Trends in Self-Reporting of Marijuana Consumption in the United States

ABSTRACT To adjust for underreporting of marijuana use, researchers multiply the proportion of individuals who reported using marijuana by a constant factor, such as the US Office of National Drug Control Policy’s 1.3. Although the current adjustments are simple, they do not account for changes in reporting over time. This article presents a novel way to explore relative changes in reporting from one survey to another simply by using data already available in a self-reported survey, the National Survey on Drug Use and Health. Using domain estimation to examine the stability in reported marijuana use by age 25 in individuals older than 25, this analysis provides estimates of the trends in marijuana reporting and standard errors, as long as the survey weights properly account for sampling variability. There was no significant evidence of an upward or downward trend in reporting changes from 1979 to 2016 for all birth cohorts, although there were significant differences in reporting between years and a slight downward trend in later years. These results suggest that individuals have become increasingly less willing to report their drug use in recent years, and thus the ONDCP likely underestimated the already drastic increase in use from 1992 to 2016.


Introduction
Policy makers need to know how many people are using marijuana in the United States. Policy decisions, such as funding for public health programs and law enforcement, as well as how to legislate for the growing billion dollar marijuana industry (Gettman 2007), require properly estimating the prevalence of marijuana use. Self-reported surveys are the most common sources of information that researchers use to estimate how many individuals consume marijuana. However, when individuals respond to surveys about illicit drug use, they tend to underreport their consumption (Harrison and Hughes 1997).
Underreporting is common in self-reported surveys on topics as varied as rape and sexual assault to HIV/AIDS (Kruttschnitt, Kalsbeek, and House 2014), and it is particularly common in surveys about drug use (Harrison and Hughes 1997). A report by ONDCP on drug consumption in the United States, for instance, stated that users are likely to underreport socially disapproved behaviors, even when those behaviors are legal, and that "they would seem to have even more incentive to underreport illegal behaviors" (Kilmer et al. 2011). Various authors have estimated the percentage of respondents that were honest about their marijuana use in self-reports and have found values ranging from 60% to 87.7% (Harrison et al. 2007).
Estimates of the difference between the true and reported number of users have been translated into multiplicative adjustments that correct for underreporting and help researchers obtain a more accurate percentage of users. For instance, the adjustments for certain studies range from to 1.15 to 1.6 (Harrison and Hughes 1997;Harrison et al. 2007 2007; Kilmer et al. 2011). Some researchers compared the self-reported marijuana use of a sample of individuals to urine test results from those individuals, and determined that the adjustment should be 1.4 (Harrison and Hughes 1997). Others slightly increased the adjustments used to account for underreporting in legal drugs like alcohol and tobacco (Kilmer et al. 2011). Of course, by using the legal drug adjustments for illegal drugs, researchers must make an assumption about the similarity of underreporting for legal and illegal drugs, and this assumption is difficult to test. The selection of an adjustment factor matters because even slight changes in adjustments could drastically affect estimates of prevalence, which would in turn affect the estimates of the funding required to handle public health issues related to marijuana use and estimates of the size of the marijuana market.
Although the adjustments provided by researchers are simple and easy to use when correcting for underreporting, they do not allow for heterogeneities in the sample. This is problematic because, for instance, younger respondents have been found to overreport their marijuana use while older respondents tend to underreport their use (McAllister and Makkai 1991), African Americans are less likely to report their marijuana use than Whites, Hispanics, and Asians (Fendrich et al. 2004), and opiate users have been found to overreport their drug use (Zullino et al. 2008). The heterogeneities that might arise from changes in reporting over time could be significant as well.
If reporting trends are changing over time, then using the adjustment factors that have been derived at a specific point in time could lead to overestimating or underestimating the Figure . Fictitious example. If underreporting of marijuana decreases over time, using a constant adjustment for underreporting that was obtained in -which is  in this case-could lead to over or overestimating the proportion of individuals who used marijuana (by age ) in , by about %.
true value. Furthermore, if people are reporting differently over time, reports that "marijuana use in the United States has remained fundamentally unchanged in the last decade and a half" (Gettman 2007), could be invalid. Recent statements about marijuana use drastically increasing from 1992 until today, as the ONDCP reports (Kilmer et al. 2011), could also be invalid because it is possible that the increase in reported use does not reflect an increase in true use.
