Analysis of Deterrent Events on Cocaine Prices

Methodology

Transfer Function Models: Technical Exposition
Our objective was to estimate the effect of interdictions on prices, controlling for confounding variables and correlation in the data. The effects of interdictions may be delayed or dynamic (or both), and the price series itself is a correlated time series. These possibilities were allowed for by transfer function models (Box et al. (1994), Mills (1990)) of the form

where is the combined effect of control variables (time, weather, and Coast Guard and Department of Defense force laydown), is the combined effect of interdictions, and N(L)w_t represents a correlated noise series. Thus, aside from the control variables, the transfer function model containing p interdictions is

where price_t is the price, x_jt is the value of the jth interdiction, and w_t is a white noise error term, all at month t. The effect of the jth interdiction is assumed to transfer its effect on price via the ratio of two polynomials in the lag operator L, where Lx_t = x_t-1, L²x_t = x_t-2, L³x_t = x_t-3, and so on,

Thus, R(L) is a possibly infinite polynomial in L that can flexibly and parsimoniously model the dynamic relationship between price_t and x_t, x_t-1, x_t-2, …. The factor L^k allows for the possibility of pure delay of k months, the time period before the dynamic effect of x_t on pricet takes effect.

The correlated noise series n_t = N(L)w_t is modeled in a similar fashion, by a transfer function of a white noise series

If n_t = N(L)w_t then n_t is said to be generated by an autoregressive moving average process of order (u,v), a flexible and parsimonious way of accounting for the relationship between pricet and price_t-1, price_t-2, price_t-3, …. In the above expression, n_t is assumed to be stationary, possibly after detrending or differencing.

Transfer Function Models: Our Approach
All of our price series were stationary, and ARMA representations were either AR(1) or the product of AR(1) and AR(12), and thus respectively. The parameters were typically positive, representing positive month-to-month correlation in the AR(1) case, and positive year-to-year correlation (seasonality) in the AR(12) case.

Our transfer functions for interdictions usually incorporated delays, and occasionally incorporated simple geometric dynamic components, "Dynamic" simply means that the total effect was distributed over time. If then an interdiction commencing operation at month t (a unit increase in x from (t-1) to t) changes prices at time (t + k). That is, there is a k-period delay. When the interdiction becomes inactive, prices return to their original level k months later. In the dynamic case, and an interdiction commencing at month t changes price by at time (t + k), by a at time (t + k + 1), by a² at time (t + k + 2), and so on. The total increase in price is and the rate of adjustment is governed by the parameter a. Finally, k months after interdiction termination, prices begin an exponential decline to their original level.

For example, wholesale prices at the South West Border were influenced by three interdictions, x₁ = Orejuela Arrests, x₂ = Support Justice III, and x₃ = Support Justice IV. Orejuela Arrests had a dynamic effect with no delay, and the two source zone interdictions had non-dynamic effects each with a five month delay. In addition, the model contained a time trend, Coast Guard force laydown, and an AR(1) error term. The estimated model was

Parameters associated with control variables are interpreted in the usual way. The positive effect for CG indicates that a unit increase in the Coast Guard force laydown (a doubling from the 1991 figure) would increase wholesale prices by $3.8 (actually this was not statistically significant). The trend parameter implies that, had the other variables in the model held constant, wholesale prices would have dropped by about $6 over the course of the 96 months study.

Turning to the time series components of the model, the AR(1) error term estimates a positive month-to-month correlation of 0.14, but the absence of an AR(12) term implies no seasonal effects. The effect of Support Justice III (November 1991 through April 1992) and Support Justice IV (January 1993 through March 1994) was to increase wholesale prices five months later by $5.5 and $3.1 respectively. The effect of the Orejuela Arrests (June 1995 through August 1995) was to increase wholesale prices by $3.8 in the month of the arrest, $3.8(0.7) = $2.66 the next month, and $3.8(0.7)² = $1.862 the following month. Conceptually, if a king pin figure equivalent to the Orejuela brothers could have been arrested every month, prices would have eventually increased by $3.8/(1 . 0.7) = $12.67,and 90% of this would have been attained by the sixth month after the first arrest. In fact, the last Orejuela arrest was in August 1995, so prices began an exponential decline to their original level starting in September 1995.

Model Selection
For all four price series (retail and wholesale, national and South West Border), we used a common baseline model containing a time trend (t), Weather (the proportion of hurricane days in the month t), Coast Guard and Department of Defense force laydown (CG_t and DoD_t) and the product of an AR(1) and AR(12) term allowing month-to-month, and year-to-year correlation. Thus the baseline model was

All such models appeared to be suitably stationary; the Dickey-Fuller test rejected the unit root null hypothesis in favor of stationarity for all four price series (p<0.006).

For a given price series, model selection involved adding interdictions, with appropriate delays, to the baseline model. The motivation for a careful selection process arises because several interdictions occurred at, or around, the same time as other interdictions. Two interdictions that are active over exactly the same time period are impossible to untangle. They cannot both be entered into the model, and if one interdiction is chosen, it actually represents both. More generally, a set of interdictions with considerable overlap presents similar problems. Estimates are either extremely imprecise (if the entire set is included in the model), or the partial set is to some extent a surrogate for the entire set (if a partial set is included). Statisticians say that such effects are collinear, a problem that will plague all the analyses reported here.

Many of our 18 interdictions were either extended impulses (Gateway, Support Justice IV) or step functions (Laser Strike, Border Shield, River Sweep) with considerable overlap. One way of measuring the degree of overlap is by the R² obtained by regressing one interdiction on all the others (table xx). The variance inflation factor, VIF = 1/(1 - R²), then measures the resulting imprecision. For example, 96% of the variation in Laser Strike is explained by the other interdictions, and the inflation of the variance of Laser Strikes parameter estimate increases 27 fold as a result of including the other interdictions in the model. Although these remarks strictly apply to linear models, they are broadly applicable to transfer function models also. For this reason, we were cautious about including highly correlated interdictions in the same model, and in particular, we excluded Laser Strike when analyzing South West Border prices.

Interdiction R-squares and Variance Inflation Factors

Interdiction Type	Interdiction	R2	VIF
Source Zone	Shoot-down Laser Strike Support Justice III Support Justice IV Stand Down	0.65 0.96 0.18 0.34 0.22	2.88 27.21 1.22 1.52 1.29
Arrests/Deaths	Orejuela Arrests Arrests/Deaths Columbus	0.34 0.22 0.11	1.52 1.28 1.12
Arrival/Transit Zone	Zorro II Border Shield Brass Ring Hard Line River Sweep Conjuntos I Conjuntos II Frontier Lance Frontier Shield Gateway	0.67 0.90 0.78 0.87 0.94 0.41 0.06 0.67 0.29 0.94	3.03 10.45 4.46 7.49 17.68 1.69 1.06 3.04 1.41 15.50

Model development involved the selection of an interdiction as well as the selection of its optimal delay. The ranges of delay that were considered were 4 to 6 for the five source zone interdictions, and 0 to 2 for all other interdiction types. Thus with 18 interdictions, there were (3)(18) = 54 candidate variables. Since the high correlation among interdictions favored a forward selection method, our approach was to add the best variable (interdiction-lag combination) to the baseline model. Given the large number of tests involved (about 50), we used a reasonably stringent entry criteria: at each iteration of the selection procedure, the best variable entered the model provided its p-value was less than 0.075. At the end of the selection procedure we then considered excluding insignificant variables, and explored dynamic specifications for the transfer functions. The following tables illustrate this procedure for retail prices at the South West Border.

