Kodera & Kuroda , 2002

2. Theoretical solar influences on climate Kodera & Kuroda , 2002

3. Yu, 2002 Ion-aerosol clean-air mechanism Shiptracks

4. Global electric circuit Ion-aerosol near-cloud mechanism Carslaw, Harrison et al., 2002 Makino and Ogawa, 1984

5. Cosmics Leaving OUtdoor Droplets • Laboratory experimentwith a special cloud chamber to study the possible link between galactic cosmic rays and cloud formation. • Almeida et al., 2013, Nature • new results show generally small contribution of ion-induced aerosol formation(small influence of GCR on cluster formation - except at low overall formation rates) • acid-amine nucleation can explain faster particle formation rates (as observed in the atmosphere) CERN CLOUD experiment

6. ISCCP (International Satellite Cloud Climatology Project) • - D1 dataset (from 1983 to 2008), intercalibrated radiance measurements from a fleet of polar and geostationary satellites • temporal resolution: 3h (IR data) • spatial resolution: 2.5° x2.5° (280 x 280km2) • distinguishes clouds at different altitude levels: e.g. high (>6.5km), middle (3.2 – 6.5km) and low (0 – 3.2km) Cloud datasets • MODIS (MODerate Resolution Imaging Spectroradiometer) • - views in 36 channels from Visible to thermal IR, on board two polar orbiting satellites Aqua, and Terra, operational since 2000 • temporal resolution: 12h, spatial resolution: 1° x 1°

7. The hypothesized link between cosmic ray flux and cloud cover Long-term studies • Svensmark and Friis-Chistensen (1997) • analyzed one solar cycle and reported that global cloud cover changed in phase with the GCR flux by 2-3%. MarshandSvensmark, 2000 Many critics for a found correlation and heavy debates in the scientific community: e.g. Kernthaler et al., 1999 Sun & Bradley, 2002; Laut, 2003; Kristjansson et al., 2002; 2003; Sloan and Wolfendale, 2008… low clouds (0-3.2km)

8. Long-term cloud data doesn’t support GCR-cloud link Low clouds (<3.2km), global GCR flux (%, Climax & Moscow NM) MODIS low cloud anomaly (%) ISCCP low cloud anomaly (%) • Correlation only in low (<3.2km) ISCCP cloud (1983–1995) • High correlation from 12-month smoothed data (df=4) • Low (non-significant) correlation from unsmoothed data Laken, Pallé, Čalogović & Dunne, 2012, SWSC

9. Correlations between CR flux and clouds are artificial ISCCP cloud data show clear indications of artificial trends conforming to the geostationary satellite footprint areas If linear trends in CR and cloud data are removed correlation becomes weak Laken, Pallé, Čalogović & Dunne, 2012, SWSC

10. Timeline of geostationarysatellite operation at equator over ISCCP observation period Cloud data should not be used for long-term analysis (Stubenrauch et al. 2012, 2013) from: NOAA/NCDC, satellite data, ISCCP resource

11. Artificial anti-correlation exists between lowand high/middle troposphere cloud • Low cloud obscured by overlying cloud (measurements are non-cloud penetrating). • Additionally, errors in identifying cloud height can contribute to shifts between low and high cloud. • Satellite cloud issues well known: e.g. Hughes, 1984; Minnis, 1989, Tian & Curry, 1989; Rozendall et al. 1995; Loeb & Davies, 1996; Salby & Callaghan, 1997, Campbell, 2004 Evidence for CR – cloud link is based on low level clouds: these data are not reliable! Many additional problems of long-term analysis (e.g. ENSO, volcanic eruptions...) Laken, Pallé, Čalogović & Dunne, 2012, SWSC

12. Short-term studies - opportunityto test GCR-cloud hypothesis • Short-term changes in cosmic rays (Forbush decreases) are comparable to variations during the solar cycle. • Upper limit cloud response to ionization around one week (e.g. Arnold, 2007). Čalogović et al., 2010 • But: • Meteorological variability (noise) in clouds has to be reduced to be able to detect the solar-related changes(signal) • Limitednumber of high-magnitude Forbush decreases

