The Winter 2017 edition of Foresight magazine includes my commentary on the article Changing the Paradigm for Business Forecasting by Michael Gilliland from SAS. Both are behind a paywall (though a longer version of Michael’s argument can be read on his SAS blog), but here is a brief summary.
According to Gilliland, business forecasting is currently dominated by an “offensive” paradigm, which is “characterized by a focus on models, methods, and organizational processes that seek to extract every last fraction of accuracy from our forecasts. More is thought to be better—more data, bigger computers, more complex models—and more elaborate collaborative processes.”
He argues that our “love affair with complexity” can lead to extra effort and cost, while actually reducing forecast accuracy. And while managers have often been seduced by the idea that “big data was going to solve all our forecasting problems”, research shows that even with complex models, forecast accuracy often fails to beat even a no-change forecasting model. His article therefore advocates a paradigm shift towards “defensive” forecasting, which focuses on simplifying the forecasting process, eliminating bad practices, and adding value.
My comment on this (in about 1200 words) is … I agree. But I would argue that the problem is less big data, or even complexity, than big theory.
Our current modelling paradigm is fundamentally reductionist – the idea is to reduce a system to its parts, figure out the laws that govern their interactions, build a giant simulation of the whole thing, and solve. The resulting models are highly complex, and their flexibility makes them good at fitting past data, but they tend to be unstable (or stable in the wrong way) and are poor at making predictions.
If however we recognise that complex systems have emergent properties that resist a reductionist approach, it makes more sense to build models that only attempt to capture some aspect of the system behaviour, instead of reproducing the whole thing.
An example of this approach, discussed earlier on this blog, relates to the question of predicting heart toxicity for new drug compounds, based on ion channel readings. One way to predict heart toxicity based on these test results is to employ teams of researchers to build an incredibly complicated mechanistic model of the heart, consisting of hundreds of differential equations, and use the ion channel inputs as inputs. Or you can use a machine learning model. Or, most complicated, you can combine these in a multi-model approach. However Hitesh Mistry found that a simple model, which simply adds or subtracts the ion channel readings – the only parameters are +1 and -1 – performs just as well as the multi-model approach using three large-scale models plus a machine learning model (see Complexity v Simplicity, the winner is?).
Now, to obtain the simple model Mistry used some fairly sophisticated data analysis tools. But what counts is not the complexity of the methods, but the complexity of the final model. And in general, complexity-based models are often simpler than their reductionist counterparts.
I therefore strongly agree with Michael Gilliland that a “defensive” approach makes sense. But I think the paradigm shift he describes is part of, or related to, a move away from reductionist models, which we are realising don’t work very well for complex systems. With this new paradigm, models will be simpler, but they can also draw on a range of techniques that have developed for the analysis of complex systems.
In a previous blog post we highlighted the pitfalls of applying null hypothesis testing to simulated data, see here. We showed that modellers applying null hypothesis testing to simulated data can control the p-value because they can control the sample size. Thus it’s not a great idea to analyse simulations using null hypothesis tests, instead modellers should focus on the size of the effect. This problem has been highlighted before by White et al. which is well worth a read, see here.
Why are we blogging about this subject again? Since that last post, co-authors of the original article we discussed there have repeated the same misdemeanour (Liberos et al., 2016), and a group of mathematical oncologists based at Moffitt Cancer Center has joined them (Kim et al., 2016).
The article by Kim et al., preprint available here, describes a combined experimental and modelling approach that “predicts” new dosing schedules for combination therapies that can delay onset of resistance and thus increase patient survival. They also show how their approach can be used to identify key stratification factors that can determine which patients are likely to do better than others. All of the results in the paper are based on applying statistical tests to simulated data.
The first part of the approach taken by Kim et al. involves calibrating a mathematical model to certain in-vitro experiments. These experiments basically measure the number of cells over a fixed observation time under 4 different conditions: control (no drug), AKT inhibitor, Chemotherapy and Combination (AKT/Chemotherapy). This was done for two different cell lines. The authors found a range of parameter values when trying to fit their model to the data. From this range they took forward a particular set, no real justification as to why that certain set, to test the model’s ability to predict different in-vitro dosing schedules. Unsurprisingly the model predictions came true.
