The Connecticut Crackdown on Speeding: Time-Series Data in Quasi-Experimental Analysis

Author(s): Donald T. Campbell and H. Laurence Ross

Source: Law & Society Review, Vol. 3, No. 1 (Aug., 1968), pp. 33-54
Published by: Blackwell Publishing on behalf of the Law and Society Association
Stable URL: http://www.jstor.org/stable/3052794
Accessed: 04/10/2009 22:02

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=black.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Blackwell Publishing and Law and Society Association are collaborating with JSTOR to digitize, preserve and
extend access to Law & Society Review.

http://www.jstor.org
 THECONNECTICUT
CRACKDOWN
ON SPEEDING
Time-SeriesData in
Quasi-Experimental
Analysis

DONALDT. CAMPBELL

Northwestern University

and

H. LAURENCERoss

University of Denver College of Law

SOCIALBRFSFARCHFREQuENTLY encounters the task of evaluating change

produced in nonrandomlyselected groups by events which are beyond
the researcher'scontrol. The social scientist must verify that there has in

AUTHORS' NOTE: The preparation of this paper has been supported in part
by the National Science Foundation (Grant GS 1309x), the U. S. Ofice of
Education (Project C-998, Contract 3-20-001), the U. S. Bureau of Public
Roads (CPR 11-5981), the National Institutes of Health, the U. S. Public
Health Service (RG-5359), and the Automotive Safety Foundation (as
an aspect of Experimental Case Studies of Traffic Accidents conducted at
Northwestern University). A brief version of it appears as H. L. Ross &
D. T. Campbell, The Connecticut Speed Crackdown: A Study of the
Effects of Legal Change, in PERSPECTIVES ON THESOCIALORDER:READ-
INGSIN SOCIOLOGY 30-35 (2d ed. H. L. Ross ed. 1968).

.33 .
LAW AND SocIETYREVIEW

fact been a change, and that the indicated event is its cause. Illustrations
are manifold: a state terminatescapital punishment,and proponentsof
this type of punishmentpredict an increasein the murderrate; a school
is integrated,and supportersof the reformexpect to find an increase in
the positive self-evaluation of Negro pupils; a natural disaster occurs
in a community,and altruisticbehavioris expected to increase. Because
in these situationsthe investigatorhas no control over the assignmentof
individuals or groups to "experimental"and "control"situations, the
logic of the classical experiment must be reexamined in a search for
optimal interpretativeprocedures.
This paper introduces,in the context of a problem in applied soci-
ology and the sociology of law, a mode of analysisdesigned to deal with
a common class of situations in which research must proceed without
the benefitof experimentalcontrol. The generalmethodologyexpounded
here is termed "quasi-experimentalanalysis." The specific mode of
analysis is the "interruptedtime-seriesdesign." Perhaps its fundamental
credo is that lack of control and lack of randomizationare damaging to
inferences of cause and effect only to the extent that a systematic con-
sideration of alternative explanationsreveals some that are plausible.
More completeexplicationsof quasi-experimentalanalysishave appeared
elsewhere;1this paper will merely illustrateits use in a situation where
a series of observationshas been recordedfor periods of time both prior
and subsequent to the experience of the specific event to be studied.
Such data are quite commonlyavailable,yet they are seldom fully utilized
and investigators often confine themselves unnecessarily to much less
satisfactorymethodologies. The 1955 crackdownon speeding in the State
of Connecticut furnishes an apt example of the potentialities of such
quasi-experimentalanalysis.

1. E.g., D. T. Campbell & J. S. Stanley, Experimental and Quasi-Experimental

Designs for Research on Teaching, in HANDBOOK OF RESEARCH ON TEACHING171-246
(N. L. Gage ed. 1963) reprinted as EXPERIMENTAL AND QUASI-EXPERIMENTAL DESIGNS
FOR RESEARCH(1963); D. T. Campbell, From Description to Experimentation: Inter-
preting Trends as Quasi-Experiments, in PROBLEMS IN MEASURING CHANGE(C. W. Harris
ed. 1963); D. T. Campbell & K. N. Clayton, Avoiding Regression Effects in Panel
Studies of Communication Impact, in STUDIESIN PUBLICCOMMUNICATION 99-118 (Dept.
of Sociology. University of Chicago, No. 3, 1961) reprinted in Bobbs-Merrill Reprints
in Sociology as S-353. For an application of this type of analysis to legal impact, see
R. Lempert, Strategies of Research Design in the Legal Impact Study, 1 L. & Soc'Y REV.
111 (1966).

