Confidence Intervals

Faye Jones, Resource Geologist

0 11 September, 2006
Introduction
Conditional Simulation currently accepted method for estimating Confidence
Intervals. However..

Time consuming methods often taking days to generate simulations

Complex method to use
Not viable in the majority of cases

Need to find alternative method which produces the reliability of simulation but
within much shorter time frame.

1 11 September, 2006
Confidence Intervals
Confidence Interval (CI) reflect the inability to exactly define an unknown value;
CI = 0: Value is known exactly
CI > 0: Value is not known exactly and the uncertainty increases with magnitude of CI

If the CI is linked to probability, it is possible to estimate the chance of the unknown estimate lying
within a given grade range e.g. 50% probability that the value lies within the range 3g/t +/- 0.8 g/t

Grade estimation methods calculate the value of a given block or node. What Confidence do we
have in that value?

2 11 September, 2006
Calculating Confidence Intervals
Calculated using the following equation:
CI = Upper Limit Lower Limit

Upper and lower limits usually defined in terms of Standard Deviations (SD):
Upper Limit = Mean + nSD
Lower Limit = Mean nSD

1 SD: Defines 60% Confidence Limits

1.96 SD: Defines 95% Confidence Limits

It should therefore be possible to estimate a confidence interval for each block.

However.

3 11 September, 2006
Assumptions
Block estimates are normally distributed This is rarely the case!!
U3O8 (Gaussian): Area Code = 2 U3O8
-4. -3. -2. -1. 0. 1. 2. 3. 4. 0. 1. 2.

0.08 0.08

0.125 0.125
0.07 0.07

0.06 0.06 0.100 0.100

Frequencies
Frequencies

Frequencies
0.05 0.05

Frequencies
0.075 0.075

0.04 0.04

0.03 0.03 0.050 0.050

0.02 0.02

0.025 0.025

0.01 0.01

0.00 -4. -3. -2. -1. 0. 1. 2. 3. 4. 0.00 0.000 0.000

0. 1. 2.
U3O8 (Gaussian): Area Code = 2 U3O8
Isatis Isatis

Requirement Actual

Calculating CIs for skewed data will result in Negative Grades

4 11 September, 2006
 Ordinary Kriging
Provides an estimate of a block value and an indication of the local precision
kriging variance

Kriging variance can be converted in SD and CIs determined. However negative

lower limits are produced.

Kriging Variance is based upon sample distance and does not take into account
the effects of sample distribution

Alternative Method required

5 11 September, 2006
 Lower Confidence Limit via Ordinary Kriging
Lower confidence = kriged estimate 1.96 KSD

X (Meter)
500. 1000. 1500. 2000. 2500. 3000.
1200. 1200. >=1
1100. 1100. 0.9375
0.875
1000. 1000. 0.8125
0.75

Y (Meter)
Y (Meter)

900. 900. 0.6875

800. 800. 0.625
0.5625
700. 700. 0.5
0.4375
600. 600. 0.375
500. 500. 0.3125
0.25
400. 400. 0.1875
0.125
300. 300. 0.0625
500. 1000. 1500. 2000. 2500. 3000. <0
N/A
X (Meter)
Isatis

Note: Many blocks with ve lower values

6 11 September, 2006
Kriging Variance
Consider the 2 blocks, whose value is estimated from the samples located at each corner.
Both blocks have same sample configuration;
The same variogram is used in both cases;
Block 1 Block 2
1.00 2.00 1.00 2.00

2.00 1.00 50.00 1.00

Block 1 estimate = 1.50 Block 2 estimate = 13.25

Both blocks have same kriging variance same confidence!

7 11 September, 2006
Alternative Methods
Conditional Simulation currently accepted method for estimating Confidence
Intervals. However..

Time consuming methods often taking days to generate simulations

Complex method to use
Not viable in the majority of cases

Need to find alternative method which produces the reliability of simulation but
within much shorter time frame.

8 11 September, 2006
Direct Confidence Interval Method
Idea proposed by Armstrong and Roth;

Based upon Simple Kriging but with modifications;

Isatis program modified by Geovariances;

Hard rock metal mine reserve definition drilling used for study;

Estimates generated by DCIM and compared against Conditional Simulation

9 11 September, 2006
 Comparison of Confidence Intervals
X (Meter)
500. 1000. 1500. 2000. 2500. 3000.
1200. 1200.
1100. 1100.

Scatterplot of Simulated Kriged 1000. 1000.

Y (Meter)
Y (Meter)
900. 900.
800. 800.

CI vs DCIM CI 700.
600.
700.
600.
500. 500.
400. 400.
300. 300.
500. 1000. 1500. 2000. 2500. 3000.
U3O8 (New Method) 95% Confidence Interv
X (Meter)
0. 1. 2. Isatis

X (Meter)
500. 1000. 1500. 2000. 2500. 3000.
1200. 1200.
U3O8 (Simulated) 95% Confidence Interva
Simulated 1100.
1000.
1100.
1000.
U3O8 (Simulated) 95% Confidence Interva

Y (Meter)
Y (Meter)
2. 2. 900. 900.
800. 800.
700. 700.
600. 600.
500. 500.
400. 400.
300. 300.
500. 1000. 1500. 2000. 2500. 3000.
X (Meter)
1. 1. Isatis

X (Meter)
500. 1000. 1500. 2000. 2500. 3000.
1200. 1200.
1100. 1100.

