Sie sind auf Seite 1von 20

A Statistical Analysis of Solicitation

on Sproul Plaza
Murray, Vo, Castillejo-Munoz, Kaplun, Figus
May 8, 2014
Abstract
Our group set out to empirically study a form of solicitation commonly
referred to as yering on Sproul Plaza. We observed over 180 interactions
between volunteers and walkers, and performed analysis using both logistic
and linear regression. Of the eight independent variables we observed, only
two (whether or not the volunteer smiled, and the organization doing the y-
ering) proved to be statistically signicant in predicting the success rate for an
organization yering on Sproul Plaza.
1
Contents
1 Problem and Method of Data Collection 3
2 Raw Data: A Bernoulli Model 4
2.1 Model Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Paired T-Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3 Averaged Data: A Scaled Binomial Model 9
3.1 Model Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.4 Analysis of Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 Conclusion 19
4.1 Signicance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Repeatability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2
Preface
This report contains two distinct types of analysis. Readers will likely nd the
section assuming a Bernoulli model to be easier to follow than the section assuming
a scaled binomial model.
For the Bernoulli model, the T-tests are much more useful than the logistic
regression (in terms of statistical signicance and ease of understanding).
For the scaled binomial model, single-variate regression produces some results
that are consistent with the Bernoulli model, while other results conict with the
Bernoulli model.
Whenever there is a conict between the two models, the reader should consider
the Bernoulli model to be more reliable.
1 Problem and Method of Data Collection
In order to better understand yering on Sproul Plaza, we identied eight measur-
able factors that we believed would inuence whether a volunteer could successfully
hand o a yer to a walker.
1
If there were multiple walkers in a group, we only
recorded data on the walker that the volunteer reached most directly towards. We
recorded these factors for each attempted hando made by a volunteer (see the
Repeatability section of the Conclusion for more information on data gathering).
x
1
= number of walkers in group approached by volunteer
x
2
=

1 if the walker was wearing headphones


0 otherwise
x
3
=

1 if the walker was wearing sunglasses


0 otherwise
x
4
=

1 if the volunteer smiled


0 otherwise
x
5
=

1 if the volunteer appeared female


0 otherwise
x
6
= Number of volunteers yering for organization under observation
1
A walker is dened as any individual walking down Sproul Plaza.
3
x
7
=

1 if we categorized the organization as aperforming-arts organization


2 if ... service
3 if ... social
4 if ... identity-cultural
5 if ... professional
x
8
= Total number of organizations yering on Sproul
We dene success as an attempted hando that resulted in the walker (the
volunteers target) receiving the yer and not discarding it immediately thereafter.
Mathematically...
Y =

