alcohol consumption and fatal accidents in canada, 1950–98

RESEARCH REPORT

© 2003 Society for the Study of Addiction to Alcohol and Other Drugs

Addiction,

98

, 883–893

Blackwell Science, Ltd

Oxford, UK

ADDAddiction

1360-0443© 2003 Society for the Study of Addiction to Alcohol and Other Drugs

98

883893

Original Article

Alcohol and fatal accidents in CanadaOle-Jørgen Skog

Correspondence to:

Ole-Jørgen SkogCentre for Advanced StudyThe Norwegian Academy of Science and

LettersDrammensveien 100271 OsloNorwayE-mail: [email protected]

Submitted 27 August 2002;initial review completed 16 December 2002;

final version accepted 3 February 2003

RESEARCH REPORT

Alcohol consumption and fatal accidents in Canada, 1950–98

Ole-Jørgen Skog

Centre for Advanced Study, The Norwegian Academy of Science and Letters, and Department of Sociology, University of Oslo, Norway

ABSTRACT

Aims

To evaluate the effects of changes in aggregate alcohol consumption onoverall fatal accidents, motor vehicle accidents, fatal falling accidents anddrowning accidents in Canadian provinces after 1950.

Design

Time-series analysis of annual mortality rates (15–69 years) in relationto per capita alcohol consumption, utilising the Box–Jenkins technique. Allseries were differenced to remove long-term trends.

Measurements

Gender-specific and age-adjusted mortality rates for the agegroup 15–69 years were calculated on the basis of mortality data for 5-year agegroups, using a standard population. Data on per capita alcohol consumptionwas converted to consumption per inhabitant 15 years and older. In the anal-ysis of motor vehicle accidents, the number of motor vehicles was used as a con-trol variable.

Findings

Statistically significant associations between alcohol consumptionand overall fatal accident rates were uncovered in all provinces for males, and inall provinces except Ontario for females. For Canada at large, an increase in percapita alcohol consumption of 1 litre was accompanied by an increase in acci-dent mortality of 5.9 among males and 1.9 among females per 100 000 inhab-itants. Among males there was a significant association with alcohol for bothfalling accidents, motor vehicle accident and other accidents, but the associa-tion was insignificant for drowning accidents. Among females, the associationwith falling accidents and other accidents was significant.

Conclusion

Changes in alcohol consumption have had substantial effects onmost of the main types of fatal accidents in Canada during the second half of the20th century. The size of the association is comparable to the one previouslyreported from Northern Europe.

KEYWORDS

Alcohol consumption,

fatal accidents,

mortality,

time series

analysis,

traffic accidents,

violent deaths.

INTRODUCTION

A large body of literature demonstrates that alcohol playsan important role in relation to both fatal and non-fatalaccidents in many countries (for a review, cf. Brismar &Bergman 1998). Most of the evidence comes from indi-vidual-level studies where the individual risk for acci-dents is seen in relation to the individual’s alcohol intakeor blood alcohol concentration (e.g. Andreasson

et al

.

1988; Cherpitel, Parés & Rhodes 1993; McLeod

et al

.1999; Theobald

et al

. 2001). Traffic accidents are clearlythe most heavily researched area, and alcohol seems to bea causal factor of considerable importance (e.g. BritishMedical Association 1988; National Highway TrafficSafety Administration 1989). Furthermore, in manycountries, alcohol also plays a significant role in acciden-tal falls (Hingson & Howland 1987), in accidents causedby fire, probably due to correlation with smoking


Addiction,

98

, 883–893

884

Ole-Jørgen Skog

(Glucksman 1994; Leth, Gregersen & Sabroe 1998) andin accidental drowning (Lunetta, Penttila & Sarna 1998).Some studies also suggest that the presence of alcoholmay be associated with greater severity of injury (Holt

et al

. 1980; Evans & Frick 1993), although this findingremains controversial (Li

et al

. 1997).The number of fatal accidents per 100 000 population

varies considerably within each population over time, aswell as between regions and countries. These aggregatelevel changes and differences need to be explained. Whyare accident mortality rates increasing in certain periodsand decreasing in other periods, and why do some coun-tries or regions experience higher rates than others? Thefull explanation for these changes and differences is likelyto be complex and many-faceted, but given the results ofthe above-mentioned individual-level studies, one wouldexpect that alcohol consumption is part of the explana-tion. For instance, many European countries experiencedincreasing accident rates during the 1960s, while rateswere decreasing in the 1980s. During the same periodlevels of alcohol consumption in these countries werechanging, and the latter changes seem to have been partof the causal web that produced the observed changes inaccident rates (see below).

However, there are far fewer empirical studies of theaggregate level relationships, i.e. of the determinants ofthe overall number of accidents in society. This is some-what surprising, because this is an issue of considerablepublic health interest. Furthermore, individual level stud-ies will generally not give a sufficient basis for answeringquestions such as ‘Why do overall rates of fatal accidentincrease or decrease over time, and what are the effects ofchanges in specific risk factors such as alcohol consump-tion on the overall volume of such accidents?’ In order toanswer these questions, aggregate level analyses arerequired.

Although it seems likely that changes in the consump-tion of alcoholic beverages may have a bearing on theaggregate number of accidents in society, the relationshipis probably not a simple and straightforward one. Whenaggregate alcohol consumption increases in a popula-tion, the reason could be that people drink more fre-quently or that they drink more per drinking occasion. Inboth cases we could expect,

ceteris paribus

, an increase inthe number of high-risk occasions due to intoxication.Under these circumstances, increased aggregate con-sumption would be expected to bring about an increase inthe aggregate number of accidents, provided everythingelse remains unchanged.