After “validating” their model against a set of in-vitro experiments the authors proceed to using the model to analyse retrospective clinical data; a study involving 24 patients. The authors acknowledge that the in-vitro system is clearly not the same as a human system. So to account for this difference they perform an optimisation method to generate a humanised model. The optimisation is based on a genetic algorithm which searched the parameter space to find parameter sets that replicate the clinical results observed. Again, similar to the in-vitro situation, they found that there were multiple parameter sets that were able to replicate the observed clinical results. In fact they found a total of 3391 parameter sets.
Having now generated a distribution of parameters that describe patients within the clinical study they are interested in, the authors next set about generating stratification factors. For each parameter set the virtual patient exhibits one of four possible response categories. Therefore for each category a distribution of parameter values exists for the entire population. To assess the difference in the distribution of parameter values across the categories they perform a students t-test to ascertain whether the differences are statistically significant. Since they can control the sample size the authors can control the standard error and p-value, this is exactly the issue raised by White et al. An alternative approach would be to state the difference in the size of the effect, so the difference in means of the distributions. If the claim is that a given parameter can discriminate between two types of responses then a ROC AUC (Receiver Operating Characteristic Area Under Curve) value could be reported. Indeed a ROC AUC value would allow readers to ascertain the strength of a given parameter in discriminating between two response types.
The application of hypothesis testing to simulated data continues throughout the rest of the paper, culminating in applying a log-rank test to simulated survival data, where again they control the sample size. Furthermore, the authors choose an arbitrary cancer cell number which dictates when a patient dies. Therefore they have two ways of controlling the p-value. In this final act the authors again abuse the use of null hypothesis testing to show that the schedule found by their modelling approach is better than that used in the actual clinical study. Since the major results in the paper have all involved this type of manipulation, we believe they should be treated with extreme caution until better verified.
Liberos, A., Bueno-Orovio, A., Rodrigo, M., Ravens, U., Hernandez-Romero, I., Fernandez-Aviles, F., Guillem, M.S., Rodriguez, B., Climent, A.M., 2016. Balance between sodium and calcium currents underlying chronic atrial fibrillation termination: An in silico intersubject variability study. Heart Rhythm 0. doi:10.1016/j.hrthm.2016.08.028
White, J.W., Rassweiler, A., Samhouri, J.F., Stier, A.C., White, C., 2014. Ecologists should not use statistical significance tests to interpret simulation model results. Oikos 123, 385–388. doi:10.1111/j.1600-0706.2013.01073.x
Kim, E., Rebecca, V.W., Smalley, K.S.M., Anderson, A.R.A., 2016. Phase i trials in melanoma: A framework to translate preclinical findings to the clinic. Eur. J. Cancer 67, 213–222. doi:10.1016/j.ejca.2016.07.024
A lot has changed in Seattle in the last 15 years. The area around South Lake Union, near where I lived, has been turned into a major hub for biotechnology and the life sciences. Amazon is constructing a new campus featuring giant ‘biospheres’ which look like nothing I have ever seen.
Attending the conference, though, was like a blast from the past – because unlike the models used by architects to design their space-age buildings, the models used in pharmacology have barely moved on.
While there were many interesting and informative presentations and posters, most of these involved relatively simple models based on ordinary differential equations, very similar to the ones we were developing at the ISB years ago. The emphasis at the conference was on using models to graphically present relationships, such as the interaction between drugs when used in combination, and compute optimal doses. There was very little about more modern techniques such as machine learning or data analysis.
There was also little interest in producing models that are truly predictive. Many models were said to be predictive, but this just meant that they could reproduce some kind of known behaviour once the parameters were tweaked. A session on model complexity did not discuss the fact, for example, that complex models are often less predictive than simple models (a recurrent theme in this blog, see for example Complexity v Simplicity, the winner is?). Problems such as overfitting were also not discussed. The focus seemed to be on models that are descriptive of a system, rather than on forecasting techniques.
The reason for this appears to come down to institutional effects. For example, models that look familiar are more acceptable. Also, not everyone has the skills or incentives to question claims of predictability or accuracy, and there is a general acceptance that complex models are the way forward. This was shown by a presentation from an FDA regulator, which concentrated on models being seen as gold-standard rather than accurate (see our post on model misuse in cardiac models).
Pharmacometrics is clearly a very conservative area. However this conservatism means only that change is delayed, not that it won’t happen; and when it does happen it will probably be quick. The area of personalized medicine, for example, will only work if models can actually make reliable predictions.
As with Seattle, the skyline may change dramatically in a very short time.
The title of this blog entry refers to a letter published in the journal entitled, CPT: Pharmacometrics & Systems Pharmacology. The letter is open-access so those of you interested can read it online here. In this blog entry we will go through it.