1 if the walker accepted and did not immediately discard the yer
0 otherwise
Y
i
Bern (p
i
) where we suspect p
i
= f (x) (1)
We dene Sproul Plaza as what is formally known as Upper Sproul Plaza.
We dene the length of Sproul Plaza to be the region from the sidewalk running
parallel to Bancroft to the far side of the bridge that crosses Strawberry Creek (just
beyond Sather Gate). All data was gathered during periods of high trac on Sproul
Plaza (on weekdays, between 11:30am and 3pm). Each group member observed 30
or more attempted handos.
2 Raw Data: A Bernoulli Model
2.1 Model Overview
As described in the previous section, our response variable follows a Bernoulli dis-
tribution in that each interaction was deemed either a success (Y = 1) or a failure
(Y = 0). As a result, scatter plots are not introduced for this model at any point
in this report. Similarly, ANOVA cannot be performed with this data. We provide
paired T-tests for an analysis qualitatively similar to ANOVA. Descriptive statistics
include various bar charts, and average success rates conditional upon the state of
input variables.
Readers may wonder why we include organization type in logistic regression
but not in any other point in our analysis in this section. We omit organization
type from our analysis so frequently because several organizations could not be
easily categorized (e.g. Improv for Charity). Because any categorization of these
organizations would have been nearly arbitrary, we concluded that analysis on this
variable would not make for justied conclusions.
4
2.2 Descriptive Statistics
Because our response variable was Bernoulli, with a global mean less than 0.5, we
can conclude that the minimum, median, and mode of the data set are all 0. The
global mean and variance is given at the bottom of Table 1.
Figure 3 shows that very few interactions were observed in which the walker was
wearing headphones. This may have two explanations: (1) a small proportion of stu-
dents walking down Sproul Plaza wore headphones, and (2) volunteers intentionally
avoided walkers wearing headphones.
Figure 1 demonstrates that we observed a nearly equal number of male and
female volunteers. Moreover, the dierence in success rate is very dicult to assess
visually.
Figure 2 shows the indicates that we observed many more smiling interactions
than non-smiling interactions. This has may imply that most volunteers on Sproul
Plaza do in fact smile. However, this can also be because we observed more active
volunteers than passive volunteers; if smiling volunteers tended to be more active,
then our data would be showing an articially high prevalence of smiling among
volunteers on Sproul Plaza. If this explanation is true, then organizations yering
on Sproul Plaza may still be able to gain a competitve edge by smiling.
Finally, Figure 4 shows that most volunteers approach only single walkers. Al-
though the success rate for single walkers is somewhat higher than other group sizes,
it is a possibility that a single interaction with a larger group could be of greater
benet to the organization than a single interaction with a smaller group. Although
the bar for 4 walkers seems to be entirely black, this is the result of a black outline.
In fact, we only observed a single interaction between a volunteer and a group of
four, and the attempt was unsuccessful.
Table 1: Descriptive Statistics for Raw Data
Given Success Rate Variance Count
x
2
= 0 0.4125 0.2423 160
x
2
= 1 0.3103 0.2140 29
x
3
= 0 0.4079 0.2415 152
x
3
= 1 0.3484 0.2270 37
x
4
= 0 0.2273 0.1756 44
x
4
= 1 0.4483 0.2473 145
x
5
= 0 0.3763 0.2347 93
x
5
= 1 0.4167 0.2431 96
Nothing 0.3968 0.2394 189
5
Figure 1: Gender Bar Graph
0
20
40
60
80
100
120
X5 = 0 (Male) X5 = 1 (Female)
!"##$%&"' )*%"##$%%&"'
Figure 2: Smiling Bar Graph
0
20
40
60
80
100
120
140
160
X4 = 0 X4 = 1
Succesful Unsuccessful
6
Figure 3: Headphones Bar Graph
0
20
40
60
80
100
120
140
160
180
X2 = 0 X2 = 1
Succesful Unsuccessful
Figure 4: Number of Walkers Bar Graph
0
20
40
60
80
100
120
140
160
X1 = 1 X1 = 2 X1 = 3 X1 = 4
Succesful Unsuccessful
7
2.3 Logistic Regression
Because Bernoulli random variables only take on values of zero and one, matching
a curve to such data is unlikely to meaningfully describe the phenomenon under
consideration. Additionally, linear regression assumes the existence of independent
and normally distributed error terms (e
i
N(0,
2
)). Logistic regression is the
most common analysis performed on data that follows a Bernoulli distribution. Ex-
plaining logistic regression is beyond the scope of this report, but many explanations
and tutorials can be found online. We proceed to our analysis.
Table 2: Multivariate Logistic Regression (all variables)
Variable -2 Log Likelihood Chi-Square df Sig.
Intercept 123.2 0 0
x
1
123.852 0.695 2 0.707
x
2
123.404 0.247 1 0.619
x
3
123.596 0.438 1 0.508
x
4
128.514 5.3569 1 0.021*
x
5
124.713 1.556 1 0.212
x
6
124.713 1.555 3 0.670
x
7
127.852 4.694 4 0.320
x
8
125.604 2.447 2 0.294
With the likelihood ratio tests, we can see that only x
4
(whether or not the
volunteer smiles) is signicant. Therefore only x
4
may be used to explain variation
observed in the output. Table 3 shows the results of binary logistic regression.
Table 3: Single-Variate Logistic Regression (smiling only)
Variable B S.E. Wald df Sig. Exp(B)
x
4
1.016 0.397 6.564 1 0.010 2.762
Constant -1.224 0.360 11.573 1 0.001 0.294
Table 4: Fit Assessments for Binary Logistic Regression
-2 Log likelihood Cox & Snell R Square Nagelkerke R Square
246.623 0.038 .051
Unfortunately, although we determined at least one signicant variable, our R
2
values were incredibly low. Table 4 shows that our model is not accurate enough to
usefully predict the results of any particular interaction. The following section on
T-tests will conclude our analysis using a Bernoulli model.
8
2.4 Paired T-Tests
A common way to make inferences on binary data is to run a test on proportions
for dierent groups or categories. In each case, we make pairwise comparisons
between two proportions, with a null hypothesis that the true proportions are
equal between the two categories or groups being compared. It is important to note
that we cannot make conjectures that rely on the results of multiple tests (this is
because the possible error for each test would compound and leave the evidence for
the conjecture unreliable).
That established, Table 5 shows that only x
4
(smiling) has a statistically signif-
icant eect on the probability of success.
Table 5: Tests on Proportions for Binary Variables
Category Headphones Sunglasses
Variables x