An example of why using a fixed number to adjust for underreporting could be misleading if there are heterogeneities in reporting over time is shown in Figure 1. Suppose the true number of individuals born in 1975 who used marijuana by age 25 is constant at 40, a number unknown to the researcher. Suppose that in 2000 the researcher correctly estimates an adjustment to underreporting of marijuana use of 2.0 by using known underreporting rates from other drugs, such as alcohol and tobacco. In 2000, 20 individuals report they used marijuana by age 25, so using the adjustment of 2.0, the researcher correctly adjusts the estimate to 40. Later in 2010, close to 30 individuals say they used marijuana by age 25, so, using the adjustment from 2000, the researcher adjusts the value to about 60. However, underreporting decreased in this example (i.e., people were more willing to report their use), so the true value was actually 20 lower than the adjusted value of 60. Thus, using a constant adjustment could greatly overestimate the true value of the proportion of individuals who use marijuana.
Over the past 25 years public opinion about whether marijuana should be legalized has changed dramatically, and thus it is likely that individuals are underreporting their marijuana use less over time. As shown in Figure 2, a General Social Survey showed that in 1990, only 16% of the U.S. population said they thought marijuana should be legal, and in 2016 this value had risen to 55%. Furthermore, Wish et al. (1997) showed that as drug use becomes less stigmatized, individuals may become more willing to disclose past use.
The objective of this study is to quantify the time trends in reporting of marijuana consumption in the United States. The question of interest is, Are people changing the way in which they report their marijuana use in surveys over time? If so, by how much? To study this question, the National Survey on Figure . General Social Survey responses to whether individuals think marijuana should be legalized with % confidence intervals. An increase in support for marijuana legalization suggests a decrease in the stigma for admitting marijuana use. (The data were acquired from the GSS data explorer website online at https://gssdataexplorer.norc.org/trends/Civil%Liberties?measure=grass. Last accessed April , .) Drug Use and Health, a cross-sectional household survey of the population of the United States sponsored by the Department of Health and Human Services, is analyzed over the years 1979 to 2016. Instead of focusing on past-month or past-year use, this analysis studies the stability in reported ever-use by age 25, since this quantity should stay constant over time for an individual older than 25. Retrospective questions are asked in each survey, and thus cohort-specific responses about marijuana use can be compared year to year. If there were no changes in reporting, these responses should remain constant, and if reporting were indeed increasing, then the proportion of reported marijuana use should increase, independently of a true increase in use, provided one accounts for changes in the population such as attrition due to death, immigration, and emigration. To clarify, this analysis does not estimate the measurement error, that is, the difference between the true value and the reported value of use, nor does it propose using a specific factor over others to adjust for underreporting. Instead, it provides updates to currently used adjustments to correct for changes in reporting over time.

Data
The National Survey on Drug Use and Health (NSDUH), formerly the National Household Survey on Drug Abuse, is an ongoing nationwide survey that was started in 1971 by the Public Health Service of the U.S. Department of Health and Human Services. It was created to measure the use of illicit drugs, alcohol, and tobacco within the general population of Americans aged 12 and above (NSDUH 2017). Today, NSDUH is the only study that annually produces estimates of drug use among civilian members of the noninstitutionalized population in the United States (NSDUH 2017). The institutionalized population, such as the incarcerated and the homeless, is not included in the survey, and thus it is not included in this analysis. Between 1979 and 2016 the survey was performed almost every year, with a total of 27 years of available data. Data prior to 1979 is not available on the NSDUH website. NSDUH has Figure . Simplified example of the analysis. Suppose in , when the four individuals born in  turn , one of them reports having used by age . In , this number is two, in  it is three, and so on. Since the individuals are older than  throughout the survey, the true answer to this question cannot change. Thus, the changes observed are merely changes in reporting.
interviewed an increasing number of individuals each year, starting with 7224 in 1979 and ending with 67,942 in 2016.
The NSDUH is implemented via home visits. Initially surveyors used paper-and-pencil interviewing, and later they switched to a combination of paper-and-pencil and computer-assisted interviewing. Starting in 2002, the individuals interviewed are given $30 for participating, which has resulted in an increase in response rates. The survey oversamples adolescents aged 12 to 17 years and young adults aged 18 to 25 years, and it has differential sampling rates used based on race/ethnicity groups as well. For the NSDUH survey from 1979 to 2016, the average interview response rate (i.e., number of respondents divided by the number of selected individuals) was 76%. Much of the data are available publicly online, but there are some variables that are restricted for confidentiality, which can be accessed after an application process.