The letter discusses a rather strange modelling practice which is becoming the norm within certain modelling and simulation groups in the pharmaceutical industry. There has been a spate of publications citing that tumour re-growth rate (GR) and time to tumour re-growth (TTG), derived using models to describe imaging time-series data, correlates to survival [1-6]. In those publications the authors show survival curves (Kaplan-Meiers) highlighting a very strong relationship between GR/ TTG and survival. They either split on the median value of GR/TTG or into quartiles and show very impressive differences in survival times between the groups created; see Figure 2 in  for an example (open access).
Do these relationships seem too good to be true? In fact they may well be. In order to derive GR/TTG you need time-series data. The value of these covariates are not known at the beginning of the study, and only become available after a certain amount of time has passed. Therefore this type of covariate is typically referred to as a time-dependent covariate. None of the authors in [1-6] describe GR/TTG as a time-dependent covariate nor treat it as such.
When the correlations to survival were performed in those articles the authors assumed that they knew GR/TTG before any time-series data was collected, which is clearly not true. Therefore survival curves, such as Figure 2 in , are biased as they are based on survival times calculated from study start time to time of death, rather than time from when GR/TTG becomes available to time of death. Therefore, the results in [1-6] should be questioned and GR/TTG should not be used for decision making, as the question around whether tumour growth rate correlates to survival is still rather open.
Could it be the case that the GR/TTG correlation to survival is just an illusion of a flawed modelling practice? This is what we shall answer in a future blog-post.
 W.D. Stein et al., Other Paradigms: Growth Rate Constants and Tumor Burden Determined Using Computed Tomography Data Correlate Strongly With the Overall Survival of Patients With Renal Cell Carcinoma, Cancer J (2009)
 W.D. Stein, J.L. Gulley, J. Schlom, R.A. Madan, W. Dahut, W.D. Figg, Y. Ning, P.M. Arlen, D. Price, S.E. Bates, T. Fojo, Tumor Regression and Growth Rates Determined in Five Intramural NCI Prostate Cancer Trials: The Growth Rate Constant as an Indicator of Therapeutic Efficacy, Clin. Cancer Res. (2011)
 W.D. Stein et al., Tumor Growth Rates Derived from Data for Patients in a Clinical Trial Correlate Strongly with Patient Survival: A Novel Strategy for Evaluation of Clinical Trial Data, The Oncologist. (2008)
 K. Han, L. Claret, Y. Piao, P. Hegde, A. Joshi, J. Powell, J. Jin, R. Bruno, Simulations to Predict Clinical Trial Outcome of Bevacizumab Plus Chemotherapy vs. Chemotherapy Alone in Patients With First-Line Gastric Cancer and Elevated Plasma VEGF-A, CPT Pharmacomet. Syst. Pharmacol. (2016)
 J. van Hasselt et al., Disease Progression/Clinical Outcome Model for Castration-Resistant Prostate Cancer in Patients Treated With Eribulin, CPT Pharmacomet. Syst. Pharmacol. (2015)
 L. Claret et al., Evaluation of Tumor-Size Response Metrics to Predict Overall Survival in Western and Chinese Patients With First-Line Metastatic Colorectal Cancer, J. Clin. Oncol. (2013)
I recently published a letter with the above title in the journal of Clinical Pharmacology and Therapeutics; unfortunately it’s behind a paywall so I will briefly take you through the key point raised. The letter describes a specific prediction problem around drug induced cardiac toxicity mentioned in a previous blog entry (Mathematical models for ion-channel cardiac toxicity: David v Goliath). In short what we show in the letter is that a simple model using subtraction and addition (pre-school Mathematics) performs just as well for a given prediction problem as a multi-model approach using three large-scale models consisting of 100s of differential equations combined with machine learning approach (University level Mathematics and Computation)! The addition/subtraction model gave a ROC AUC of 0.97 which is very similar to multi-model/machine learning approach which gave a ROC AUC of 0.96. More detail on the analysis can be found on slides 17 and 18 within this presentation, A simple model for ion-channel related cardiac toxicity, which was given at an NC3Rs meeting.
The result described in the letter and presentation continues to add weight within that field that simple models are performing just as well as complex approaches for a given prediction task.
The standard way of answering this question is to ask whether the effect could reasonably have happened by chance (the null hypothesis). If not, then the result is announced to be ‘significant’. The usual threshold for significance is that there is only a 5 percent chance of the results happening due to purely random effects.