.34
 THE CONNECTICUT CRACKDOWN ON SPEEDING

A PROGRAM FOR REDUCING HIGHWAY FATALITIES

In 1955, 324 people were killed in automobile accidents on the high-

ways of Connecticut. Deaths by motor vehicle accidents had reached
a record high for the decade of the fifties as the usually hazardous
Christmas holidays approached. Two days before Christmas, Governor
Abraham Ribicoff of Connecticut initiated an unparalleled attempt to
control traffic deaths by law enforcement, and announced his crackdown
on speeders in that state.
Ribicoff believed, along with many safety specialists, that excess speed
was the most common contributing factor in traffic deaths, and that con-
trol of speed would result in diminished fatalities. He believed that
previous efforts to control speeding under the usual court procedures and
by the existing "point system" had been inadequate. In a study of three
months' records of the police court in Hartford, it was noted that no
more than half the persons originally charged with speeding were so
prosecuted, the charge often being diminished to a less serious one.
Ribicoff wanted to initiate a program with reliable procedures and strong
sanctions as a means to control speeding and thus to reduce traffic
deaths.
On December 23, 1955, Governor Ribicoff announced that in the
future all persons convicted of speeding would have their licenses sus-
pended for thirty days on the first offense. A second violation was to
mean a sixty-day suspension, and a third conviction for speeding would
result in indefinite suspension of the driver's license, subject to a hearing
after ninety days.
The decree was put into force through the Governor's power of
appointment over local judges. Under Connecticut practice, the Motor
Vehicle Department was suspending licenses on the recommendation of
police court judges. The judges were appointed by the Governor, who
threatened loss of reappointment in 1957 to judges who appeared lax in
the conviction of speeders, or who did not recommend suspension of
licenses to the Motor Vehicle Department.
In the first three months of 1956, license suspensions for speeding
numbered 2,855, an increase of almost 2,700 over the corresponding
period in 1955. There were ten fewer fatalities, and 765 fewer arrests
for speeding. The Governor was reported "encouraged" by the drop in
violations and in fatalities. The press quoted him as saying, "This is

. 35 ?
LAW AND SocITY REVIEW

positive proof that operators are not only driving slower, but are driving
better."
By late May, deaths had declined from 122 in 1955 to 107 in 1956.
Suspensions for speeding numbered 4,559, as against 209 in 1955. Speed-
ing arrests had dropped 53 per cent. The Governor received a telegram
of commendation for the program from the National Safety Council.
At the end of June there were twenty-two fewer fatalities than in the
first six months of 1955, representing a 15 per cent reduction. Suspen-
sions for speeding in the first six months of the year had risen from 231
to 5,398, and arrests had declined from 4,377 to 2,735. Ribicoff an-
nounced:

Connecticut has succeeded in stopping the upward surge in highway

deaths, and in the first six months of this year, contraryto the national
trend, we have saved lives. Fewer people died on the highways this year
than in the same period last year, in Connecticut. We did it by enforcing
the law, something the safety experts said couldn't be done because the
people wouldn'tbe behind it.

In July, a new State Police program, using unmarked police cars and
making extensive use of radar, was inaugurated. The police issued a
report stating that 2 per cent of the cars observed by radar on July 4
were found to be speeding; at a later date, it was claimed that no speeders
were found among 53,000 cars similarly observed.
In the late summer, however, Connecticut experienced a very high
number of traffic fatalities. By the beginning of September, 194 people
had been killed, a number almost equal to the 195 of the comparable
period in the previous year. The accident "epidemic" was embarrassing
to the authorities, who retreated to defending the speeding crackdown
on the grounds (a) that the fatality rate remained low in comparison
with the national trend, which showed a 7 per cent increase; (b) that
exposure to accidents in the State had increased by 100 million vehicle
miles without an increase in deaths; and (c) that the total accident rate
had risen, thereby lowering the proportion of fatal accidents to total
accidents.
Fatalities were fewer in the fall of 1956, and by the end of the year
Connecticut could count 284 deaths in traffic as against 324 in 1955. The
Governor stated, "With the saving of forty lives in 1955, a reduction of
12.3 per cent from the 1955 motor vehicle death toll, we can say the
program is definitely worthwhile."

* 36 ?
 THE CONNECTICUT CRACKDOWN ON SPEEDING

The crackdownon speeding is still in effect in Connecticut,although

it is no longer the subject of newsworthycomment. It was not entirely a
political asset for the Democratic Governor. From the start, there were
problems with neighboring states, which originate a substantial share
of Connecticuttraffic,and which at first refused to suspend licenses of
drivers convicted of speeding in Connecticut. More important, many
powerful individuals and groups within Connecticutresented the direct
effects of the crackdown. Membersof the RepublicanParty wanted the
program "temperedwith justice." The Teamsters sponsored a bill to
eliminate compulsory license suspension on a first offense, and other
cases was
legislation granting restricted driving permits for '"hardship"
introduced. These effortswere not successful in officiallymoderatingthe
crackdownpolicy.
The people of Connecticut and their officials are paying what in
many instances appears to be a high price for the continuationof the
crackdownon speeding. Few will feel the price is too high if it can be
shown that as many as forty lives per year are being saved. However,
the question must be raised as to whether the results claimed for the
program in 1956 are valid in the light of both formerly and more
recently available statistics on highway fatalities.