However, in principle one can also imagine that peoplestart to drink more often, but drink less per occasion. Thiswould imply a qualitative transformation of the drinkingculture, where drinking to reach high levels of intoxica-tion gradually disappears, and is replaced with frequent

consumption of smaller quantities. If this should occur,the number of high-risk occasions might actuallydecrease, even though the overall level of consumptionincreases, and we should not expect increased aggregateconsumption to be accompanied by more accidents.

The latter mechanism may play a role in some cross-cultural comparisons. The traditional wine countries inSouthern Europe typically have higher per capita con-sumption of alcoholic beverages than the drinking cul-tures of Northern Europe and North America. At the sametime, drunkenness may be more common in the lattercountries. Hence, cross-cultural correlations between percapita consumption and aggregate rates of accidents andother acute problem due to alcohol may be weak or evennegative in cross-cultural comparisons (cf. Mäkelä 1978).

When we study change over time, however, the formerpattern may be more common. In fact, not many exam-ples are known where increased per capita consumptionof alcoholic beverages has been accompanied by decliningrates of drunkenness. This is due probably to cultural iner-tia with respect to drinking patterns. Traditions of drunk-enness probably tend to change very slowly, while thefrequency of drinking tends to change much more rapidly,e.g. in response to changes in access, prices, income, etc.(cf. Edwards

et al

. 1994). Therefore, it appears much morelikely that one could expect to find a positive relationshipbetween per capita consumption and aggregate levels offatal accidents in studies of change.

Recent analyses of the European experience afterWorld War II has demonstrated that changes in per capitaalcohol consumption is in fact related to changes in acci-dent mortality rates (Skog 2001a). Increasing alcoholconsumption was accompanied by increasing accidentmortality rates, both in Northern, Central and SouthernEurope. However, a clear north–south gradient was alsouncovered: an increase in per capita consumption of 1litre was associated with a larger increase in accidentmortality rates in Northern Europe than in SouthernEurope, with Central Europe at an intermediate level. Thisis according to expectations, given the qualitative differ-ences between these drinking cultures. Furthermore, itwas demonstrated that the association between aggre-gate alcohol consumption and rates of fatal accidents isdue mainly to traffic accidents in Central and SouthernEurope, and to falls and other accidents in NorthernEurope (Skog 2001b).

The present work is an attempt to expand the empiri-cal basis for this relationship, and to see where Canada fitsinto this pattern. Have changes in the overall level of alco-hol consumption been accompanied by parallel changesin accident mortality during the period 1950–98, and ifso, how large are these changes? Is the strength of theassociation in Canada comparable to Northern, Centralor Southern Europe?

Alcohol and fatal accidents in Canada

885


Addiction,

98

, 883–893

The research task of the present paper is thus to inves-tigate if changes in per capita alcohol consumption areassociated with changes in accident mortality rates inCanadian provinces during the second half of the 20thcentury, and to compare the size of this association withresults from Europe. As accidents among the very youngand very old are seldom alcohol-related, the analysis willfocus on fatal accidents in the age range 15–69 years. Inaddition to overall accident mortality rates, we will alsoanalyse specific subgroups of accidents, namely fatal fall-ing accidents, drowning accidents and fatal trafficaccidents.

DATA AND METHODS

The statistical technique to be used requires that mortal-ity rates are not very small. Therefore, analyses wererestricted to the larger Canadian provinces, namelyBritish Columbia, Alberta, Saskatchewan, Manitoba,Ontario and Quebec. The Atlantic provinces were col-lapsed into one region (the Maritimes). The remainingregions are too small and were included only in the anal-yses of Canada at large.

Annual data (1950–98) on accidental deaths for allprovinces were obtained from Statistics Canada. All E-codes (external causes of injury and poisoning) wereincluded, with the exception of suicide and homicide. Thecategories E980–989, introduced in 1969 (ICD-8) andintended for use when it has not been determinedwhether the injuries are accidental, suicidal or homicidal,were included. The mortality figures were converted toage-specific (5-year brackets) mortality rates per 100 000inhabitants. Age-adjusted mortality rates for the agegroup 15–69 years were calculated using a standard pop-ulation (Waterhouse

et al

. 1976).Annual data on sales of alcoholic beverages per capita

was also obtained from Statistics Canada, and was con-verted to litres of pure alcohol per inhabitant 15 yearsand older per year. Both on-premise and off-premise salesare included. The statistics seem to be a combination ofretail sale and wholesale. Unfortunately, the fiscal yeardoes not coincide with the calendar year, and the fiscalyear runs from 1 April to 31 March. For the present anal-yses, the sales data have been coded so that sales in agiven year actually refer to sales 1 April

-

31 Decemberthe same year, plus 1 January

-

31 March the followingyear.

The fact that the mortality and the sales series refer todifferent parts of the year could represent a methodolog-ical problem. If the causal effect of alcohol on accidents iswithout delay, biased estimates of the effect of alcohol onmortality should be expected as a result of this mismatch.However, it is possible to calculate the approximate mag-

nitude of this error with the aid of a simple error-in-variable model (cf. Appendix for details). It appears thatone should expect a moderate deflation of the parameterestimate, and the effect of increasing alcohol sales onaccident mortality would in effect be slightly underesti-mated. If the differenced alcohol series contains no tem-poral structure (white noise), then the parameter wouldbe underestimated by 25%. However, lower alcohol salesduring the first quarter, compared to the remaining quar-ters, and positive autocorrelation in the (differenced)sales series would pull in the direction of reducing thiserror. The first assumption is reasonable on a priorigrounds. Moreover, it appears empirically that there arepositive autocorrelation left in the alcohol series after dif-ferencing. Therefore we could expect that the error is less,and probably considerably less, than 25%.