This sounds sensible, and has the advantage of being easy to compute. Which is perhaps why statistical significance has been adopted as the default test in most fields of science. However, there is something a little confusing about the approach; because it asks whether adopting the opposite of the theory – the null hypothesis – would mean that the data is unlikely to be true. But what we want to know is whether a theory is true. And that isn’t the same thing.
As just one example, suppose we have lots of data and after extensive testing of various theories we discover one that passes the 5 percent significance test. Is it really 95 percent likely to be true? Not necessarily – because if we are trying out lots of ideas, then it is likely that we will find one that matches purely by chance.
While there are ways of working around this within the framework of standard statistics, the problem usually gets glossed over in the vast majority of textbooks and articles. So for example it is typical to say that a result is ‘significant’ without any discussion of whether it is plausible in a more general sense (see our post on model misuse in cardiac modeling).
The effect is magnified by publication bias – try out multiple theories, find one that works, and publish. Which might explain why, according to a number of studies (see for example here and here), much scientific work proves impossible to replicate – a situation which scientist Robert Matthews calls a ‘scandal of stunning proportions’ (see his book Chancing It: The laws of chance – and what they mean for you).
The way of Bayes
An alternative approach is provided by Bayesian statistics. Instead of starting with the assumption that data is random and making weird significance tests on null hypotheses, it just tries to estimate the probability that a model is right (i.e. the thing we want to know) given the complete context. But it is harder to calculate for two reasons.
One is that, because it treats new data as updating our confidence in a theory, it also requires we have some prior estimate of that confidence, which of course may be hard to quantify – though the problem goes away as more data becomes available. (To see how the prior can affect the results, see the BayesianOpionionator web app.) Another problem is that the approach does not treat the theory as fixed, which means that we may have to evaluate probabilities over whole families of theories, or at least a range of parameter values. However this is less of an issue today since the simulations can be performed automatically using fast computers and specialised software.
Perhaps the biggest impediment, though, is that when results are passed through the Bayesian filter, they often just don’t seem all that significant. But while that may be bad for publications, and media stories, it is surely good for science.
A common critique of biologists, and scientists in general, concerns their occasionally overenthusiastic tendency to find patterns in nature – especially when the pattern is a straight line. It is certainly notable how, confronted with a cloud of noisy data, scientists often manage to draw a straight line through it and announce that the result is “statistically significant”.
Straight lines have many pleasing properties, both in architecture and in science. If a time series follows a straight line, for example, it is pretty easy to forecast how it should evolve in the near future – just assume that the line continues (note: doesn’t always work).
However this fondness for straightness doesn’t always hold; indeed there are cases where scientists prefer to opt for a more complicated solution. An example is the modelling of tumour growth in cancer biology.
Tumour growth is caused by the proliferation of dividing cells. For example if cells have a cell cycle length td, then the total number of cells will double every td hours, which according to theory should result in exponential growth. In the 1950s (see Collins et al., 1956) it was therefore decided that the growth rate could be measured using the cell doubling time.
In practice, however, it is found that tumours grow more slowly as time goes on, so this exponential curve needed to be modified. One variant is the Gompertz curve, which was originally derived as a model for human lifespans by the British actuary Benjamin Gompertz in 1825, but was adapted for modelling tumour growth in the 1960s (Laird, 1964). This curve gives a tapered growth rate, at the expense of extra parameters, and has remained highly popular as a means of modelling a variety of tumour types.
However, it has often been observed empirically that tumour diameters, as opposed to volumes, appear to grow in a roughly linear fashion. Indeed, this has been known since at least the 1930s. As Mayneord wrote in 1932: “The rather surprising fact emerges that the increase in long diameter of the implanted tumour follows a linear law.” Furthermore, he noted, there was “a simple explanation of the approximate linearity in terms of the structure of the sarcoma. On cutting open the tumour it is often apparent that not the whole of the mass is in a state of active growth, but only a thin capsule (sometimes not more than 1 cm thick) enclosing the necrotic centre of the tumour.”
Because only this outer layer contains dividing cells, the rate of increase for the volume depends on the doubling time multiplied by the volume of the outer layer. If the thickness of the growing layer is small compared to the total tumour radius, then it is easily seen that the radius grows at a constant rate which is equal to the doubling time multiplied by the thickness of the growing layer. The result is a linear growth in radius. This translates to cubic growth in volume, which of course grows more slowly than an exponential curve at longer times – just as the data suggests.
