Mercury in Florida's Bass

The State of Florida has set a standard of 1/2 part per million as the unsafe level of the mercury concentration in edible foods. What percentage of the lakes in this study have standardized mercury concentrations that would be considered unsafe by the state of Florida? Would it be reasonable to use this as an estimate of the percentage of all the lakes in Florida that exceed the declared safety level? Why or why not?

 

• Understand the distinction between a population and a sample.

Twenty-four of the 53 (45.3%) lakes in the study had standardized values of at least 1/2 part per million (this includes the 10 lakes for which the ages of the fish were not determined). It is not reasonable to use this as an estimate of the parameter because probability methods were not used to pick the sample. It is also not appropriate to estimate the standard error of the estimated percentage for this judgment sample.

In order to study the accuracy of the technique that was used to analyze mercury concentration, the researchers measured 117 "spiked" fish with a known amount of mercury. The results were an average of 3.9% above the known values with an SD of 9.9%.

Suppose a single fish with a 1 part per million mercury concentration is measured by this method. This measurement is likely to be around ______ give or take ______. Next, suppose this same fish is measured 25 times. The average of these 25 measurements should come out around ______ give or take ______. Fill in the blanks and explain.

• Understand that repeated measurements cannot help you judge the level of bias—for that you need an external standard.
  

The researchers’ study of the quality of their measurement technique shows that it has a bias of about 3.9% and a chance error of about 9.9% of the actual value. A single measurement on a fish with a 1 part per million mercury concentration is likely to come out around 1.039 parts per million give or take about 0.099 parts per million. The average of 25 independent measurements of this same fish should come out around 1.039 parts per million give or take about 0.0198 (= 0.099/25^1/2) parts per million.

The smallest level of mercury concentration that the measurement technique can detect is 40 parts per billion. Concentrations below this detection limit were reported as 40 parts per billion. How will this affect the standardized mercury concentrations that were used for the final analysis?

• Understand that repeated measurements cannot help you judge the level of bias—for that you need an external standard.

Since these extremely low observations are most likely to occur on younger than average fish, using 40 parts per billion instead of the true lower value will cause the regression line to have a more shallow slope (this is the effect of moving a point in the bottom left of the scatter diagram vertically upward). Thus, young fish with very low mercury concentrations will bring the standardized value closer to the average value for the fish sampled from the lake.

For each of the 53 lakes studied, the researchers reported the minimum and maximum mercury concentration found in the sampled fish. Without looking at the data, suggest a function of these two quantities that should have a strong relation with the number of fish taken from the lake. Explain the logic behind your suggestion and then check if it is verified in this data set.

 

• Be able to use your software to make a scatterplot.
  Be able to interpret the display (for example, recognizing patterns or spotting outliers).
• Be able to use software to calculate the correlation coefficient.  Be able to interpret this value.

For a single lake, as more fish are sampled, new minimums and maximums will be achieved causing the range of observations to grow. Also, lakes with generally higher values will have both higher minimums and higher maximums. Thus, one reasonable suggestion is that the range divided by the standardized level should be strongly associated with the sample size. Checking this suggestion with the data, it is seen that the association is not strong (the rank correlation is about 0.45).

Make scatterplots of the standardized mercury concentration versus the water chemistry variables. Which of the water chemistry variables has the strongest association with the mercury concentration of the fish in the lakes? Which of the associations appear to be linear? Which of the relationships are homoscedastic?

• Be able to use your software to make a scatterplot.
  Be able to interpret the display (for example, recognizing patterns or spotting outliers).
• Be able to interpret scatterplots including identifying:
• the distinction between response and explanatory variables.
• the form of association and correlation between the response and explanatory variables.

In the plots that follow, alkalinity appears to have the strongest association with the standardized mercury values, although the relationship is non-linear. Only the association with pH appears linear. Only the relationship with alkalinity appears reasonably homoscedastic (particularly on a log scale).

Examine the regression of the logarithm of the standardized mercury concentration on alkalinity.

a) Are the residuals homoscedastic? Do they look like they follow a normal distribution?

b) Plot the residuals versus the other water chemistry variables. Are there any strong relationships? What does this tell you?

c) Does this regression allow you to say anything about the ages of the fish in the 10 lakes where age wasn't measured?

• Understand how to check the key model assumptions that are made in regression using residual plots and other model diagnostics.
• Understand that simple linear regression can be extended to multiple regression and be able to interpret the following in the context of multiple regression (y = response variable, = estimated response, and x’s = the explanatory variables).

 

a) The regression shows that the Alkalinity level of the lake helps to predict the standardized mercury concentration of the fish for the lakes in this study. The residuals are somewhat heteroscedastic with three or four outliers whose residuals are larger in magnitude than would be expected for a normal distribution.

 

b) The residual plots below indicate the residuals do not appear to be related to the pH or calcium levels in the lakes but there does seem to be a moderate negative association with the chlorophyll level. This tells us that we might improve our prediction of the standardized mercury concentration by including chlorophyll as an explanatory variable in our regression equation.

In fact the value of R² rises from 54% to 75% with the inclusion of chlorophyll. [Note: interestingly, the paper by Lange, Royals, and Connor (1993) found these same two explanatory variables (alkalinity and chlorophyll) as the best two-variable model but did not recognize that using the log of the standardized values gave a superior fit (R² of 75% instead of 45%).]

 

c) Yes, some of the variability left in the standardized values for the ten lakes where age was not measured is likely to be due to using the mean mercury concentration instead of the regression-based estimate for a 3 year old fish. Thus, there is likely to be a correlation between the age of the fish in the lakes and the residuals from our best model. It is likely that lake #2 (with the largest positive residual) has older fish than lake #15 (with the smallest negative residual).

In which lakes was the average age of the sampled fish greater than the three-year-old standard? In which lakes was it less?

• Be able to use your software to make a scatterplot.
  Be able to interpret the display (for example, recognizing patterns or spotting outliers).
• Be able to interpret scatterplots including identifying:
• the distinction between response and explanatory variables.
• the form of association and correlation between the response and explanatory variables.

Since the correlation between mercury concentration and age is positive, the lakes with a sample average concentration that is larger than the 3 year standard concentration (a regression estimate) must have an average age that is larger than 3 years. Similarly when the sample average concentration is less than the standardized value, the average age must be below 3 years. The ID numbers of the lakes with the older fish (blue o's) are: 6, 7, 8, 9, 14, 18, 19, 20, 21, 26, 27, 31, 35, 36, 40, 43, 44, 45, 49, 51, and 53. The ID numbers of the lakes with the younger fish (red x's) are: 1, 5, 12, 13, 22, 24, 25, 28, 30, 32, 37, 39, 41, 42, 46, 48, and 52.

Researchers Canfield and Hoyer (1988) found that pH and alkalinity generally increase in Florida's lakes as you go from the Northwest to the Southeast and from highland to coastal areas of the state. Does this claim seem to be verified by the 53 lakes (see map below) studied in the article by Lange, Royals and Connor?

• Understand Numerical Properties:
• Correlation has a value between –1 and 1.
• When the correlation is 1 the points on the scatterplot fall exactly on a straight line with positive slope.
• Understand that positive correlation means that when one variable increases so does the other on average.
• When the correlation is –1 the points on the scatterplot fall exactly on a straight line with negative slope.
• Understand that negative correlation means that when one variable increases the other decreases on average.
• Correlation has no units and does not depend on the choice of units for x and y.

The claim seems to be generally true, although this association with geography is not strong.