ANALYSIS
QUASI-EXPEIMENTAL

Before-and-AfterMeasures
Trafficfatalities in Connecticut for 1956, compared with 1955, are
presentedin Figure 1. These are the data upon which GovernorRibicoff
relied in claiming success for the crackdownon speeding. Skillfullypre-
sented, such results can look impressive,but can also be fundamentally
misleading.
We can speak of the evidence presentedin Figure 1 as a quasi-experi-
ment: there is a "pretest"(the 1955 figures), an "experimentaltreatment"
(the crackdown),and a "posttest"(the 1956figures). A substantialchange
is noted which one would like to ascribeto the "experimentaltreatment."
In quasi-experimentalanalysisthis interpretationis held to be legitimate,
provided considerationis given to plausible rival explanations of the
differences,with supplementaryanalyses being added to eliminate these
where possible. In the language of quasi-experimentalanalysis,the data
of Figure 1 constitute a One-GroupPretest-PosttestDesign. This design

37.
LAW AND SOCIETYREVIEW

320 - \

\
310-\

\
300-

~~\

290 \

280 -

'T I.... I
BEFORE AFTER
CRACKDOWN CRACKDOWN
(1955) (1956)
Figure 1. ConnecticutTrafficFatalities, 1955-1956
fails to control for the six common threats to the validity of experiments
specified below:
1. History. This term denotes specific events, other than the experi-
mental treatment, occurring between the pretest and posttest, which might
account for the change. It furnishes a "rival hypothesis" to the experi-
mental hypothesis, a competing explanation of the before-to-after change
that must be eliminated as implausible, by one means or another, before
full credence can be given to the experimental hypothesis. For instance,
1956 might have been a particularly dry year, with fewer accidents due

38 .
 THE CONNECTICUT CRACKDOWN ON SPEEDING

to rain and snow, or there might have been a dramaticimprovementof

the safety features on the 1956-modelcars. In fact, neither of these is a
particularly plausible rival hypothesis in this instance, and we have
not encountered more likely ones, so this potential weakness may not
be crucial here.
2. Maturation. This term originatesin studies of individuals,where
it refers to regularchanges correlatedwith the passage of time, such as
growing older, more tired, more sophisticated, etc. It is distinguished
from history in refering to processes, rather than to discrete events.
Thus, one could classify here the general long-term trend toward a
reduction in automobile mileage death rates, presumablydue to better
roads, increased efficacy of medical care, etc. The better designs dis-
cussed below provide evidence concerning this trend in Connecticut in
previous years, and in other states for the same year.
3. Testing. A change may occur as a result of the pretest, even
without the experimentaltreatment. In the present instance, the assess-
ment of the trafficdeath rate for 1955 constitutesthe pretest. In this case
it is conceivablethat the measurementand publicizingof the trafficdeath
rate for 1955 could change driver caution in 1956.
4. Instrumentation.This term refers to a shifting of the measuring
instrument independent of any change in the phenomenon measured.
In the use of public records for time-series data, a shift in the govern-
ment agency recording the fatality statistics could account for such a
shift. For example, suicide statistics increased a dramatic 20 per cent
in Prussiabetween 1882 and 1883, when record keeping was transferred
from the local police to the national civil service.2 Similarly, Orlando
Wilson's reforms of the police system in Chicago led to dramatic in-
creases in rates for most crimes, due presumably to more complete
reporting.3 In earlier versions of the present study, the death rate per
hundred million vehicle miles is computed by using the number of
gallons of gasoline sold in the state to estimate the number of miles
driven. The latter figure is obtained by multiplying the former by an
empirically-derivedconstant. A decrease in the actual miles obtained
2. Cited in C. SELLTIZ,M. JAHODA,M. DEUTSCH, & S. W. COOK,RESEARCH METHODS
IN SOCIALRELATIONS 323 (1959).
3. J. Sween & D. T. Campbell, A Study of the Effect of Proximally Autocorelated
Error on Tests of Significance for the Interrupted Time Series Quasi-Experimental
Design 31-32, Figs. 11 & 12 (mimeographed Research Report, Dept. of Psychology,
Northwestern University, 1965). These figures also will appear in D. T. Campbell,
Reforms as Experiments, AM. PSYCHOLOGIST (to be submitted).