2
=0 x

2
=1 x

3
=0 x

3
=1
Proportion 0.4125 0.3103 0.4079 0.3514
Count 160 29 152 38
Test Statistic 1.0346 0.6372
Sig. 0.3009 0.5238
Category Smiling Gender
Variables x

4
=0 x

4
=1 x

5
=0 x

5
=1
Proportion 0.2273 0.4483 0.3763 0.4167
Count 44 145 93 96
Test Statistic -2.6246 -0.5665
Sig. 0.0087* 0.5711
3 Averaged Data: A Scaled Binomial Model
3.1 Model Overview
By averaging all variables for each organization we observed, we were left with ten
data points. Averaging Bernoulli outcomes eectively creates new random variables
with binomial distributions (scaled down by the number of trials). These random
variables measure the success rate for each organization. Mathematically...

P
i
=

n
i
jS
i
Y
j
n
i
where S
i
= {k | x
7,k
= i}

P
i
N

p
i
,
p
i
(1 p
i
)
n
i

where

P
i


P
j
i = j (2)
9
Equation (2) is an approximation that is increasingly accurate as n
i
increases
and p
i
approaches 1/2. Although our n
i
s are small (less than 15 in three cases),
our global success rate is very near 1/2. This leads us to believe that

P
i
s may in
fact be normal (subsequent analysis supports this belief). Moreover, the fact that
the variance is so small may justify the following assumption.
2

P
i
p
i
i (3)
After a brief section on descriptive statistics, we will consider regression models
of the following form.

P
i
=

i
+ e
i
where e
i
N(0,
2
) (4)
3.2 Descriptive Statistics
In Table 6, the given mean is the arithmetic average of the

P
i
over all organizations
reported in Table 7.
3
The variance reported in Table 6 uses sample variance of the
following form.
S
2
=
n

i=1
(

P
i

P)
2
/(n 1) (5)
Table 6: Descriptive Statistics for Averaged Data
Minimum Maximum Median Mean Variance
0.000 0.800 0.455 0.455 0.060
2
We will continue to refer to the the averaged success rate as

Pi, but readers should be aware
of this fact.
3
all variables reported as primes are averaged over every data point that applied to that
organization (in a way very similar to the denition of

Pi.)
10
Table 7: Averaged Data
Organization

P
i
x

1
x

2
x

3
Asian American Assn. 0.615 1.231 0.077 0.308
Blood drive 0.286 1.238 0.143 0.214
Business Fraternity 0.286 1.286 0.286 0.143
Fraternity activity 0.000 1.333 0.333 0.000
Fraternity recruiting 0.800 1.400 0.100 0.000
Improv for Charity 0.350 1.400 0.133 0.217
ISAB 0.455 1.091 0.318 0.182
Kappa Gamma Delta 0.800 1.200 0.000 0.000
Ticket sales 0.500 1.500 0.100 0.300
Theatre Rice 0.455 1.364 0.182 0.091
Organization x

