"The Apple iPod iTunes Anti-Trust Litigation"

Filing 754

Administrative Motion to File Under Seal Opposition to Plaintiffs' Daubert Motion 737 filed by Apple Inc.. (Attachments: # 1 Declaration of Kiernan ISO Admin Motion to Seal, # 2 Exhibit 1 of Kiernan ISO Admin Motion to Seal, # 3 Exhibit 2 of Kiernan ISO Admin Motion to Seal, # 4 Proposed Order Granting Motion to Seal, # 5 Apple's Opp to Pls' Daubert Motion (Redacted), # 6 Apple's Opp to Pls' Daubert Motion, # 7 Declaration of Kiernan ISO Apple's Opp to Pls' Daubert Motion, # 8 Exhibit 1-4 (Redacted), # 9 Exhibit 5-12 (Redacted), # 10 Exhibit 1-2, 6, 9-11, # 11 Proposed Order Denying Plfs' Daubert Motion)(Kiernan, David) (Filed on 1/14/2014)

1
2
3
4
5
6
7
8
9

Robert A. Mittelstaedt (State Bar No. 60359)
[e-mail]
Craig E. Stewart (State Bar No. 129530)
[e-mail]
David C. Kiernan (State Bar No. 129530)
[e-mail]
Amir Q. Amiri (State Bar No. 271224)
[e-mail]
JONES DAY
555 California Street, 26th Floor
San Francisco, CA 94104
Telephone:
+1.415.626.3939
Facsimile:
+1.415.875.5700
Attorneys for Defendant
APPLE INC.

10

UNITED STATES DISTRICT COURT

11

NORTHERN DISTRICT OF CALIFORNIA

12

OAKLAND DIVISION

13
14
15
16
17
18
19

THE APPLE IPOD ITUNES ANTITRUST
LITIGATION

Case No. 4:05-cv-00037 YGR
APPLE’S OPPOSITION TO
PLAINTIFFS’ DAUBERT MOTION
TO EXCLUDE CERTAIN OPINION
TESTIMONY OF KEVIN M.
MURPHY AND ROBERT H. TOPEL
Date:
Time:
Courtroom:

February 18, 2014
2:00 PM
5

20
21
22
23
24
25
26
27
28
Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

TABLE OF CONTENTS

2

Page

3

INTRODUCTION ............................................................................................................................ 1 

4

BACKGROUND .............................................................................................................................. 1 

5

STATISTICAL SIGNIFICANCE..................................................................................................... 3 

6

A. 

Regression Models and Statistical Significance. ............................................................ 3 

7

B. 

Independence Assumption In Calculating Standard Errors. ........................................... 5 

C. 

How to Test for and Correct Correlation of Residuals. .................................................. 7 

D. 

Application of These Principles Here ............................................................................. 8 

8
9
10
11
12

ARGUMENT .................................................................................................................................... 8 
I. 

APPLYING THESE GENERALLY ACCEPTED PRINCIPLES, MURPHY AND
TOPEL SHOW THAT THE RESIDUALS ARE HIGHLY CORRELATED AND
THAT NOLL’S RESULTS ARE STATISTICALLY INSIGNIFICANT WHEN
THE CORRELATION IS CORRECTED. .............................................................................. 8 

II. 

PLAINTIFFS’ EFFORT TO REHABILITATE NOLL BY ATTACKING APPLE’S
EXPERTS IS WITHOUT MERIT AND PROVIDE GROUNDS TO EXCLUDE
NOLL AND WOOLDRIDGE. .............................................................................................. 11 

13
14
15
16
17
18

A. 

The Claim That Clustering Is Only A Problem When Sampling Is Objectively
Wrong And Should Be Excluded Under Daubert ........................................................ 12 

B. 

Plaintiffs’ Claim That Clustering Overestimates The Standard Errors Is
Baseless ......................................................................................................................... 15 

C. 

Clustering With Or Without Time Period Demonstrates That Noll’s Results
Are Not Statistically Significant ................................................................................... 18 

19
20
21
22

III.  WOOLDRIDGE’S REPORT SHOULD BE STRICKEN. .................................................... 18 

23

IV.  PLAINTIFFS’ ASSERTION THAT MURPHY’S OPINIONS HAVE BEEN
REJECTED ELSEWHERE IS UNFOUNDED. .................................................................... 19 

24

CONCLUSION ............................................................................................................................... 20 

25
26
27
28
-i-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

TABLE OF AUTHORITIES

2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

Page(s)
Cases 
Cabrera v. Cordis Corp.,
134 F.3d 1418 (9th Cir. 1998).................................................................................................... 15
Daubert v. Merrell Dow Pharm., Inc.,
43 F.3d 1311 (9th Cir. 1995)...................................................................................................... 15
In In re High-Tech Employee Antitrust Litig.,
289 F.R.D. 555 (N.D. Cal. 2013) ............................................................................................... 19
Jeffries v. Centre Life Ins. Co.,
No. 1:02-cv-351, 2004 WL 5506494 (S.D. Ohio Jan. 28, 2004). .............................................. 19
Johnson v. Manitowoc Boom Trucks, Inc.,
484 F.3d 426 (6th Cir. 2007)................................................................................................ 12, 18
Lust v. Merrell Dow Pharmaceuticals, Inc.,
89 F.3d 594 (9th Cir. 1996)........................................................................................................ 15
Moore v. Napolitano,
926 F. Supp. 2d 8 (D.D.C. 2013) ............................................................................................... 19
Reed v. Smith & Nephew, Inc.,
527 F. Supp. 2d 1336 (W.D. Okla. 2007) .................................................................................. 19
United States v. Apple, Inc.,
Nos. 12 Civ. 2826 (DLC), 12 CIV 3394 (DLC),
2013 U.S. Dist. LEXIS 96424 (S.D.N.Y. Oct. 4, 2013) ............................................................ 19
Wagner v. County of Maricopa,
673 F.3d 977 (9th Cir. 2012)................................................................................................ 11, 18
Rules 
Federal Rule of Evidence 702 ........................................................................................................ 14

23
24
25
26
27
28
- ii -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

INTRODUCTION

2

Apple has moved for summary judgment, relying in part on reports by two economists,

3

and moved to exclude plaintiffs’ economist. Plaintiffs moved at the same time to exclude one

4

aspect of the reports of Apple’s economists—the portion dealing with “clustering.” In this

5

opposition brief, Apple shows why plaintiffs’ motion to exclude should be denied. At bottom,

6

plaintiffs are trying to create the false impression of a battle of experts over statistical

7

significance. But as plaintiffs’ new expert all but admits, Apple’s experts are correct, and no

8

credible basis exists for the contrary arguments by plaintiffs’ experts.