Not all alcohol is consumed during the same year itwas sold. This also produces a slight time-lag, and a mod-erate underestimation of the alcohol effect, particularlyin the case of wholesale data (Wagenaar 1985). However,this would be true even for other countries, and wouldnot affect comparisons.

Data on the total number of motor vehicle registra-tions were also obtained from Statistics Canada. Theyears up to 1975 were obtained from the historical sta-tistics, while the remaining years were obtained fromCansim II. The figures were converted to number of vehi-cles per population 15 years and older.

Both regional and temporal variation could beexploited in order to estimate the association betweenmortality and consumption. However, cross-sectionalcovariation is a rather shaky testing-ground for a causalhypothesis, because regional differences could have amultiplicity of causes which are difficult to control statis-tically. Temporal covariation is a better criterion, providedwe use techniques that allow us to test if a cause-event (achange in consumption) is followed by the anticipatedeffect-event (a change in mortality). This implies thatwhen testing causal hypotheses, one should not study thecovariation of trends, but the covariation of annualchanges (cf. Skog 1988).

Consequently, it was decided to use a conservativetechnique that avoids exploiting both regional covaria-tion and covariation of trends, and to rely only on cova-riation of annual changes. For each region the series wereanalysed with the technique suggested by Box & Jenkins(1976). The series were differenced in order to removenon-stationary trends. The noise terms (representingmeasurement error and causal factors besides thoseincluded in the equation) were modelled with the aid ofmoving average (MA) and autoregressive (AR) terms. Dif-ferencing often tends to induce a first-order MA term inthe residual, and a model with only a first-order MA termwas generally estimated as a first step. In most cases this


Addiction,

98

, 883–893

886

Ole-Jørgen Skog

turned out to offer an adequate fit. The fit was evaluatedwith the aid of the Box–Ljung portmanteau test of thefirst 10 autocorrelations, Q(10), and a visual inspectionof the residuals. In the few cases where the first-order MAwas insufficient, a model including both MA and ARterms was estimated (see below). The models are identi-fied as ARIMA(

x

,

y

,

z

) where

x

is the number of AR terms,

y

the number of differentiations and

z

the number MAterms.

The removal of trends from the data (differencing)reduces the risk of spurious correlation, but unfortu-nately at the price of increasing the standard errors of theparameter estimates. Hence, there is a risk of Type 2error, i.e. being unable to uncover relationships that areactually present. In order to reduce standard errors, theparameter estimates from individual provinces werepooled, in order to obtain a reliable parameter estimatefor Canada at large. The provinces were weighted accord-ing to population (

p

i

), and the standard error of the result-ing average was estimated according to the formula:

(1)

where

SE

(

i

) denotes the standard errors for each region.Furthermore, Canada at large was also analysed and theresult compared with the pooled estimate.

One could, perhaps, expect that annual changes inmortality rates are correlated across regions. If these cor-relations are large, they could be taken into considerationby applying a so-called ‘seemingly unrelated regression’technique, in order to obtain more efficient estimates(Kmenta 1990). In order to test this possibility, correla-tions between residuals in different regions were calcu-lated. The correlations turned out to be very small (forinstance, in the case of males, the average squared corre-lation was 0.07). Hence, the potential gain in statisticalefficiency would have been negligible.

During the period of observation there were multiplechanges in the classification of causes of death. Canadaimplemented ICD-7 in 1958, ICD-8 in 1969 and ICD-9 in1979. As these changes could produce artificial jumps inthe series, dummy variables (coded zero prior to thechange and one after the change) were used to control forthese changes. Furthermore, other special events alsorequired the use of dummy variables. There was a strikeat the alcohol monopoly in Alberta in 1980 and alcoholsales decreased substantially. It appears probable thatconsumption was not equally affected, and hence thatthe sales figure is a poor indicator of consumption in thisparticular year. Consequently, a dummy variable (codedone in 1980 and zero all other years) was inserted intothe model for Alberta. The sales statistics for all Canadaprior to 1952 do not include British Columbia and New-foundland. Hence, a dummy variable (coded one in 1950

SE p SE bi ipool i= Â ◊ ( )[ ]2 2ˆ

b̂

and 1951, and zero thereafter) was included in the anal-yses of all Canada. There is also a temporary drop in theaccident mortality rate in Alberta in 1985 and in Ontarioin 1990. According to official documents [Statistics Can-ada, Causes of Death Documentation. Supplementalinformation taken from Causes of Death (84–208) andMortality—Summary List of Causes (84–209)], this isdue to more deaths being classified to ‘other and unspec-ified and unknown causes’ and fewer deaths were classi-fied to ‘accidents, poisonings and violence’ in these twoinstances. Therefore, dummy variables were inserted intothe models for Ontario and Alberta for the respectiveyears.