39 -
LAW AND SocmIEY REVIEW

per gallon, as throughengines of larger horsepoweror driving at higher

speeds could masqueradeas a lower mileage death rate throughinflating
the estimate of miles driven. Conversely, if the crackdown actually
reduced drivingspeeds, this would increasethe miles-per-gallonactually
obtained, leading to an underestimateof mileage driven in the post
crackdownperiod, and consequentlyan overestimateof the fatality rate.
5. Instability.4 A ubiquitous plausible rival hypothesis is that the
change observedis due to the instabilityof the measuresinvolved. Were
Figure 1 to show fatality rates for a single township,with the same 12.3
per cent drop, we would be totally unimpressed,so unstable would we
expect such rates to be. In general,as is made explicit in the models for
tests of significance,the smaller the population base, the greater the
instability. In the uncontrolledfield situation sample size is only one of
many sources of instability. Much instabilitymay be due to large num-
bers of change-producingevents of the type which, taken individually,
we have called history.
6. Regression. Where a group has been selected for treatmentjust
because of its extremeperformanceon the pretest, and if the pretest and
posttestareimperfectlycorrelated,as they almostalwaysare,it followsthat
on the average the posttest will be less extreme than the pretest. This
regressionis a tautologicalrestatementof the imperfect correlationbe-
tween pretest and posttest, as it relates to pretest scores selected for
their extremity. The r of the correlationcoefficient actually stands for
the percentage of regressiontoward the mean. An analogousregression
problem exists for time-seriescorrelations.
Selection for extremity (and resultantretest regression) can be seen
as plausibly operatinghere in two ways: (a) of all states in 1955, this
treatmentwas most likely to be applied to one with an exceptionallyhigh
trafficcasualty rate; (b) for Connecticut,the most likely time in which
a crackdownwould be appliedwould be following a year in which traffic
fatalitieswere exceptionallyhigh.
In the true experiment,the treatmentis applied randomly,without
relationto the prior state of the dependent variable: the correlationbe-
tween pretest scores and exposureto treatmentis zero. Likwise, in the

4. Instability has not been singled out as a specific threat to validity in previous dis-
cussions of quasi-experimental design, although the discussion of tests of significance in
such situations has implied it. Tests of significance obviously do not provide "proof"
relevant to the many other sources of invalidity, but they are relevant to this one plausible
rival hypothesis even where randomization has not been used.

.40
 THE CONNECMICUT
CRACKDOWNON SPEEDING

most interpretableof quasi-experiments,the treatmentis applied without

systematicrelationshipto the priorstatus of the group. Thus, an analysis
of the effects of a tornadoor an earthquakecan be made with confidence
that the pretreatmentvalues did not cause the tornadoor the earthquake.
Not so here: the high 1955 rates can plausiblybe argued to have caused
the treatment. That 1956 was less extreme would then be expected be-
cause of regression.5

Interrupted Times-SeriesAnalysis
Figure 2 plots trafficfatalities for five years before and four years
after the crackdown. This mode of quasi-experimentalanalysishas been
labeled "InterruptedTime-Series"to distinguish it from the time-series
analysis of economics. In the latter, the exogenous variable to which
cause is imputed is a continuouslypresent variable,occurringin different
degrees. In the Interrupted Time-Series, the "causal"variable is ex-
amined as an event or change occurringat a single time, specified inde-
pendently of inspection of the data.
The Interrupted Time-Series design represents a use of the more
extensivedata which are often availableeven when only before-and-after
measures are reported. Some potential outcomes of such a time-series
analysis greatly reduce the plausibility of certain threats to validity. If
the preexposureseries shows but minor point-to-pointfluctuationsand
no trend anticipatinga big transtreatmentshift, then maturationmay not
be plausible, for in most instances the plausible maturationhypothesis
would have predicted shifts of the same order as the transtreatmentshift
in each of the pretreatmentstages. Reasonable models of the testing
effect would have the same implications. (In our instance, this would
be on conditionthat the annualfatality rates had been given equal pub-
licity.) The outcome in Figure 2 is not of this readily interpretablesort,
although the trend is perhaps generally upward prior to the treatment,
and steadily downward subsequently.
Judgmentsof the plausibilityof instrumentationeffects must be based
upon other than time-seriesdata. However, notice should be taken here
of a frequent unfortunateconfounding:the administrativereform which
is meant to produce a social change very frequently is accompaniedby
5. This issue is extremely complex. In ordinary correlation, the regression is tech-
nically toward the mean of the second variable, not to the mean of the selection
variable, if these means differ. In time-series, the regression is toward the general
trend-line, which may of course be upward or downward or unchanging. A more ex-
panded analysis of the regression problem in correlation across persons is contained in
Campbell & Clayton (1961) and in Campbell & Stanley (1963), both supra note 1.