Data Preparation
Trends in reported marijuana lifetime-use by age 25 are estimated by tracking birth cohorts (who are older than 25) and comparing their responses in different surveys. The key idea is that since the individual respondents cannot change the reality about whether they had used marijuana by age 25, the true number of individuals who used by age 25 cannot change after they are 26 and older. Thus, any changes seen in their responses about use by age 25 reflect solely changes in reporting. Figure 3 explains this process.
The reason for choosing age 25 as the cutoff was that 97% of the U.S. population who used marijuana started doing so before age 25, as shown in Figure 4(b). The sensitivity to this cutoff age was checked by changing the age to 30 and 35, and the results did not change significantly, as could be expected from looking at the figure. Also, according to the NSDUH, individuals born later in time tend to have higher proportions of individuals who report that they used marijuana by age 25. Figure 4(a) shows this increase in the cohorts born from 1942 to 1991. However, the analysis focuses on relative changes in reporting, so the fact that some cohorts are more likely to report using marijuana than others in general, although interesting, is not directly relevant.
Before calculating the weighted counts of individuals who reported using by age 25, four issues with the data had to be addressed: (1) the public data did not contain the individuals' date of birth, (2) the survey is cross-sectional, not longitudinal, so following individuals over time was not possible, (3) NSDUH does not ask individuals whether they used marijuana by age Figure . Trends in the NSDUH survey from  to . (a) Weighted number of individuals who reported using marijuana by age  for each birth cohort, with % confidence intervals. (b) Histogram of the age of first marijuana use from the  survey. This distribution has scarcely changed since . Only .% of individuals try marijuana after age . 25, which is the key variable of interest for this analysis, and (4) changes in the variable of interest (i.e., number of individuals who reported using marijuana by age 25) could be due to changes in reporting or in other variables, such as sampling variability from survey to survey, attrition due to death of cohort members, immigration of individuals from the same birth cohort, etc.
First, a "year of birth" variable was generated by using the "year of first use" and "age of first use" variables, which are publicly available. Across the surveys from 1979 to 2016, 77% of the U.S. population, on average, reported using at least one of the following substances: cigarettes, cigars, smokeless tobacco, chewing tobacco, snuff, alcohol, marijuana, cocaine, hallucinogens, LSD, PCP, ecstasy, inhalants, stimulants, and methamphetamines. NSDUH has information about more drugs, but these were the ones with most commonly reported use. The rest of the respondents reported never having used any substance, including alcohol and tobacco. The "age of first use" variables were reported for 99% of the respondents who reported having used the drug. Once an individual has reported using the drug at some point, the computer software directs the individual to provide a response for the age of first use for that drug. Since the respondents only get the cash incentive when they have completed the survey, the item nonresponse rate is very low. Hence, it was possible to impute year of birth for most respondents, including all the ones who reported using marijuana, by subtracting "age of first use" from "year of first use, " to obtain "year of birth. " One issue with this approach is that our analysis cannot distinguish between an individual who is a member of the birth cohort but never reported using any drug and an individual who is not a member of a birth cohort, for those individuals who reported never having used any drug. Since the analysis focuses on detecting changes in the number of individuals who reported using by age 25, however, if there is a negative answer this is relevant to the results, regardless of whether it was because they were not included in the cohort or because they reported not using.
Second, NSDUH is a cross-sectional survey, so it does not track individuals over time. However, NSDUH asks retrospective questions about marijuana use, and this allows for the comparison of response over time. Respondents are picked at random from the same population birth cohort for each survey, so the individuals vary from year to year. This has the benefit of eliminating individual-specific fixed effects. Cohorts have an unweighted sample size of 951 on average, with the smallest one being 80, thanks to the large sample size of the NSDUH survey. Longitudinal data tend to have smaller sample sizes, it suffers from attrition, and it is more costly to implement, but it could in theory be used for this type of analysis.
Third, the outcome was generated as a binary variable, equal to one if the person reported using marijuana by age 25, and equal to zero otherwise. The NSDUH survey asks the question, "How old were you the first time you used marijuana or hashish?" From the response to this question and the "year of birth" variable, NSDUH calculates a "year of first use" variable, which is available in the public use file. Using this variable, a new indicator was generated for all the individuals in the survey. The NSDUH asks respondents a sequence of questions that checks for consistency and eliminates logical mistakes.
The responses to these questions are then used to generate imputed variables, although the survey provides the raw version as well. The NSDUH codebook (2016) recommends that "Where imputed or recoded variables are provided, users are encouraged to use them to produce estimates rather than raw or edited variables from the interview. " Thus, this analysis uses the imputed variables.