Accidents due to alcohol (as opposed to somatic dis-eases) ought to occur shortly after the alcohol has beeningested. Hence, one should not expect a time lag in thepresent case. Furthermore, there are no theoretical rea-sons for expecting the relationship to be non-linear (alsoas opposed to somatic diseases). Therefore, a linear modelwas chosen for the relationship between alcohol con-sumption and mortality rates. As trends in mortalityrates were typically non-stationary (cf. Figure 1), a con-stant (a) was included in the model after differencing. Themodel to be estimated can therefore be written:

—

M

t

=

a

+

b

·

—

A

t

+

S

i

c

i

·

—

D

it

+

—

N

t

, (2)

where

M

denotes age adjusted mortality rates per100 000 inhabitants,

A

denotes per capita alcohol salesin litres of pure alcohol per inhabitants 15 years andolder,

D

i

denotes dummy variables and

N

the noise term.The operator

—

denotes differencing. In the analyses oftraffic accidents, the number of motor vehicles per inhab-itants 15 years and older were used as an additional con-trol variable (also differenced). The parameter

b

measuresthe effect of alcohol on mortality.

In one case, namely drowning accidents, mortalityrates declined by approximately 80% during the period ofobservation and the annual number of deaths towardsthe end of the period was fairly small. As a result, theresiduals turned out to be strongly heteroscedastic. Forthis reason, a logarithmic transformation of the mortalityseries was used.

Because males typically drink much more thanfemales, one would expect that in increase in per capitaalcohol consumption of 1 litre per year would have largereffects on males than females. Hence one should expectthe effect parameter

b

to be substantially larger for males.If males drink three to four times as much as females, thedifference between the parameter estimates should be ofthe same order of magnitude.

In addition to analysing overall accident rates, weshall also analyse a few specific subcategories where alco-hol is known to be an important factor. These includemotor vehicle accidents, falling accidents and drowning


887


Addiction,

98

, 883–893

Figure 1

Trends in per capita alcohol consumption 15 years and older (—) and age-adjusted male (---) and female (–

-

–) overall accidentmortality (15–69 years) in Canadian provinces and all Canada 1950–98. Logarithmic scale

200

1008060

40

20

1086

50 55 60 65 70 75 80 85 90 95

Alberta200

1008060

40

20

108

50 55 60 65 70 75 80 85 90 95

Quebec

6

200

1008060

40

20

108

50 55 60 65 70 75 80 85 90 95

Ontario

6

200

1008060

40

20

108

50 55 60 65 70 75 80 85 90 95

British Columbia

6

200

1008060

40

20

108

50 55 60 65 70 75 80 85 90 95

Manitoba

6

200

1008060

40

20

108

50 55 60 65 70 75 80 85 90 95

Maritimes

6

4

200

10080

60

40

20

108

50 55 60 65 70 75 80 85 90 95

Saskatchewan

6

4

200

1008060

40

20

108

50 55 60 65 70 75 80 85 90 95

Canada

6


Addiction,

98

, 883–893

888

Ole-Jørgen Skog

accidents. Motor vehicle accidents constitute approxi-mate one-half of all fatal accidents in Canada in thisperiod. Fatal falling accidents account for 8–9%, whiledrowning accidents constitute 7% among males and only3% among females. The remaining accidents were alsoanalysed. This residual category also includes causeswhere alcohol plays an important role, such as fire (cf.Glucksman 1994; Leth

et al.

1998), and poisoning (cf.Poikolainen 1977; Mäkelä 1998).

RESULTS

Trends in per capita alcohol sales and total accident mor-tality rates in Canada as a whole and in the largest prov-inces are shown in Fig. 1. Male mortality rates were fairlystable in Canada during the period 1950–70, whilefemale mortality rates were increasing. However, duringthe last third of the 20th century, mortality rates havebeen decreasing at a fairly steady rate for both genders.The pattern is similar in most of the provinces. On theother hand, alcohol sales and consequently alcohol con-sumption was increasing until the middle of the 1970s.From 1975 a decreasing trend in consumption is foundin all provinces, followed by a slight increase towards theend of the century.

Results of the time-series analysis of the annualchanges in overall accidents are reproduced in Table 1.For males, a significant relationship between alcohol con-sumption and total accident mortality is found in allprovinces, as well as for Canada as a whole. For females,there are significant relationships in all provinces exceptOntario. The (weighted) average parameter indicates thatan increase in alcohol consumption of 1 litre pure alcoholwas accompanied by an increase in accident mortality

per 100 000 inhabitants of 5.9 among males and1.9 among females. The parameter estimates aresomewhat larger than average in British Columbia andSaskatchewan, and smaller than average in Ontario forboth males and females. These regional differences areinsignificant among males, but not among females. Theanalysis of Canada at large produced parameter estimatesthat are close to the average values, but with considerablylarger standard errors (as expected). The male parameterestimate is approximately three times as large as thefemale parameter estimate, as expected.

Trends in drowning accidents, motor vehicle acci-dents, falling accidents and other types of accidents forCanada at large are shown in Fig. 2. Because the absolutenumber of accidents becomes fairly small in some of thesecategories, only Canada at large was analysed. As pointedout earlier, there has been a substantial and steadydecline in drowning accidents during the whole period ofobservation. Even falling accidents have been decliningduring the whole period, but most strongly after 1975.On the other hand, fatal motor vehicle accidentsincreased until the early 1970s and have been decreasingafter that time. The number of motor vehicles registeredhas been increasing during the whole period, but at aslower rate in recent years. The remaining fatal accidentshave been decreasing during the last quarter of the cen-tury for both genders. From 1950 to the early 1970s,female mortality was increasing, while males experi-enced a decrease, followed by an increasing trend in the1960s (Fig. 2).