9

BACKGROUND

10

As shown by Apple’s motion for summary judgment and motion to exclude plaintiffs’

11

expert, plaintiffs offer Professor Roger Noll as the sole purported basis for an incoherent,

12

unsubstantiated theory of antitrust harm and damages. Plaintiffs assert that an Apple software

13

update in 2006, which prevented digital music from one insignificant source of music

14

(RealNetworks music store or RMS) from being played directly on iPods, somehow resulted in

15

RMS customers buying so much more music from Apple that demand for iPods increased so

16

much that Apple was able to, and did, inflate iPod prices.

17

The real-world facts refute this theory. There is no evidence that anyone switched from

18

RMS to Apple’s music store as a result of the update or, that if anyone did, the incremental

19

amount of music they bought from Apple led them to buy an iPod rather than a competing player.

20

The named plaintiffs do not fit that description. They have identified no one who does. As Noll

21

admitted, RealNetworks was a minor player, with minimal sales at the time of Apple’s software

22

update.

23

Lacking any real-world evidence that the 2006 update raised iPod prices, plaintiffs rely on

24

two regression models offered by Noll (one for Apple’s sales to resellers like

25

for Apple’s direct sales to consumers at Apple retail and online stores). See Noll Rebuttal

26

Declaration (ECF 740-14) at Exs. 3-A, 3-B. But, as Apple’s motion shows, the regressions are

27

replete with fundamental errors and inconsistencies that render them unreliable and unable to

28

support any finding of impact. The regressions lack statistical significance. Mot. Summ. J. and
-1-

the other

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

to Exclude Noll (ECF No. 738) at 20-21. They predict counterfactually that Apple would have

2

made incremental price changes of a kind Apple has never made. Id. at 11-12. They fail to

3

separate out the purported effect of a previous update that this Court determined was lawful. Id.

4

at 14-16. And they do not account for other factors affecting iPod prices including factors

5

considered by Apple’s Price Committee when setting iPod prices. Id. at 18-19.

6

To support its showing of these and other fundamental defects in Noll’s models, Apple

7

submitted reports from two highly regarded antitrust economists, Dr. Kevin Murphy and Dr.

8

Robert Topel, who have extensive experience estimating impact and damages in antitrust cases.1

9

Murphy and Topel showed that Noll vastly exaggerated the statistical precision of his models in

10

that, as explained more fully below, he calculated “standard errors” as if the “residuals” or “error

11

terms” were independent when in fact they are highly correlated within groups or “clusters” of

12

observations. When the “standard errors” are properly calculated using standard econometric

13

methods known as “clustering”—a technique recommended by the ABA treatise, Proving

14

Antitrust Damages, by plaintiffs’ new expert’s textbook, and the authorities Noll cites—Noll’s

15

regression results are not statistically significant. This is the expected result given the lack of any

16

evidence that the 2006 update had any impact on demand for iPods or that Apple considered it in

17

determining iPod prices.

18

In response, plaintiffs have moved to exclude this one aspect of the Murphy and Topel

19

opinions, contending that their opinion does not “fit” the facts of this case. Plaintiffs’ motion to

20

exclude is based largely on the declaration of a belatedly disclosed expert, Dr. Jeffrey M.

21

Wooldridge. As shown below, his declaration should be stricken under Daubert as unreliable.

22

Additionally, it should be stricken because plaintiffs did not disclose him by the April 1, 2013

23

deadline for merits experts or the November 25, 2013 deadline for rebuttal experts. This case has

24

1

25
26
27
28

Murphy holds a PhD in economics from the University of Chicago, where he has taught
since 1983. Topel holds a PhD in economics from the University of California, Los Angeles, and
also teaches at the University of Chicago. Both have lectured, written and testified extensively on
the use of econometric methods, including regression models in antitrust cases. Further, both are
experts on the topic of clustering standard errors. See Declaration of David C. Kiernan in Support
of Apple’s Opposition Brief (“Kiernan Decl.”) at Ex. 1 (Murphy Dep., taken January 8, 2014),
256:15-23 & Ex. 2 (Topel Dep., taken January 8, 2014), 239:10-14.
-2-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

been pending for over eight years. Whether Noll should cluster the standard errors has been at

2

issue in the case for over two years. No justification exists for plaintiffs now trying to salvage

3

Noll’s testimony and avoid summary judgment by presenting another expert at the eleventh hour.

4

In any event, Woolridge’s declaration provides no basis to exclude Apple’s experts—and

5

instead only confirms that Apple’s pending for motion for summary and to exclude Noll’s

6

testimony should be granted. Wooldridge’s theories are contrary to generally accepted

7

econometric theory and practice including the literature he cites, have not been peer reviewed, are

8

untested, and were manufactured for this litigation. Woolridge’s declaration directly contradicts

9

his textbook and research, leading to his deposition admission that if his new opinion on

10

“clustering” were correct, he would need to revise the textbook that he has been using for years.

11

Reflecting his own doubts about his new opinion, after submitting his report Wooldridge started

12

work on theoretical calculations and simulations to test his newly formed opinions. But he has

13

not produced them and says he is not relying on them.

14

Plaintiffs hope to create the false impression of a battle of the experts by taking advantage

15

of the highly-technical nature of accounting for clustering in calculating regression standard

16

errors and hence statistical significance. Given the complete lack of real-world evidence to

17

support their claim of antitrust impact, and the numerous other defects in Noll’s opinions, that

18

effort would fail even if plaintiffs’ current motion had any merit. The absence of merit in the

19

motion is further reason to finally bring this long-running case to a close.

20

STATISTICAL SIGNIFICANCE

21

This section explains the technical concepts relevant to this motion.

22

A.

23

Generally speaking, a regression model is a statistical method used to try to estimate or

Regression Models and Statistical Significance.

24

predict whether some conduct caused an impact on a “dependent” variable like price, student

25

performance, compensation, etc. Expert Report of Topel (ECF No. 740-10), ¶¶ 43-52 (“Topel

26

Report”).2 Because it is a statistical tool, the reliability of the estimates in a regression is

27
28

2

A regression measures the average relationship between the “dependent” variable and the
(continued)
-3-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

measured by statistical significance: “the degree of confidence they have that the estimated value

2

of the coefficient did not arise by chance when the true effect of the variable in question is

3

actually zero.” Topel Report ¶¶ 62-63.3 Here, it is the degree of confidence that the impact of

4

estimated by Noll’s regression did not arise merely by chance when, in fact, the

5

true effect was zero (i.e., no damages). Only where a regression’s results are statistically

6

significant are the results deemed reliable.