Fourth, to ensure that the changes observed in the outcome variable were due to changes in reporting and not to factors like sampling variability or changing size of birth cohort, the analysis relies on the NSDUH survey weights. The survey weights are used to obtain unbiased estimates for survey outcomes in the population represented by the survey. The person-level analysis weights (called ANALWT_C) are the product of 16 weight components, which include household weights for nonresponse at the screener level, poststratification of household weights to meet population controls for various household-level demographics by state, adjustment of household weights for extremes, among many others. In earlier surveys the weights had fewer components, especially before 2002, after which the survey methodology barely changed. Each weight component accounts for either a selection probability at a selection stage or an adjustment factor adjusting for nonresponse, coverage, or extreme weights. 1 For variance estimation, suitable software was used to account for the multistage stratified cluster sample design. In this analysis, the survey package in R was used for the analysis.
Finally, there have been several changes in the NSDUH sampling design over the years. Notably, in 1999 the survey changed from paper-and-pencil to computer-assisted interviewing, and in 2002 NSDUH introduced incentives to counteract declining response rates. The population of the US is also changing. For example, the population has a higher percentage of immigrants, it is aging as a whole, and some of the respondents from the earlier years might have passed away in the later years. The NSDUH weights reflect the number of people that each respondent represents at the time of the survey, so the weights are consistent with population control totals obtained from the U.S. Census Bureau (NSDUH 2017). These controls are based on the most recently available decennial census. The Census Bureau updates these control totals annually to account for population changes after the census.

Method
The quantity of interest in the analysis isn cs , the number of individuals who reported using marijuana by age 25 from each birth cohort c and each survey s. This number is denoted by n cs , which is defined as i∈Population domain y i , where the population domain is the number of people in the population who are in that birth cohort in a given year when the survey was performed. The variable y i is an indicator defined as one if the respondent is in cohort c and survey s, and reported using marijuana by age 25 and zero otherwise.   A natural estimator of n cs isn cs = i∈Sample domain y i , which could be estimated simply by using the sample count. However, until the sample is drawn, it is unknown which individuals in the population belong to which cohort, so the number of individuals who fall into each cohort is a random variable with unknown value at the moment when the survey is designed. The standard errors forn cs were derived as for ratio estimators, following the procedure in Lohr (2009). From Lohr (2009), the estimated variance used in this analysis was var(n cs ) = k j=1 V ( i∈S w i y i j ) + 2 k−1 j=1 k l= j+1 cov( i∈S w i y i j , i∈S w i y il ), where w i are survey weights. Domain estimation was performed using the survey package in R, written by Thomas Lumley, which helps account for the complex survey design in domain estimation. The formula svyby() was used to compute estimates for the set of subpopulations. Under an approximation (that sampling units are identified by the variance estimation cluster replicates and are sampled with replacement, within strata identified by the variance estimation stratum), the weighted total of the variable for each variance estimation cluster replicates was computed, then the ordinary variance of these totals within each stratum was computed, the variances from each stratum were added up to obtain the variance of the total, and finally the square root was taken to get the standard error of the total.
Estimates of the 1,392 quantities of interest were obtained, one estimate for each birth cohort and survey year combination, along with their corresponding standard errors. The birth cohorts included were from 1942, because earlier cohorts had sample sizes that were too small, to 1990, because after that year no one in the surveys is older than 25 in 2016, which is the latest survey. The surveys were conducted from 1979 to 2016, omitting some years, for a total of 29 surveys. This process accounts for the complex survey design (stratification and clustering), including the proper weighting by the survey's probability weights.

Results
The full set of weighted counts and standard errors can be found in Tables 1, 2, and 3. In the tables, the counts are presented after being leveled (by subtracting the minimum value per year of birth) and normalized (by dividing by the maximum value of the leveled counts). These tables can be used to plot the trends in reporting for each birth cohort with corresponding confidence intervals.