SWB Retail Prices: (AIC=1015.07, Obs=96)
Model = Baseline +

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) AR(12) Time Trend Coast Guard DoD Weather		131.02 0.36 0.03 –0.76 140.42 –16.56 –38.11	56.89 0.10 0.11 0.91 96.35 86.32 55.56	0.0236 0.0005 0.8114 0.4062 0.1485 0.8483 0.4945

SWB Retail Prices: (AIC=961.74, Obs=92)
Model = Baseline + SJIII

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) AR(12) Time Trend Coast Guard DoD Weather Support Justice III	4	153.69 0.15 –0.11 –1.07 180.80 –64.59 –34.56 96.41	42.33 0.11 0.12 0.75 77.74 66.57 52.76 30.74	0.0005 0.1772 0.3530 0.1583 0.0224 0.3348 0.5143 0.0024

SWB Retail Prices: (AIC=954.40, Obs=92)
Model = Baseline + SJIII + Arrests/Deaths

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) AR(12) Time Trend Coast Guard DoD Weather Support Justice III Arrests/Deaths	4 1	142.77 0.08 –0.16 –0.89 158.79 –47.10 –27.33 101.81 74.70	38.37 0.11 0.12 0.68 70.47 60.60 49.88 28.18 24.89	0.0004 0.4519 0.1830 0.1982 0.0269 0.4393 0.5852 0.0005 0.0036

SWB Retail Prices: (AIC=949.63, Obs=92)
Model = Baseline + SJIII + Arrests/Deaths + Hard Line

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) AR(12) Time Trend Coast Guard DoD Weather Support Justice III Arrests/Deaths Hard Line	4 1 2	114.20 –0.01 –0.20 –0.87 159.87 –16.85 –44.28 98.48 64.56 33.99	36.13 0.11 0.12 0.62 63.58 56.07 47.90 25.74 24.52 12.76	0.0022 0.9207 0.0808 0.1658 0.0139 0.7645 0.3579 0.0003 0.0101 0.0093

SWB Retail Prices: (AIC=946.71, Obs=92)
Model = Baseline + SJIII + Arrests/Deaths + Hard Line + Conjuntos II

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) AR(12) Time Trend Coast Guard DoD Weather Support Justice III Arrests/Deaths Hard Line Conjuntos II	4 1 2 1	121.18 –0.06 –0.16 –0.99 167.50 –25.21 –39.23 98.91 66.35 35.37 86.48	34.18 0.11 0.12 0.59 60.09 52.86 46.89 24.38 24.20 12.18 41.68	0.0007 0.6291 0.1781 0.0933 0.0066 0.6347 0.4052 0.0001 0.0075 0.0048 0.0412

Thus the final iteration of the selection procedure chose Support Justice III with lag 4 (p<0.0001), Arrests/Deaths with lag 1 (p=0.0075), Hard Line with lag 2 (p=0.0048), and Conjuntos II with lag 1 (p=0.0412). The likelihood ratio test for the inclusion of all four variables was 29.9 on 4 degrees of freedom (p<0.0001) and the improvement in AIC was 21.9. Given these four interdictions, we were able to do an exhaustive search for the optimum delays.

Results

Having illustrated our approach, we proceed to apply the forward selection procedure to all four price series. The results follow.

National Wholesale Prices
The final iteration of the forward selection procedure for national wholesale prices added Support Justice III with lag 5 (p=0.006), Orejuela Arrests with no lag (p=0.004), Shoot-down with lag 6 (p=0.001), Border Shield with lag 2 (p=0.002), and Brass Ring with a lag of 2 (p=0.020) to the baseline model. The likelihood ratio test for the inclusion of all five variables was 28.1 on 5 degrees of freedom (p<0.0001), and the improvement in AIC was 18.1.

Fine-tuning the above model resulted in the removal of Brass Ring, DoD, Coast Guard, and AR(12), which had the effect of slightly reducing the AIC from 471.2 to 466.5. The final model is shown in the following table

Final Model for National Wholesale Prices (AIC=466.5, Obs=102)

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) Time Trend Weather Support Justice III Orejuela Arrests Shoot-down Border Shield	5 0 6 2	38.96 0.22 –0.17 4.03 3.53 4.71 3.58 4.52	0.84 0.11 0.02 1.30 1.24 1.60 1.09 1.15	<.0001 0.0384 <.0001 0.0026 0.0055 0.0042 0.0014 0.0002

The effects of all four interdictions were highly statistically significant (p<0.006). The effect of the two source zone interdictions, Support Justice III (November 1991 through April 1992) and Shoot Down (March 1995 through November 1995) was to increase wholesale prices by about $3.5 five to six months later (p<0.006). The effect of Orejuela Arrests (June 1995 through August 1995) was to increase wholesale prices by $4.7 in the month of the arrest (p=0.004), and the effect of long active Border Shield (March 1997 onwards) was to increase wholesale prices by $4.5 with a delay of two months.

Both national wholesale and national retail prices (presented below) estimated a positive month-to-month correlation [0.22 (p=0.038) for wholesale, and 0.37 (p=0.0002) for retail], but no seasonal effect could be detected in the final models. In addition, both wholesale and retail prices were affected by weather, $4.0 (p=0.003) for wholesale and $17.2 (p=0.134) for retail. Thus wholesale prices in months with hurricanes every day (e.g. April, May and June 1997) are expected to be $4 higher than months with no hurricanes (e.g. the four months preceding April 1997). Similarly, retail prices are expected to be $17 higher in hurricane-saturated months than in hurricane-free months. The effects of weather and month-to-month correlation are evident from the figure depicting the final models for national wholesale and national retail prices. In particular, high prices in the spring of 1997 appear to be largely attributable to the high frequency of hurricanes during that period.

National Retail Prices
The first four interdictions added to the baseline model for national retail prices by the forward selection procedure were the same for national wholesale prices: Support Justice III, Orejuela Arrests, Shoot Down, and Border Shield. In the case of retail prices, the optimal lags were Support Justice III with lag 4 (p=0.006), Orejuela Arrests with lag 1 (p=0.004), Shoot-down with lag 6 (p=0.001), and Border Shield with lag 1 (p=0.002). The likelihood ratio test for the inclusion of all four variables was 19.3 on 4 degrees of freedom (p=0.0007), and the improvement in AIC was 11.3

Fine-tuning the above model resulted in the removal of Weather, DoD, Coast Guard, and AR(12), which had the effect of slightly reducing the AIC from 907.3 to 903.9. The final model is shown in the following table.