.41 .
LAW AND SocIETY REVIEW

325

00 -

275

z
225- E
4
hi

200

-r I I . I I I
'51 '52 '53 '54 '55 '56 '57 '58 '59

Figure 2. ConnecticutTrafficFatalities, 1951-1959

a coincident reform of the record keeping, ruling out valid inferences as

to effects. The Chicago police reform cited above is a case in point.
In the present instance, we have found no evidence of a change in record
keeping or index computing of the type that would produce a pseudo-
effect.
The likelihood of regression, or of selection for "treatment" on a
basis tending to introduce regression, is supported by inspection of the
time-series data. The largest change of any year is not the one after
the crackdown, but is instead the upswing in the series occurring in
1954-55, just prior to the crackdown. In terms of crude fatality rates,
1955 is strikingly the highest point reached. It thus seems plausible

*42 .
 THE CONNECsICUTCRACKDOWNON SPEEDING

that the high figure of 1955 caused the crackdown,and hence it seems
much less likely that the crackdowncaused the low figure of 1956, for
such a drop would have been predicted on regression grounds in any
case.
The graphic presentation of the precrackdownyears provides evi-
dence of the general instability of the accidental death rate measure,
against which the 1955-56shift can be compared. This instabilitymakes
the "treatmenteffect"of Figure 1 now look more trivial. Had the drop
following the treatment been the largest shift in the time series, the
hypothesis of effect would have been much more plausible. Instead,
shifts that large are relatively frequent. The 1955-56 drop is less than
half the magnitude of the 1954-55gain, and the 1953 gain also exceeds
it. It is the largest drop of the series, but it exceeds the drops of 1952,
1954, and 1958 by trivial amounts. Thus the unexplainedinstabilitiesof
the series are of such a magnitude as to make the 1955-56 drop under-
standableas more of the same. On the other hand, it is noteworthythat
after the crackdown,there are no year-to-yeargains, and in this respect,
the character of the time-series has changed. The plausibility of the
hypothesisthat instabilityaccountsfor the effect can be judged by visual
inspection of the graphed figures, or by qualitative discussion, but in
addition it is this one threat to validity which can be evaluated by tests
of significance. These will be discussed later, and they do find some
evidence of change exceeding that which the pretreatmentinstability
would lead one to expect.

Multiple Time-Series
In many situations, time-series involving but a single experimental
unit will be all that are available. In these situations, analyses on the
above model are a great improvementover the usual before-and-after
study. However,it is in the spirit of quasi-experimentalanalysisto make
use of all available data that could help to rule out or confirm any
plausiblerival hypothesis. In a setting such as this, no randomlyassigned
control group is available. But in quasi-experimentation,even a non-
equivalent control group is helpful. It provides the only control for his-
tory (for those extraneous change agents that would be expected to
affect both the experimentaland control group), and assists in control-
ling maturation,testing, and instrumentation. For Connecticut,it was
judged that a pool of adjacent and similar states-New York, New
Jersey, Rhode Island and Massachusetts-provideda meaningful com-
parison. Figure 3 plots the death rates for the control states alongside
* 43
LAW AND SocmIEr Rivuvw

Connecticut,all data being expressed on a per 100,000population base

to bring the figuresinto proximity. The control data are much smoother,
due to the much larger base, i.e., the canceling out of chance deviations
in the annual figures for particularstates.

17 Connecticut

Control Stotes--
16

16
LU
wS
14

13
1-
12

U. II

10

7
1 I I I I I 1 I I I
'51 '52 '53 '54 '55 '56 '57 '58 '59

YEAR
Figure 3. Connecticutand ControlStates TrafficFatalities, 1951-1959
(per 100,000 population)
While in general these data confirm the single time-series analysis, the
differences between Connecticut and the control states show a pattern
supporting the hypothesis that the crackdown made a difference. In the
pretest years, Connecticut's rate is parallel or rising relative to the con-

.44 .
 THE CONNECTICUT CRACKDOWN ON SPEEDING

trol, exceeding it in 1955. In the posttest years, Connecticut'srate drops

faster than does the control, steadily increasing the gap. While the re-
gression argument applies to the high point of 1955 and to the subse-
quent departurein 1956, it does not plausibly explain the steadily in-
creasing gap in 1957, 1958, and 1959.
Figure 4 shows the comparisonstates individually. Note that four
of the five show an upward swing in 1955, Connecticut having the
largest. Note that all five show a downward trend in 1956. Rhode
Island is most similarto Connecticutin both the 1955 upswing and 1956

17-
? Connectlout -

16 mm OAS
.~
\
18 1-

4; 14
QL:
13

3-
4
Ii. II

10 _, aS /\ %\
I \
9 % /.o
_ \ .-
fA- ;40 16~~~~...
_~ --dx
%
7

I I I I I I I I I
'51 '52 '53 '54 '55 '56 '57 '58 '59

YEAR
Figure 4. TrafficFatalitiesfor Connecticut,New York,New Jersey,
Rhode Island, and Massachusetts(per 100,000 persons)