The first result is that the estimates of the outcome variable do not seem to increase over time, which contradicts the hypothesis suggested by Figure 2. The second result is that the number of individuals who reported using marijuana by age 25 for each birth cohort over the NSDUH surveys , within each cohort, did vary significantly from some surveys to others, but there was no significant monotonic upward or downward trend for all birth cohorts. However, the third result is that there is a slight nonmonotonic downward trend in the later years between 2006 and 2016. It is statistically significant for some birth cohorts but not for others, and it is not clear for which types of birth  cohorts it is significant. This trend can be seen in Figure 5. The downward trend implies that individuals have increased their underreporting over time, however slightly. All of this holds only as long as the NSDUH survey weights account for other changes in the measure of reported marijuana use by age 25. Table 1-3 can also be used to generate new adjustments to a factor that was already proposed by other researchers. For instance, the factor of 1.4 suggested by Harrison et al. (2007) was derived by using a costly survey that compared contents of marijuana in urine and self-reported responses. If this is used as the base adjustment factor, then it can be adjusted further by survey year and by birth cohort, according to the reporting trends found in the Table 1-3. Thus, it is possible to use Tables 1-3 to derive new survey-and cohort-adjustments to that underreporting-adjustment factor such that there is no need to perform a costly survey each year to determine the amount of underreporting.
For instance, Table 2 shows adjustments for a 2007 factor. For the year 2008, one could use this table by first multiplying the number of people who self-reported using marijuana by 1.4 (the 2007 adjustment), and then again by 1.26 to account for changes in reporting from 2007 to 2008. To derive these values, all the years were shifted to be at the same level on the y-axis by subtracting the minimum of each birth cohort across all survey years. Since the trends in reporting are the quantity of interest in the analysis, only the changes between surveys matter, and not the y-intercept (i.e., the level of the plot). Then the estimates were normalized (divided) by the 2007 values for each birth cohort. For birth cohorts that were too young in 2007 to be included, the average of all the 2007 cohorts was used. Finally, the values were averaged over all the birth cohorts to arrive at a single value per survey, shown in Table 4. As Tables 1-3 show, there is clearly a heterogeneity in reporting by birth cohort as well as by survey, so it is somewhat simplistic to do this last averaging step, but it is a way to simplify the values so researchers can adjust their estimates just by survey year and not also by birth cohort. This analysis can be repeated to find an adjustment for each birth cohort, as well as an adjustment for each birth cohort for each survey. Table . Estimated weighted count (with standard error in parentheses) of individuals who reported using marijuana by age , leveled and normalized. The rows correspond to different birth cohorts and the columns to surveys. Unfortunately, the standard errors derived for the adjustments in Table 4 were quite large, and thus it is possible that the variability between surveys or between birth cohorts is too large to get a simple one-number adjustment for each survey. To have smaller standard errors, it is possible to have a larger sample size by pooling data from several birth cohorts. However, combining individuals of different ages could allow the changes to be due to real differences in consumption between birth cohorts rather than just changes in reporting, especially if many cohorts are pooled together. Thus, we did not include pooling in this  analysis. The downward trend observed in Figure 5 is not visible here, possibly because the variability across cohorts is too large. Nevertheless, Table 4 shows that there is a large amount of variability in reporting from survey to survey, and therefore it is important to account for changes in reporting over time, not just for one specific year.

Discussion
This article presents a novel way to explore relative changes in measurement error from one survey to another simply by using data already available in a self-reported survey, the National Survey on Drug Use and Health (NSDUH). To do this, the consistency of responses to the question "Did you use marijuana by age 25?" over time was evaluated. If individuals were underreporting at a constant rate (i.e., if the measurement error were constant), then the responses to this question for individuals who are older than 26 should be constant. This article presents evidence, however, that responses to this question are not constant over time in the NSDUH, which suggests that there are changes in reporting across different surveys (as long as the NSDUH survey weights properly accounted for sampling variability). It is not clear whether there is a monotonic upward or downward shift in reporting over time for all birth cohorts, althoughcontrary to our hypothesis that a decrease in stigma would increase people's willingness to report over time-there seems to be a downward trend in reporting over the decade of 2006 to 2016. For instance, the individuals born in 1980 were more likely to report their marijuana use in 2006 than in 2016, and this is a statistically significant difference, as shown in Figure 5.
This suggests that in recent years individuals have become increasingly less willing to report their drug use over time.
Our results have two primary implications for policy. First, if the results are correct in suggesting that there is a downward trend in reporting in recent years, this would imply that the ONDCP (Kilmer et al. 2011) likely underestimated the already drastic increase in marijuana use from 1992 to 2016. Second, although this analysis does not assess whether the adjustment factors used by various researchers are correct in accounting for underreporting, it does show how to further adjust the currently used adjustment factors according to changes in reporting over time, as well as changes across birth cohorts. This is a low-cost method to update adjustments required to translate reported marijuana prevalence to true marijuana prevalence. In other words, previously, researchers would need to perform costly physical tests on the population or extrapolate from alcohol and tobacco reporting trends. Now, researchers can find the changes in reporting for new years cheaply and quickly by using this method to update previous factors.