Final Model for National Retail Prices (AIC=903.9, Obs=102)

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) Time Trend Weather Support Justice III Orejuela Arrests Shoot-down Border Shield	4 1 6 1	131.34 0.37 -0.43 17.15 29.81 33.12 25.83 29.73	8.66 0.10 0.19 11.35 12.01 14.61 10.54 11.67	<.0001 0.0002 0.0229 0.1341 0.0149 0.0257 0.0161 0.0125

The effects of all four interdictions were statistically significant (p<0.026). The effect of the two source zone interdictions, Support Justice III (November 1991 through April 1992) and Shoot Down (March 1995 through November 1995) was to increase wholesale prices four and six months later by $30 and $26 respectively (p<0.017). The effect of the Orejuela Arrests (June 1995 through August 1995) was to increase wholesale prices by $33 (p=0.026), and the effect of long active Border Shield (March 1997 onward) was to increase wholesale prices by $30 (p=0.013). In both cases the delay was by one month.

For wholesale prices, the effects of Shoot Down and Orejuela Arrests were strictly adjacent, giving rise to two plateaus of elevated prices lasting 12 months. For retail prices, in contrast, the two interdictions were additive for the month of September 1995, and this combined effect appears to largely account for the unusually large spike in retail prices that month. The final model predicts a still higher price because of the high hurricane percentage in August (48% or 15 days) and September (32% or 10 days), the effect of hurricanes in August on prices in September being explained by the positive month-to-month correlation.

South West Border Wholesale Prices
The data for the South West Border prices differed from that for the National prices in two ways. First, to avoid collinearity problems with interdictions less relevant to the South West Border, we excluded Frontier Lance, Frontier Shield, Gateway and Laser Strike. Second, in order to obtain stationarity (the mean and variance of the series were unusually high during this period), we did not use the first year of data. Based on the resulting 1992 through 1999 series, the Dickey-Fuller test easily rejected the unit root null hypothesis in favor of stationarity for both wholesale prices (p=0.0001) and retail prices (p= 0.0061).

The final iteration of the forward selection procedure for South West Border wholesale prices added the following interdictions to the baseline model: Orejuela Arrests with lag 1 (p=0.0004), Support Justice III with lag 5 (p=0.007), and Support Justice IV with lag 5 (p=0.024). The likelihood ratio test for the inclusion of all four variables was 19.7 on 3 degrees of freedom (p<0.0001), and the improvement in AIC was 13.7. A subsequent exhaustive search for the optimum delays for these three interdictions showed that the forward selection model was actually the optimal model. The top ten of the 3³ = 27 models (delay choices) are listed below. As expected, similar models produced similar estimates.

Top delay choices for the three selected interdictions: SWB Wholesale

Delay								Improvement	Model	Baseline	Diff
d1 1 1 2 2 1 1 1 0 2 1	d2 5 5 5 5 4 4 5 5 4 6	d3 5 4 5 4 5 4 6 5 5 5	Or/A 6.77 6.82 6.87 6.84 6.90 6.92 6.79 5.02 6.92 6.99	SJIII 5.57 5.69 5.64 5.72 4.29 4.25 4.32 5.69 4.16 2.17	SJIV 2.48 2.29 2.51 2.25 2.25 2.04 1.53 2.51 2.27 2.29	p-LRT .0002 .0003 .0003 .0006 .0013 .0019 .0021 .0023 .0025 .0029	LRT 19.7 18.8 18.5 17.4 15.7 14.9 14.7 14.5 14.3 14.0	in AIC 13.7 12.8 12.5 11.4  9.7  8.9  8.7  8.5  8.3  8.0	AIC  450.7 451.6 451.9 453.0 454.7 459.5 449.5 455.9 456.1 450.2	AIC  464.4 464.4 464.4 464.4 464.4 468.4 458.2 464.4 464.4 458.2	param 3 3 3 3 3 3 3 3 3 3

Fine-tuning the above model resulted in the removal of Weather, DoD, and AR(12), and the addition of a geometric lag for Orejuela Arrests starting at lag 0 (table xx). As a result, the AIC was reduced from 450.7 to 443.8.

Final Model for South West Border Wholesale Prices (AIC=443.8, Obs=91)

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) Time Trend Coast Guard Orejuela Arrests Orej/Arr: Geom Lag Support Justice III Support Justice IV	0 0 5 5	24.42 0.14 -0.06 3.81 3.84 0.73 5.47 3.14	1.48 0.11 0.03 3.04 1.11 0.11 1.90 1.09	<.0001 0.1958 0.0268 0.2136 0.0009 <.0001 0.0051 0.0052

The estimates in the table correspond to the parameters of the transfer function model represented by

where x₁ = Orejuela Arrests, x₂ = Support Justice III, and x₃ = Support Justice IV. The model indicates that wholesale prices at the South West Border were influenced by the three interdictions with Orejuela Arrests having a dynamic effect with no delay, and the two source zone interdictions having non-dynamic effects each with a five month delay.

The effect of Support Justice III (November 1991 through April 1992) and Support Justice IV (January 1993 through March 1994) was to increase wholesale prices five months later by $5.5 and $3.1 respectively. The effect of the Orejuela Arrests (June 1995 through August 1995) was to increase wholesale prices by $3.8 in the month of the arrest, $3.8(0.7) = $2.66 the next month, and $3.8(0.7)² = $1.862 the following month. Since the last Orejuela arrest was in August 1995, prices began an exponential decline towards their original level starting in September 1995.

The other parameters in the model are unremarkable. The AR(1) error term is suggestive of a positive month-to-month correlation of 0.14, but this result is not statistically significant. The positive effect for CG indicates that a unit increase in the Coast Guard force laydown (a doubling from the 1991 figure) would increase wholesale prices by $3.8, but again, this was not statistically significant. The trend parameter implies that, had the other variables in the model held constant, wholesale prices would have dropped by about $6 over the course of the 96 months study.

South West Border Retail Prices
The final iteration of the forward selection for South West Border retail prices procedure added Support Justice III with lag 4 (p<0.0001), Hard Line with lag 2 (p=0.005), Arrests/Deaths with lag 1 (p=0.008), and Conjuntos II with lag 1 (p=0.041) to the baseline model. The likelihood ratio test for the inclusion of all four variables was 29.9 on 4 degrees of freedom (p<0.0001), and the improvement in AIC was 21.9. Given the four interdictions, an exhaustive search for the optimum delays again showed that the optimal model coincided with that chosen by the forward selection method. The top ten of the 3₄ = 81 models (delay choices) are listed below. Again, it is encouraging to find that similar models produced similar estimates.