As mentioned above, drowning accidents required alogarithmic transformation due to heteroscedasticity. Agood fit was obtained with an ARIMA(1,1,1) model(Q10

=

7.9). The parameter estimate for males was equalto 0.050 (SE

=

0.031), which is not significant (

P

=

0.11).

Table 1

Estimated effect of alcohol consumption on all accidents (age-adjusted mortality rates, 15–69 years), controlled for ICD-7, -8 and-9. All models are ARIMA(0,1,1), except in the case of females in Saskatchewan, where the model is ARIMA(1,1,1).

Males Females

Estimate SE Q(10) Estimate SE Q(10)

British Colombia 11.60** 3.00 13.0 4.34** 0.96 14.8Alberta

a

5.33** 1.82 6.6 1.62** 0.42 3.9Manitoba 5.95** 1.36 8.8 2.28* 1.06 4.1Saskatchewan 8.40** 1.51 5.9 4.15** 0.80 16.7Ontario

b

3.88* 1.83 10.1 0.66 NS 0.65 7.7Quebec 5.92* 2.40 10.7 2.28* 0.94 12.3Maritimes 6.72* 2.69 5.6 2.22** 0.29 7.2F-test 1.43 NS – – 2.87** – –Weighted average 5.94** 1.02 – 1.94** 0.37 –All Canada

c

7.01** 2.05 9.5 2.01** 0.66 6.9

a

Dummy variables for 1980 and 1985;

b

dummy variable for 1990;

c

dummy variable for 1950–51.**

P

<

0.01; *

P

<

0.05; NS

P

>

0.05.


889


Addiction,

98

, 883–893

It should be noted that this estimate has a different metricthan the parameters obtained from linear models. Wecan, however, convert this estimate into the linear metric.The mean rate of drowning accidents was 4.96 per100 000 inhabitants during this period, and at this levelthe estimate obtained corresponds to a parameter in lin-ear metric of 0.25 deaths per 100 000 inhabitants. Forfemales the estimated relationship was even smaller andalso not significant.

In the case of fatal falling accidents (as well as theother categories reported below) no transformation wasnecessary. An ARIMA(1,1,1) model gave a good fit(Q10

=

14.9) for males. The parameter estimate wassignificant at the 5% level, and was equal to 0.31(SE

=

0.11). For females the same model resulted in aparameter estimate of 0.14 (SE

=

0.03), which is signifi-cant at the 1% level (Q10

=

3.3).Regarding the relationship between alcohol and fatal

motor vehicle accidents, traffic density is a potential con-founder. Annual changes in the number of registeredmotor vehicles are correlated with annual changes inalcohol consumption (

r

=

0.45), presumably because

both variables are influenced by economic factors. Con-sequently, the number of registered motor vehicles wasused as a control variable in this analysis. For alcoholconsumption we obtained the parameter estimate 3.61(SE

=

1.74) for males, which is significant at the 5% level.The model was ARIMA(0,1,1) and Q(10)

=

8.2. Thisparameter estimate is only slightly less than the estimateobtained without the control variable (4.08). In the caseof females, the parameter estimate was small (0.51) andinsignificant.

For the remaining fatal accidents an ARIMA(1,1,1)gave a good fit for males (Q10

=

7.9). The parameter esti-mate was 1.95 (SE

=

0.61), which is significant at 1%level. For females the same model (Q10

=

5.9) resulted inthe parameter estimate 0.98 (SE

=

0.26), significant atthe 1% level.

Hence, the analyses of subgroups of fatal accidentsdemonstrate that an increase in per capita alcohol con-sumption of 1 litre is associated with an increase in themale mortality rate per 100 000 population of 0.25 dueto drowning (insignificant), of 0.31 due to falling acci-dents, of 3.61 due to motor vehicle accidents and 1.95

Figure 2

Trends in per capita alcohol consumption 15 years and older (—) and age-adjusted male (---) and female (–

-

–) mortality (15–69 years) from drowning, falling accidents, motor vehicle accidents and other accidents in Canada 1950–98. In the motor vehicle accident dia-gram, the number of registered motor vehicles per population 15 years and older (— - — - —) is also shown. Logarithmic scale

20

1086

4

1

50 55 60 65 70 75 80 85 90 95

0.80.60.4

0.2

2

Drowning50

3020

10

5

1

50 55 60 65 70 75 80 85 90 95

0.50.40.30.2

432

40

Motor vehicle accidents

20

1086

4

1

50 55 60 65 70 75 80 85 90 95

0.8

2

Accidental falls50

5

50 55 60 65 70 75 80 85 90 95

6789

10

4

40

30

20

Other accidents


Addiction,

98

, 883–893

890

Ole-Jørgen Skog

due to other types of accidents. In sum this is 6.1, whichis reasonably close to the estimate obtained in the analy-ses of all accidents.

DISCUSSION

The results of the analyses offer a fairly coherent picture,well in line with previous results from Europe (Skog2001a, 2001b). Significant associations betweenchanges in alcohol consumption and changes in accidentmortality have been found in all Canadian provinces andfor both genders, with the exception of females in Ontario.The parameter estimates are approximately three times aslarge among males as among females, which is accordingto expectations, given the known differences betweenmale and female drinking. The numerical size of the asso-ciation—5.9 deaths per 100 000 inhabitants per litre ofpure alcohol for males—is similar to that found in North-ern Europe (5.2). Furthermore, the effects of alcohol con-sumption could be observed for three types of accidents(falling accidents, motor vehicle accidents and other acci-dents), and the sum of these partial effects correspondswell with the estimated overall effect.

