of OLS regressions with individual, departmentand year fixed effects, where the dependent variable is bonus (logged), and frontstage and
20
H1: Perceptual Accuracy Predicts Behavioral Fit
Perceptual accuracy is related to behavioral fit, but value congruence is not
Lifting the Curtain
TABLE 3
C����-S�������� ��� L����������� F���� E������ R���������� �� B��������� F��
Cross-Sectional Longitudinal
Model 1† Model 2† Model 3† Model 4 Model 5 Model 6Perceptual Accuracy‡ 0.122⇤⇤⇤ 0.149⇤⇤⇤ 0.046⇤⇤ 0.046⇤⇤
(3.56) (3.37) (2.81) (2.79)Value Congruence‡ -0.008 -0.040 0.013 0.012
(-0.17) (-0.86) (1.35) (1.29)Manager 0.613⇤⇤⇤ 0.599⇤⇤⇤ 0.555⇤⇤⇤ 0.293⇤⇤⇤ 0.297⇤⇤⇤ 0.292⇤⇤⇤
(6.73) (4.20) (3.92) (5.42) (5.47) (5.40)First Year -0.246⇤⇤ -0.351⇤⇤⇤ -0.317⇤⇤ -0.074⇤ -0.082⇤⇤ -0.074⇤
(-3.20) (-3.49) (-3.13) (-2.54) (-2.81) (-2.53)Female 0.043 -0.033 -0.065
(0.62) (-0.35) (-0.68)Age -0.003 -0.002 0.001
(-0.84) (-0.30) (0.10)Constant 0.345⇤ 0.223 0.183 -0.142 -0.145 -0.145
(2.37) (1.13) (0.93) (-1.14) (-1.11) (-1.17)Individual FE No No No Yes Yes YesDepartment FE Yes Yes Yes Yes Yes YesYear FE No No No Yes Yes YesObservations 386 209 202 24215 24215 24215R2 0.275 0.235 0.279 0.107 0.075 0.107
t statistics in parentheses; standard errors clustered by individual when individual fixed e�ects are used† Behavioral Fit is averaged over 3 months, ‡ Imputed measures in Models 4-6⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001
42
Cross-sectional OCP Longitudinal imputed
w/ Person Fixed Effects
H2: Interpersonal Transmission We assess peer influence effects using an instrumental variable approach: • We exploit a companywide reorganiza(on, which shiaed
peers quasi-‐exogenously • We use a two-‐stage instrumental variable framework with
two instruments: Waldinger 2012 – Induced change in peers’ cultural fit – Change in number of peers
• First stage: es(mate the change, induced by the reorganiza(on, in peers’ cultural fit and number of peers
• Second stage: es(mate focal actor’s cultural fit post-‐reorganiza(on with person and department fixed effects
Lifting the Curtain
TABLE 4
OLS ��� I����������� V�������� F���� E������ R���������� �� B��������� F��
OLS Instrumental Variable
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6Behav. Behav. Percep. Value Percep. Value
Fit Fit Accuracy Congr. Accuracy Congr.Peer Behavioral 0.221⇤⇤⇤ 0.266⇤⇤⇤ 0.068⇤⇤ -0.020Fit† (12.68) (6.38) (3.03) (-0.47)
Peer Perceptual 0.064Accuracy† (0.63)
Peer Value 0.073Congruence† (0.83)
Num. Peers† 0.001⇤⇤ -0.013⇤ 0.001 0.008⇤ 0.024 -0.004(3.11) (-2.50) (0.27) (2.14) (1.36) (-0.38)
Manager 0.365⇤⇤⇤ 0.555⇤⇤⇤ 0.042 -0.096 -0.430 0.136(7.67) (4.34) (0.77) (-0.95) (-1.18) (0.68)
First Year -0.154⇤⇤⇤ -0.204⇤⇤⇤ -0.163⇤⇤⇤ 0.028 -0.013 -0.043(-6.72) (-4.12) (-6.28) (0.65) (-0.12) (-0.64)
Constant -0.065 0.648⇤⇤ 0.259⇤⇤ -0.257 -0.756 0.257(-1.23) (2.67) (2.67) (-1.45) (-0.99) (0.63)
Individual FE Yes Yes Yes Yes Yes YesDepartment FE Yes Yes Yes Yes Yes YesYear FE Yes Yes Yes Yes Yes YesN 22080 21998 21998 21998 21985 21985Num. Individuals 1515 1508 1508 1508 1504 1504R2 0.28Kleibergen-Paap F 8.99 8.99 8.99 0.85 1.79
t statistics in parentheses; standard errors clustered by individual† lagged variables, instrumented endogenous variables in Models 2-6⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001
43
H2: Interpersonal Transmission
• Peers’ behavior affects actor’s perceptual accuracy and behavioral fit, but not value congruence
• Actors only observe peers’ behaviors
H2: Interpersonal Transmission
Lifting the Curtain
3 6 9 12 15 18 21 24 27 30 33 36Tenure
-0.6
-0.4
-0.2
0
0.2
0.4
Within-person
Perceptual AccuracyValue Congruence
3 6 9 12 15 18 21 24 27 30 33 36Tenure
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
OLS
Perceptual AccuracyValue Congruence
FIGURE 3: OLS and fixed e�ect regressions of perceptual accuracy and value congruence, with indicatorsfor each tenure month up to 36 months in the company.
-4 -3 -2 -1 0 1 2 3 4 5 6 Month
-0.2
0
0.2
0.4
Beha
vior
al F
it
Reorganization
+0.5 Change in Peer Behavioral Fit
-0.5 Change in Peer Behavioral Fit
FIGURE 4: Marginal e�ects, estimated by monthly consecutive instrumental variable models, of change inpeer behavioral fit on individual behavioral fit. The two lines correspond to individuals who experienced a 0.5increase (blue) or decrease (red) in peer behavioral fit. Shaded areas correspond to 95% confidence intervals.
39
Peer effects are substan(al
Conclusion
Why do some people vary in their cultural fit? • Perceptual accuracy is more consequen(al for the ability to read
the cultural code and behave compliantly than is value congruence
• Peers ma1er: culture is learned from those with whom one interacts
Using text to infer culture: • Interac(onal Language-‐Use Model: to infer behavioral cultural
fit (using LIWC) • Amplified asking: Machine learning to impute backstage cultural
fit over (me