22.07.2013 Views

Principles of Fluorescence Spectroscopy

Principles of Fluorescence Spectroscopy

Principles of Fluorescence Spectroscopy

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

132 TIME-DOMAIN LIFETIME MEASUREMENTS<br />

predict the probability for obtaining a value <strong>of</strong> χ R 2 because<br />

<strong>of</strong> random errors. These values can be found in standard<br />

mathematical tables <strong>of</strong> the χ R 2 distribution. Selected values<br />

are shown in Table 4.2 for various probabilities and numbers<br />

<strong>of</strong> degrees <strong>of</strong> freedom. Suppose you have over 200 datapoints<br />

and the value <strong>of</strong> χ R 2 = 1.25. There is only a 1%<br />

chance (P = 0.01) that random errors could result in this<br />

value. Then the model yielding χ R 2 = 1.25 can be rejected,<br />

assuming the data are free <strong>of</strong> systematic errors. If the value<br />

<strong>of</strong> χ R 2 is 1.08, then there is a 20% chance that this value is<br />

due to random deviations in the data. While this may seem<br />

like a small probability, it is not advisable to reject a model<br />

if the probability exceeds 5%, which corresponds to χ R 2 =<br />

1.17 (Table 4.2). In our experience, systematic errors in the<br />

data can easily result in a 10–20% elevation in χ R 2. For<br />

example, suppose the data does contain systematic errors<br />

and a two-component fit returns a value <strong>of</strong> χ R 2 = 1.17. The<br />

data are then fit to three components, which results in a<br />

decreased χ R 2. If the three-component model is accepted as<br />

the correct model, then an incorrect conclusion is reached.<br />

The systematic errors in the data will have resulted in addition<br />

<strong>of</strong> a third component that does not exist in the experimental<br />

system.<br />

While we stated that assumptions 1 through 6 (above)<br />

are generally true for TCSPC, we are not convinced that<br />

number 5 is true. Based on our experience we have the<br />

impression that the TCSPC data have fewer independent<br />

observations (degrees <strong>of</strong> freedom) in the TCSPC data than<br />

the number <strong>of</strong> actual observations (channels). This is not a<br />

criticism <strong>of</strong> NLLS analysis or TCSPC. However, if the<br />

effective number <strong>of</strong> independent datapoints is less than the<br />

number <strong>of</strong> datapoints, then small changes in χ R 2 may not be<br />

as significant as understood from the mathematical tables.<br />

Complete reliance on mathematical tables can lead to<br />

overinterpretation <strong>of</strong> the data. The absolute value <strong>of</strong> χ R 2 is<br />

<strong>of</strong>ten <strong>of</strong> less significance than its relative values. Systematic<br />

errors in the data can easily result in χ R 2 values in excess<br />

Table 4.2. χ R 2 Distribution a<br />

Probability (P)/<br />

degrees <strong>of</strong> freedom 0.2 0.1 0.05 0.02 0.01 0.001<br />

10 1.344 1.599 1.831 2.116 2.321 2.959<br />

20 1.252 1.421 1.571 1.751 1.878 2.266<br />

50 1.163 1.263 1.350 1.452 1.523 1.733<br />

100 1.117 1.185 1.243 1.311 1.358 1.494<br />

200 1.083 b 1.131 1.170 1.216 1.247 1.338<br />

aFrom [1, Table C-4].<br />

b Mentioned in text.<br />

<strong>of</strong> 1.5. This small elevation in χ R 2 does not mean the model<br />

is incorrect. We find that for systematic errors the χ R 2 value<br />

is not significantly decreased using the next more complex<br />

model. Hence, if the value <strong>of</strong> χ R 2 does not decrease significantly<br />

when the data are analyzed with a more complex<br />

model, the value <strong>of</strong> χ R 2 probably reflects the poor quality <strong>of</strong><br />

the data. In general we consider decreases in χ R 2 significant<br />

if the ratio decreases by tw<strong>of</strong>old or more. Smaller changes<br />

in χ R 2 are interpreted with caution, typically based on some<br />

prior understanding <strong>of</strong> the system.<br />

4.9.4. Autocorrelation Function Advanced Topic<br />

Another diagnostic for the goodness <strong>of</strong> fit is the autocorrelation<br />

function. 3 For a correct model, and the absence <strong>of</strong><br />

systematic errors, the deviations are expected to be randomly<br />

distributed around zero. The randomness <strong>of</strong> the deviations<br />

can be judged from the autocorrelation function. Calculation<br />

<strong>of</strong> the correlation function is moderately complex,<br />

and does not need to be understood in detail to interpret<br />

these plots. The autocorrelation function C(t j ) is the extent<br />

<strong>of</strong> correlation between deviations in the k and k+jth channel.<br />

The values <strong>of</strong> C(t j ) are calculated using<br />

C(tj) ( 1<br />

m ∑m DkDkj ) / (<br />

k1<br />

1<br />

n ∑n D<br />

k1<br />

2 k )<br />

(4.23)<br />

where D k is the deviation in the kth datapoint and D k+j is the<br />

deviation in the k+jth datapoint (eq. 4.23). This function<br />

measures whether a deviation at one datapoint (time channel)<br />

predicts that the deviation in the jth higher channel will<br />

have the same or opposite sign. For example, the probability<br />

is higher that channels 50 and 51 have the same sign<br />

than channels 50 and 251. The calculation is usually<br />

extended to test for correlations across half <strong>of</strong> the data<br />

channels (m = n/2) because the order <strong>of</strong> multiplication in

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!