. 45
LAW AND SocmIEa REVIEW

downswing, actually exceeding Connecticut in the latter-in a striking

argumentagainst the hypothesis of a crackdown effect. However, the
trendin 1957,1958, and 1959is steadily upwardin Rhode Island, steadily
downward in Connecticut,supportingthe concept of effect.
The list of plausiblerival hypothesesshould include factors disguising
experimentaleffects as well as factors producing pseudo-effects. Thus,
to the list should be added diffusion,the tendency for the experimental
effect to modify not only the experimentalgroup, but also the control
group. Thus the crackdownon speeding in Connecticutmight well have
reduced trafficspeed and fatalities in neighboringstates. Dodd reports
such an effect in his classic experimenton communityhygiene in Syria.6
The comparisonof posttreatmentlevels of Connecticut and the neigh-
boring states might thus be invalid, or at least underestimatethe effects.
Conceivablyone might for this reasonpreferthe single time-seriesanaly-
sis to the multiple time-seriesone. If highly similar remote states were
available, these would make better controls, but for matters of either
weather or culture adjacencyand similarityare apt to be strongly asso-
ciated.

Tests of Significance
Our position in regard to tests of significanceis an intermediateone.
On the one hand, we would agree that they are overly honored and are
often useful in ruling out that one threat and should be used for that
purpose. They are appropriateeven where randomizationhas not been
used because even there it is a relevant threat to validity to' be able to
argue that even had these data been assigned at random,differencesthis
large would be frequent.7
The simplest tests conceptually are those testing for a difference in
slope or interceptbetween pretreatmentand posttreatmentobservations.
As applied here these assume linearity and independence of error. It
has been shown that the "proximallyautocorrelated"error typical of
natural situations (in which adjacent points in time share more error
than non-adjacentones) biases the usual tests in the directionof finding
too many significantdifferences.8Unaffected by this bias is a t-test by
Mood which compares a single posttreatmentpoint with a value ex-

6. S. C. DODD,A CONTROLLED EXPERIMENT ON RURALHYGIENE IN SYRIA(1934).

7. D. T. Campbell, Quasi-Experimental Design in 5 INT'L ENCYC.Soc. SCI. 259
(Sills ed. 1968).
8. J. Sween & D. T. Campbell, supra note 3. The tests thus biased include tests of
slope and intercept provided by H. M. WALKER& J. LEV,STATISTICAL INFERENCE 390-95,

?46 ?
 THE CONNECTICUT CRACKDOWN ON SPEEDING

trapolatedfrom the pretreatmentseries.9 None of these approachedany

interesting level of significance.
Glass'?has introduced into the social sciences a more sophisticated
statistical approach,based upon the work of Box and Tiao.11 This has
the advantagesof realisticallyassumingthe interdependenceof adjacent
points and estimating a weighting parameter thereof, of avoiding the
assumptionof linearity (at least in a simple or direct manner), and of
weighting more heavily the observationscloser to the point of treatment.
A number of assumptionsabout the nature of the data must be made,
such as the absence of cycles, but these can be examinedfrom the data.
Applying this test to monthly data, he finds a drop in fatalities not quite
reaching the P<.10 level of significance. Using a monthly difference
between Connecticut'srate and that of the pool of the four controlstates,
still less of a significanteffect is found. In what he regards as the most
powerful analysis available, he computes an effect parameterfor each
of the four comparison states and compares the effect parameter of
Connecticutwith this. Connecticutshows more effect, with a significance
level somewhere between P<.05 and P<.07, with a one-tailed test. A
more detailed description of the method and analysis of these data is
given in Glass' article, immediately following in this issue of Law &
Society Review.
Thus on the graphic evidence of steadily droppingfatality rates, and
on these marginal statistical grounds, there may be an effect. This
effect, it must be restated, could be due to the crackdown,or could be
due to the regressioneffect. (Regression effects can of course produce
"statisticallysignificant"results.)