Top delay choices for the three selected interdictions: SWB Retail

Delay										Improvement	Model	Baseline	Diff
d1 4 4 4 4 4 4 5 6 5 5	d2 1 1 1 1 1 1 1 1 1 1	d3 2 1 2 0 1 0 2 2 1 2	d4 1 1 0 1 0 0 1 1 1 0	SJIII 98.91 98.41 98.72 97.91 98.11 97.53 94.00 77.37 93.30 93.29	A/D 66.35 67.90 67.34 70.01 68.91 71.06 65.02 67.22 66.41 65.92	HL 35.37 33.82 36.25 32.33 34.84 33.46 35.58 34.30 33.90 36.40	CII 86.48 84.35 79.05 82.24 76.26 73.34 86.65 91.94 84.50 76.06	p-LRT .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000	LRT 29.9 28.9 28.8 27.9 27.8 26.8 26.8 26.3 25.8 25.4	in AIC 21.9 20.9 20.8 19.9 19.8 18.8 18.8 18.3 17.8 17.4	AIC  946.7 947.8 947.8 948.7 948.9 949.8 939.3 927.5 940.3 940.6	AIC  968.6 968.6 968.6 968.6 968.6 968.6 958.0 945.8 958.0 958.0	param 4 4 4 4 4 4 4 4 4 4

Fine-tuning the outcome of the forward selection procedure resulted in the removal of Weather, DoD, and AR(12), and the addition of geometric lags for Support Justice III starting at lag 4 and Hard Line starting at lag 0 (table xx). As a result, the AIC was reduced from 946.7 to 930.3.

Final Model for South West Border Retail Prices (AIC=930.32, Obs=91)

Interdiction	Delay	Estimate	Std Error	p-value
Intercept AR(1) Time Trend Coast Guard Support Justice III SJIII Geom Lag Hard Line Hard Line: Geom Lag Arrests/Deaths Conjuntos II	4 4 0 0 1 1	107.58 0.00 -0.37 97.26 59.36 0.57 5.72 0.87 67.69 95.46	15.28 0.12 0.50 55.02 25.66 0.19 4.19 0.13 23.47 40.30	<.0001 0.9699 0.4682 0.0809 0.0232 0.0043 0.1753 <.0001 0.0050 0.0202

The estimates in the table correspond to the parameters of the transfer function model

where x₁ = Support Justice III, x₂ = Hard Line, x₃ = Arrests/Deaths of Mexican traffickers, and x₄ = Conjuntos II. Thus the changes in South West Border retail prices were dynamic for Support Justice and Hard Line, and were non-dynamic for Arrests/Deaths and Conjuntos II.

The effect of Arrests/Deaths (January 1996, February 1997, and July 1997) and Conjuntos II (December 1998) was to increase South West Border retail prices one month later by $68 and $95 respectively.

The effect of Support Justice III (November 1991 through April 1992) was to increase wholesale prices by $59 four months after initiation of the operation, $59(0.57) = $34 in the next month, and $59(0.57)2 = $19 in the following month, up to an asymptote of $59/(1 – 0.57) = $137. However, what is more evident from the figure is the subsequent exponential decline starting in September 1992.

Operation Hard Line (February 1995 through July 1997) had a similar dynamic effect while it was in full operation from January 1996 through August 1996. The initial increase in prices was about $6 approaching a total increase of $6/(1 – 0.87) = $46, half of which would be achieved by the fifth month. This effect was attenuated as Hard Line ramped up (February 1995 through December 1995) and ramped down (September 1996 through July 1997). These effects induced an S-shape to the modeled curve from February 1995 through July 1997.

The other parameter in the model that was close to statistically significance was Coast Guard force laydown (p=0.081). The magnitude of this parameter suggests that a unit increase in the Coast Guard force laydown (a doubling from the 1991 figure) would increase wholesale prices by $97, other factors held constant.

Analysis of Deterrent Events on Cocaine Movement

Characterizing Trafficker Behavior

Data from the interagency Consolidated Counterdrug Database (CCDB) was used as the indicator of trafficker behavior in the Transit Zone. This database contains drug smuggling event data dating back to 1991 and supports two primary types of analyses: (1) evaluations of drug flow from areas of production to consumption countries, and (2) assessments of the performance of counterdrug forces against that movement³².

The CCDB was founded by the Interagency Counterdrug Performance Assessment Working Group (ICPAWG). The ICPAWG—established in 1992 to measure the performance of counterdrug forces—maintained a database of known drug shipments in the transit zone whose final destination was either the U.S. or Canada. Known events are distinguished by (1) seizure or observation of drugs; (2) observation of activity that could not be reasonably attributed to anything other than drug smuggling; (3) highly reliable intelligence. In 1996, the CCDB was expanded to include events that would support Interagency Assessment of Cocaine Movement (IACM) evaluations. In order to allow a wide spectrum of cocaine flow information, ICPAWG accepted a higher degree of uncertainty for events inclusion in the database. This concern for drug flow data resulted in a dramatic increase in the total number of CCDB events from 1995 to 1996 (see Figure 1). For our purposes, this modification disrupted what otherwise would be a consistently measured time-series.

Figure 1 CCDB Event Data—1991 to 1999 All Purposes (ICPAWG and Flow) * Only 9 months of data

Fortunately, during the transformation, ICPAWG added a field to distinguish ICPAWG events from flow events. By examining only those events with an ICPAWG application, we constructed a consistently defined 1991–1999 event data set (see Figure 2). Although there is a large loss in the number of data points that entered the flow data, the value gained by expanding the time span of the study to 1991 outweighs this loss. We felt it was particularly important to ensure the model covered the 1994/1995 reductions in drug interdiction funding.

Figure 2 ICPAWG Only Event Data—1991 to 1999 * Includes only 9 months of data

The ability to reach back to 1991, however, is limited to non-commercial air and maritime events. Land, commercial air events and commercial maritime events are non-existent or, at best, sparse prior to 1996—whether ICPAWG, flow or otherwise. Excluding them from the Transit Zone data set, nevertheless, does not seem to be problematic. The CCDB Users Guide defines commercial events as port-to-port movements. As such, they are neither targeted nor impacted by interdiction efforts in the transit zone, but rather, by U.S. Customs operations in the Arrival Zone. As far as land events are concerned, the data set of events would not be large enough to conduct any meaningful analysis of how trafficking via land conveyance in the Transit Zone has been influenced by enforcement operations.

Events identified as 'subsequent' or 'other' movements were also excluded from our data set. 'Subsequent' movements are defined as events that originate from locations in the transit zone. 'Other' movements include shipments of any drug other than cocaine. For consistency and integrity of the data set, we felt it appropriate to exclude these types of events.