13. positive correlations: • Tinsley & Deen, 1991; Pudovkin & Vertenenko, 1995; Todd & Kniveton, 2001; 2004; Kniveton, 2004; Harrison & Stephenson, 2006; Svensmarket al., 2009; Solovyev & Kozlov, 2009; Harrison & Ambaum, 2010; Harrison et al. 2011; Okike & Collier, 2011; Dragićet al. 2011; 2013; Svensmark et al., 2012; Zhou et al. 2013 • negative correlations: • Wang et al., 2006; Troshichevet al., 2008 • no correlations or inconclusive results: • Pallé & Butler, 2001; Lam & Rodger, 2002 ; Kristjánssonet al., 2008 ; Sloan & Wolfendale, 2008; Lakenet al., 2009; Čalogovićet al., 2010; Laken & Kniveton 2011; Laken et al., 2012 Short-term studiesusing Forbush decreasesshow also conflicting results: • Why? • Improper use of statistical tools / wrong statistical assumptions • “quality” and properties of cloud datasets

14. Svensmark et al. 2012, ACPD Laken, Čalogović, Beer and Pallé (2012), ACPD 95 percentile(2σ) MODIS CF (5 EVENTS) 99 percentile (3σ) GCR (Climax NM) Big variability in the clouds can be often mixed with the expected signal! Dashed/dotted lines show correctly adjusted 2 and 3σ level – calculated from 10,000 MC simulations Proper statistical tests (MC simulations ) are needed to asses the correct statistical significance! Data NORMALIZED between period of day -15 and day -5

15. Calculate thresholds for statistical significance with Monte Carlo approach By generating large populations of random events identical in design to a composite with real events, the probability (p) of obtaining a given value by chance in a composite with real eventscan be accurately known. Distribution of daily anomalies This has advantages over traditional tests (e.g. T/U tests), as it requires no minimum sample size or specific distribution, and it doesn’t need adjustment for autocorrelation. Laken& Čalogović, SWSC, 2013

16. Size of sample area and number of events impact the noise Noise levels of data govern detectability of a signal. The noise varies with both the spatial area (a) considered by the data, and the number of composite events (n). ‘Noise’ indicated by 97.5th percentile values from 10,000 random composites of varying aand n size. Each point of grid represents another independent set of 10,000 MC simulations Laken& Čalogović, SWSC, 2013 possible to see how large a and n would need to be at minimum to see a hypothesized effect.

17. Majority of Fd studies use less than 50 events (n<50) Size of Europe (2%) Studies using only strong Fd events have usually less than 10 events

18. There are numerous issues that may affect results of studies issues of signal-to-noise ratios connected to spatio-temporal restrictions(e.g. by decreasing analyzed region size the searched signal may be buried in noise) interference from variability in data attime scales greater than those concerning hypothesis testing, which may not necessarily be removedby accounting for linear trends over the composite periods normalization procedures which affect both the magnitude of anomalies in composites, and estimations of their significance theapplication of statistical tests unable to account for autocorrelated data and biases imposed by theuse of improper normalization procedures

19. Conclusions • Minor methodological differences in composite analysis can produce conflicting results. These are the likely source of discrepancies between cosmic ray – cloud composite studies. • Present cloud datasets are limited to detect a small changes in cloud cover as well to detect the regional cloud changes (<several thousand km) due to the big natural cloud variability (noise) • No compelling evidence to support acosmic ray cloud connection hypothesis using the satellite cloud data (ISCCP, MODIS) with long- or short-term (Fd) studies.

20. Thank you! This work received support from the European COST Action ES1005 (TOSCA) and COMESEP FP7 project. PICTURE BY: J. ČALOGOVIĆ, HVAR, MAY 2013

21. DTR shows response to Fd events? Surface level diurnal temperature range (DTR) an effective proxy for cloud cover • Compositeof 35 Fd events (>7%) show significant increase in DTR - support for GCR-cloud link (Dragić et al., 2011) • LargerFd events producelarger effects (Dragić et al., 2011;2013).

22. ...and also to GLE events

23. Analysis of Dragic et al. results Dragic et al. normalization from day -10 to day -5 & significance levels (3σ) Analysis of the same data as in Dragic et al. (DTR station data and 37 Fd events ) shows that Dragic et al. overestimated the statistical significance of their result by using just t-test and some statistical assumptions. confidence intervals calculated from 100,000 MC simulations No normalization, 21-day running mean Laken, Čalogović, Shahbaz and Pallé (2012), JGR Laken & Čalogović, 2013, ERL

24. ISCCP cloud data show clear indications of artificial trends conforming to the geostationary satellite footprint areas linear trend of each ISCCP D1 VIS-IR pixel over the 1983 to 2010 period Laken, Pallé, Čalogović & Dunne, 2012, SWSC

25. Changes in satellites contributing to ISCCP Start of the dataset Arbitrary colors represent data collected by different satellite instruments. 20 years later When available ISCCP preferentially uses only center of geostationary region.