Some limitations of this study are, first, that it relies on the survey weights to pick up any changes over time that are not generated by changes in reporting. It is possible that the survey weights were not calculated properly, or that they are derived in a way that precisely counteracts the changes in reporting over time. Without knowing all the details about the generations of the weights, as well as the steps taken to protect the respondents' confidentiality, this is difficult to do. Second, this method only tests whether individuals have ever used marijuana, so it does not work for questions about recent use, and this is perhaps affected more drastically by changes in stigma. Third, although it might be interesting to see if there are also heterogeneities in terms of race, socioeconomic background, and gender, this analysis does not account for changes in demographics other than birth year. Fourth, this analysis makes the assumption that respondents are unlikely to forget whether they had used the drug by age 25. It is possible that as time passes, individuals are more prone to forget facts about the past, and perhaps this is a greater problem with aging populations. Memory issues could affect the analysis if individuals simply forgot that they used marijuana at all, if they remembered using it at an age greater than 25 whereas in fact they had used it by age 25, or if after some age individuals began remembering that they had used by age 25. Testing this assumption would strengthen the analysis.
Furthermore, an alternative explanation for our finding of a decrease in reporting is that more individuals are dying, and thus the number of individuals who report using marijuana by age 25 is decreasing, especially if the ones dying are the ones who used marijuana. Another alternative is that immigrants are coming into the country, and they report never having used marijuana, so the outcome of interest also decreases. However, the decrease is seen in all the birth cohorts from 2005 to 2016 (this can be seen in Figure 5). Mortality would affect older cohorts more, and immigration would affect younger cohorts more. If both mortality and immigration work together, they might create the effect of a downward trend in the outcome of interest for all birth cohorts. However, this explanation is more complex than simply that underreporting increased in the past decade, and thus, by Occam's razor, the latter explanation is preferred.
Some of the advantages of this method include that it could be used to study trends in reporting for other drugs. The NSDUH survey contains information about numerous drugs (e.g., tobacco, alcohol, cocaine, opioids), and this analysis could be repeated by researchers interested in assessing trends in reporting from 1979 to 2016, as long as the sample sizes of respondents are large enough. So this might not work with a drug with fewer self-reporters such as hallucinogens. As more NSDUH surveys are performed, new years could be added to the analysis. Furthermore, the method could be used to study reporting trends in other lifetime behaviors as well (e.g., sexual assault, abuse, other illegal behaviors). The method requires that the first time individuals engage in the activity be early in life, on average, so as to allow for the analysis to contain more years.
This method could also be used to study reporting trends in different surveys. The analysis requires that three conditions be satisfied: (1) the survey makes it possible to separate the individuals into birth cohorts, that is, knowing year of birth, (2) the survey is large enough to have more than 100 individuals per birth cohort (unweighted) in each survey for the purposes of inference, and (3) the survey contains the variable for first year or age of use of marijuana. Some other surveys are good candidates to analyze the trends in reporting of marijuana or other drugs and some are not. For instance, the Behavioral Risk Factor Surveillance System (BRFSS), a telephone survey that studies behavioral risk factors, asks about marijuana use in the past 30 days but not age of first use. Therefore, it is not a good candidate for the analysis proposed by this article. The National Youth Tobacco Survey (NYTS) and the National Adult Tobacco Survey (NATS), national surveys on indicators for tobacco prevention and control, focus primarily on tobacco, not marijuana. In theory their tobacco reporting trends could be analyzed with the analysis proposed here. Monitoring the Future (MTF), an ongoing longitudinal survey that surveys trends in legal and illegal drug use as well as other topics, has nationally representative samples of eighth-, tenth-, and twelfth-grade students. The National Longitudinal Survey of Youth 1979 (NLSY79), a nationally representative sample of 12,686 young men and women who were 14-22 years old when they were first surveyed in 1979, also has information about drug use. The Youth Risk Behavior Surveillance System (YRBSS) monitors health-risk behaviors, and one of its purposes is to determine the prevalence of marijuana use among the youth and adults in the United States. The MTF, NLSY, and YRBSS satisfy all three points required for the analyzing reporting trends with the method proposed in this article. Household surveys (called "general population surveys") of drug use are performed in many countries outside the United States. If their data satisfy the three requirements, it might be possible to analyze trends in reporting of drug use or other behaviors in international settings.