4
x

5
x

6
x
8
Asian American Assn. 0.846 0.231 4.000 9.000
Blood drive 0.286 0.952 0.286 3.357
Business Fraternity 0.429 0.286 4.000 9.000
Fraternity activity 0.333 0.000 2.000 9.000
Fraternity recruiting 0.700 0.500 2.000 11.000
Improv for Charity 0.583 0.500 3.900 7.350
ISAB 1.000 1.000 2.000 2.000
Kappa Gamma Delta 1.000 0.800 1.000 9.000
Ticket sales 0.900 1.000 2.000 11.000
Theatre Rice 1.000 0.727 1.364 9.000
11
3.3 Linear Regression
Multivariate linear regression results are given in Table 8 and Table 9. Although
the results indicate that no variable explained a statistically signicant amount of
variation in success rate
4 5
, single-variate linear regression suggests otherwise.
Scatter plots including trend lines for each x

i
are given in Figures 5 through
11. We can see that x

2
, x

4
, and x

5
each show promise in explaining variation in
success rate for the ten organizations we monitored. With these results in mind, we
performed multivariate linear regression on combinations of these variables.
Table 8: Fit Assessments for Multivariate Linear Regression (all variables)
R R Square Adjusted R Square Std. Error of the Estimate
0.944 0.890 0.507 0.17214
Table 9: ANOVA for Multivariate Linear Regression (all variables)
Variation Source Sum of Squares df Mean Square F Sig.
Regression 0.482 7 0.069 2.323 0.334
Residual 0.059 2 0.030
Total 0.541 9
Detailed results of each regression are included in Table 10. Table 10 shows
that for multivariate linear regression on x

2
, x

4
, and x

5
, there is only statistical
signicance for x

2
. This is especially surprising given that T-tests performed on raw
data (Table 5) indicate signicance for smiling (here, x

4
), and non-signicance for
headphones and gender (here, x

2
and x

5
). This apparent contradiction implies x

2
and x

4
interact with one another.
6
In order to test for interaction, we included another variable (x

2
x

4
) in our
model. However, multivariate linear regression on this model indicates that this
new variable is not statistically signicant (Table 11). Because this new variable is
not statistically signicant, the nature of the interaction cannot be determined.
As per the remarks in the Preface, we rely on the results of the Bernoulli model
in the event that the models contradict one another. Although multivariate linear
regression suggests that x

2
is more signicant than x

4
, we know from Table 5 that
smiling plays a more signicant role in aecting yering success than whether or not
4
Note that the null hypothesis for ANOVA for linear regression loosely states that the response
of the model does not depend on any input variable. Failure to reject this null hypothesis would
seem to imply that none of our variables were useful in predicting success rate.
5
Also note that this occurs in spite of a high R
2
value. This can occur because of the very high
ratio of data points to input variables, and is called overtting.
6
Note that this does not imply that x2 and x4 interact with one another. The analysis in this
section concerns x

i
, which is an aggregated value and should not be considered a replacement for
analysis on any xi.
12
Figure 5: Scatter Plot for

P versus x

1
! # $%&''() * %&(%+,
!" $ %&%%'()
%
%&'
%&-
%&.
%&/
%&+
%&(
%&0
%&,
%&1
' '&' '&- '&. '&/ '&+ '&(
Figure 6: Scatter Plot for

P versus x

2
! # $%&'(%)* + ,&(-.%
!" $ %&'(')
,
,&/
,&.
,&'
,&0
%
, ,&,) ,&% ,&%) ,&/ ,&/) ,&- ,&-)
Figure 7: Scatter Plot for

P versus x

3
! # $%&'(()* + %&,--.
!" $ %&%%'()
%
%&.
%&,
%&)
%&/
'
% %&%( %&' %&'( %&. %&.( %&0 %&0(
13
Figure 8: Scatter Plot for

P versus x

4
! # $%&''() * $%$+&'
!" $ %&'''(
$
$%,
$%-
$%'
$%.
+
$%, $%/ $%- $%& $%' $%0 $%. $%( + +%+
Figure 9: Scatter Plot for

P versus x

5
! # $%&''() * $%(+,-
!" $ %&'()*'
$
$%(
$%-
$%'
$%.
/
$ $%( $%- $%' $%. / /%(
Figure 10: Scatter Plot for

P versus x

6
! # $%&%'()* + %&',()
!" $ %&%'()*
%
%&,
%&-
%&'
%&.
/
% %&) / /&) , ,&) 0 0&) - -&)
14
Table 10: Multivariate Linear Regression of