26. MarshandSvensmark, 2003

27. Čalogović et al. 2010, GRL • 6 largest Forbush decreases (1989 - 1998) • calculated cosmic ray induced ionization rate • independent analysis of all grid cells for each lag (total 5.8 million correlations calculated) Example of short-term study

28. No indication of any significant response of the cloud cover to Forbush decreases. • Control events and sensitivity tests show that our method is capable to detect the changes in the clouds. Results Čalogović et al. 2010, GRL

29. Laken et. al., 2011 (JGR) ISCCP clouds GCR flux TSI • Epoch superpositional (composite) analysis of 123 Fd events. • cloud cover decreases about 2 days before the onset of Fd TSI influences the cloud cover?

30. TSI composites (superposed epoch analysis) • Active Cavity Radiometer Irradiance Monitor (ACRIM) reconstruction, 1978-present, daily values • ISCCP D1 data • area-averaging was applied for different regions: global, low latitudes (<45°), high latitudes (>45°) regions over land, regions over ocean TSI, GCR & clouds Laken & Čalogović, 2011, GRL

31. GCR composites • all composites (TSI, GCR, F10.7) correlated with corresponding cloud data using a lag of 20 days. Robust Monte Carlo techniques are used to determine the statistical significance. • Results • no widespread detectable changes in cloud cover at any tropospheric level or analyzed region within a 20 day period of the solar forcing clearly associated with solar activity changes (TSI/GCR/F10.7). F10.7 (EUV) composites Clouds doesn’t show any related changes to TSI/GCR/F10.7 • TSI or UV changes occurring during reductions in the GCR flux are not masking a solar – cloud response. Laken & Čalogović, 2011, GRL

32. Possible reasons • there is no relationship between cosmic rays and clouds • other solar parameters may interfere with the results (e.g. TSI, UV) – Laken & Čalogović, 2011, JGR • a relationship is too weak to detect (signal to noise ratio) • relationship exists but it is constrained by the atmospheric conditions at the time So how to test the CR-cloud link reliably ?

33. Composites and significance intervals daily averaged CR flux global cloud cover anomalies ±1.96 Standard error of mean (SEM) Anomalies (daily mean – 21-day running mean) Confidence intervals at p0.05and p0.01 levels (obtained from PDF of 10,000 Monte Carlo simulations)

34. How to obtain a false positive • Cookbook… • Identify a base or ‘undisturbed’ period before the key events, that represent ‘normalconditions’(e.g. figure uses t-10- t-5) • Calculate deviations against this ‘undisturbed’ period (i.e. subtract every t point from mean of ‘normal conditions’) • Statistically compare the data to the ‘undisturbed’ period (e.g. T-test, or even MC from the base period [red lines p0.05 p0.01]) Normalization to base period reduces population variability towards base period, narrowing confidence intervals.

35. How to avoid a false positive • Overcoming biaswith • Monte Carlo (MC): • Use confidence intervalsfrom PDFs obtained with MCs, calculated independently for each t point • Autocorrelation effects are automatically taken in to account (random samples in the MC all treated with an identical approach to the analyzed composite[bluelines p0.05 p0.01]) Two different results for t+5 (the above with a mean p<0,05 and the earlier, with a mean 0.01>p<0.05 so which is correct?

36. Impact of sample restrictions Fd events are quasi-random with respect to climate Restricting Fd events (e.g. by magnitude) diminishes subsample, making it less likely a random draw from the parent population may represent the subsample. Strongly restricted samples need this consideration added to the MC…

37. Meteorological noise seriously limits the detection of GCR induced cloud signal • Careful selection of analyzed region size and enough big number of a significant events in a composite is essential to be able to detect the searched signal in clouds. • For composites smaller than 10 events (depending on the event size) and area sizes smaller than approx. 15x15 deg. (~ size of Europe) the noise is probably much bigger than the searched signal in clouds. • MC simulations with static normalization period prior Fd events (e.g. -10 to -5 days) showed that their variability is increased by factor 1.5 to 8 times. Such normalization can be avoided just by using proper low pass filtering (e.g. 21 day moving average).