Methodology

This section is divided into three parts. In E1.0, the dependent variable is shipments confirmed by intelligence data. In E2.0, the dependent variable is all shipments. A short section, E3.0, reports results when the dependent variable is the total amount shipped.

Regardless of the definition of the dependent variable, the independent variables are the same:

CYCLE1 CYCLE2	cosine(2MONTH/12) sine(2MONTH/12)	Where MONTH is the month of the year, coded 1 through 12. Taken together, the cosine and sine functions account for seasonality, at least so far as seasonality is approximated by a cosine function.
VM1–VM17	A series of seventeen dummy variables coded one when the observation came from a specific vector/mode combination and coded zero otherwise.
TVM1–TVM17	A series of seventeen variable created by multiplying the VM1–VM17 dummy variables by time. For this purpose, time was coded as zero (the earliest month) to one (the last month).
INT1 NINT1 INT2 NINT2	As explained earlier, for each vector/mode targeted by a specified interdiction program, we coded an indicator variable as one when the interdiction was on and zero when it was off. For the period when the interdiction was on, we created a complement variable that was one in those vector/mode combinations that were not the target of that interdiction. INT1 is the sum of all indicator variables pertaining to border interdictions for each vector/mode combination during each month; and NINT1 is the sum of the complement variables for border interdiction programs. INT2 and NINT2 are the counterparts for transit zone interdiction programs.

E1.0 Analysis based on Shipments Corroborated by Intelligence

This section reports analysis that is based on shipments corroborated by intelligence. There are three subsections. In subsection E1.1, we report a baseline model. In section E1.2, we add specific interdiction programs (and their complements) one at a time to the baseline model. A Wald test identifies programs that had statistically significant effects on trafficker behavior. Section E1.3 includes the baseline variables and the significant effects from section E1.2 in a final regression model.

There are disadvantages to incremental development of a regression model.³³ Those disadvantages notwithstanding, we adopted this incremental approach to overcome serious collinearity problems. In short, we found it impossible to include all interdictions programs in a single model, so we adopted a search procedure to identify the most important candidate programs.

We adopted a liberal test of statistical significance. First, because we could predict the direction for these interventions, we used a one-tailed test of statistical significance. Second, we adopted a critical value of P=0.10. Of course, readers who prefer a more conservative test can use their judgment, and we provide P-values for those so inclined.

E1.1 Baseline Model: Shipments Corroborated by Intelligence
Table E1 reports regression results for the baseline model. The dependent variable is the number of shipments by vector/mode for each month in the study. The independent variables were identified above. The estimation procedure was a Poisson regression.

We are not much interested in the parameter estimates from this baseline model, so little discussion is necessary. Based on the coefficients associated with the variables CYCLE1 and CYCLE2, shipments do seem to follow a seasonal pattern. Based on the variables TVM1 through TVM17, there appear to be linear trends, and the directions of those trends vary by vector and mode. The parameters associated with the variables VM1 through VM17 imply that some vector/mode combinations support more traffic than other vector/mode combinations.

The parameters associated with the remaining independent variables are of substantive interest. If border interdiction were effective, we would expect the difference between the parameters associated with NINT1 and INT1 to be positive and statistically significant. We use similar tests throughout this study, so the reasoning behind this inference needs to be explained. If interdiction were effective, and if we had controlled for all other factors that affect shipping decisions, then the parameter associated with INT1 should be negative. This implies that traffickers ship less frequently through vector/mode combinations that are interdiction targets. If traffickers transferred their loads to vector/modes that were not targeted by interdiction, then the parameters associated with NINT1 should be positive. Thus, the value of the parameter associated with NINT1 minus the value of the parameter associated with INT1 should be positive. We adopt a less demanding test, however. Quite possibly, we have not controlled sufficiently for underlying trends, so the parameter associated with INT1 could be positive, or the parameter associated with NINT2 could be negative, even though interdiction was effective. The more lenient test is that the difference between the two parameter estimates be positive, because this implies a relative shift from vector/modes that are targeted by interdiction to vector/modes that are not targeted by interdiction. This is the test used throughout this report.

The difference between the parameter estimates for NINT1 and INT1 is negative, but it does not approach statistical significance. If transit area interdiction were effective, we would expect the difference between the parameters associated with NINT2 and INT2 to be positive and statistically significant. That difference is positive but not statistically significant (P=0.17). If this were the only evidence at our disposal, we would probably conclude that neither border interdiction programs nor transit zone interdiction programs disrupted traffickers. In fact, however, we can look at individual interdiction programs using this baseline model as background. We do this in the next section dealing with the enhanced baseline model.

Table E1: Baseline Model based on Shipments Confirmed by Intelligence

Poisson Regression Maximum Likelihood Estimates Dependent variable Weighting variable Number of observations Iterations completed Log likelihood function Restricted log likelihood Chi-squared Degrees of freedom Significance level N_SHIP_I ONE 1836 10 –2094.570 –2867.591 1546.041 39 .0000000 Chi- squared = 2781.93922 RsqP= .4479 G - squared = 2294.76503 RsqD= .4025

Variable	Coefficient	Standard Error	b/St.Er.	P[|Z|>z]	Mean of X
CYCLE1 CYCLE2 TVM1 TVM2 TVM3 TVM4 TVM5 TVM6 TVM7 TVM10 TVM11 TVM12 TVM13 TVM14 TVM15 TVM16 TVM17 TVM18 TVM19 VM1 VM2 VM3 VM4 VM5 VM6 VM7 VM10 VM11 VM12 VM13 VM14 VM15 VM16 VM17 VM18 VM19 INT1 NINT1 INT2 NINT2	–.1500526630 –.9939249148E-01 1.000027673 –1.441172945 –.5795320633 –2.090834448 –.8925335561 –2.690843073 1.192119176 2.495930061 3.176922259 3.251232916 1.360851818 3.962670355 3.081394984 2.804221258 2.039931377 .5890692988 2.711534100 –1.378330194 1.204863240 .4172125587 1.595449154 .2099930851 1.511223572 –2.549738868 –3.487084848 –1.473230381 –1.968400011 –.3229995088 –1.766712630 –1.110548906 –.6640140747 –.8061104337 –2.942762471 –4.534233549 –.2370591623 –.2541116115 .1522977365 .2756067348	.35806385E-01 .35873467E-01 .64202907 .35993054 .41818764 .34680529 .49700812 .40155541 1.0580415 1.2142762 .43801073 .51429727 .38985939 .41942376 .38719632 .35091487 .48421925 1.4837441 1.9247698 .35657943 .13503590 .18021664 .11891265 .20662410 .13047791 .61983533 .81259772 .27422004 .33827139 .20402925 .27230011 .23232914 .19674777 .24480057 .81782887 1.3210413 .11025857 .50608554E-01 .12019474 .62320496E-01	–4.191 –2.771 1.558 –4.004 –1.386 –6.029 –1.796 –6.701 1.127 2.055 7.253 6.322 3.491 9.448 7.958 7.991 4.213 .397 1.409 –3.865 8.923 2.315 13.417 1.016 11.582 –4.114 –4.291 –5.372 –5.819 –1.583 –6.488 –4.780 –3.375 –3.293 –3.598 –3.432 –2.150 –5.021 1.267 4.422	.0000 .0056 .1193 .0001 .1658 .0000 .0725 .0000 .2599 .0398 .0000 .0000 .0005 .0000 .0000 .0000 .0000 .6914 .1589 .0001 .0000 .0206 .0000 .3095 .0000 .0000 .0000 .0000 .0000 .1134 .0000 .0000 .0007 .0010 .0003 .0006 .0316 .0000 .2051 .0000	.64376427E-13 .17360728E-13 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .12908497 1.2690632 .39215686E-01 .26851852