P on x

2
, x

4
, and x

5
ANOVA for Linear Regression of

P on x

2
, x

4
, and x

5
Variation Type SS df Mean Square F Sig.
Regression 0.371 3 0.124 4.375 0.059
Residual 0.17 6 0.028
Total 0.541 9
ANOVA for Linear Regression of

P on x

4
, and x

5
Regression 0.193 2 0.096 1.941 0.214
Residual 0.348 7 0.05
Total 0.541 9
ANOVA for Linear Regression of

P on x

2
, and x

4
Regression 0.342 2 0.171 6 0.03*
Residual 0.199 7 0.028
Total 0.541 9
ANOVA for Linear Regression of

P on x

2
, and x

5
Regression 0.371 2 0.186 7.647 0.017*
Residual 0.17 7 0.024
Total 0.541 9
Coecients of Linear Regression of

P on x

2
, x

4
, and x

5
Variable B Std. Error Test Statistic t Sig.
Constant 0.555 0.276 2.009 0.091
x

2
-1.466 0.584 -0.665 -2.51 0.046*
x

4
0.024 0.35 0.025 0.069 0.947
x

5
0.236 0.231 0.329 1.023 0.346
Coecients of Linear Regression of

P on x

4
, and x

5
Constant 0.045 0.248 0.183 0.86
x

4
0.425 0.414 0.433 1.028 0.338
x

5
0.15 0.303 0.208 0.495 0.636
Coecients of Linear Regression of

P on x

2
, and x

4
Constant 0.478 0.267 1.792 0.116
x

2
-1.377 0.579 -0.625 -2.377 0.049*
x

4
0.268 0.258 0.273 1.037 0.334
Coecients of Linear Regression of

P on x

2
, and x

5
Constant 0.571 0.139 4.117 0.004
x

2
0.247 0.157 0.344 1.573 0.16
x

5
-1.484 0.481 -0.674 -3.084 0.018*
15
Figure 11: Scatter Plot for

P versus x

8
! # $%$&'() * $%&+,,
!" $ %&%'()*
$
$%&
$%(
$%+
$%,
-
$ & ( + , -$ -&
the walker wears headphones. Therefore, as we wrap up this section, we consider
single-variate regression on x

4
in more detail.
7
The results of single-variate linear regression (see Table 12) indicate R
2
values
signicantly higher than those seen in logistic regression. A quadratic model exhibits
the highest R
2
, but we believe this is due to over tting. The next best model is
logarithmic on x

4
.
8
Tables 13 and 14 show the signicance of x

4
beyond < 0.05. Table 13 contains
the 95 percent condence interval for the coecient on ln(x

4
). Although far from
complete, an accurate model of yering on Sproul Plaza is likely to expand upon
the following:

P
i
= 0.581 + 0.919 ln

4,i

(6)
7
We move on to single-variate linear regression because no instances of multivariate linear re-
gression which we consider reliable demonstrate statistically signicant ANOVA.
8
Because

P contains non-positive values (the minimum is 0.00), exponential and power models
cannot be applied to this data.
16
Table 11: Multivariate Linear Regression Results with Interaction
Variation Type SS df MS F Sig.
Regression 0.36 3 0.12 3.984 0.071
Residual 0.181 6 0.03
Total 0.541 9
Variable B Std. Error Test Statistic t Sig.
Constant 0.901 0.606 1.488 0.187
x

2
-2.979 2.131 -1.352 -1.398 0.212
x

4
-0.226 0.684 -0.231 -0.331 0.752
(x

2
x

4
) 1.936 2.477 0.676 0.782 0.464
Table 12: ANOVA of Nonlinear Regressions on x

4
Model Type R Square F df1 df2 Sig.
Linear 0.333 4.001 1 8 0.080
Logarithmic 0.407 5.498 1 8 0.047*
Quadratic 0.564 4.522 2 7 0.055
Table 13: Linear Regression of

P on ln(x

4
)
Term Coecients Standard Error t Stat
Intercept 0.581 0.0831 6.989
x Variable 0.919 0.392 2.345
Term P-value Lower 0.95 Upper 0.95
Intercept 0.0001 0.389 0.772
x Variable 0.0471* 0.0152 1.823
Table 14: ANOVA for Linear Regression on ln(x