7

Statistical significance for a coefficient estimate is calculated using the standard error

8

(standard deviation) of the coefficient estimate, which reflects the precision with which the

9

coefficient is estimated. The smaller the standard error, the more precise the estimate. Topel

10

Report ¶¶ 49, 62; Kiernan Decl., Ex. 3 (ABA Section of Antitrust Law, Proving Antitrust

11

Damages: Legal and Economic Issues, (2010)) at 144-147. Specifically, statistical significance is

12

measured by taking the ratio of a coefficient estimate to its standard error (called the “t-ratio” or

13

“t-statistic”), which measures the distance in standard deviations (errors) between the estimated

14

value of the coefficient and zero. Topel Report ¶¶ 62-63. The larger the t-statistic, the further

15

away the estimate is from zero, which increases the confidence that effect of the variable being

16

measured is not zero.4

17
18
19
20
21
22
23
24
25
26
27
28

conduct at issue represented by an “indicator” or “dummy” variable, while controlling for other
“explanatory” or “independent” variables that could also affect the dependent variable. Id. Noll
claims that his dummy variable (“iTunes_7.0_rev”) measures the impact on the dependent
variable (iPod prices) of the challenged update
controlling for
other explanatory variables.
3

Statistical significance is often represented at various thresholds (e.g., 1%, 5%, or (most
leniently) at 10% levels). The smaller the significance level used, the greater the confidence in
the results. For example, at the 1% level, there is no more than 1% probability of getting the
result merely by chance. Topel Report ¶¶ 62-63.
4

A common benchmark for statistical significance is a t-statistic that is 2.0 or larger (t ≥
2.0), which means the estimated coefficient is at least twice its standard error (two standard
deviations). Id. A t-statistic of 2.0 corresponds to approximately a 5% level of statistical
significance—i.e., there is a 5% probability that a coefficient estimate as large as the one obtained
could have arisen by chance if the true value of that coefficient is zero. Id. A t-statistic of t =
2.58 corresponds to approximately a 1% level of statistical significance—i.e., a 1% probability of
arising by chance; and a t-statistic of t = 4.9 corresponds to a .0001% probability of arising by
chance—i.e., a one-in-one million chance. Topel Report ¶ 62-63.
-4-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

B.

2

The calculation of the standard error is based on the “residuals” or “error terms”

Independence Assumption In Calculating Standard Errors.

3

calculated when estimating the price for each observation in the regression. Kiernan Decl., Ex. 3

4

at p. 144-145. The residual represents all the unmeasured factors or unobserved factors that affect

5

the price of the observation that are left out of the regression either because a variable was not

6

included for that factor or because the factor is “unobserved” (i.e., not recorded in the data—e.g.,

7

unobserved economic factors or market conditions). Topel Report ¶ 46; Kiernan Decl., Ex. 3 at

8

pp. 144-145. As a general rule, the larger the number of independent observations and thus

9

residuals used to estimate the standard errors the smaller will be the estimated standard errors.

10

A fundamental assumption in calculating the standard errors and thus the statistical

11

significance of the coefficient estimates is that the residuals are statistically independent

12

(uncorrelated)—knowledge of the residual for one transaction provides no information about the

13

residual for another transaction. Noll Rebuttal pp. 38, 40; see also Expert Report of Murphy

14

(ECF No. 740-8), ¶ 98 (“Murphy Report”); Topel Report ¶ 75; Kiernan Decl., Ex. 3 at 144-45 &

15

Ex. 4 (Angrist, Mostly Harmless Econometrics: An Empiricist’s Companion (2009)) at pp. 293-

16

294 & Ex. 7 (Cameron, A Practitioner’s Guide to Cluster-Robust Inference (2013)) at p. 7

17

(“Intuitively, if errors are positively correlated within cluster then an additional observation in the

18

cluster no longer provides a completely independent piece of new information.”).5 6 Residuals

19

are considered independent when they are not correlated to each other. Id.

20

Residuals can be correlated for various reasons. The typical reason is that, within the

21

population of observations (here, iPod transactions), certain groups of observations are affected

22

by common factors that are not captured by the variables in the model because they were left out

23
24
25
26
27
28

5

As Noll puts it, “The standard assumption about the error term [residual] is that it is
independent and identically distributed with a mean of zero and finite variance, which means that
the value of [the residual] does not depend on any of the other variables in the equation and that
the variance of the error term [the residual] is the same for each observation.” Noll Rebuttal p.
40; see also Kiernan Decl., Ex. 3 at p. 144.
6

Noll relies on both the Angrist book and Cameron article in his rebuttal report as do
Murphy and Topel. See Noll Rebuttal fns. 14, 18, 21 (Angrist) and 11, 14-15 (Cameron)
-5-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

or because they are unobserved. Murphy Report ¶ 79; Topel Report ¶¶ 48, 85; Noll Rebuttal at p.

2

41; Expert Declaration of Jeffrey M. Wooldridge, filed as Ex. 1 to Pls.’ Daubert Motion, (ECF

3

737) at p. 3 (“Wooldridge Report”).7 Wooldridge and Noll illustrate when correlation arises with

4

an example taken from Wooldridge’s graduate textbook, Econometric Analysis of Cross Section

5

and Panel Data (2010).8 The example involves estimating the effect of class size on test scores

6

of students within a state.9 As Wooldridge explains, because “student outcomes within a school

7

are likely to be influenced by common factors determined at the school level, such as

8

(unmeasured) teacher or principal quality,” there will be “cluster correlation” at the school level.

9

Wooldridge Report at p. 3. Failure to cluster at the school level will “over-estimate the

10

magnitude and statistical significance of the effect of class size;” without accounting for the

11

correlation, the “regression under-estimated the standard error of the regression coefficient as

12

well as over-estimated the value of the coefficient.” Noll Rebuttal at p. 42.

13

If the residuals in the regression are correlated within a group of observations (i.e., not

14

independent), that correlation must be accounted for in calculating standard errors. Topel Report

15

¶ 74-76; Supplemental Report of Murphy & Topel (ECF No. 740-23) at ¶ 5 (“Murphy/Topel

16

Supp.”); Kiernan Decl., Ex. 3 at 145-146. If no correction is made, the standard errors will be

17

miscalculated and generally lead to overstating the level of statistical significance; i.e., the

18

standard errors will be too small. Topel Report ¶¶ 74-76; Murphy/Topel Supp. ¶ 5; Kiernan

19

Decl.., Ex. 7 at p. 4 & Ex. 3 at pp. 145-46. As Cameron (relied on by Noll) explains, “Failure to

20

control for within-cluster error correlation can lead to very misleadingly small standard errors,

21

and consequent misleadingly narrow confidence intervals, large t-statistics and low p-values. It is

22
23

7

24

8

25
26
27
28

See also Kiernan Decl., Ex. 3 at 145-146 (collecting examples).