38. DTR anomalies in Fd, and GLE composites with CIs corrected for subsample effects Laken & Čalogović, ERL, accepted

39. For deviations at timescales of aprox. 1/3rd the width of the smooth filter, disturbance attenuation is neglible. Overshoot / undershoot effects by filtering the data with different smooths (running mean)

40. Extension to longer analysis periods reveals no unusual variability in clouds during Fd events MODIS Liquid cloud fraction changes using 5 biggest Fd events from Svensmark et al. (2012) Values are anomalies from 21-day moving averages (i.e. mean of each day subtracted from 21-day moving average). Dashed and dotted lines indicate the 95th and 99th (two-tailed) percentile confidence intervals respectively calculated from 100,000 Monte Carlo simulations. ±20 day analysis period ±100 day analysis period Laken, Čalogović, Beer and Pallé (2012), ACPD

41. Just one event (and eventually outlier) can influence the whole composite MODIS cloud fraction composite for Fd events 1, 3, 4, 5, 6 ranked by Svensmark et al. 2012 By replacing the event 2 with event 6 there are no significant changes in the composite! Individual 5 Fd events plotted against event 2 (19.1.2005) where is clear that all significance in Svensmark composite comes from event 2. Laken, Čalogović, Beer and Pallé (2012), ACPD

42. Noise in clouds can be reduced with bigger composite sizes! Example for ISCCP low clouds (0-3.2km) Calculated as a 97.5 percentile value from 100,000 MC simulations, no normalization applied, 41 day analysis period. Similar results are obtained for ISCCP total, middle and high clouds and MODIS cloud fraction and optical depth. Composite sizes and cloud variability • std decrease to 50% -> composite sample sizes of approx. 32 events • std decrease to 20% -> approx. 126 events • (calculated from difference between 10 & 1,000 events)

43. By decreasing the analyzed region size noise in clouds is increased! Example for ISCCP low clouds (0-3.2km) Calculated as 97.5 percentile value from 100,000 MC simulations, no normalization applied, 41 day analysis period with 20 day analysis period. Area size is expressed as part (%) of Earth’s surface analyzed (exceptions are missing measurements). Analyzed region size and cloud variability • std decrease to 50% -> by area sizes of approx. 17˚x17˚(valid for equator regions) • std decrease to 20% -> approx. 42˚x42˚ (valid for equator regions) • (calculated from difference between 0.08% and 100% of Earth’s surface)

44. DTR shows response to Fd events? • Surface level Diurnal Temperature Range (DTR) – effective proxy for cloud cover • Dragić et al. (2011) – composite of 35 Fd events (>7%) show significant increase in DTR - support for GCR-cloud hypothesis • interesting approach worth investigating further

45. Extended analysis of DTR data doesn’t show any response to Fd events n=267 n=29 NCEP/NCAR reanalysis data (60°N – 60°S, land-area pixels only) DTR from 210 meteorological stations (77.7°N – 34.7°N, 179.4°W – 170.4°E) TSI flux from the PMOD reconstruction 99th and 95th percentile confidence intervals (dotted and dashed lines) are calculated from 100,000 MC simulations Climax/Moscow NM Laken, Čalogović, ShahbazandPallé (2012), JGR

46. DTR shows no response to GCR or solar activity Spatial distribution of DTR anomalies between day +3 and +6 Long term analysis (60 years of data) shows alsothat there is no significant periodicities in DTR data connected to the solar periodicities (e.g. 11-year, 1.68-year ). In conclusion, we find no evidence to support claims of a link between DTR and solar activity. Laken, Čalogović, ShahbazandPallé (2012), JGR

47. How to isolate the signal using composite Successive averaging of events (in time or space) Used to increase signal-to-noise ratio (SNR) Enable detection of small amplitude signal against large variability

48. Composites should be made with anomalies rather than raw data... ... to minimize variations in data unconnected with hypothesis testing (high-pass filtering) Comparision of twomethods toremove long-termvariations (n=20): linear trends removed (black), and a only 21-day running mean removed a) one random composite over t±40 Differences (b), indicates remaining non-linear variations in composite from synoptic scale variability. If not removed, this will bias results of composites. b) 100,000 composites at t0 Proper selection of smooth filter width is needed to prevent signal attenuation (duration of searched signal is 1/3rd width of smooth filter)

49. How many Monte Carlo iterations are enough to get reliable significance intervals? Laken& Čalogović, SWSC, 2013

50. What is the best procedure? We should find the optimum procedure for generating and assessing composites: Createproper anomalies for testing Accurately identify their significance (p-values) with robust Monte Carlo methods