In other words, rather than use a modified exponential curve to fit volume growth, it may be better to use a linear equation to model diameter. This idea that tumour growth is driven by an outer layer of proliferating cells, surrounding a quiescent or necrotic core, has been featured in a number of mathematical models (see e.g. Checkley et al., 2015, and our own CellCycler model). The linear growth law can also be used to analyse tumour data, as in the draft paper: “Analysing within and between patient patient tumour heterogenity via imaging: Vemurafenib, Dabrafenib and Trametinib.” The linear growth equation will of course not be a perfect fit for the growth of all tumours (no simple model is), but it is based on a consistent and empirically verified model of tumour growth, and can be easily parameterised and fit to data.
So why hasn’t this linear growth law caught on more widely? The reason is that what scientists see in data often depends on their mental model of what is going on.
I first encountered this phenomenon in the late 1990s when doing my D.Phil. in the prediction of nonlinear systems, with applications to weather forecasting. The dominant theory at the time said that forecast error was due to sensitivity to initial condition, aka the butterfly effect. As I described in The Future of Everything, researchers insisted that forecast errors showed the exponential growth characteristic of chaos, even though plots showed they clearly grew with slightly negative curvature, which was characteristic of model error.
A similar effect in cancer biology has again changed the way scientists interpret data. Sometimes, a straight line really is the best solution.
Collins, V. P., Loeffler, R. K. & Tivey, H. Observations on growth rates of human tumors. The American journal of roentgenology, radium therapy, and nuclear medicine 76, 988-1000 (1956).
Laird A. K. Dynamics of tumor growth. Br J of Cancer 18 (3): 490–502 (1964).
W. V. Mayneord. On a Law of Growth of Jensen’s Rat Sarcoma. Am J Cancer 16, 841-846 (1932).
Stephen Checkley, Linda MacCallum, James Yates, Paul Jasper, Haobin Luo, John Tolsma, Claus Bendtsen. Bridging the gap between in vitro and in vivo: Dose and schedule predictions for the ATR inhibitor AZD6738. Scientific Reports, 5(3)13545 (2015).
Yorke, E. D., Fuks, Z., Norton, L., Whitmore, W. & Ling, C. C. Modeling the Development of Metastases from Primary and Locally Recurrent Tumors: Comparison with a Clinical Data Base for Prostatic Cancer. Cancer Research53, 2987-2993 (1993).
Tumour modelling has been an active field of research for some decades, and a number of approaches have been taken, ranging from simple models of an idealised spherical tumour, to highly complex models which attempt to account for everything from cellular chemistry to mechanical stresses. Some models use ordinary differential equations, while others use an agent-based approach to track individual cells.
A disadvantage of the more complex models is that they involve a large number of parameters, which can only be roughly estimated from available data. If the aim is to predict, rather than to describe, then this leads to the problem of overfitting: the model is very flexible and can be tuned to fit available data, but is less useful for predicting for example the effect of a new drug.
Indeed, there is a rarely acknowledged tension in mathematical modelling between realism, in the sense of including lots of apparently relevant features, and predictive accuracy. When it comes to the latter, simple models often out-perform complex models. Yet in most areas there is a strong tendency for researchers to develop increasingly intricate models. The reason appears to have less to do with science, than with institutional effects. As one survey of business models notes (and these points would apply equally to cancer modelling) complex models are preferred in large part because: “(1) researchers are rewarded for publishing in highly ranked journals, which favor complexity; (2) forecasters can use complex methods to provide forecasts that support decision-makers’ plans; and (3) forecasters’ clients may be reassured by incomprehensibility.”
Being immune to all such pressures (this is just a blog post after all!) we decided to develop the CellCycler – a parsimonius “toy” model of a cancer tumour that attempts to capture the basic growth and drug-response dynamics using only a minimal number of parameters and assumptions. The model uses circa 100 ordinary differential equations (ODEs) to simulate cells as they pass through the phases of the cell cycle; however the equations are simple and the model only uses parameters that can be observed or reasonably well approximated. It is available online as a Shiny app.
The CellCycler model divides the cell cycle into a number of discrete compartments, and is therefore similar in spirit to other models that for example treat each phase G1, S, G2, and mitosis as a separate compartment, with damaged cells being shunted to their own compartment (see for example the model by Checkley et al. here). Each compartment has its own set of ordinary differential equations which govern how its volume changes with time due to growth, apoptosis, or damage from drugs. There are additional compartments for damaged cells, which may be repaired or lost to apoptosis. Drugs are simulated using standard PK models, along with a simple description of phase-dependent drug action on cells. For the tumour growth, we use a linear model, based like the Checkley et al. paper on the assumption of a thin growing layer (see also our post on The exponential growth effect).