Noll relies on an outdated (2002) version of Wooldridge’s textbook. As Wooldridge
points out in the Preface to the 2010 edition, the first edition relied on by Noll “was hardly
perfect” due in part to “gaps in coverage” and “some important developments in econometrics.”
Kiernan Decl., Ex. 5 at p. xvii. He revised the chapter Noll relies on “the most,” including adding
new material on clustering reflecting current research ignored by Noll.

9

Noll Rebuttal at p. 42; Wooldridge Report at p. 3; Kiernan Decl., Ex. 5 (Wooldridge,
Econometric Analysis of Cross Section and Panel Data (2010)) at p. 864.
-6-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

not unusual to have applications where standard errors that control for within-cluster correlation

2

are several times larger than default standard errors that ignore such correlation.” Kiernan Decl.,

3

p. 4.

4

C.

5

Standard procedures exist to test whether the errors within groups are correlated. Id., Ex.

6

6 (Noll Dep. Tr., taken Dec. 18, 2013) at 24:9-14; see also Ex. 3 at 147, n. 73. One test is to see

7

“whether the mean residual errors … are statistically significantly different from zero, which

8

would have to be the case if the errors within a cluster are correlated.” Noll Rebuttal at p. 34. If

9

there is reason to suspect that the residuals are correlated, one should test the assumption.

How to Test for and Correct Correlation of Residuals.

10

Kiernan Decl., Ex. 7 at p. 20; see also Ex. 3 at 147 & n. 73. As discussed below and in Apple’s

11

motion to exclude, although there is reason to suspect that the residuals are correlated, Noll

12

refused to perform any tests.

13

Standard methods also exist to correct for correlation of residuals. Id., Ex. 6 at 24:9-14;

14

see also Ex. 3 at p. 146; Murphy Report ¶ 98; Topel Report ¶ 81; Wooldridge Report at p. 3. The

15

method used by Murphy and Topel is called “clustering” standard errors. It has “become widely

16

used” in antitrust cases and is “easily implemented” using standard statistical software.10 Kiernan

17

Decl., Ex. 3 at p. 146 & Ex. 7 at pp. 4-5; see also Murphy/Topel Supp. at ¶ 9; Murphy Report ¶

18

98; Topel Report ¶ 81; Kiernan Decl., Ex. 1 at 266:7-267:1; Kiernan Decl., Ex. 8 (Solon, Haider,

19

Wooldridge, What are we Weighting For? (2013)) at pp. 9-10. Wooldridge recognizes this. In

20

his school test scores example, he identifies “a couple of appropriate responses” to correct for

21

correlation at the school level, including “comput[ing] standard errors that allow for correlation in

22

the errors within a school”—i.e., clustering. Wooldridge Report at p. 3; see also Kiernan Decl.,

23

Ex. 5 at p. 864. A recognized benefit of clustering is that, in general, it performs well even when

24

there is no correlation among residuals. Topel Report at n.76; Kiernan Decl., Ex. 1 at 277:6-19;

25

Ex. 3 at 147.

26
27
28

10

For example, “Stata, a popular econometrics software package, includes a ‘cluster’ option
for calculating standard errors assuming unspecified within-group (cluster) correlation between
the error terms.” Kiernan Decl., Ex. 3 at p. 146, n. 71.
-7-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

D.

2

Noll claims that nearly all his coefficients are statistically significant at the 1% level,

Application of These Principles Here

3

including his variable for the

4

significant in fact that they are not believable. Exhibits 3-A and 3-B of Noll’s rebuttal report

5

(ECF No. 740-14), sets forth the coefficient estimate for each variable and its purported standard

6

error and level of statistical significance.

They are actually much more significant than that—so

7
8
9
10
11
12
13
14

Noll’s standard errors are unbelievably small because he estimates them based on 2.1

15

million observations in the reseller regression and 36.9 million observations in the direct sales

16

regression, assuming that he has 2.1 million and 36.9 million independent residuals. Noll

17

Rebuttal Exs. 3A, 3B.

18
19

ARGUMENT
I.

APPLYING THESE GENERALLY ACCEPTED PRINCIPLES, MURPHY AND

20

TOPEL SHOW THAT THE RESIDUALS ARE HIGHLY CORRELATED AND

21

THAT NOLL’S RESULTS ARE STATISTICALLY INSIGNIFICANT WHEN THE

22

CORRELATION IS CORRECTED.

23

As Cameron explains, if there is reason to suspect that the residuals are correlated, one

24

should test the assumption. Kiernan Decl., Ex. 7 at p. 20; see also id., Ex. 3 at p. 147.

25
26

11

27
28
-8-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
12

17
18
19
20
21
22
23
24

13

iPods are in hierarchical categories defined by Apple—model, generation, family (like
taxonomic categories in biology, from Kingdom to Species). The iPod model (classic, nano,
mini, shuffle, touch) describes the highest level of commonality (e.g., shuffles are small without
screens). Generation (e.g., 1st, 2nd, 3rd, 4th, etc.) refers to the next more narrow level of
commonality, and include iPod models that have features common for that generation (e.g., iPod
nano 2nd generation had iTunes 7.0, but iPod nano 1st generation did not.) Family is the most
specific, referring to an iPod of a certain model and generation with a specific feature set (e.g.,
iPod nano 2nd generation 4GB vs iPod nano 2nd generation 6GB).

14

25
26
27
28

15

See, e.g., Murphy Report ¶ 110-112, Murphy Dep. Tr. at 135:7-136:4 (attached as Exhibit
10 to Sweeney Decl. in Supp. of Pls.’ Mot. (ECF No. 737)); see also Kiernan Decl., Ex. 10
(various Apple Price Committee Documents).
-9-

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1
2
3
4
5
6

And if such correlation is not corrected, the regression “will lead to very

7

misleadingly small standard errors, and consequent misleadingly narrow confidence intervals,

8

large t-statistics and low p-values”17 and “over-estimate[] the magnitude and statistical

9

significance of the effect of [

10

].”18

Noll admitted that his model should include any product attribute so long as “prices

11

plausibly could be affected by it.” Noll 2d Supp. Decl. (ECF No. 685) at p. 9. Nevertheless, he

12

has excluded such features. Topel Report ¶ 112, n. 94, see also Kiernan Decl., Ex. 10. Most of

13

the omitted attributes were considered by Apple’s Price Committee when setting prices for iPods.