As noted previously, there is a mismatch between mor-tality statistics and alcohol sales statistics, as the latterfollows the fiscal year. We also noted that if the truecausal effect of alcohol is instantaneous, the result couldbe underestimation of the effect of alcohol. Nevertheless,the parameter estimates are of the same magnitude as inNorthern Europe and even slightly higher. However, thestandard errors of these estimates are large and the dif-ference is not significant. In fact, even if we assume thatthe Canadian parameter is underestimated by as much as25%, the corrected estimate would not be significantlydifferent from the one obtained for Northern Europe.

In Europe, the association with alcohol was quitestrong in Central and Southern Europe with respect totraffic accidents, but weaker for other types of accidents.In Northen Europe it was the other way around. Theexplanation for the weaker association between trafficaccidents and per capita alcohol consumption in North-ern Europe could be stronger compliance with nationalblood alcohol concentration (BAC) laws (cf. Snortum,Hauge & Berger 1986). The Canadian data seem to sug-gest that alcohol is quite strongly associated with bothtypes of accidents. This might suggest that Canadiandrinking patterns are closer to the ones found in North-ern Europe, while at the same time Canada does not profitfrom the same level of compliance with BAC laws. If this iscorrect, Canada may be combining ‘the worst of twoworlds’ with respect to alcohol related accidents. Itshould be added that this suggestion is speculative, andneeds independent evidence.

In Europe the association between alcohol and acci-dents varied considerably between countries, due pre-sumably to differences in drinking patterns. There aredifferences with respect to drinking patterns betweenCanadian provinces as well, although these differencesare probably smaller. Surveys suggest that high-intakedrinking occasions are more prevalent in the Atlanticregion (the Maritimes) and that light frequent drinking isless common in this region. Ontario and Quebec have asomewhat lower rate of high-intake drinking occasionsthan the prairie provinces (Kellner 1997). Furthermore,a comparative study of emergency room patients inAlberta and Quebec (Cherpitel

et al

. 1999) found ahigher alcohol involvement in accidents in Alberta. Inlight of these differences, one could perhaps haveexpected regional differences in the parameter estimatesin Canada. However, for males these differences are notsignificant. As can be seen from Table 1, the standarderrors of individual parameter estimates are large andthis makes it difficult to unravel any coherent pattern. Forinstance, the exceptionally large parameter estimate forBritish Colombia may very well be due to estimationerror, and the difference between the latter estimate andthe one obtained for Manitoba (which is close to the aver-age) is not significant; nor is the difference betweenOntario and Manitoba significant. For females, theregional differences are significant. However, no clear-cutand easily interpretable pattern has been uncovered andone should probably avoid drawing conclusions aboutparticular regions from these results.

Although the present study was motivated by causalhypotheses, and even though the results are in accor-dance with the hypotheses, the empirical evidence doesnot prove that these hypotheses are correct. As always inthis type of study, causality is imputed, not proved. Thecovariation observed could, in principle at least, be spu-rious. This would occur if annual changes in other causesfor accident rates were closely correlated with the annualchanges in alcohol consumption. In the case of fatal traf-fic accidents, traffic density is a potential confounderbecause alcohol consumption and traffic density are cor-related. However, as we have seen, controlling for thenumber cars did not change the result to any significantextent.

In relation the other types of accidents, it is not easy tofind obvious control variables. It should be stressed thatrelevant controls would be variables that

both

have anindependent causal effect on mortality and are correlatedclosely with changes in alcohol consumption on a year-to-year basis. Causal factors that are not correlated withalcohol consumption will

not

produce biased estimates ofthe effect of alcohol.

In the absence of obvious control variables for the lat-ter types of accidents, one could contemplate using eco-


891


Addiction,

98

, 883–893

nomic indicators as control variables. However, in thepresent case this would not be a sound strategy. What weneed to control is other causal factors. Economic indica-tors are not direct causes of accidents but may affect acci-dents indirectly, by affecting factors that themselves havea direct causal effect. Consequently, economic indicatorsare simply proxies for actual causal mechanism (e.g.changes in life-style, implemented accident preventionmeasures, etc.). It may not be a good proxy, however, asthe effects of economic development on the causal factorsin question could easily have a strongly distributed timelag. Hence, the instantaneous correlation between eco-nomic factors and the causal factors they are supposed torepresent could be modest. The main problem, however,is that the same economic indicators (for instanceincome) also tend to have a strong positive effect on alco-hol consumption (Bruun

et al

. 1975). Hence, one wouldcontrol things that should not be controlled. A regressionmodel including both alcohol consumption and eco-nomic variables would estimate the effects of changes inalcohol consumption on accident rates, given that theeconomic indicators remain fixed. However, this is notwhat we wish to know. We wish to estimate the effect of achange in alcohol consumption, whether the reason forthe change in consumption is increased income or some-thing else. (It should be noted that the effect of a 1-litreincrease in consumption on accident rates is not neces-sarily the same when the increase is due to increasedincome and when it is due to something else.) Further-more, the standard error of the parameter estimate foralcohol consumption would become substantially largerif economic indicators were controlled, as one would befacing a serious collinearity problem. Insignificant esti-mates could easily be the result (Type 2 error).