The advantages of compartmentalising
Dividing the cell cycle into separate compartments has an interesting and useful side effect, which is that it introduces a degree of uncertainty into the calculation. For example, if a drug causes damage and delays progress in a particular phase, then that drug will tend to synchronize the cell population in that state. However there is an obvious difference between cells that are affected when they are at the start of the phase, and those that are already near the end of the phase. If the compartments are too large, that precise information about the state of cells is lost.
The only way to restore precision would be to use a very large number of compartments. But in reality, individual cells will not all have exactly the same doubling time. We therefore want to have a degree of uncertainty. And this can be controlled by adjusting the number of compartments.
This effect is illustrated by the figure below, which shows how a perturbation at time zero in one compartment tends to blur out over time, for models with 25, 50, and 100 compartments, and a doubling time of 24 hours. In each case a perturbation is made to compartment 1 at the beginning of the cell cycle (the magnitude is scaled to the number of compartments so the total size of the perturbation is the same in terms of total volume). For the case with 50 compartments, the curve after one 24 hours is closely approximated by a normal distribution with standard deviation of 3.4 hours or about 14 percent. In general, the standard deviation can be shown to be approximately equal to the doubling time divided by the square root of N.
A unique feature of the CellCycler is that it exploits this property as a way of adjusting the variability of doubling time in the cell population. The model can therefore provide a first-order approximation to the more complex heterogeneity that can be simulated using agent-based models. While we don’t usually have exact data on the spread of doubling times in the growing layer, the default level of 50 compartments gives what appears to be a reasonable degree of spread (about 14 percent). Using 25 compartments gives 20 percent, while using 100 compartments decreases this to 10 percent.
Using the CellCycler
The starting point for the Shiny web application is the Cells page, which is used to model the dynamics of a growing cell population. The key parameters are the average cell doubling time, and the fraction spent in each phase. The number of model compartments can be adjusted in the Advanced page: note that, along with doubling time spread, the choice also affects both the simulation time (more compartments is slower), and the discretisation of the cell cycle. For example with 50 compartments the proportional phase times will be rounded off to the nearest 1/50=0.02.
The next pages, PK1 and PK2, are used to parameterise the PK models and drug effects. The program has a choice of standard PK models, with adjustable parameters such as Dose/Volume. In addition the phase of action (choices are G1, S, G2, M, or all), and rates for death, damage, and repair can be adjusted. Finally, the Tumor page (shown below) uses the model simulation to generate a plot of tumor radius, given an initial radius and growing layer. Plots can be overlaid with experimental data.
We hope the CellCycler can be a useful tool for research or for exploring the dynamics of tumour growth. As mentioned above it is only a “toy” model of a tumour. However, all our models of complex organic systems – be they of a tumor, the economy, or the global climate system – are toys compared to the real things. And of course there is nothing to stop users from extending the model to incorporate additional effects. Though whether this will lead to improved predictive accuracy is another question.
Stephen Checkley, Linda MacCallum, James Yates, Paul Jasper, Haobin Luo, John Tolsma, Claus Bendtsen. “Bridging the gap between in vitro and in vivo: Dose and schedule predictions for the ATR inhibitor AZD6738,” Scientific Reports.2015;5(3)13545.
Green, Kesten C. & Armstrong, J. Scott, 2015. “Simple versus complex forecasting: The evidence,” Journal of Business Research, Elsevier, vol. 68(8), pages 1678-1685.
This blog entry will focus on a rather long standing debate around model complexity and predictivity for a specific prediction problem from drug development. A typical drug project starts off with 1000’s of drugs for a certain idea. All but one of these drugs is eventually weened out through a series of experiments, which explore safety and efficacy, with the final drug being the one that enters human trials. The question we will explore is around a toxicity experiment performed rather early in the development (weening out) process, which determines the drug’s effect on the cardiac system.
Many years of research has identified certain proteins, ion-channels, which if a drug were to affect could lead to dire consequences for a patient. In simple terms, ion-channels allow ions, such as calcium, to flow in and out of a cell. Drugs can bind to ion-channels and disrupt their ability to function, thus affecting the flow of ions. The early experiment we are interested in basically measures how many ions flow across an ion-channel with increasing amount of drug. The cells used in these experiments are engineered to over-express the human protein we are interested in and so do not reflect a real cardiac cell. The experiment is pretty much automated and so allows one to screen 1000s of drugs a year against certain ion-channels. The output of the system is an IC50 value, the amount of drug needed to reduce the flow of ions across the ion-channel by 50 percent.