399-400 (1953). Note that this invalidates the discussion of tests of significance in
Campbell, From Description to Experimentation, supra note 1, at 220-30. The "Clayton
test" presented there was found in the Monte Carlo simulation by Sween & Campbell
to have additional errors leading it to be too optimistic.
9. A. M. MOOD,INTRODUCTION TO THE THEORYOF STATISTICS 297-98 (1950).
10. G. V. Glass, Analysis of Data on the Connecticut Speeding Crackdown as a
Time-Series Quasi-Experiment, 3 L. & SOC'YREV. 55-76 (1968); T. O. Maguire & G. V.
Glass, A Program for the Analysis of Certain Time-Series Quasi-Experiments, 27 EDUCA-
TIONALAND PSYCHOLOGICAL MEASUREMENT 743-50 (1967); G. V. Glass, G. C. Tiao, &
T. O. Maguire, Analysis of Data on the 1900 Revision of German Divorce Laws as a
Time-Series Quasi-Experiment, 3 L. & SOC'YREV. (1969) (in press).
11. G. E. P. Box & G. C. Tiao, A Change in Level of a Non-stationary Time Series,
52 BIOMETRIKA 181-92 (1965); G. E. P. Box, Bayesian Approaches to Some Bothersome
Problems in Data Analysis in IMPROVING EXPERIMENTALDESIGNANDSTATISTICAL ANALY-
SIS (J. C. Stanley ed. 1967).

? 47 ?
 I

LAW AND SOCIaY REVIEW

SupplementaryAnalyses
In this section, we will present data that will further illustrate time-
series analysis and, substantively,both indicate that the crackdownwas
put into effect and that it had some unanticipatedand, to the policy-
makers,probably undesired consequences.
Figure 5 presents evidence that the crackdownwas put into effect,
as indicated by a great increase in suspensionsof licenses for speeding.
Unfortunately,we have not been able to get control state data for this
and the following variables,but the single state time-seriesis quite con-
vincing in itself. We regard it as confirmingthe appropriatenessof the
statisticaltests that they indicate significantdifferences. The single-point-

33

S0 L.
II
27 I
I.- I
z
24 w II
2
I
w I
21 II
I

18

Is

12

6 I I I I I I I I
'51 '52 *'3 '54 '55 '56 '57 '58 '59

Figure 5. Suspensionsof Licensesfor Speeding, as a Per Cent

of All Suspensions

.48 .
 THE CoNNEcrIcuT CRACKDOWNON SPEEDING

extrapolation t is 4.33 with 4 degrees of freedom, where 3.75 is signifi-

cant at the P<.02 level.
Figure 6 plots the percentage which speeding violations constitute
of all traffic violations. This shows a decline, due presumably to greater
conformity to speed limits, although it is possible that policemen and
prosecutors were more willing, in the light of severe sanctions for speed-
ing, to overlook minor infractions or to charge them as something else.
While the graphic portrayal of declining speeding violations is convincing
of a genuine effect, the statistical tests are not so emphatic. The single-
point-extrapolation t is 2.66 with 4 degrees of freedom, not reaching the
P<.05 level of 2.78.

19 1
18

IT

12 w

16 a-

12

10 -
J-
T I I I I I I I I I

'51 '52 '53 '54 '55 '56 '57 '58 '59

Figure 6. Speeding Violations, as Per Cent of All Traffic Violations

From Figure 5 and the reports cited in the first section of this paper
it is clear that a real change in enforcement behavior resulted. It seems
likely that the proportion of drivers exceeding the speed limits on Con-
I 49 ?
LAW AND SocIETYR;vmi;

necticut highways actually decreased. However, over and above these

desired effects there are signs of unforeseen and unwanted reactions.
Figure 7 concerns persons whose licenses were further suspended be-
cause they were convicted of drivingwith a suspendedlicense, expressed
as a percentageof all suspensions. This jumpsfrom an almost consistent
zero to some 4 to 6 per cent. Tests of significanceconfirmthe effect. The
single-point-extrapolationt reaches an incredible 130.75, due to the
very small errorterm which the negligible variance of the pretest scores
produces. (While one feels uneasy with a practicallyzero variance,the
consistent pretest zero does genuinely make the later values unlikely.)

7-

6 -z

4 I

3 -
/

I
I
2_ I
I

I I I I I I I I I

'51 '52 '53 '54 '55 '56 '57 '58 '59

Figure 7. Arrestedwhile Drivingwith a Suspended License,as

Per Cent of Suspensions

. 50
 THE CONNECICUT CRACKDOWNON SPEEDING

Our interpretationof this phenomenon is that automobile transporta-

tion has become a virtual necessity for many residents of the diffusely
settled megalopolitanregion that includes Connecticut,and these people
are willing to risk very severe sanctions in order to continue daily rou-
tines that involve driving. Since they are willing to drive with a sus-
pended license, suspension does not have the desired restrictive effect
on this group of drivers,which is probablymuch larger than the number
apprehendedand appearingin these statistics would indicate. Alterna-
tively, of course, the increase could result, in whole or part, from more
vigorouseffortsat enforcementboth in the crackdownitself and in special
efforts at inspection comprisinga followup of the crackdowneffort.
Figure 8 shows a reactionon the part of the legal system. Even with
fewer speeding violationsreachingthe courts (Figure 6), the courtswere
more lenient in their handling of these cases as expressedby the propor-