E1.2 Enhanced Baseline Model based on Shipments Confirmed by Intelligence
In this section, we begin with the baseline model, add individual interdictions, and test whether or not those individual interdictions seem to affect traffickers. For example, to test whether Operation Gateway caused shippers to alter their shipping procedures, we added variables GATEWAY (turned on when and where Operation Gateway was operational) and NGATEWAY (the complement of GATEWAY) to the baseline model.³⁴ If Gateway were effective, we would expect the parameter associated with GATEWAY to be negative (fewer shipments) and the parameter associated with NGATEWAY to be positive (more shipments). In fact, we infer the presence of deterrence when the parameter associated with NGATEWAY minus the parameter associated with GATEWAY is positive and statistically significant. This difference implies a relative shift in trafficking from where Operation Gateway was operational to where it was not. We used a standard Wald test to determine statistical significance. Because we could predict direction, we employed one-tailed test and an . value of 0.10. Table E2 summarizes our findings.

Table E2: Adding Interdictions Programs to the Baseline Model—Shipments Confirmed with Intelligence

	Contrast from Adding Variables	P Value	Contrast from Regression	P-Value
Hard Line Two Dozen Zorro II Gateway White Shark I White Shark II Frontier Shield Border Shield Gulf Shield White Shark III Brass Ring River Sweep Frontier Lance Conjuntos I and II	0.35 1.43 0.43 0.29 –0.60 –0.95 –0.53 *** 1.26 *** –0.27 0.72 0.21 –0.26	0.08 0.08 0.02 0.06 0.02 0.14 0.09 *** 0.00 *** 0.53 0.00 0.61 0.58	0.27 1.01 0.28 0.09 –0.70 1.21 1.21 0.62	0.14 0.17 0.07 0.39 0.12 0.00 0.00 0.01

Column 1 identifies the interdiction programs considered in this study. The second column reports the contrast (e.g. the difference between the NGATEWAY parameter and the GATEWAY parameter), and the third column reports the P-value from the Wald test. We will discuss columns four and five in section E1.3.

For two programs (marked with ***), we could not compute the Wald statistic, presumably because the data were too sparse to provide estimates. For six of the interventions, the contrast was in the expected direction (positive) and was significant at p<0.10. These tests suggest that interdiction has caused traffickers to shift vectors and modes. However, the parameters for White Shark I and Frontier Shield are negative and might be judged to be statistically significant if we had not employed a one-tailed test. Although the evidence would seem to be on the side of interdictions effectiveness at disrupting trafficker behaviors, the evidence is equivocal. A regression analysis may be more convincing.

E1.3 Enhanced Regression Model based on Shipments Confirmed by Intelligence
Building on the results reported in section AE1.2, we identified interdiction programs that were significant (from Table AE2) and added them as a group to the baseline regression. As before, we tested for the statistical significance of the contrast using a Wald test. The regression results appear as Table E3.

The values and statistical significance of the contrasts are most important to this study. They are summarized in table E2. Column four reports the value of the contrast. Column five reports the P-value of the Wald test.

Four of the interdiction programs seem to have caused traffickers to change their shipment methods. (We combined Border Shield and Gulf Shield into single variables. Otherwise, when included treated separately in the same regression, they were collinear.) Two programs still have unexpected negative signs, but neither is statistically significant. We might accept this as modest evidence in favor of a finding that interdiction has disrupted trafficking behavior. Next, we repeat the analysis with a different dependent variable, all known shipments regardless of whether or not they had intelligence verification.

Table E3: Adding Interdiction Programs to the Baseline Regressions—Shipments Confirmed with Intelligence

Poisson Regression Maximum Likelihood Estimates Dependent variable Weighting variable Number of observations Iterations completed Log likelihood function Restricted log likelihood Chi-squared Degrees of freedom Significance level N_SHIP_I ONE 1836 10 –2004.052 –2867.591 1727.077 53 .0000000 Chi- squared = 2489.64066 RsqP= .5059 G - squared = 2113.72905 RsqD= .4497

Variable	Coefficient	Standard Error	b/St.Er.	P[|Z|>z]	Mean of X
CYCLE1 CYCLE2 TVM1 TVM2 TVM3 TVM4 TVM5 TVM6 TVM7 TVM10 TVM11 TVM12 TVM13	–.1467272238 –.1128102009 2.317709869 –.6710934168 .2851088598 –1.306475995 .1599118158E-02 –2.013531988 2.576579200 4.084014116 5.972371563 5.069425468 2.706757319	.37023890E-01 .36935424E-01 .72780797 .40582667 .47691572 .40511684 .56243747 .45728761 1.1860748 1.3130657 .67191322 .60073563 .46348191	–3.963 –3.054 3.185 –1.654 .598 –3.225 .003 –4.403 2.172 3.110 8.889 8.439 5.840	.0001 .0023 .0015 .0982 .5500 .0013 .9977 .0000 .0298 .0019 .0000 .0000 .0000	.64376427E-13 .17360728E-13 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01