A series of IC50 values are generated for each drug against a number of ion-channels. (We are actually only interested in three.) The reason why a large screening effort is made is because we cannot test all the compounds in an animal model nor can we take all of them into man! So we can’t measure the effect of these drugs in real cardiac systems but we can measure their effect on certain ion-channel proteins which are expressed in the cardiac system we are interested in. The question is then: given a set of IC50 values against certain ion-channels for a particular drug can we predict how this drug will affect a cardiac system?
As mentioned earlier, drug development involves performing a series of experiments over time. The screening experiment described above is one of many used to look at cardiac toxicity. The next experiment in the pipeline, which could occur one or maybe two years later, is exploring the remaining drugs in an intact cardiac system. This could be a single cardiac cell taken from a dog, a portion of the ventricular wall, or something else entirely. After which, even less compounds are taken into dog studies before entering human trials. So the prediction question could be related to any one of these cardiac systems. The inputs into the prediction problem are the set of IC50 values, three in the cases we will look at, whereas the output, which we want to predict, are certain measures from the cardiac systems described.
At this point some of you may be thinking, well if we want to predict what will happen in a real cardiac system then why don’t we build a virtual version of the system using a large mathematical model (biophysical model)? Indeed people have done this. However, others (especially those who follow this blog) might also be thinking, I have three inputs and one output and given we screen lots of these compounds surely the dynamics are not that difficult to figure out, such that I can do something simpler and more cost effective! Again people have done this too. If I were to refer to the virtual system (consists of >100 parameters) as Goliath and the simple model (3 parameters) as David some of you can guess what the outcome is! A paper documenting the story in detail can be found here and the model used is available online here. I will just give a brief summary of the findings in the main paper.
The data-sets explored in the article involve making predictions in both animal studies and human. Something noticeable about the biophysical models used in the original articles was that a different structural model was needed for each study. This was not the case for the simple model which uses the same structure across all data sets. Given that the simple model gave the same if not better performance than the biophysical models it raises a question: why do the biophysical modelling community need a different model for different studies? In fact for two human studies, A and B, different human models were used, why? The reason may be that the degree of confidence in those models by people using them is actually quite low, hence the lack of consistency in the models used across the studies. Another issue not discussed by any of the biophysical modeling literature is the reproducibility of the data used to build such models. Given the growing skepticism of the reproducibility of preclinical data in science this adds further doubt to the suitability of such models for industrial use.
Given the points raised here (as well as a previous blog entry highlighting the misuse of these models by their own developers) can the biophysical modelling community be trusted to deliver a modelling solution that is both trustworthy and reliable? This is an important question as regulatory agencies are now also considering using these biophysical models together with some quite exciting new experimental techniques to change the way people assess the cardiac liability of a new drug.
Peer to peer lending is an option people are increasingly turning to, both for obtaining loans and for investment. The principle idea is that investors can decide who they give loans to, based on information provided by the loaner, and the loaner can decide what interest rate they are willing to pay. This new lending environment can give investors higher returns than traditional savings accounts, and loaners better interest rates than those available from commercial lenders.
Given the open nature of peer to peer lending, information is becoming readily available on who loans are given to and what the outcome of that loan was in terms of profitability for the investor. Available information includes the loaner’s credit rating, loan amount, interest rate, annual income, amount received etc. The open-source nature of this data has clearly led to an increased interest in analysing and modelling the data to come up with strategies for the investor which maximises their return. In this blog entry we will look at developing a model of this kind using an approach routinely used in healthcare, survival analysis. We will provide motivation as to why this approach is useful and demonstrate how a simple strategy can lead to significant returns when applied to data from the Lending Club.
In healthcare survival analysis is routinely used to predict the probability of survival of a patient for a given length of time based on information about that patient e.g. what diseases they have, what treatment is given etc. It is routinely used within the healthcare sector to make decisions both at the patient level, for example what treatment to give, and at the institutional level (e.g. health care providers), for example what new healthcare policies will decrease death associated with lung cancer. In most survival analysis studies the data-sets usually contain a significant proportion of patients who have yet to experience the event of interest by the time the study has finished. These patients clearly do not have an event time and so are described as being right-censored. An analysis can be conducted without these patients but this is clearly ignoring vital information and can lead to misleading and biased inferences. This could have rather large consequences were the resultant model applied prospectively. A key part of all survival analysis tools that have been developed is therefore that they do not ignore patients who are right censored. So what does this have to do with peer to peer lending?