4
)
Variation Type df SS MS F Signicance
Regression 1 0.220 0.220 5.498 0.0471*
Residual 8 0.321 0.0401
Total 9 0.541
17
3.4 Analysis of Residuals
Figure 12 shows that heteroskedasticity may be violated, but with so few data points
it is dicult to tell. Although several data points had n
i
< 15, Figure 13 shows that
our approximation of the scaled binomial response as a normal random variable was
highly accurate.
Figure 12: Standardized Residual Plot for

P on ln(x

4
)
-2
-1.3
-1
-0.3
0
0.3
1
1.3
2
2.3
-0.6 -0.3 -0.4 -0.3 -0.2 -0.1 0
S
L
a
n
d
a
r
d
l
z
e
d

8
e
s
l
d
u
a
l

Figure 13: Q-Q Plot for

P on ln(x

4
)
y = 0.9678x - 6*10^-16
-2.3
-2
-1.3
-1
-0.3
0
0.3
1
1.3
2
2.3
-2.3 -2 -1.3 -1 -0.3 0 0.3 1 1.3 2 2.3
A
c
L
u
a
l

S
L
a
n
d
a
r
d
l
z
e
d

8
e
s
l
d
u
a
l

LxpecLed SLandardlzed 8esldual
LxpecLed Lrend llne for normal reslduals
AcLual Lrend llne for reslduals
18
4 Conclusion
4.1 Signicance
Both of these models agree that the following do not signicantly aect yering
success.
x
1
and x

1
: Number of walkers in group approached by the volunteer.
x
3
and x

3
: Indicator and rate of walkers wearing sunglasses.
x
6
and x

6
: Number of volunteers for organization under observation.
x
8
and x

8
: Total number of organizations yering on Sproul.
x
7
and x

7
: Organization type.
Signicance unsupported by the Bernoulli model.
9
Signicance not assessed by the scaled binomial model.
10
The models are in disagreement over whether the following signicantly aect
yering success.
x
2
and x

2
: Indicator and rate of walkers wearing headphones
Signicance supported by the scaled binomial model.
Signicance unsupported by the Bernoulli model.
x
5
and x

5
: Indicator and proportion of female volunteers.
Signicance supported by the scaled binomial model.
Signicance unsupported by the Bernoulli model.
The models agree that the following signicantly aects yering success.
x
4
and x

4
: Volunteer smiling indicator and rate.
Signicance strongly supported by the Bernoulli model.
Signicance supported by the scaled binomial model.
11
4.2 Repeatability
Although every eort was made to collect accurate and reliable data, real time
observation of interactions on Sproul Plaza proved very dicult. Some interactions
occurred too quickly to gather information on each variable under consideration,
while other interactions were obscured by other walkers (recall that we gathered
9
This conclusion is based only on the results from logistic regression
10
As per remarks in Section 2.1.
11
When considered independently of other variables, especially x

2
.
19
data during a period of high trac on Sproul Plaza). In general, these interactions
were removed from our dataset.
Some variables were recorded only at the beginning of a round of data collection.
This is the case for all instances of x
8
, and some instances of x
6
.
The only variable that could not be reliably captured for each interaction was
whether or not the volunteer smiled. We addressed this issue with the following
policy: Record the actual value of x
4
whenever the volunteers face is clearly visible.
In the event that the volunteer turns around and their face is not visible, default to
the value of x
4
in the immediately preceding interaction.
Although this uncertainty lies in our most important variable, we do not believe
that this fact introduces much inaccuracy into our model. In the overwhelming
majority of cases, volunteers tended strongly to consistently smile or consistently
not smile for each interaction they initiated.
In addition to this observer bias, our data was subject to sample bias in the fol-
lowing form: most observations were of active volunteers. Active volunteers provide
a large number of useful interactions for statistical analysis, but these volunteers
may also have a higher success rate than passive volunteers.
Finally, our data may exhibit time bias since it was collected during ASUC elec-
tions.
12
ASUC (the undergraduate student government) elections are characterized
by high activity on Sproul Plaza, but do not necessarily signify campus activity. We
cannot conclude in what way ASUC elections bias our data, but we believe that
yering during the beginning of the academic year is likely to be met with more
success.
12
That is, the nal weeks of the semester.
20

Das könnte Ihnen auch gefallen