Presumably, causal factors of importance in relationfalling accidents, drowning accidents, accidents causedby fire, intoxication, etc. would typically not increase anddecrease in exactly the same years as does alcohol con-sumption. If this is correct, failing to control for these fac-tors would not cause any bias in the alcohol effectparameter. Needless to say, the results should be inter-preted with this limitation in mind.

As pointed out earlier, accident mortality rates werestable among Canadian males until the middle of the1970s, and were decreasing after that time. This changein trends can now be given at least a partial explanation,provided that the observed association between alcoholconsumption and fatal accidents is in fact a causal rela-tionship. Until the middle of the 1970s, alcohol consump-tion was increasing and pulled in the direction ofincreasing accident mortality rates. This force apparentlywas counterbalanced by other forces, pulling in the oppo-site direction, towards reduced mortality (e.g. differentsorts of accident prevention measures). Consequently,

fatal accident rates did not change greatly. When alcoholconsumption started to decrease towards the end of the1970s both forces were pulling in the same direction, andaccident mortality rates started to decrease. The changein trend for accident mortality in the middle of the 1970smay thus be connected closely to the changes thatoccurred in alcohol consumption. However, other factorsmay certainly have also contributed to this change.

In conclusion, it appears that alcohol consumption isan important factor when we wish to explain changes inaccident rates over time, and presumably also acrossregions and countries. In this particular study we haveevaluated only the effects of changes in aggregate con-sumption levels. We have not had the possibility of eval-uating other important aspects of drinking patterns. Thisdoes not, of course, mean that drinking patterns areunimportant (cf. the previously mentioned North–Southgradient in Europe). Neither have we evaluated the effectsof preventive measures, e.g. legal and economic limita-tions on the access to alcoholic beverages for particulargroups or the population at large. It does not follow auto-matically from the results presented here that specificmeasures would have a desired effect on accident rates.This would require special studies where such measureshave been implemented. For instance, Adrian, Ferguson& Her (2001) have published results from Canada sug-gesting that prices of alcohol has a significant effect inreducing alcohol-related motor vehicle accidents. Morestudies of this sort are clearly needed in order to draw anyconclusions about the effectiveness of policy measures.

ACKNOWLEDGEMENTS

I am very grateful to Mats Ramstedt for assistance withthe data files. I also obtained helpful suggestions fromThor Norström, Mats Ramstedt, Robin Room, IngeborgRossow, and the reviewers and editors of

Addiction

.

REFERENCES

Adrian, M., Ferguson, B. S. & Her, M. (2001) Can alcohol pricepolicies be used to reduce drunk driving? Evidence from Can-ada.

Substance Use and Misuse

,

36

, 1923–1957.Andreasson, S., Allebeck, P. & Romelsjö, A. (1988) Alcohol and

mortality among young men, a longitudinal study of Swedishconscripts.

British Medical Journal

,

296

, 1021–1025.Box, G. E. P. & Jenkins, G. M. (1976)

Time-series Analysis:Forecasting and Control

, rev. edn. San Francisco, CA: Holden-Day.

Brismar, B. & Bergman, B. (1998) The significance of alcohol forviolence and accidents.

Alcoholism: Clinical and ExperimentalResearch

,

22

, 2995–3065.British Medical Association (1988)

The Drinking Driver.

Report ofthe Board for Science and Education. London: British MedicalAssociation.


Addiction,

98

, 883–893

892

Ole-Jørgen Skog

Bruun, K., Edwards, G., Lumio, M., Mäkelä, K., Pan, L., Popham,R. E., Room, R., Schmidt, W., Skog, O.-J., Sulkunen, P. &Österberg, E. (1975)

Alcohol Control Policies in Public HealthPerspective

. Helsinki: Finnish Foundation for Alcohol Studies.Cherpitel, C. J., Giesbrecht, N. & Macdonald, S. (1999) Alcohol

and injury: a comparison of ER populations in two Canadianprovinces.

American Journal of Alcohol and Drug Abuse

,

25

,743–759.

Cherpitel, C. J., Parés, A. & Rodes, J. (1993) Prediction of alcohol-related casualties in the emergency room: a US–Spain com-parison.

Journal of Studies on Alcohol

,

54

, 308–314.Edwards, G., Anderson, P., Babor, T., Casswell, S., Ferrence, R.,

Giesbrecht, N., Godfrey, C., Holder, H., Lemmens, P., Mäkelä,K., Midanik, L., Norström, T., Österberg, E., Romelsjö, A.,Room, R., Simpura, J. & Skog, O.-J. (1994)

Alcohol and the Pub-lic Good

. Oxford: Oxford University Press.Evans, L. & Frick, M. C. (1993) Alcohol effect on fatality risk

from a physical insult.


,

54

,441–449.

Glucksman, E. (1994) Alcohol and accidents.

British MedicalBulletin

,

50

, 76–84.Hingson, R. & Howland, J. (1987) Alcohol as a risk factor for

injury or death resulting from accidental falls: a review ofthe literature.


, 48, 212–219.Holt, S., Stewart, I. C., Dixon, J. M. J., Elton, R. A., Taylor,

T. V. & Little, K. (1980) Alcohol and the emergency ser-vice patient. British Medical Journal, 281, 638–640.

Kellner, F. (1997) Alcohol. In: MacNeil, P. & Webster, I., eds.Canada’s Alcohol and Other Drugs Survey 1994: a Discussion ofthe Findings, pp. 17–42. Ontario: Health Canada.

Kmenta, J. (1990) Elements of Econometrics. 2nd edn. New York:Macmillan.