14

Id. By omitting variables that impact prices of iPods, Noll has introduced correlation in the

15

residuals that allows him to vastly underestimate the standard errors.

16

Because there is ample reason to suspect correlation, Murphy and Topel used standard

17

tests to determine whether the residuals are highly correlated at the family and quarter level.

18

They are. Topel Report at ¶ 80 & Ex. 14a thereto; Murphy Report ¶ 97 & Ex. 13a thereto;

19

Murphy/Topel Supp. at n. 14 and Exhibits JT-3a and JT-3b thereto.19 And after correcting for the

20
16

21
22

Or in Noll’s terms, the “outcome for a member of a group [here prices for iPods within a
family] may be affected by factors other than the treatment variable [here factors other than
iTunes 7.0] that are common to all group members [here all iPods within a family].” Noll
Rebuttal at p. 41.

23

17

Kiernan Decl., Ex. 7 at p. 4.

24

18

Noll Rebuttal at p. 42; see also Wooldridge Report at p. 3; Kiernan Decl., Ex. 5 at p. 864.

25
26
27
28

19

Murphy and Topel calculated the estimated residual (error term) for each of the
transactions. Within each family and quarter, they divided the residuals into two equal size
groups and calculated the average residual within the group. If the residuals are independent, the
mean should be zero--the residuals should be grouped around zero. The results, however, show
the opposite. The mean residuals are not grouped around zero, they range from -.485 to .347 for
the reseller regression and from -.692 to .273 for the direct regression. And they are strongly
(continued)
- 10 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

high correlation using clustering, Murphy and Topel demonstrate that Noll’s results are

2

statistically meaningless; i.e., the results are indistinguishable from zero. Murphy/Topel Supp. ¶

3

9.20

4

Despite having reasons to suspect correlation at the family level including the results of

5

the tests run by Murphy and Topel, Noll refused to employ any of the standard, accepted

6

procedures (including the one he referred to in his Rebuttal Report) to test the independence

7

assumption.21 Refusing to apply generally accepted methods and ignoring facts that undermine

8

his opinion are grounds alone to exclude his opinions.

9

II.

PLAINTIFFS’ EFFORT TO REHABILITATE NOLL BY ATTACKING APPLE’S

10

EXPERTS IS WITHOUT MERIT AND PROVIDE GROUNDS TO EXCLUDE

11

NOLL AND WOOLDRIDGE.

12

Plaintiffs attack Murphy’s and Topel’s testimony on statistical significance and clustering

13

on the grounds that (a) clustering is never appropriate when the regression uses the entire

14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

positively correlated. The 45 degree line in the graph shows the place on the graph at which the
residuals are perfectly correlated. Here, the residuals lie so close to the 45-degree line showing
that the residuals are highly correlated within family and quarter. See Topel Report ¶ 80; Murphy
Report ¶ 97.
20

For the first time in this case, in a footnote, Plaintiffs assert that regression results need
not be statistically significant to be admissible. Mot. at 8 n.12. That issue has no bearing on the
current motion to exclude Murphy and Topel’s criticism of Noll’s methodology in calculating
statistical significance. But plaintiffs’ assertion is not true. Their sole authority for the assertion
is addressed to epidemiological studies, not to regression models proffered as the sole basis for a
claim of antitrust impact and measure of damages. Cook v. Rockwell Inter. Corp., 580 F. Supp.
2d 1071, 1102-03 & n.29 (D. Col. 2006). Noll offers no opinion that his regression models could
be relied upon to find impact or damages if their results are not significantly different from
zero—and no such assertion would make sense. Recognizing as much, Noll asserts that the
results should be accepted because the coefficients are “highly significant.” Noll Report, p. 90.
In his Rebuttal Report, he never made plaintiffs’ argument. Instead, he vigorously defends his
reported levels of statistical significance. Nor did Murphy say that statistical significance is
irrelevant. He said only that there is no bright-line for statistical significance and a T-statistic of
1.95 is not materially different from one of 2.05. See Ex. 10 to Pls.’ Mot. at 51:9 to 52:9.

21

This is not the first time the clustering issue has been raised. During the class certification
phase, Apple’s experts demonstrated that Noll’s regressions, which are based on the same data,
overstated the standard errors by calculating them assuming the residuals were independent when
in fact they were not. See, e.g., Second Supp. Expert Report of Dr. Michelle Burtis, (ECF No.
692) at ¶ 42-46.
- 11 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

population of all transactions as opposed to a sample of the data, (b) Murphy and Topel engaged

2

in “ex post clustering”; and (c) even if clustering is appropriate when using entire population, the

3

technique used by Murphy and Topel is inappropriate because there are too many observations

4

per cluster.

5

For each of these points, plaintiffs rely on Woolridge, who spent 12 hours on his report

6

and largely parrots the arguments Noll made in his rebuttal report. Wooldridge Report at p. 2;

7

Kiernan Decl., Ex. 11 (Wooldridge Dep. Tr., taken Jan. 6, 2014) at 8:16-20, 10:15, 53:8. But

8

when confronted at deposition with generally accepted econometrics and his own writings,

9

Wooldridge was forced to admit that his (and by extension Noll’s) theories have no support, have

10

not been peer reviewed, and are contrary to generally accepted econometrics and literature

11

including Wooldridge’s own writings on the subject. Wagner v. County of Maricopa, 673 F.3d

12

977, 982 (9th Cir. 2012) (an expert’s analysis should be “supported by the typical Daubert factors

13

‒ testing, peer review and general acceptance”). Such theories and opinions obviously

14

“conceived, executed, and invented solely in the context of th[e] litigation” are per se

15

inadmissible. Johnson v. Manitowoc Boom Trucks, Inc., 484 F.3d 426, 434-35 (6th Cir. 2007).

16
17

A.

The Claim That Clustering Is Only A Problem When Sampling Is Objectively
Wrong And Should Be Excluded Under Daubert

18

Plaintiffs’ principal argument is that it is appropriate to account for within-group

19

correlation (clustering) only when the data were drawn as a so-called “cluster sample,” not when

20

using the entire population of transactions. Pls.’ Mot. 9-10; Noll Rebuttal at p. 10; 39-43. They

21

assert that, because Noll used virtually all iPod transactions and not a sample of transactions,

22

there can be no clustering problem. They have no support for this theory.