DCIM 1000. 1000.

Y (Meter)
Y (Meter)
900. 900.
800. 800.
0. 0. 700. 700.
600. 600.
500. 500.
0. 1. 2.
400. 400.
U3O8 (New Method) 95% Confidence Interv
Isatis 300. 300.
500. 1000. 1500. 2000. 2500. 3000.
X (Meter)
Isatis

10 11 September, 2006
 Comparison of Lower Confidence Limit
X (Meter)
500. 1000. 1500. 2000. 2500. 3000.
1200. 1200.
1100. 1100.
1000. 1000.
Kriged

Y (Meter)

Y (Meter)
900. 900.

Scatterplot of Simulated 800.

700.
600.
800.
700.
600.

CL vs DCIM CL 500.
400.
500.
400.
300. 300.
500. 1000. 1500. 2000. 2500. 3000.
U3O8 (New Method) Kriged Est - 1.96SD ( X (Meter)
0.00 0.25 0.50 0.75 Isatis

X (Meter)
500. 1000. 1500. 2000. 2500. 3000.
1200. 1200.
0.75 0.75
1100. 1100.

Simulation 1000. 1000.

Y (Meter)

Y (Meter)
U3O8 (Simulated)_Z{2.500000}
U3O8 (Simulated)_Z{2.500000}

900. 900.
800. 800.
700. 700.
0.50 0.50
600. 600.
500. 500.
400. 400.
300. 300.
500. 1000. 1500. 2000. 2500. 3000.
0.25 0.25 X (Meter)
Isatis
X (Meter)
500. 1000. 1500. 2000. 2500. 3000.

0.00 0.00
DCIM Y (Meter)
1200.
1100.
1000.
1200.
1100.
1000.

Y (Meter)
0.00 0.25 0.50 0.75
900. 900.
U3O8 (New Method) Kriged Est - 1.96SD (
Isatis 800. 800.
700. 700.
600. 600.
500. 500.
400. 400.
300. 300.
500. 1000. 1500. 2000. 2500. 3000.
X (Meter)
Isatis

11 11 September, 2006
Reliability of Grade Estimates
Simulation and DCIM used to generate Confidence Limits

Block model grade estimates generated by the mine

How do the grade estimates relate to the confidence intervals?

12 11 September, 2006
 Grade vs Confidence Interval
Confidence Interval increases with grade estimate U3O8 - L01 Insitu
U3O8 - L01 Insitu
0.0 0.5 1.0 1.5
3. 3. 0.0 0.5 1.0 1.5
3. 3.

U3O8 (New Method) 95% Confidence Interv

U3O8 (New Method) 95% Confidence Interv

U3O8 (Simulated) 95% Confidence Interva

U3O8 (Simulated) 95% Confidence Interva
2. 2.
2. 2.

1. 1.
1. 1.

0. 0. 0. 0.

0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5

U3O8 - L01 Insitu U3O8 - L01 Insitu
Isatis Isatis

DCIM Confidence Interval vs Simulation Confidence Interval vs

Block Estimate Block Estimate

13 11 September, 2006
 Iron Ore
Wide spaced exploration data available
Conditional Simulations already in existence

14 11 September, 2006
95% Confidence Limits

Simulated DCIM

15 11 September, 2006
 Probabilities
Simulations can be used to estimate the probability of a block grade
exceeding a certain value requires at least 100 simulations
DCIM adapted to allow probabilities to be calculated directly
Simulation Direct Approach

16 11 September, 2006
Generating Confidence Intervals
Process:
Convert the input sample data into Gaussian values;
Calculate the change of support for the points/blocks;
Calculate the gaussian block variogram from the raw point variogram and model;
Convert the point values into pseudo block gaussian values via the change of support
and migrate the points to the required block centres;
Krige the gaussian block values and determine the gaussian kriging variance
Back transform the gaussian values into raw values

17 11 September, 2006
Confidence Intervals - Program
Input requirements
Gaussian values
Block anamorphosis
Block gaussian variogram

18 11 September, 2006
Calculation of Probabilities
Input gaussian values derived from
CI program (Kriged estimate & KSD)

Indicators to be supplied

Gaussian back transformation to

convert gaussian to raw values

19 11 September, 2006
 Direct calculation of Confidence intervals
Advantages & Disadvantages
Advantages
Provides indication of potential reliability of estimates
Direct approach is much quicker to run than simulation
Produces results which are similar to those generated by simulation

Disadvantages
Reliance upon a good variogram model
Number of assumptions made about the data during the calculation process

However all tests to date provide comparable results

20 11 September, 2006
Further Work

Additional studies completed

Iron Ore Deposit, Western Australia (MSc Thesis Faye Jones)

Hard Rock Metal Deposit, Middle East (OTX & MSc Thesis George Gestrich)
Comparing confidence intervals from simulations and new method
Application to impact on mine planning
Pit optimisation using Whittle.

21 11 September, 2006
New Developments
Recent addition of bivariate
confidence intervals calculation

Probability that two variables lie

within given limits, e.g. Fe and
Silica

Input for each variable:

kriged block gaussian values
block anamorphosis

Future development of use for

multivariate case

22 11 September, 2006