Leth, P., Gregersen, M. & Sabroe, S. (1998) Fatal residential fireaccidents in the municipality of Copenhagen. Preventive Med-icine, 27, 444–451.

Li, G. H., Keyl, P. M., Smith, G. S. & Baker, S. P. (1997)Alcohol and injury severity: reappraisal of the continuingcontroversy. Journal of Trauma–Injury Infection and CriticalCare, 42, 562–569.

Lunetta, P., Penttila, A. & Sarna, S. (1998) Water traffic acci-dents, drowning and alcohol in Finland 1969–95. Interna-tional Journal of Epidemiology, 27, 1038–1043.

Mäkelä, K. (1978) Levels of consumption and social conse-quences of drinking. In: Israel, Y., Glaser, F. B., Kalant, H.,Popham, R. E., Schmidt, W. & Smart, R. G., eds. ResearchAdvances in Alcohol and Drug Problems, Vol. 4. New York: Ple-num Press, pp. 303–348.

Mäkelä, P. (1998) Alcohol-related mortality by age and sex andits impact on life expectancy. European Journal of Public Health,8, 43–51.

McLeod, R., Stockwell, T., Stevens, M. & Phillips, M. (1999)The relationship between alcohol consumption patternsand injury. Addiction, 94, 1719–1734.

National Highway Traffic Safety Administration (NHTSA)(1989) Alcohol Involvement in Fatal Traffic Crashes 1987.Washington, DC: NHTSA.

Poikolainen, K. (1977) Alcohol Poisoning Mortality in FourNordic Countries. Helsinki: Finnish Foundation for AlcoholStudies.

Skog, O.-J. (1988) Testing causal hypotheses about correlatedtrends: pitfalls and remedies. Contemporary Drug Problems, 15,565–607.

Skog, O.-J. (2001a) Alcohol consumption and overall accidentmortality in 14 European countries. Addiction, 96, S35–S47.

Skog, O.-J. (2001b) Alcohol consumption and mortality ratesfrom traffic accidents, accidental falls, and other accidents in14 European countries. Addiction, 96, S49–S58.

Snortum, J. R., Hauge, R. & Berger, D. E. (1986) Deterring alco-hol-impaired driving. Justice Quarterly, 3, 139–165.

Theobald, H., Johansson, S. E., Bygren, L. O. & Engfeldt, P.(2001) The effects of alcohol consumption on mortality andmorbidity: a 26-years follow-up study. Journal of Studies onAlcohol, 62, 783–789.

Wagenaar, A. C. (1985) Alcohol consumption and the incidenceof acute alcohol-related problems. British Journal of Addiction,80, 173–180.

Waterhouse, J., Correa, P., Muir, C. & Powell, J. (1976) CancerIncidence in Five Continents, Vol. III, p. 456. Lyon: IARC.

Alcohol and fatal accidents in Canada 893

© 2003 Society for the Study of Addiction to Alcohol and Other Drugs Addiction, 98, 883–893

APPENDIX

Let At,q and Dt,q denote, respectively (differenced) per cap-ita alcohol consumption and (differenced) mortality ratesin quarter q and year t, and let At. and Dt. denote the cor-responding annual figures, e.g.:

At. = At,1 + At,2 + At,3 + At,4.

Consider the causal model:

Dt. = a + b·At. + Nt..

where N denotes the noise term, which is assumed to beindependent of alcohol consumption A. The Canadianstatistics give, instead of At.:

A*t. = At,2 + At,3 + At,4 + At+1,1

= At. - At,1 + At+1,1.

By substituting the latter expression into the causalmodel, we obtain:

Dt. = a + b·A*t. + N*t.,

where N*t. = Nt. – b·(At+1,1 – At,1). This new equationexpresses the observed mortality rate in terms of theobserved consumption figure.

The estimation problem is now easily diagnosed: a biasis introduced as the noise term N* is no longer indepen-dent of the input variable A* as At+1,1 is part of both A* andN*. In order to evaluate this bias we look at the condi-tional expectation of D, given A*. We obtain:

E[Dt.|A*t.] = a + b·A*t. - b·K,

where K = E[At+1,1 – At,1|A*t.]. The correction term K willtypically be different from zero, and dependent on A*t.

(as A*t. includes At+1,1).Consider the case where At,q varies as white noise, and

assume without loss of generality that E[At,q] = 0. ThenA*t. is independent of At,1 and the correction term reducesto K = E[At+1,1|A*t.]. It follows from the standard theory ofindependent normal variates that K = ·A*t., and weobtain:

E[Dt.|A*t.] = a + b·A*t..

Hence, in this case we would obtain a parameter estimatethat is only -th of its actual value: i.e. an underestima-tion of 25%.

The white noise assumption implies that variations inconsumption are the same for all quarters. If consump-tion is typically smaller in the first quarter, compared tothe remaining quarters (which appears likely), the errorcould easily be smaller than 25%.

Next, consider the case where the (differenced) con-sumption series has a positive autocorrelation structure.This means that even At,1 would be correlated with A*t.,because At,1 would be correlated with At,2 which isincluded in A*t. The difference At+1,1 – At,1 measures thechange in consumption in the first quarter from one yearto the next. The more regular the (differenced) consump-tion series, i.e. the higher the autocorrelation, the smallerthis change will be. Hence, positive autocorrelation in the(differenced) consumption series would tend to reducethe bias.

14/

3 4/

3 4/

alcohol consumption and fatal accidents in canada, 1950–98

Documents