23

As Murphy and Topel explain, accounting for clustering is the norm in applied

24

econometrics whether dealing with a population or sampling. “The issue is not whether you have

25

‘all’ of the data or merely a sample, but whether the residuals in the statistical model assumed by

26

Professor Noll are correlated.” Murphy/Topel Supp. at ¶ 11(b) (citing sources); see also Kiernan

27

Decl., Ex. 3 at 144-147. Indeed, in antitrust cases, the parties usually have the entire population

28

of transactions from defendants—not random samples. Kiernan Decl., Ex. 1 at 266:21-267:1. In
- 12 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

these situations, the ABA Section of Antitrust Law, Proving Antitrust Damages: Legal and

2

Economic Issues (2010), warns that within-group correlation of regression residuals must be

3

considered and corrected. Id. at 144-147. Otherwise “the incorrectly estimated standard errors

4

generally will be biased downward, making the regression coefficients seem to be more precisely

5

estimated than they really are. As a result, a statistical test on the coefficients may yield what

6

appears to be a statistically significant result but is not.” Kiernan Decl., Ex. 3 at pp. 145-146.

7

And none of the authorities cited by Noll or Wooldridge support their argument that correlation of

8

residuals within group is limited to cluster samples.

9

Given this authority, it is no surprise that Noll retreated at his deposition and admitted that

10

the independence assumption applies when using the entire population of transactions and that

11

residuals could be correlated within groups. Kiernan Decl., Ex. 6 at pp. 44-45. But as discussed

12

above, he refused to employ any procedures to determine whether the residuals are correlated.

13

And now Woolridge has effectively done the same thing. In his declaration, Wooldridge

14

restates Noll’s opinion that the clustering problem only arises when sampling, and never arises

15

when the regression utilizes the entire population of data. He asserts that any clustering that

16

occurs after collecting the data (what he terms “ex post clustering”) is inappropriate. Wooldridge

17

Report at p. 9. But when confronted at deposition with his own writings, he admitted that he

18

could point to no authorities that support his opinions, and that they are untested and have not

19

been peer reviewed, and were formed after being retained in this litigation. Kiernan Decl, Ex. 11

20

at 91:22-92:7; 62:21-24.22 In fact, it appears that Wooldridge coined the term “ex post clustering”

21

specifically for this litigation. He can point to no other source in the field of econometrics that

22

uses the term. Id. at 90:21-24.

23

Although he allegedly first began “thinking” about the theory several years ago (id. at

24

62:18-21; 64:8-19), it was only after being retained in this litigation in December 2013 for

25

$500/hour and after submitting his declaration that he “worked out a little bit of theory as well as

26
27
28

22

The closest academic work he could point to was his article on stratified sampling. But he
admitted that it did not address whether “ex post clustering is inappropriate when dealing with an
entire population.” Kiernan Decl., Ex. 11 at 92:11-14.
- 13 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

the simulation” to test his and Noll’s opinions. Id. at 62:21-24. However, he was explicit that he

2

is not relying on any of this preliminary work for support. Id. at 13:8-11, 18:5-12.

3

His new theory contradicts the economics literature on when clustering problems arise and

4

contradicts how econometricians have dealt with the issue “for a long time.” Id. at 90:24-91:6. It

5

also contradicts Wooldridge’s own work. Before being retained, Wooldridge recognized that

6

clustering can be a problem even when using the entire population of data. In his recent graduate

7

textbook on econometrics, for example, he explains that the “cluster sampling” problem refers to

8

the problem where groups are drawn from a population rather than individual units, which in turn

9

cause “the units within the cluster [to be] correlated through unobserved ‘cluster effects.’”

10

Kiernan Decl., Ex. 5 at p. 853. Later in the chapter he is clear that the problem is not limited to

11

sampling: “it is probably a sensible rule to at least consider the data as being generated as a

12

cluster sample whenever [variables] at a level more aggregated than the individual units are

13

included in an analysis. For example, in analyzing firm-level data, if industry-level [variables]

14

are included then we should treat the data as a cluster sample, with each industry acting as a

15

cluster.” Id. at p. 864. Indeed, in his Example 20.3 to show “Cluster Correlation in Teacher

16

Compensation,” he uses “virtually” the entire population of data of average compensation at the

17

school level for teachers in Michigan and states that he “view[s] this as a cluster sample of school

18

districts, with the schools within districts representing the individual units” Id. at p. 868; Ex. 11

19

at 127:15-128:23.

20

Directly contradicting his current assertions, his textbook also recommends clustering

21

after collecting the data (i.e., what he now derides as “ex post clustering”) if the researcher has

22

reason to believe that there are common unmeasured factors that affect outcomes. Using the

23

school example, Wooldridge’ textbook and declaration note that, if the researcher originally

24

sampled fourth-grade classrooms, the researcher would cluster at the classroom level—that is the

25

“cluster sample.” Wooldridge’s declaration (pp. 3-4) states that it would be improper later to

26

cluster at the school level because such “ex post clustering” is done after the data is collected. He

27

accuses Murphy and Topel of effectively doing this. However, his textbook acknowledges that, if

28

the researcher is “worried about correlation in student performance not just within class but also
- 14 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

within school, then [the researcher] can define the clusters to be the schools.” Id., Ex. 5 at p. 864.

2

That is the equivalent of what Murphy and Topel did in this case. Because they considered that

3

iPod prices could be correlated within family, they defined the clusters to be family.

4

When asked about the discrepancy between his declaration and textbook, Wooldridge said

5

he “might rethink” what he previously wrote in his textbooks and “re-examine [certain]

6

statement[s] in light of [the] recent research that [he] has done.” Id., Ex. 11 at 117:6-15.

7

Ultimately, he conceded that he would need to rewrite his text and claimed that the ABA’s

8

treatise, Proving Antitrust Damages, should be rewritten in light of his new, unsupported theory.

9

Id. 159:9-11.

10

In short, Wooldridge’s opinions have been custom-manufactured for this litigation, are

11

supported only by his own ipse dixit and run contrary to his own previous works on the subject

12

and the articles he cites in his declaration. Neither Daubert nor Rule 702 requires a district court

13

to admit such opinion evidence. See Daubert v. Merrell Dow Pharm., Inc., 43 F.3d 1311, 1318

14

(9th Cir. 1995) (Daubert II) (requiring expert to “point to some objective source” to show

15

conclusions are scientifically valid). Accordingly, the Court should exclude Wooldridge’s

16

declaration.23 And it provides no support for plaintiffs motion to exclude.