Variable	Coefficient	Standard Error	b/St.Er.	P[|Z|>z]	Mean of X
TVM14 TVM15 TVM16 TVM17 TVM18 TVM19 VM1 VM2 VM3 VM4 VM5 VM6 VM7 VM10 VM11 VM12 VM13 VM14 VM15 VM16 VM17 VM18 VM19 INT1 NINT1 INT2 NINT2 TWODOZ NTWODOZ HARDLNE WHITESI BORGLF RIVSWP NHARDLNE NWHI NBORGLF NRIVSWP GATEWAY NGATEWAY SHIELD NSHIELD	5.900911540 4.853254423 4.517353018 6.425450913 3.172937606 4.400938468 –1.789159972 1.026773521 .1874096669 1.441687835 –.1022476886E-01 1.387924735 –2.991682453 –4.000411548 –2.628244836 –2.672704838 –.7768329996 –2.570984579 –1.775501144 –1.317490937 –2.500443009 –3.754525590 –5.158771869 .3110889125 .1734423999 .1931763316 .2468156891 –.5689010411 .4371016851 –.4138179349 –.1499855917E-01 –1.570947893 –1.754805825 –.1457431178 .7087233715E-01 –.3593930368 –1.135668465 –.7474104510 –.4683676083 .5499718757E-01 –.6453586667	.50641212 .46951719 .43423327 .97502496 1.9466202 2.1151093 .39040177 .14573128 .19629061 .12805624 .22403601 .13901570 .67654087 .86420724 .37570632 .37636039 .23028723 .31038373 .26428070 .22926556 .43726432 1.0067593 1.4304390 .14734896 .86683244E-01 .30654103 .17713148 1.0085545 .25636325 .24365322 .40493803 .46641634 .39191370 .95320297E-01 .29071813 .14713832 .31593503 .33763610 .29670468 .40743396 .27373697	11.652 10.337 10.403 6.590 1.630 2.081 –4.583 7.046 .955 11.258 –.046 9.984 –4.422 –4.629 –6.995 –7.101 –3.373 –8.283 –6.718 –5.747 –5.718 –3.729 –3.606 2.111 2.001 .630 1.393 –.564 1.705 –1.698 –.037 –3.368 –4.478 –1.529 .244 –2.443 –3.595 –2.214 –1.579 .135 –2.358	.0000 .0000 .0000 .0000 .1031 .0375 .0000 .0000 .3397 .0000 .9636 .0000 .0000 .0000 .0000 .0000 .0007 .0000 .0000 .0000 .0000 .0002 .0003 .0348 .0454 .5286 .1635 .5727 .0882 .0894 .9705 .0008 .0000 .1263 .8074 .0146 .0003 .0269 .1144 .8926 .0184	.29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .28867102E-01 .12854031 .10893246E-01 .63180828E-01 .21786492E-02 .92592593E-02 .31045752E-01 .70806100E-02 .37037037E-01 .25054466E-01 .24673203 .11328976 .59259259 .18790850 .26143791E-01 .19607843 .65359477E-02 .21241830E-01

E2.0 Analysis Based on All Known Shipments

Section E1 was based on an analysis of movements for which there were corroborating intelligence reports. In this section, we use data based on all sources. Otherwise, this section has the same structure as the last one.

E2.1 Baseline Model: All Shipments
The first step is to develop a baseline model. The regression results are reported in Table E4.

The baseline model requires little discussion, because it raises few new issues that were not discussed in section E1.1. However, in this model, the summary interdiction variable have the expected sign and both are statistically significant—at P = 0.025 for transit zone interdictions and at 0.01 for border interdictions.

Statistical modeling based on all movement events gives a different picture of how interdiction disrupts traffickers. We are uncertain about the reasons for these differences. Both have unique biases, but the biases associated with “all known movements” would seem to work against the findings reported here. At any rate, we now ask what happens as we add individual interdiction programs (and their complements) to the baseline model.

Table E4: Baseline Model based on All Shipments

Poisson Regression Maximum Likelihood Estimates Dependent variable Weighting variable Number of observations Iterations completed Log likelihood function Restricted log likelihood Chi-squared Degrees of freedom Significance level N_SHIP ONE 1836 9 –2705.301 –3769.244 2127.886 39 .0000000 Chi- squared = 3101.22822 RsqP= .4916 G - squared = 2798.30398 RsqD= .4320

Variable	Coefficient	Standard Error	b/St.Er.	P[|Z|>z]	Mean of X
CYCLE1 CYCLE2 TVM1 TVM2 TVM3 TVM4 TVM5 TVM6 TVM7 TVM10 TVM11 TVM12 TVM13 TVM14 TVM15 TVM16 TVM17 TVM18 TVM19 VM1 VM2 VM3 VM4 VM5 VM6 VM7 VM10 VM11 VM12 VM13 VM14 VM15 VM16 VM17 VM18 VM19 INT1 NINT1 INT2 NINT2	–.7314810134E-01 –.8670401309E-01 3.627079071 –.5451966037 –.4391948251 –2.151267528 –.6553497859 –2.259043312 .3734234213 2.275626220 2.967443231 3.135859298 .3681322160 2.894876567 2.955244816 2.392078334 1.925173109 .7304181048 .8143265570 –1.940373885 1.255213076 .7314593647 1.977469640 .4933610021 1.908320234 –1.359187413 –3.064293462 –1.361807101 –1.732672415 .3930657130 –1.025271643 –.8483104009 –.1972843768 –.1920736328 –2.137837711 –3.233501541 –.3307793842 –.1173706063 .5948420239E-01 .2529261888	.28468042E-01 .28334912E-01 .42099198 .26527633 .31534570 .26936526 .37048270 .28231601 .65712883 .95894578 .39907456 .42692994 .31031886 .33906906 .31407062 .27723297 .37624410 .92093838 1.4687860 .29745830 .11661263 .14783769 .96742995E-01 .17023955 .10094571 .37306002 .66556793 .26034184 .29415610 .15841300 .21884004 .20040433 .16042235 .18523863 .53128271 .89457981 .88126654E-01 .38560643E-01 .94865464E-01 .46650226E-01	–2.569 –3.060 8.616 –2.055 –1.393 –7.986 –1.769 –8.002 .568 2.373 7.436 7.345 1.186 8.538 9.409 8.628 5.117 .793 .554 –6.523 10.764 4.948 20.440 2.898 18.904 –3.643 –4.604 –5.231 –5.890 2.481 –4.685 –4.233 –1.230 –1.037 –4.024 –3.615 –3.753 –3.044 .627 5.422	.0102 .0022 .0000 .0399 .1637 .0000 .0769 .0000 .5699 .0176 .0000 .0000 .2355 .0000 .0000 .0000 .0000 .4277 .5793 .0000 .0000 .0000 .0000 .0038 .0000 .0003 .0000 .0000 .0000 .0131 .0000 .0000 .2188 .2998 .0001 .0003 .0002 .0023 .5306 .0000	.64376427E-13 .17360728E-13 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .12908497 1.2690632 .39215686E-01 .26851852

E2.2 Enhanced Baseline Model based on All Shipments
The next step is now familiar. We start with the baseline model, add individual interdiction programs to that model, and test for the statistical significance of these added programs. Findings are summarized in Table E5, which is the counterpart to Table E2. Differences arise because of the change in the dependent variables, from shipments with intelligence in Table E2 to all shipments in Table E5.

Eight individual programs now pass the test for statistical significance. That is, for those eight, the contrast is positive and statistically significant at better than P=0.10. We find a perverse negative contrast (statistically significant) in two other programs.

E2.3 Enhanced Regression Model based on All Shipments
As a final step, we estimated a regression model that included all the variables from the baseline model and all the variables that resulted in statistically significant contrast as reported in table E6. Table E5 summarizes the Wald tests in columns four and five.