The data on the loans available through sites such as the Lending Club contain loans that are current and most modelling methods described in other blogs have simply ignored these loans when building models to maximise investor’s returns. These loans described as being current are the same as our patients in survival analysis who have yet to experience an event at the time the data was collected. Applying a survival analysis approach will allow us to keep people whose loans are described as being current in our model development and thus utilise all information available. How can we apply survival analysis methods to loan data though, as we are interested in maximising profit and not how quickly a loan is paid back?
We need to select relevant dependent and independent variables first before starting the analysis. The dependent variable in this case is whether a loan has finished (fully repaid, defaulted etc.) or not (current). The independent variable chosen here is the relative return (RR) on that loan, this is basically the amount repaid divided by the amount loaned. Therefore if a loan has a RR value less than 1 it is loss making and greater than 1 it is profit making. Clearly loans that have yet to have finished are quite likely to have an RR value less than 1 however they have not finished and so within the survival analysis approach this is accounted for by treating that loan as being right-censored. A plot showing the survival curve of the lending club data can be seen in the below figure.
The black line shows the fraction of loans as a function of RR. We’ve marked the break-even line in red. Crosses represent loans that are right censored. We can already see from this plot that there are approximately 17-18% loans that are loss making, to the left of the red line. The remaining loans to the right of the red line are profit making. How do we model this data?
Having established what the independent and dependent variables are, we can now perform a survival analysis exercise on the data. There are numerous modelling options in survival analysis. We have chosen one of the easiest options, Cox-regression/proportional hazards, to highlight the approach which may not be the optimal one. So now we have decided on the modelling approach we need to think about what covariates we will consider.
A previous blog entry at yhat.com already highlighted certain covariates that could be useful, all of which are actually quite intuitive. We found that one of the covariates FICO range high (essentially is a credit score) had an interesting relationship to RR, see below.
Each circle represents a loan. It’s strikingly obvious that once the last FICO Range High score exceeds ~ 700 the number of loss making loans, ones below the red line decreases quite dramatically. So a simple risk adverse strategy would be just to invest in loans whose FICO Range High score exceeds 700, however there are still profitable loans which have a FICO Range High value less than 700. In our survival analysis we can stratify for loans below and above this 700 FICO Range High score value.
We then performed a rather routine survival analysis. Using FICO Range High as a stratification marker we looked at a series of covariates previously identified in a univariate analysis. We ranked each of the covariates based on the concordance probability. The concordance probability gives us information on how good a covariate is at ranking loans, a value of 0.5 suggests that covariate is no better than tossing a coin whereas a value of 1 is perfect, which never happens! We are using concordance probability rather than p-values, which is often done, because the data-set is very large and so many covariates come out as being “statistically significant” even though they have little effect on the concordance probability. This is a classic problem of Big Data and one option, of many, is to focus model building on another metric to counter this issue. If we use a step-wise building approach and use a simple criterion that to include a covariate the concordance probability must increase by at least 0.01 units, then we end up with a rather simple model: interest rate + term of loan. This model gave a concordance probability value of 0.81 in FICO Range High >700 and 0.63 for a FICO Range High value <700. Therefore, it does a really good job once we have screened out the bad loans and not so great when we have a lot of bad loans but we have a strategy that removes those.
This final model is available online here and can be found on the web-apps section of the website. When playing with the model you will find that if interest rates are high and the term of loan is low then regardless of the FICO Range High value all loans are profitable, however those with FICO Range High values >700 provide a higher return, see figure below.
The above plot was created by using an interest rate of 20% for a 36 month loan. The plot shows two curves, the one in red represents a loan whose FICO Range High value <700 and the black one a loan with FICO Range High value >700. The curves describe your probability of attaining a certain amount of profit or loss. You can see that for the input values used here, the probability of making a loss is similar regardless of the FICO Range High Value; however the amount of return is better for loans with FICO Range High value >700.
Using survival analysis techniques we have shown that you can create a relatively simple model that lends itself well for interpretation, i.e. probability curves. Performance of the model could be improved using random survival forests – the gain may not be as large as you might expect but every percentage point counts. In a future blog we will provide an example of applying survival analysis to actual survival data.