14

12
II
I

I0

I-
4 2

t-
2 I-

I I I I I I I I I

'61 '52 '53 '54 '55 '56 '57 '58 '59

Figure 8. Per Cent of Speeding ViolationsJudged Not Guilty

* 51 ?
LAW AND SOCIETY
REVIE

tion of not guilty decisions. Tests of significance are borderline. The

single-point-extrapolationt is 2.42, which with but 4 degrees of freedom
fails to reach significanceat the P<.05 level, for which 2.78 would be
required. Larger proportionsof not guilty judgments could be the re-
sult of more cases getting to court because of tightening of precourt
standards,more generous handling by judges and prosecutors,or more
vigorous defenses by the accused because more is at stake. The two
effects shown in Figures 7 and 8 indicate a vitiation of the punitive ef-
fects of the crackdownin operation in a society where dependence on
automobile transportationis acknowledged.

CONCLUSION
On the substantive side, the analysis has demonstrated that the
Connecticutcrackdownon speeding was a substantialenforcementeffort,
although some of its most punitive aspects were mitigated in practice.
As to fatalities, we find a sustained trend toward reduction,but no un-
equivocal proof that they were due to the crackdown. The likelihood
that the very high prior rate instigated the crackdown seriously com-
plicates the inference.
We have, however,learned somethingabout the response of the legal
system to a reform bearing a harsh penal sanction. The courts, and
probably also the police, are apparentlyunwilling to invoke penalties
that might seem severe and unfamiliarin context. Moreover,the force
of such penalties as are inflicted is vitiated by the willingness of the
public to evade them. As in the case of white-collarcrime, the effective
punishmentvaries with the criminal.12
More important,we believe, than the specific findings of the study
is the methodology here explored. While the social scientist cannot as
a rule experimenton a societal scale, societal "experimentation"
or abrupt
focused social change is continuallygoing on, initiated by government,
business, natural forces, etc. The social scientist adds to his tools for
understandingthe social system when he attends to these events and
documents their effects in as thorough a fashion as is possible. Insofar
as correlationalapproaches differ from experimental analysis, it adds
depth to the social scientist's work when he examines the fit of an
experimentalinterpretationwith full attention to the uncontrolledcom-
peting hypotheses.
12. The classic reference is E. H. SUTHERLAND, WHITE COLLARCRIME (1959).
See also H. L. Ross, Traffic Law Violation: A Folk Crime, 8 SOCIALPROBLEMS
231-41
(1961).

52
 THE CONNECrICUTCRACKDOWNON SPEEDING

The methodologyfor such quasi-experimentalanalysishas a long but

unsystematichistory, and offers much room for development. It should
be rememberedthat not only are the raw materialsshaped by the tools,
but in the long run the tools are shaped by the materialsupon which they
work. We should not passivelyaccept a methodologyas a revealed truth,
but rather should test it in use with our materials. Methodologyhas in
fact an empiricalhistory and its constituentshave the status of empirical
discoveries. The classical control group experimentis not typical of the
physical sciences, but instead emerged from psychologicallaboratoryre-
search,and is peculiar to the social sciences and their problems.l3 Medi-
cal research has the placebo control group, and neurophysiologythe
sham operation control, as achievementsof specific research traditions,
not as logical dispensationsfrom the philosophy of science or mathe-
matical statistics. So too the methods for quasi-experimentationin set-
tings like the present will emerge from an iteration of effort and criti-
cism, in which many approacheswill be rejected.
A final note on the treatment of uncontrolledvariables is in order.
On the one extreme there is that attitude often unwittingly inculcated
in courses on experimentaldesign, which looks askance at all efforts to
make inferences where some variables have been left uncontrolled or
where randomizationhas not taken place. In contrast,the quasi-experi-
mental approach takes a radically different posture: any experimentis
valid until proven invalid. The only invalidation comes from plausible
rival explanationsof the specific outcome. Regression effects and test-
retest effects are such in many settings. An absence of randomization
may in some specific way plausibly explain the obtained results. But
unless one can specify such a hypothesis and the direction of its effects,
it should not be regarded as invalidating. Subsequentconsiderationmay
uncoverplausible rival hypotheseswhich have been overlooked,but such
transitoryvalidity is often the fate of laboratoryexperimentstoo.
At the other extreme is the naive attributionof cause which blithely
fails to consider any explanationsother than the author'sfavorite can-
didate. Such an orientationis likewise opposed. The quasi-experimen-
talist is obliged to search out and consider the available plausible rival
hypotheses with all the vigilance at his command. While our coverage
in this regard has been incomplete,we hope that we have at least illus-
trated such an approach.

13. E. G. Boring, The Nature and History of Experimental Control, 67 AM. J.

PSYCHOLOGY 573-89 (1954).

.53 .