17

B.

18

Plaintiffs’ Claim That Clustering Overestimates The Standard Errors Is
Baseless

19

Relying on Noll and Wooldridge, plaintiffs contend that, even if clustering is appropriate,

20

it leads to biased results when the ratio of the number of clusters (N) and the number of

21

observations per cluster (T) is large. Motion at 12. They argue that, as the number of

22

observations per cluster increases relative to the number of clusters, clustering will overstate

23

standard errors and understate statistical significance of regression results. Id. (citing Wooldridge

24
25
26
27
28

23

See also Cabrera v. Cordis Corp., 134 F.3d 1418, 1423 (9th Cir. 1998) (explaining expert
opinion unreliable where developed “expressly for the purpose of testifying” and the expert
cannot “identify any peer-reviewed research justifying his conclusions”); Lust v. Merrell Dow
Pharmaceuticals, Inc., 89 F.3d 594, 597 (9th Cir. 1996) (“When a scientist claims to rely on a
method practiced by most scientists, yet presents conclusions that are shared by no other scientist,
the district court should be wary that the method has not been faithfully applied.”).
- 15 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

Report at p. 9). This new theory, like the others, was manufactured for this litigation, has no

2

support in peer reviewed or any other literature, is untested, and contradicts the very authorities

3

Noll and Wooldridge cite. Indeed, at deposition, Wooldridge admitted he had no authority for it.

4

Kiernan Decl., Ex. 11 at 131:4-132:7. And none of the authorities they cite support them.

5

Noll cites the book by Angrist and Pischke, but that work shows the opposite of Noll’s

6

opinion. Angrist and Pischke establish that the conventional estimators used to estimate standard

7

errors become “increasingly misleading” when the number of observations per group and the

8

amount of within-group correlation increase. Kiernan Decl., Ex. 4 at p. 310; see also pp. 319-

9

320.24 This is precisely when clustering should be performed, because it leads to more accurate

10

results. Id.; see also Topel Report at ¶ 75; Kiernan Decl., Ex. 12 (Hansen, Asymptotic Properties

11

of A Robust Variance Matrix Estimator For Panel Data When T is Large (2007)) at pp. 612-615.

12

And as Noll admits, Angrist and Pischke state that clustering performs well when the number of

13

clusters is larger than about 50.25 In this case, there are over 375 clusters when clustering by

14

family and quarter, and over 68 clusters when clustering by family.26

15

Wooldridge confirmed that Noll has no support for this novel theory. At deposition, he

16

admitted that no authority supports the opinion that the clustering estimator used by Murphy and

17

Topel produces unreliable results or that it performs more poorly as the number of observations

18
19
20
21
22
23
24
25
26
27
28

24

“Conventional [OLS] standard errors become increasingly misleading as n [the number of
observations per group] and p [the amount of within-group correlation] increase.” Kiernan Decl.,
Ex. 4 at p. 310; see also pp. 319-20.

25

Noll Rebuttal at n. 18 (“Angrist and Pischke [ ] whimsically state that 42 is the magic number
of clusters . . . [o]f course, no magic solution exists, but the effects of violating the standard
assumptions about the distribution of errors declines as the number of clusters increases.”); see
also Kiernan Decl., Ex. 4 at § 8.2.3; see also p. 313: “The [CCM estimator] is consistent as the
number of groups gets large under any within-group correlation structure and not just the
parametric model in (8.2.30).”

26

Noll makes the added argument that there is no clustering problem where there are large
number of clusters. Wooldridge does not support this part of Noll’s opinion, presumably
recognizing that it is wrong. Nor do Plaintiffs rely on it in their motion. Noll has it exactly
backwards. The text Noll cites states that a large “number of clusters” is precisely the condition
that makes clustering a reliable and accurate method for calculating standard errors than the
estimator used by Noll. See Kiernan Decl., Ex. 4 at § 8.2.3, p. 313 & Ex. 12 at 612-615 & Ex. 11
at 135:17-139:13.
- 16 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

per cluster increase relative to the number of clusters. Kiernan Decl., Ex. 11 at 145:19-147:9.

2

Wooldridge claimed that he started working on “simulations” to test his new theory after

3

submitting his report, id. at 145:16-146:23, but is not relying on those simulations. Id. at 13:8-11.

4

The one paper Wooldridge does cite, Christian Hansen’s Asymptotic Properties of A

5

Robust Variance Matrix Estimator For Panel Data When T is Large, shows exactly the opposite

6

of what Noll and Wooldridge claim. According to Hansen, it is the number of clusters (N) that

7

matter, not the ratio between the number of observations per cluster (T) and the number of

8

clusters (N): “[I]t is the N dimension and not the size of N relative to T that matters for

9

determining the properties of the CCM estimator.” Kiernan Decl., Ex. 12 at p. 598; see also p.

10

611. Dr. Hansen also includes a number of simulations, which Wooldridge admitted show that,

11

as the number of clusters increase, the standard estimator performs more poorly while clustering

12

leads to much more accurate standard errors. Kiernan Decl., Ex. 11 at 135:23-139:13. This is

13

exactly the opposite of what Noll and Wooldridge claim.27 Hansen’s results are consistent with

14

Angrist and Pichke’s opinion that clustering performs well when, as here, the number of clusters

15

is larger than about 50. Id. at Ex. 12, p. 612-615 (showing that clustering performs better as

16

number of clusters increases from 10 to 50; it also shows that clustering performs better as the

17

number of observations per cluster increases).

18

Finally, Plaintiffs and their experts assert that clustering “in the absence of any need for

19

one is not harmless” because it bias the results. Wooldridge’s declaration (p.5) cites no authority

20

for that proposition, and at deposition he admitted he was unaware of any. And as demonstrated

21

above and recognized in the ABA treatise Proving Antitrust Damages, clustering produces

22

accurate estimates of standard errors even when there is no within-group correlation among the

23

residuals. Kiernan Decl., Ex. 3 at p. 147.

24
25
26
27
28

27

At deposition, Wooldridge admitted that he had not carefully reviewed Hansen’s work. Early
in the deposition, he claimed that Hansen supported his opinion that ratio matters and that
clustering performs more poorly as the number of observations increase. Kiernan Decl., Ex. 11 at
131:4-132:2, 135:9-16. As discussed above, when confronted with the paper, Wooldridge was
forced to admit that Hansen’s paper states the opposite.
- 17 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

C.