Seven of the eight contrasts that were positive and significant in section E2.2 remain positive and significant in these regressions. The eight remains positive but does not quite reach statistical significance. Although two of the contrasts are negative, neither is statistically significant. These regression results seem to lend weight to a conclusions that interdiction has caused traffickers to shift activity away from vectors/modes that are the focus of interdiction to vectors/modes that are not targets.

Table E5: Adding Interdictions Programs to the Baseline Model—All Shipments

	Contrast from Adding Variables	P Value	Contrast from Regression	P-Value
Hard Line Two Dozen Zorro II Gateway White Shark I White Shark II Frontier Shield Border Shield* Gulf Shield* White Shark III Brass Ring River Sweep Frontier Lance Conjuntos I and II	0.69 2.08 –0.24 0.21 –0.44 –0.97 –0.34 3.01 1.33 1.14 –0.08 1.03 0.52 0.11	0.00 0.02 0.15 0.07 0.02 0.04 0.15 0.01 0.00 0.03 0.80 0.00 0.08 0.38	0.67 1.69 0.20 0.08 –0.32 –0.57 0.93 0.93 0.97 0.91 0.43	0.00 0.05 0.09 0.36 0.50 0.20 0.00 0.00 0.06 0.00 0.13

Table E6: Adding Interdiction Programs to the Baseline Regressions—All Shipments

Poisson Regression Maximum Likelihood Estimates Dependent variable Weighting variable Number of observations Iterations completed Log likelihood function Restricted log likelihood Chi-squared Degrees of freedom Significance level N_SHIP ONE 1836 9 2612.317 3769.244 2313.854 59 .0000000 Chi- squared = 2845.29768 RsqP= .5335 G - squared = 2612.33601 RsqD= .4697

Variable	Coefficient	Standard Error	b/St.Er.	P[|Z|>z]	Mean of X
CYCLE1 CYCLE2 TVM1 TVM2 TVM3 TVM4 TVM5 TVM6 TVM7 TVM10 TVM11 TVM12 TVM13 TVM14 TVM15 TVM16 TVM17 TVM18 TVM19 VM1 VM2 VM3 VM4 VM5 VM6 VM7 VM10 VM11 VM12 VM13 VM14 VM15 VM16 VM17 VM18 VM19 INT1 NINT1 INT2	–.5640615585E-01 –.1479483199 4.457587986 –.2748442050E-01 .9475163611E-02 –1.704266089 –.2329633852 –1.818364138 .9845031877 2.868978150 4.857922644 4.013029303 .9047154776 3.732514166 3.751430735 3.175529131 4.573977668 2.215618002 1.437452861 –2.267235978 1.116537858 .6012671359 1.900828964 .3755144049 1.833650258 –1.542086790 –3.204949585 –2.164431723 –2.067059149 .2209788722 –1.380887350 –1.174544739 –.5175372470 –1.222098927 –2.609563540 –3.428420810 .3706183697 .1672939725 .3707689356E-01	.29706813E-01 .31431251E-01 .46212290 .29681085 .34780173 .30105202 .40466252 .31311502 .69885317 .97679877 .56777459 .46922700 .34623356 .38557568 .36147027 .32569314 .65624128 1.1089964 1.5385191 .31224141 .12339469 .15548342 .10048114 .17821032 .10461724 .38847407 .67094037 .33260189 .30915981 .16793808 .23640654 .21749955 .17670733 .28433774 .60044753 .92673218 .13054511 .86117880E-01 .34592276	–1.899 –4.707 9.646 –.093 .027 –5.661 –.576 –5.807 1.409 2.937 8.556 8.552 2.613 9.680 10.378 9.750 6.970 1.998 .934 –7.261 9.049 3.867 18.917 2.107 17.527 –3.970 –4.777 –6.508 –6.686 1.316 –5.841 –5.400 –2.929 –4.298 –4.346 –3.699 2.839 1.943 .107	.0576 .0000 .0000 .9262 .9783 .0000 .5648 .0000 .1589 .0033 .0000 .0000 .0090 .0000 .0000 .0000 .0000 .0457 .3501 .0000 .0000 .0001 .0000 .0351 .0000 .0001 .0000 .0000 .0000 .1882 .0000 .0000 .0034 .0000 .0000 .0002 .0045 .0521 .9146	.64376427E-13 .17360728E-13 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .29684096E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .58823529E-01 .26143791E-01 .84967320E-01 .76252723E-02

Variable	Coefficient	Standard Error	b/St.Er.	P[|Z|>z]	Mean of X
NINT2 TWODOZ NTWODOZ HARDLNE WHITESI WHITEII WHITIII BORGLF RIVSWP NHARDLNE NWHI NWHII NWHIII NBORGLF NRIVSWP GATEWAY NGATEWAY SHIELD NSHIELD FTRLANC NFTRLANC	.1964373102 –1.434637652 .2526764629 –.7165552386 .3906454009 –.6283123549 –.7922272362 –.9902991675 –1.617492723 –.4478952241E-01 .4749202867 –.9528289574 .1799395180 –.6261011307E-01 –.7097157269 –.6500331426 –.4505055424 .3461122629 –.2241198839 –.1150491404 .3173690780	.15732225 1.0042811 .21213183 .19485292 .27508759 .43500540 .61166226 .32608436 .28849429 .77432264E-01 .18289456 .21420069 .13952638 .93145441E-01 .20893499 .22464778 .18734259 .41005949 .22290290 .38554620 .16936055	1.249 –1.429 1.191 –3.677 1.420 –1.444 –1.295 –3.037 –5.607 –.578 2.597 –4.448 1.290 –.672 –3.397 –2.894 –2.405 .844 –1.005 –.298 1.874	.2118 .1531 .2336 .0002 .1556 .1486 .1952 .0024 .0000 .5630 .0094 .0000 .1972 .5015 .0007 .0038 .0162 .3986 .3147 .7654 .0609	.38671024E-01 .21786492E-02 .92592593E-02 .31045752E-01 .70806100E-02 .10893246E-02 .16339869E-02 .37037037E-01 .25054466E-01 .24673203 .11328976 .17429194E-01 .26143791E-01 .59259259 .18790850 .26143791E-01 .19607843 .65359477E-02 .21241830E-01 .32679739E-02 .24509804E-01

E3.0 The Amount of Drugs Transported

The final part of this analysis was to treat the amount shipped as the dependent variable. For this analysis, we excluded vectors/modes that had zero shipments. Here the question is whether or not shipments, measured as the bulk amount of cocaine transported, changed conditional on the number of shipments that were actually made.

We do not show details. Four of the interdiction events seemed to have reduced the bulk amount of cocaine moving through targeted vector/mode combinations. A Fifth program almost reached statistical significance (P=0.103). One program had a perverse negative effective. These findings reinforce the earlier findings that were based on the number of movements.

Download Adobe Acrobat Reader