2

Clustering With Or Without Time Period Demonstrates That Noll’s Results
Are Not Statistically Significant

3

Plaintiffs also criticize Murphy and Topel for clustering at the family and quarter level,

4

contending that there is no justification to cluster by time period. That assertion is wrong on two

5

counts. First, even if the standard errors are clustered only by family and not by time period,

6

Noll’s results still are not statistically significant. Murphy/Topel Supp. ¶¶ 8-10, n. 14 & Exhibits

7

JT3a and JT3b thereto. Second, whether or not a time dimension should be used, there clearly is

8

correlation of residuals which has not been properly corrected in calculating standards errors.

9

When corrected, with or without time period, Noll’s results are not statistically significant.28

10

****

11

In short, Plaintiffs’ motion should be denied. Murphy and Topel have applied generally

12

accepted econometric methods used in antitrust cases and showed that Noll inflated the statistical

13

significance of his regressions by calculating standard errors on the assumption that the errors are

14

independent when in fact they are not.

15

III.

16

WOOLDRIDGE’S REPORT SHOULD BE STRICKEN.
For the reasons shown above, Wooldridge’s opinions on clustering are unreliable and do

17

not pass muster under Daubert. His theories are that (i) clustering is never appropriate when

18

using the entire population, (ii) what he calls “ex post clustering” is inappropriate, and (iii)

19

clustering is inappropriate where ratio between number of clusters and number of observations is

20

large. These theories have no support. They are contrary to generally accepted econometrics and

21

28

22
23
24
25
26
27
28

Although not relied on by plaintiffs, Noll makes the added argument that no clustering is
necessary because his regressions include all statistically significant variables, which control for
any possible correlation within family. The theory is that, if all the important variables that affect
price of the various types of iPods are included in the regression, clustering may not be necessary
because the variables in the model capture the common factors. Noll is wrong on at least two
counts. First, whatever variables Noll included, the critical assumption that needs to be tested is
whether the residuals remain correlated or not. If they are, they must be corrected. As Murphy
and Topel demonstrated, the tests show that the residuals remain highly correlated within family
as would be expected. Second, Noll omitted statistically significant variables. Murphy/Topel
Supp. ¶¶ 8-11, 17-18. This not only renders his coefficient estimates unreliable by creating
omitted variable bias as discussed in Apple’s motion to exclude, but also can create correlation to
the extent that the omitted variables had common affect within family of iPods.
- 18 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

his own writings. They have not been peer reviewed. And they were manufactured for this

2

litigation. Wagner, 673 F.3d at 982 (an expert’s analysis should be “supported by the typical

3

Daubert factors ‒ testing, peer review and general acceptance”). Such theories and opinions

4

obviously “conceived, executed, and invented solely in the context of th[e] litigation” are per se

5

inadmissible. Johnson, 484 F.3d at 434-35.

6

In addition, the declaration should be stricken because plaintiffs did not timely disclose

7

Wooldridge as an expert. The schedule in this case, after several modifications, ultimately set

8

April 1, 2013 as the date for plaintiffs to disclose experts and November 25 as the date for

9

Plaintiffs to serve rebuttal reports responding to Apple’s experts, 25 days before Daubert motions

10

were due. ECF No. 735. No exception was made for a rebuttal report to be submitted with

11

Daubert motions. Id. Whether Noll should cluster the standard errors has been at issue in this

12

case for over two years. See, e.g., Second Supp. Expert Report of Dr. Michelle Burtis (ECF No.

13

692) at ¶ 42-46. Nevertheless, plaintiffs waited until after discovery was completed and disclosed

14

Wooldridge’s identity on the day Daubert motions were due. This left defendants with only two

15

weeks to respond to the new expert, who should have been disclosed months ago. This Court

16

should strike the late-filed declaration on this ground alone. Reed v. Smith & Nephew, Inc., 527

17

F. Supp. 2d 1336, 1348 (W.D. Okla. 2007) (striking declaration of undisclosed expert submitted

18

in support of Daubert motion); Moore v. Napolitano, 926 F. Supp. 2d 8, 25 n.12 (D.D.C. 2013)

19

(same); Jeffries v. Centre Life Ins. Co., No. 1:02-cv-351, 2004 WL 5506494 at *1 (S.D. Ohio Jan.

20

28, 2004).

21

IV.

PLAINTIFFS’ ASSERTION THAT MURPHY’S OPINIONS HAVE BEEN

22

REJECTED ELSEWHERE IS UNFOUNDED.

23

Plaintiffs’ assertion that Murphy’s testimony in two cases has “failed to withstand”

24

scrutiny is irrelevant and incorrect. Irrelevant, because it has no bearing on the invalidity of

25

Noll’s regression in this case—or of Murphy’s and Topel’s testimony showing that invalidity.

26

Incorrect, because it misstates the rulings in those two cases.

27
28

In In re High-Tech Employee Antitrust Litig., 289 F.R.D. 555 (N.D. Cal. 2013), Judge
Koh simply declined to resolve at the class certification stage a conflict between the experts as to
- 19 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR

1

the proper method of disaggregating results. She did not find Murphy’s opinion to be

2

“unreliable” or “inadmissible,” but stated only that she did not find his method to be “more

3

credible” than the method offered by plaintiffs’ expert—and thus that it did not require that class

4

certification be denied. Id. at 580.

5

In United States v. Apple, Inc., Nos. 12 Civ. 2826 (DLC), 12 CIV 3394 (DLC), 2013 U.S.

6

Dist. LEXIS 96424 (S.D.N.Y. Oct. 4, 2013), Murphy proposed to offer an economic opinion as to

7

whether Apple’s conduct was consistent with Apple’s independent business interests. The court

8

did not rule on the validity of Murphy’s economic analysis, but simply ruled that, as a legal

9

matter, whether Apple acted in independent interests was not relevant to finding a conspiracy and

10

that the expert evidence was unnecessary as to Apple’s motivation and intent. None of these

11

rulings on legal issues has any bearing on the soundness of Murphy’s opinions in this case on

12

econometric issues—or on validity of Noll’s regression model.

13
14
15

CONCLUSION
Plaintiffs’ motion should be denied, and Wooldridge’s declaration should be excluded.
Dated: January 13, 2014

16

Respectfully submitted,
Jones Day

17
18

By: /s/ David C. Kiernan
David C. Kiernan

19

Counsel for Defendant
APPLE INC.

20
21
22
SFI-849012v3

23
24
25
26
27
28
- 20 -

Apple’s Opposition to Plaintiffs’ Daubert
Motion 4:05-cv-00037 YGR